Landslide disaster early warning method based on LSTM-SARIMA mixed data driving model

Abstract

The invention discloses a landslide disaster early warning method based on an LSTM-SARIMA mixed data driving model. Global and local noise elimination is performed on radar displacement monitoring data by adopting a two-stage noise reduction method combining moving average and wavelet transform, so that the signal-to-noise ratio is effectively improved; the method comprises the following steps: decomposing slope displacement data into a trend term dominated by gravity and a periodic term influenced by the environment by utilizing a Hodry-Precott (HP) filtering method, and realizing differentiated analysis of a multi-factor action mechanism; a mixed prediction model combining a long short-term memory network (LSTM) and a seasonal autoregressive integrated moving average model (SARIMA) is provided, modeling and prediction are carried out on decomposed displacement components respectively, and results are fused to obtain a high-precision total displacement prediction value; an improved T-t curve is constructed based on a predicted displacement-time curve, a tangent angle index for landslide early warning is proposed according to a Saito three-stage theory, and quantitative and interpretable early warning criterion setting is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of open-pit mine side fracture control, and specifically relates to a mine slope landslide disaster early warning method, which is particularly suitable for early warning of landslide disasters on rock slopes of open-pit mines. Background Art

[0002] With the development and utilization of mineral resources, shallow resources are becoming increasingly scarce, and mining is increasingly moving deeper, accompanied by the frequent occurrence of geological disasters. Landslides are one of the most common geological hazards in mining production. In-depth research on early warning technology for slope landslide disasters in open-pit mines has important theoretical and practical significance for promoting the efficient mining of mineral resources and reducing human and property losses in open-pit mining.

[0003] Rock slope landslides are characterized by sudden and destructive nature, posing a significant threat to human life, equipment, and infrastructure. Currently, landslide early warning research primarily focuses on single tasks, such as analyzing monitoring data or studying impact mechanisms. However, landslide early warning is a complex, multi-stage technical system, each of which is closely interconnected. Focusing solely on a single component will not achieve effective monitoring and early warning.

[0004] To reduce the safety risks and property losses associated with landslides, mines have introduced a range of monitoring methods, including the Global Navigation Satellite System (GNSS), radar, anchor strain gauges, inclinometers, and microseismic monitoring. However, monitoring is merely a means to an end; early warning is the goal. Traditional mine landslide monitoring and early warning systems have certain shortcomings. First, data analysis methods lag behind and cannot keep pace with the development of monitoring technology. While the accuracy of acquired data is increasing, advanced analysis methods are lacking. Second, the increasing precision of monitoring results in the daily receipt of massive amounts of monitoring data, making traditional manual data processing methods ineffective. Monitoring is only meaningful if landslides can be successfully predicted and prevented before they occur.

[0005] Landslide early warning is a complex technical system, encompassing data preprocessing (noise reduction and data cleaning), analysis of deformation influencing factors, development of intelligent prediction algorithms, and formulation of early warning indicators. Each of these technologies is a key component of the landslide early warning system. The application of landslide early warning technology can promptly detect the precursors of landslides, assess slope hazard, and implement preventive measures in advance. This is currently considered the most effective method for reducing landslide hazards. Summary of the Invention

[0006] The purpose of the present invention is to provide a landslide disaster warning method based on an LSTM-SARIMA hybrid data-driven model that can accurately warn of rock slope landslide disasters under complex working conditions, in response to the characteristics of rock slope landslides such as strong suddenness and high destructiveness, as well as the technical difficulties of traditional landslide warning methods such as low utilization of monitoring data, poor prediction accuracy, and unreasonable warning indicators, which lead to inaccurate and untimely landslide warnings.

[0007] To achieve the above-mentioned object of the present invention, the landslide disaster early warning method based on the LSTM-SARIMA hybrid data-driven model of the present invention is implemented by the following steps:

[0008] S1 addresses the complexity and irregularities of radar displacement monitoring data from open-pit mines by combining a sliding average filter with wavelet denoising. The sliding average filter divides the radar displacement monitoring data into several windows. The data within each window is averaged to obtain an average value, which serves as the output value for that window. As the window slides forward, new data is added and old data is deleted, enabling real-time signal processing. A wavelet denoising algorithm is then used to perform localized noise reduction on the radar displacement monitoring data. By selecting an appropriate threshold, wavelet coefficients above the threshold are considered to represent signal and are retained, while those below the threshold are considered to represent noise and are set to zero. The sliding average filter effectively suppresses high-frequency noise, smoothes the signal, and enables comprehensive analysis of the global signal for overall denoising. The denoising wavelet transform allows for multi-scale analysis in the time-frequency domain, enabling better capture of local signal characteristics.

