Geological disaster time sequence monitoring data adaptive fitting early warning method

Abstract

The present application provides a kind of geological disaster timing monitoring data adaptive fitting early warning method, by combing landslide timing deformation fluctuation characteristics, quantification definition observation error and false fluctuation characteristics, multiple filtering algorithm contrast test and optimization fitting parameter, construct sliding subsection window and realize real-time buffer filtering, deformation risk classification determination and processing result output five steps;Adopt fixed 30 point sliding window, subsection least square fitting after removing 10% extreme value, point by point iteration replacement correct timing monitoring data;With deformation rate, displacement increment as parameter switching trigger index, dynamic adaptation landslide creep, uniform speed slip, disaster acceleration instability different stage deformation characteristics.The present application solves the technical problems that traditional static fitting algorithm noise reduction degree is not adjustable, real-time is poor, prone to false alarm and missing report, can filter out random noise caused by field temperature difference, wind sand, electromagnetic interference, while retaining landslide disaster acceleration deformation characteristics.Heifangtai irrigation type loess landslide GNSS measured data set verification, algorithm light weight, can be deployed in field intelligent monitoring terminal, after correction timing data continuity, reliability is high, improve landslide disaster early warning accuracy.

Description

Technical Field

[0001] This invention belongs to the technical field of big data processing, and in particular relates to an adaptive fitting early warning method for geological disaster time-series monitoring data for use in geological engineering, slope safety monitoring and disaster prevention and mitigation. Background Technology

[0002] Landslide deformation monitoring is the core foundation for judging slope deformation trends, capturing early signs of instability, and making disaster prevention decisions. Currently, the industry generally uses GNSS and displacement sensing equipment to carry out all-weather slope monitoring. By recording the entire evolution process of landslides from slow creep and uniform deformation to accelerated instability before disaster through continuous time-series displacement data, it provides data support for hazard assessment, personnel evacuation, and disaster prevention and control. It is a core link in the proactive disaster prevention and early warning system.

[0003] However, the field monitoring environment is complex and harsh. Monitoring equipment operates exposed for extended periods, continuously affected by factors such as diurnal temperature variations, wind and sand erosion, electromagnetic interference, vibrations from surrounding construction sites, and equipment zero-point drift. This results in raw monitoring data containing a large amount of systematic errors, random noise, and irregular abnormal fluctuations. The mixing of false oscillation signals with real deformation data distorts the overall deformation trend of the slope. In scenarios requiring high accuracy and timeliness for disaster early warning, this can easily lead to false alarms and missed alarms in the early warning system, seriously threatening the lives and property of people in mountainous areas.

[0004] Currently, mainstream time-series data processing algorithms both domestically and internationally have significant technical shortcomings, making them unsuitable for the real-time early warning engineering needs of landslides: Traditional batch fitting algorithms require the accumulation of complete historical data for offline computation, resulting in high computational latency. They can only be used for post-event data review and cannot achieve real-time online data stream processing; Lightweight filtering algorithms meet real-time performance standards but have poor noise reduction accuracy, and the data fluctuations after correction remain significant, interfering with subsequent deformation rate calculations and risk assessments; Conventional least squares fitting is a static global fitting mode that severs the continuous correlation of time-series data, uses fixed parameters for rigid fitting, cannot dynamically adjust the noise reduction intensity, and struggles to balance noise removal with the preservation of pre-disaster accelerated deformation characteristics. This easily leads to problems such as over-smoothing masking of instability precursors and insufficient noise reduction causing data distortion.

[0005] In summary, existing time-series data processing technologies cannot simultaneously meet the three core requirements of landslide monitoring and early warning: real-time performance, high accuracy, and feature preservation. They also lack adaptive data processing solutions that can adapt to complex field environments and match the deformation patterns of landslides throughout their entire lifecycle. There is an urgent need to develop new adaptive fitting early warning methods to overcome the shortcomings of existing technologies. Summary of the Invention

[0006] This invention aims to overcome the multiple defects of traditional static fitting and fixed parameter filtering algorithms, and solve the technical pain points of the single least squares method being unable to distinguish the causes of temporal fluctuations, having poor real-time performance, and easily losing deformation features, and provides an adaptive fitting early warning method for geological disaster temporal monitoring data.

[0007] To achieve the above objectives, this invention designs a five-step integrated adaptive fitting early warning process, combining multi-algorithm comparison and optimization, sliding window iterative correction, and risk classification and discrimination technology to realize intelligent processing of landslide monitoring data throughout the entire process.