[0009] S2 is based on the radar displacement monitoring data with noise removed in step S1, and uses the Hodrick-Prescott filter to decompose the displacement components. A smooth trend line is fitted by minimizing the trade-off between the volatility and trend of the data. The original data is decomposed into a trend part and a periodic fluctuation part, separating the trend term displacement and the periodic term displacement in the displacement monitoring data.

[0010] S3 uses the trend item displacement data separated in step S2 and employs an LSTM data-driven model to predict future trend item displacement data. The distinct pattern of trend item displacement implies a long-term dependency between the current displacement value and the previous displacement value. Long Short-Term Memory (LSTM) networks, with their ability to memorize and learn long-term dependencies, can effectively capture these dependencies and use historical data to predict future values. Furthermore, monotonically increasing trends are often accompanied by complex nonlinear relationships. As a powerful nonlinear model, LSTM can flexibly adapt to and capture nonlinear characteristics in the data, enabling more accurate predictions.

[0011] S4 uses the grey correlation method to quantitatively analyze the relationship between the influencing factors—rainfall, groundwater level, groundwater osmotic pressure—and the periodic displacement based on the periodic displacement separated in step S2. The grey correlation method determines whether the relationship is close based on the similarity of the geometric shapes of the sequence curves. The closer the curves, the greater the correlation between the corresponding sequences, and vice versa. The grey correlation analysis calculation formula is as follows:

[0012]

[0013] Where: i (k) is the resolution coefficient, x i is the data in the i-th column; k is the k-th value in the i-th column; ρ is the resolution coefficient, 0≤ρ≤1, Δ i (k)=|y(k)-x i (k)|.

[0014] S5 uses the cyclical displacement data separated in step S2 and employs the SARIMA data-driven model to predict future cyclical displacement data. The seasonality of the SARIMA model makes it particularly well-suited for predicting cyclical items. By accounting for seasonal patterns in time series data, the SARIMA model can better predict cyclical fluctuations and trends, making it superior when dealing with cyclical data.

[0015] S6 uses a combination of the LSTM data-driven model from step S3 and the SARIMA data-driven model from step S5 to predict the trend and periodic terms in slope displacement, and then performs a weighted fusion of the prediction results to produce a fused prediction. The LSTM-SARIMA hybrid model demonstrates significant advantages by combining the powerful nonlinear modeling capabilities of LSTM with the excellent handling of seasonality and trends provided by SARIMA. By combining the nonlinear modeling capabilities of LSTM with the linear modeling capabilities of SARIMA, the hybrid model can handle both short-term and long-term emergencies. When faced with complex slope displacement changes, seasonal variations can be taken into account, providing a more comprehensive description of the 13 dynamic variations in slope displacement. The LSTM-SARIMA hybrid model not only achieves high prediction accuracy but also provides more intuitive explanations for engineering practice.

[0016] S7 obtains the future spatiotemporal displacement prediction results based on the prediction of S6, and draws the displacement-time curve in combination with the radar displacement monitoring data. In order to solve the problem of low prediction accuracy in landslide disaster early warning and forecasting due to different dimensions, a landslide disaster early warning and forecasting method is proposed based on the displacement-time curve. This method uses coordinate transformation to standardize the dimensions of the two coordinates, establishes the change curve of relative time T and real time t, ensures the correct determination of the tangential angle, and thus establishes a more accurate judgment criterion for landslide disaster early warning and forecasting; according to the Tt curve obtained after the dimension transformation, the improved tangential angle α is obtained. i The expression:

[0017]

[0018] where α i is the improved tangential angle obtained according to the Tt curve; T(i) is the relative time; t i is the actual monitoring time.

[0019] S8 determines the following warning criteria based on step S7, by defining the tangential angle α i , determine the early warning indicators:

[0020] When α i When the angle is less than 45°, the slope deformation is in the initial deformation stage;

[0021] When α i When ≈45°, the slope deformation is in the uniform deformation stage;

[0022] When α i When the angle is greater than 45°, the slope deformation enters the accelerated deformation stage;

[0023] When α i When the angle is greater than 80°, the slope deformation enters the uniform acceleration deformation stage;

[0024] When α i When the angle is greater than 85°, the slope deformation is in the stage of imminent sliding and an early warning is required;

[0025] When α i When the angle is ≈89°, the slope begins to slide.

[0026] Preferably, in step S2, the calculation formula of the Hodrick-Prescott filter is as follows:

[0027] y t =τ t +c t

[0028] y t is the observed original displacement data; τ t is the trend component, indicating the long-term trend; c t is a periodic fluctuation component;

[0029] The optimization goal of the Hodrick-Prescott filter is to minimize the error of the following equation:

[0030]

[0031] where λ is a smoothing parameter that is used to balance the smoothness of the trend and the goodness of fit to the data.

[0032] Furthermore, the workflow of the Hodrick-Prescott filter is as follows:

[0033] ① Decompose the original displacement data into trend part and periodic fluctuation part.

[0034] ② Minimize an error function by adjusting the trend part to minimize the error between the original displacement data and the fitted value.

[0035] ③Determine the smoothing parameter λ based on the characteristics of the data and the purpose of analysis.

[0036] Compared with existing technologies, the comprehensive technical system for early warning of open-pit rock slope landslide disasters, constructed based on the landslide disaster early warning method of the LSTM-SARIMA hybrid data-driven model, covers the entire process from monitoring data collection and processing, impact mechanism analysis, displacement prediction based on intelligent algorithms, to the construction of landslide early warning indicators. It has the following advantages:

[0037] (1) This paper proposes a systematic early warning framework for rock slope landslides, which are characterized by strong suddenness and great destructiveness. First, a two-stage noise reduction method combining sliding average and wavelet transform is used to remove global and local noise from radar displacement monitoring data, effectively improving the signal-to-noise ratio and ensuring data quality. Second, the Hodrick-Prescott (HP) filtering method is used to decompose the slope displacement data into a trend term dominated by gravity and a periodic term affected by the environment, realizing the differentiation and analysis of the multi-factor action mechanism. Then, a hybrid prediction model combining long short-term memory network (LSTM) and seasonal autoregressive integrated moving average model (SARIMA) is proposed to model and predict the decomposed displacement components separately, and the results are fused to obtain a high-precision total displacement prediction value with a prediction accuracy of 96%. Finally, an improved T-t curve is constructed based on the predicted displacement-time curve, and a tangent angle index for landslide early warning is proposed based on Saito's three-stage theory to achieve quantitative and interpretable early warning criterion setting.

[0038] (2) The method of the present invention realizes the coordinated removal of global and local noise through a denoising algorithm that combines sliding average and wavelet transform, significantly improving the authenticity and purity of monitoring data, laying a solid foundation for subsequent analysis and prediction, and avoiding false alarms and missed alarms.

[0039] (3) The displacement component decomposition technology is used to divide the slope displacement into a gravity-dominated trend term and an environmental disturbance period term, which effectively distinguishes the specific impact of different factors (such as rainfall, deadweight, and mining activities) on the landslide process, realizes the differentiated analysis of factor responses, and is conducive to guiding precise governance.

[0040] (4) A hybrid LSTM-SARIMA data-driven model was constructed to predict the displacement trend term and the period term separately and perform a weighted fusion, achieving higher prediction accuracy than traditional models. Furthermore, an early warning indicator system based on an improved tangent angle was proposed, combined with Saito's three-stage theory, to more accurately reflect the critical state of landslides and achieve quantifiable and interpretable early warning threshold setting. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is the sliding average filter-wavelet denoising combined denoising result diagram designed by the present invention;

[0042] Figure 2 This is the result diagram of separation and decomposition of displacement monitoring data;

[0043] Figure 3 This is the result graph of trend item displacement predicted by LSTM model;

[0044] Figure 4 This is the schematic diagram of the LSTM model;

[0045] Figure 5 This is the result diagram of grey correlation analysis;

[0046] Figure 6 This is the result diagram of the periodic term displacement predicted by the SARIMA model;

[0047] Figure 7 It is a schematic diagram of the SARIMA algorithm flow;

[0048] Figure 8 It is the Saito three-stage curve diagram;

[0049] Figure 9 is the Tt curve after transformation;

[0050] Figure 10 This is the early warning indicator result diagram;

[0051] Figure 11 This is the technical roadmap of the landslide disaster early warning method based on the LSTM-SARIMA hybrid data-driven model of the present invention. DETAILED DESCRIPTION

[0052] To describe the present invention, the landslide disaster early warning method based on the LSTM-SARIMA hybrid data-driven model of the present invention is further described in detail below in conjunction with embodiments.