[0008] An adaptive fitting early warning method for geological disaster time-series monitoring data, specifically including: Step 1: Sorting out the temporal deformation and fluctuation characteristics of landslides. Integrating deformation data and original displacement time series data obtained from indoor tests and field measurements, the deformation characteristics of the slope are divided into three stages: slow creep, uniform sliding, and accelerated instability before disaster. The true deformation and false fluctuations are distinguished to provide experimental basis for error judgment and window construction. A 30-point fixed sliding window and a 10% extreme value removal preprocessing method are adopted. The preprocessing method is as follows: after removing 10% extreme values ​​from the data in the window, piecewise least squares fitting is performed. The theoretical correction displacement is calculated to replace the original measured displacement points. The corrected data is used as the input for the next round of window fitting. The real-time data stream is processed iteratively point by point to eliminate data deviation caused by differences in the environment at different times. The deformation trend of each time period is characterized by piecewise least squares fitting.

[0009] Step 2: Quantitatively define the characteristics of errors and false fluctuations. Collect data on various interferences caused by field operations. These interferences include abnormal manifestations such as instrument zero drift, temperature fluctuations, electromagnetic interference, and engineering vibrations. Record the duration, amplitude, and cycle of false fluctuations. Clarify the duration, amplitude, and cycle characteristics of false fluctuations. Delineate the distinction between the actual accelerated deformation of landslides and environmental noise. Establish quantitative discrimination criteria for effective deformation signals and interference data.

[0010] Step 3: Comparative experiments and adaptive parameter optimization of multiple filtering algorithms. Comparative experiments were conducted using multiple filtering algorithms, including moving average, piecewise least squares, Kalman filtering, and spline interpolation. The noise reduction performance of different algorithms was compared to select a dynamic parameter adjustment scheme for window length, smoothing weight, and noise reduction threshold that is suitable for the entire evolution cycle of landslides. Moving average, piecewise least squares, Kalman filtering, and spline interpolation were only used for parameter selection in the comparative experiments and were not involved in the actual data correction calculation.

[0011] Step four involves real-time buffering and filtering using a sliding segmented window. This is achieved by encrypting sampling using GNSS equipment, setting a fixed-length, point-by-point shifting data recognition interval, and implementing segmented iterative correction of the data stream using a fixed sliding window that shifts point by point. This process enables data acquisition, computation, and synchronous output, and allows for real-time online processing of the data stream through iterative correction.

[0012] Step 5: Deformation risk classification and result output. Based on the corrected high-precision time series data, extract the core indicators of deformation rate and displacement increment, compare with the quantitative discrimination standard, complete the slope state classification, and simultaneously output the noise-reduced and smoothed displacement time series curve and real-time risk level. The relevant results can be directly substituted into the landslide disaster early warning model.

[0013] In the above technical solution, the specific operation method of step one is as follows: 30 consecutive sets of monitoring data are used as the basic calculation window. The window is shifted backward by one monitoring point each time, and segmented data such as 1-30, 2-31, and 3-32 are extracted in sequence to reduce the data deviation caused by changes in environment and equipment conditions at different time periods. For all time-displacement measurement points in each segment, 10% of the extreme values ​​are removed first, and then the least squares method is used for linear fitting to obtain the segmented trend formula y=ax+b. This straight line reflects the overall deformation trend of the slope in the corresponding time period. Here, a and b are the coefficients obtained by dynamic solution through least squares fitting after removing 10% of the extreme values ​​in the corresponding sliding window. Different sliding windows correspond to a set of independent a and b coefficients.

[0014] In the above technical solution, step four iterative correction is as follows: after removing 10% of the extreme values ​​from the data within the window, piecewise least squares fitting is performed, the theoretical correction displacement is calculated to replace the original measured displacement points, and the corrected data is used as the input for the next round of window fitting, and the real-time data stream is processed iteratively point by point.

[0015] In the above technical solution, the iterative correction operation in step four is as follows: Take the horizontal coordinate x2 corresponding to the second time-series node in the current window, substitute x=x2 into the analytical expression y=ax+b of the current piecewise fitting line, calculate the theoretical correction displacement value y2, replace the measured displacement value corresponding to node x2 in the original monitoring data with this theoretically calculated value, and the replaced data will be used as the new starting point for the next set of data; after completing the single-point correction, slide backward one time-series node, extract all scattered points within the new window 2-31, remove the extreme values, and then replace the (x2, y2) values ​​to perform fitting again → take The process of value → replacement → refitting generates the second fitting analytical expression y=a2x+b2; then, the time series node x3 is extracted, substituted into the second fitting formula to solve for the correction displacement y3, the original measured data of x3 is replaced, and then (x3, y3) is used as the starting point to remove the extreme values ​​and then fit the next set of data to finally obtain y=a3x+b3; the entire set of operations of continuous sliding window truncation, removal of extreme values, piecewise least squares fitting, single-point numerical replacement and refitting correction is continuously iterated along the time series to process all real-time monitoring data point by point, and complete the adaptive smoothing filtering of the entire deformed time series.

[0016] In the above technical solution, the quantitative judgment criteria in step two are constructed as follows: the deformation rate and displacement increment are combined with the duration, amplitude, and cycle of false fluctuations for judgment; if the data fluctuation amplitude is large and the duration is short, and there is no continuous displacement increment, it is judged to be environmental interference noise; if the displacement continues to increase and the rate is stable, it is judged to be stable creep; if the rate continues to rise significantly and the displacement increment increases rapidly, it is judged to be accelerated sliding; each judgment threshold can be pre-configured according to the landslide type and on-site working conditions.

[0017] In the above technical solution, the switching logic of the dynamic parameter adjustment scheme is as follows: the deformation rate and displacement increment are used as the parameter switching trigger indicators; when the monitoring data determines that the slope is in the creep stage, the noise reduction weight is increased; when the deformation rate increases and the slope is determined to enter the pre-disaster acceleration stage, the smoothing noise reduction intensity is reduced; the judgment thresholds for creep and acceleration stages can be pre-configured according to different landslide engineering sites.

[0018] The core iterative correction logic of this invention is as follows: based on the sliding window segmented fitting data, the theoretical correction value is calculated point by point and the original measured data is replaced. The next round of fitting is carried out iteratively based on the corrected data, and the adaptive smoothing of the full time series data is realized in a loop to ensure the continuity and authenticity of the data correction.

[0019] Beneficial effects: This invention is applicable to irrigation-type loess landslides and sudden slope instability scenarios. It is adapted to complex all-weather monitoring environments in the field and can be deployed independently on GNSS field intelligent monitoring terminals. The actual data correction calculation adopts a piecewise least squares iterative correction process to remove extreme values. Moving average, Kalman filtering, and spline interpolation are only used as control test algorithms to compare and select the optimal parameters and do not participate in the actual data correction calculation. It can accurately distinguish between false data fluctuations caused by field environmental interference and the real sliding deformation of landslides before disasters, realize real-time noise reduction and adaptive optimization of monitoring data, and provide high-precision, high-continuity, and high-reliability time-series data support for landslide disaster risk assessment and emergency response, effectively avoiding the risks of false alarms and missed alarms in early warning. Attached Figure Description

[0020] Figure 1 This is a scatter plot of the original displacement-time of the present invention; Figure 2 This is a displacement-time graph of the present invention after removing 10% of the extreme values; Figure 3 This is a displacement-time diagram of the present invention for removing 10% extreme values ​​and single-point replacement correction; Figure 4 A schematic diagram of the original data introduced for fitting the straight line in this invention; Figure 5 This is a schematic diagram showing the positions of the original measuring points and the fitted line in this invention; Figure 6The original image and the schematic diagram of single-point replacement are introduced to illustrate the fitting of the straight line in this invention; Figure 7 This is a schematic diagram of the displacement time of the second sliding window interval of the present invention; Figure 8 This is a schematic diagram of the third sliding interval and the displacement time for removing extreme values ​​in this invention; Figure 9 This is a schematic diagram of the third sliding interval for removing extrema and fitting a straight line in this invention; Figure 10 This is a schematic diagram comparing displacement and time for different window lengths according to the present invention; Figure 11 This is a displacement-time diagram comparing different extreme value removal ratios of the present invention; Figure 12 This is a graph showing the original rate-time curve of Embodiment 1 of the present invention; Figure 13 This is a rate-time curve diagram of Embodiment 1 of the present invention after correction according to the present invention; Figure 14 This is a comparison of displacement-time curves for multiple algorithms in this invention; Figure 15 This is a curve comparing different window lengths in Embodiment 1 of the present invention; Figure 16 This is a graph showing the original rate-time curve of Embodiment 2 of the present invention; Figure 17 This is a rate-time curve diagram of Embodiment 2 of the present invention after correction according to the present invention; Figure 18 This is a comparison of displacement-time curves using multiple algorithms in Embodiment 2 of the present invention; Figure 19 Rate comparison curve of the present invention Figure 1 ; Figure 20 Rate comparison curve of the present invention Figure 2 ; Figure 21 Rate comparison curve of the present invention Figure 3 ; Figure 22 Rate comparison curve of the present invention Figure 4 . Detailed Implementation

[0021] The technical solution of this invention will be further described clearly and completely below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0022] All implementation cases use GNSS field measurement data of the Heifangtai irrigation type loess landslide in Yongjing County, Linxia Prefecture, Gansu Province. The loess in this area is highly collapsible and water-sensitive, and the landslides are prone to sudden occurrence, making it suitable for verifying the performance of the algorithm of this invention.