[0053] Depend on Figure 11 As shown in the technical roadmap of the landslide disaster early warning method based on the LSTM-SARIMA hybrid data-driven model of the present invention, the landslide disaster early warning method based on the LSTM-SARIMA hybrid data-driven model of the present invention is characterized by being implemented by the following steps:

[0054] In view of the complexity and irregularity of radar displacement monitoring data in open-pit mines, S1 adopts a noise reduction method that combines sliding average filtering and wavelet denoising. The sliding average filter divides the radar displacement monitoring data into several windows. The radar displacement monitoring data in each window is averaged to obtain an average value as the output value of the window; as the window slides forward, new data is added and old data is deleted, thereby realizing real-time processing of the signal; then the wavelet denoising algorithm is used to perform local key noise reduction on the radar displacement monitoring data. The basic process of wavelet threshold denoising is that after the signal is transformed by wavelet, the wavelet coefficients generated by the signal contain important information of the signal. After the signal is decomposed by wavelet, the wavelet coefficient is larger, the wavelet coefficient of the noise is smaller, and the wavelet coefficient of the noise is smaller than the wavelet coefficient of the signal. By selecting a suitable threshold, the wavelet coefficient greater than the threshold is considered to be generated by the signal and should be retained, and the wavelet coefficient less than the threshold is considered to be generated by noise and set to zero. The entire noise reduction process is shown in Figure 1 shown.

[0055] S2 is based on the radar displacement monitoring data that has been de-noised in step S1. The Hodrick-Prescott filter is used to decompose the displacement components. A smooth trend line is fitted by minimizing the trade-off between the volatility and trend of the data. The original data is decomposed into a trend part and a periodic fluctuation part, separating the trend term displacement and the periodic term displacement in the displacement monitoring data. The data separation results are shown in Figure 2 The calculation formula of the Hodrick-Prescott filter is as follows:

[0056] y t =τ t +c t

[0057] y t is the observed original displacement data; τ t is the trend component, indicating the long-term trend; c t is a periodic fluctuation component;

[0058] The optimization goal of the Hodrick-Prescott filter is to minimize the error of the following equation:

[0059]

[0060] where λ is a smoothing parameter that is used to balance the smoothness of the trend and the goodness of fit to the data.

[0061] The workflow of the Hodrick-Prescott filter is as follows:

[0062] ① Decompose the original displacement data into trend part and periodic fluctuation part.

[0063] ② Minimize an error function by adjusting the trend part to minimize the error between the original displacement data and the fitted value.

[0064] ③Determine the smoothing parameter λ based on the characteristics of the data and the purpose of analysis.

[0065] S3 uses the trend item displacement data separated in step S2 and adopts the LSTM data-driven model to predict the future trend item displacement data. The prediction results are as follows: Figure 3 As shown in the figure. The core of the LSTM network is the unit. Each unit contains three gates and a state unit (cell state), which work together to control the flow and memory of information. The three gates are the input gate, the forget gate, and the output gate. They control the input, forgetting, and output of information through learned weights. The state unit is responsible for recording and transmitting information. The structure diagram of LSTM is shown in the figure. Figure 4 shown.

[0066] S4 is based on the periodic displacement separated in step S2, and the relationship between the influencing factors - rainfall, groundwater level, groundwater osmotic pressure and periodic displacement is quantitatively analyzed using the grey correlation method. The analysis results are shown in Figure 5 The grey correlation method determines whether the relationship is close based on the similarity of the geometric shapes of the sequence curves. The closer the curves are, the greater the correlation between the corresponding sequences, and vice versa. This method is suitable for qualitative and quantitative analysis of the relationship between various factors in the system, thereby providing a scientific basis for decision makers and optimizing system performance. The grey correlation analysis calculation formula is as follows:

[0067]

[0068] Where: i (k) is the resolution coefficient, x i is the data in the i-th column; k is the k-th value in the i-th column; ρ is the resolution coefficient, 0≤ρ≤1, Δ i (k)=|y(k)-x i (k)|.

[0069] S5 uses the periodic displacement data separated in step S2 and adopts the SARIMA data-driven model to predict the future periodic displacement data. The prediction results are shown in Figure 6 SARIMA (Seasonal Autoregressive Integrated Moving Average) is a time series forecasting method that is an extension of the ARIMA model and is specifically designed to process time series data with seasonal patterns. The ARIMA model captures the trend and seasonality of time series data by considering the autoregressive term (AR), the difference (I), and the moving average term (MA). The SARIMA model adds seasonal parameters to this model to make it better adapt to seasonal data. The parameter details are shown in Table 1, and the working process is shown in Figure 7 .

[0070] Table 1 Details of SARIMA parameters

[0071]

[0072] S6 uses a combination of the LSTM data-driven model from step S3 and the SARIMA data-driven model from step S5 to predict the trend and periodic terms in slope displacement, and then performs a weighted fusion of the prediction results to produce a fused prediction. The LSTM-SARIMA hybrid model demonstrates significant advantages by combining the powerful nonlinear modeling capabilities of LSTM with the excellent handling of seasonality and trends provided by SARIMA. By combining the nonlinear modeling capabilities of LSTM with the linear modeling capabilities of SARIMA, the hybrid model can handle both short-term and long-term emergencies. When faced with complex slope displacement changes, seasonal variations can be taken into account, providing a more comprehensive description of the 13 dynamic variations in slope displacement. The LSTM-SARIMA hybrid model not only achieves high prediction accuracy but also provides more intuitive explanations for engineering practice.

[0073] S7 Scholars usually judge the sliding process of slopes based on Saito's three-stage theory, such as Figure 8 The method of the present invention obtains future spatiotemporal displacement predictions based on S6 predictions and plots displacement-time curves in combination with radar displacement monitoring data. To address the problem of low forecast accuracy in landslide disaster warning and forecasting due to different dimensions, a landslide disaster warning and forecasting method based on displacement-time curves is proposed.

[0074] A specific analysis of the uniform deformation phase reveals that the cumulative displacement S is linearly distributed with time t, i.e., s = vt, where v is the deformation rate during this phase and is a constant value. Therefore, by re-dimensioning the y-axis displacement at a constant rate to obtain the same dimension as the x-axis time coordinate, we can obtain:

[0075]

[0076] Through transformation, we can get the curve of relative time T relative to the real monitoring time t, such as Figure 9 shown.

[0077] The method of the present invention uses coordinate transformation to standardize the dimensions of the two coordinates and establish the change curve of relative time T and real time t; according to the Tt curve obtained after the dimension transformation, the improved tangential angle α is obtained. i The expression:

[0078]

[0079] where α i is the improved tangential angle obtained according to the Tt curve; T(i) is the relative time; t i is the actual monitoring time.

[0080] S8 determines the following warning criteria based on step S7, by defining the tangential angle αi , determine the early warning indicators:

[0081] When α i When the angle is less than 45°, the slope deformation is in the initial deformation stage;

[0082] When α i When ≈45°, the slope deformation is in the uniform deformation stage;

[0083] When α i When the angle is greater than 45°, the slope deformation enters the accelerated deformation stage;

[0084] When α i When the angle is greater than 80°, the slope deformation enters the uniform acceleration deformation stage;

[0085] When α i When the angle is greater than 85°, the slope deformation is in the stage of imminent sliding and an early warning is required;

[0086] When α i When the angle is ≈89°, the slope begins to slide.

[0087] The comprehensive early warning system constructed by the method of the present invention has been successfully applied to an open-pit mine in East China, verifying its accurate early warning capability for rock slope landslide disasters under complex working conditions, with an accuracy rate of nearly 100%, and has good practicality and promotion value.