[0023] The performance evaluation metrics involved in this embodiment are defined as follows: Noise Residual (RMSE): RMSE = , These are the corrected calculated values. These are the original measured values. This represents the number of sampling points.

[0024] Noise suppression rate: Noise suppression rate =

[0025] Represents the standard deviation of the original data; Deformation preservation integrity: Deformation preservation integrity =

[0026] Example 1 In this embodiment, the first landslide monitoring point in the Heifangtai area was selected, and a total of 30 sets of GNSS measured displacement-time data were obtained, and the complete processing flow was followed.

[0027] This embodiment uses the Gansu Heifangtai Dangchuan No. 4 landslide as an application example. Through statistical analysis of 120 sets of monitoring data during the stable period before landslide deformation (August 1 to August 5, 2017), the data anomalies caused by various environmental disturbances were quantitatively defined. The statistics show that instrument zero-point drift manifests as a slow, unidirectional baseline shift of approximately 0.05 to 0.10 mm, positively correlated with daily temperature changes, with a cycle of 24 hours; temperature fluctuation interference manifests as periodic sinusoidal fluctuations with an amplitude of approximately 0.08 to 0.15 mm, a phase lag of approximately 2 to 4 hours behind temperature changes, and a cycle of 24 hours; electromagnetic interference manifests as spike pulses lasting 1 to 3 seconds, with amplitude abrupt changes reaching 5 to 10 times that of normal signals; and engineering vibration interference manifests as decaying oscillations lasting 5 to 30 seconds, consistent with surrounding construction times. Based on the above statistical characteristics, the following discrimination thresholds are established: In the time dimension, low-amplitude fluctuations lasting more than 10 minutes and synchronized with temperature are classified as temperature drift interference, and amplitude abrupt changes lasting less than 1 minute are classified as electromagnetic or vibration interference; in the amplitude dimension, the standard deviation of a normal signal is approximately 0.04 mm, and single jumps exceeding 0.12 mm but recovering rapidly are judged as interference, while continuous accumulation without recovery is judged as effective deformation; in the frequency dimension, with daily as the cutoff frequency, high-frequency and 24-hour cycle energy components are removed, and low-frequency continuous accumulation components are retained as deformation signals. The distinction between real accelerated deformation and environmental noise is as follows: real deformation shows a unidirectional increasing trend, cumulative displacement increases monotonically, and multiple measuring points change synchronously in the same direction, conforming to the laws of landslide mechanics; environmental noise fluctuates around the mean, the cumulative net displacement is zero, different measuring points are unrelated, and energy is concentrated in the 24-hour cycle or high-frequency range. Based on the aforementioned distinction criteria, this embodiment determines that the cumulative displacement increase from August 1st to 20th was approximately 3.5 mm, but the changes at multiple measuring points were inconsistent, and some fluctuations coincided with the daily temperature variation cycle, thus classifying it as creep stage superimposed with environmental noise; the cumulative displacement increase from September 25th to October 1st reached 38.6 mm, with multiple measuring points changing synchronously in the same direction and the rate increasing daily, thus classifying it as pre-disaster accelerated real deformation. Combining the above statistical results and distinction criteria, a quantitative discrimination standard for effective deformation signals and interference data is established. The above thresholds and weights can be pre-configured according to the monitoring accuracy, geological conditions, and environmental interference levels at different landslide engineering sites.

[0028] A total of 30 sets of GNSS measured displacement-time data were obtained in this process, as follows: GNSS field measurement data from monitoring point Dangchuan 4, Heifangtai, Yongjing County, Linxia Prefecture, Gansu Province, with a sampling interval of 1 day; the first 8 representative original time series (time, displacement mm) are extracted: (2017 / 5 / 25, 211.11), (2017 / 5 / 26, 211.21), (2017 / 5 / 27, 211.62), (2017 / 5 / 28, 212.17), (2017 / 5 / 29, 212.37), (2017 / 5 / 30, 216.38), (2017 / 5 / 31, 217.93), (2017 / 6 / 1, 219.83)]; the complete dataset contains a total of 30 sets.