Claims

1. A landslide disaster early warning method based on LSTM-SARIMA hybrid data-driven model, characterized by Use the following steps to implement: In view of the complexity and irregularity of radar displacement monitoring data in open-pit mines, S1 adopts a noise reduction method combining sliding average filtering and wavelet denoising. The sliding average filter divides the radar displacement monitoring data into several windows. The radar displacement monitoring data in each window is averaged to obtain an average value as the output value of the window. As the window slides forward, new data is added and old data is deleted, thus achieving real-time signal processing. Then, a wavelet denoising algorithm is used to perform local noise reduction on the radar displacement monitoring data. By selecting a suitable threshold, wavelet coefficients greater than the threshold are considered to be generated by signals and should be retained, while those less than the threshold are considered to be generated by noise and set to zero. S2 uses the Hodrick-Prescott filter to decompose the displacement components based on the radar displacement monitoring data from which noise has been removed in step S1. It fits a smooth trend line by minimizing the trade-off between the volatility and trend of the data. It decomposes the original data into a trend component and a periodic fluctuation component, separating the trend term displacement and the periodic term displacement in the displacement monitoring data. S3 uses the trend item displacement data separated in step S2 and adopts the LSTM data-driven model to predict the future trend item displacement data; S4 uses the grey correlation method to quantitatively analyze the relationship between the influencing factors—rainfall, groundwater level, groundwater osmotic pressure—and the periodic displacement based on the periodic displacement separated in step S2. The grey correlation method determines whether the relationship is close based on the similarity of the geometric shapes of the sequence curves. The closer the curves, the greater the correlation between the corresponding sequences, and vice versa. The grey correlation analysis calculation formula is as follows: Where: i (k) is the resolution coefficient, x i is the data in the i-th column; k is the k-th value in the i-th column; ρ is the resolution coefficient, 0≤ρ≤1, Δ i (k)=|y(k)-x i (k)|; S5 uses the periodic displacement data separated in step S2 and adopts the SARIMA data-driven model to predict the future periodic displacement data; S6 uses the LSTM data-driven model in step S3 and the SARIMA data-driven model in step S5 to predict the trend term and period term in the slope displacement respectively, and performs weighted fusion of the prediction results to obtain a fused prediction result; S7 obtains the future spatiotemporal displacement prediction results based on the prediction of S6, and draws the displacement-time curve in combination with the radar displacement monitoring data; in order to solve the problem of low prediction accuracy due to different dimensions in landslide disaster early warning and forecasting, a landslide disaster early warning and forecasting method is proposed based on the displacement-time curve. This method uses coordinate transformation to standardize the dimensions of the two coordinates and establishes the change curve of relative time T and real time t; according to the Tt curve obtained after the dimension transformation, the improved tangential angle α is obtained. i The expression: where α i is the improved tangential angle obtained according to the Tt curve; T(i) is the relative time; t i is the actual monitoring time. S8 determines the following warning criteria based on step S7, by defining the tangential angle α i , determine the early warning indicators: When α i When the angle is less than 45°, the slope deformation is in the initial deformation stage; When α i When ≈45°, the slope deformation is in the uniform deformation stage; When α i When the angle is greater than 45°, the slope deformation enters the accelerated deformation stage; When α i When the angle is greater than 80°, the slope deformation enters the uniform acceleration deformation stage; When α i When the angle is greater than 85°, the slope deformation is in the stage of imminent sliding and an early warning is required; When α i When the angle is ≈89°, the slope begins to slide.

2. The landslide disaster early warning method based on the LSTM-SARIMA hybrid data-driven model according to claim 1, characterized in that: In step S2, the calculation formula of the Hodrick-Prescott filter is as follows: y t =t t +c t y t is the observed original displacement data; τ t is the trend component, indicating the long-term trend; c t is a periodic fluctuation component; The optimization goal of the Hodrick-Prescott filter is to minimize the error of the following equation: where λ is a smoothing parameter that is used to balance the smoothness of the trend and the goodness of fit to the data.

3. The landslide disaster early warning method based on the LSTM-SARIMA hybrid data-driven model according to claim 2, characterized in that: The workflow of the Hodrick-Prescott filter is as follows: ① Decompose the original displacement data into trend part and periodic fluctuation part. ② Minimize an error function by adjusting the trend part to minimize the error between the original displacement data and the fitted value. ③Determine the smoothing parameter λ based on the characteristics of the data and the purpose of analysis.

Images