[0029] from Figure 10 The data shows that a window that is too small significantly affects the noise reduction effect. When the window size is 10, the two peaks (285 and 300) at points 20 and 40 in the original data are completely preserved in the fitted curve, showing a clear bimodal jump phenomenon. This indicates that the weight of a single outlier on the fitting result is too large, and environmental interference cannot be effectively filtered out. When the window size is 20, the peak at point 20 still remains, and only point 40 is smoothed. When the window size is 30, only a small amount of the peak at point 40 remains, and most abnormal fluctuations have been effectively suppressed. When the window size is 50, both peaks are completely smoothed, and the curve shows a monotonically increasing trend, but this may be due to over-smoothing of the true deformation. Taking this set of data as an example, the noise residual is about 3.2 mm when the window size is 10, it decreases to about 1.6 mm when the window size is 30, and although it is further reduced when the window size is 50, the integrity of deformation preservation decreases significantly. Therefore, the window selection needs to strike a balance between noise reduction and shape preservation. A window that is too small cannot effectively filter out noise and will lead to false alarms, while a window that is too large will mask the true deformation and lead to missed alarms. Through multiple sets of comparative experiments, the 30-point window showed the best overall effect in this embodiment.

[0030] like Figure 1 , Figure 2 As shown, in step one, the temporal deformation fluctuation characteristics of the landslide are sorted out. 30 sets of monitoring data are used as the basic calculation window. The window is shifted point by point backward, and segments 1-30, 2-31, and 3-32 are extracted in sequence. 10% of the extreme values ​​of the time-displacement measurement points in the first segment 1-30 are removed. The least squares method is used for linear fitting to obtain the initial segmented trend analysis formula: y=0.3515x-14860.

[0031] Step 2: Quantify and define the characteristics of observation errors and false fluctuations. Statistically analyze the data jump amplitude and duration caused by environmental disturbances at the location. Combine deformation rate, displacement increment, duration of false fluctuations, and amplitude of fluctuations to establish a discrimination criterion that distinguishes between environmental noise and real landslide deformation. The discrimination threshold is pre-configured based on the engineering characteristics of loess landslides in the area.

[0032] Step 3: Comparative experiment of multiple filtering algorithms and optimization of fitting parameters: Comparative experiments were conducted on moving average, piecewise least squares, Kalman filtering, and spline interpolation to compare the noise reduction effects of each algorithm. Deformation rate and displacement increment were used as the trigger indicators for parameter switching. The noise reduction weight was increased for the creep stage of loess landslides, and the deformation retention capacity was reserved for the accelerated stage of disaster. A parameter scheme with a 30-point window and 10% extreme values ​​were determined.

[0033] The selection criteria for the parameter scheme to remove 10% of extreme values ​​are as follows: Figure 2 The data shows that both excessively small and excessively large outlier removal ratios affect the processing results. When the removal ratio is 5%, peaks such as the 45th (285), 50th (290), 55th (295), and 60th (300) points in the original data are still visible in the fitted curve, indicating that the removal intensity is insufficient and outliers still interfere with the fitting results. When the removal ratio is 10%, the peaks are basically effectively suppressed, the curve smoothness is significantly improved, and the overall trend remains good. When the removal ratio is 15%, the curve is further smoothed but the trend begins to deviate, and some normal data points are mistakenly removed, resulting in a lower fitted straight line. When the removal ratio is 20%, the curve is excessively smoothed and the overall trend shifts significantly downward, resulting in a large deviation from the original data magnitude. Taking this set of data as an example, the noise residual is approximately 2.1 mm at a 5% rejection rate, decreasing to approximately 1.6 mm at 10%, and further decreasing to approximately 1.4 mm at 15%, but the deformation retention rate drops to approximately 72%, and at 20%, the deformation retention rate drops to approximately 65%. Therefore, the rejection rate needs to strike a balance between noise suppression and deformation retention. Rejecting too few peaks will not effectively filter out abnormal spikes, while rejecting too many peaks will mistakenly delete normal deformation information, leading to trend distortion. Through multiple comparative experiments, a 10% rejection rate showed the best overall performance in this embodiment.

[0034] like Figures 3-10 As shown, step four, iterative correction using a sliding piecewise window: ① Take the x-coordinate of the second time-series node in the window, x=x2, and substitute x=x2 into the first fitting analytical expression y=0.3515x-14860. Calculate the theoretical correction displacement y2=211.3255039, and replace the original measured point with (x2,y2). ② Slide the window back one time point to obtain a new window 2-31. Using the corrected (x2, y2) as the new starting point, perform a least squares linear fitting again on all the corrected scattered points in window 2-31 after removing 10% of the extreme values, to obtain the second fitting analytical expression y=0.0879x-3557.2. ③ Extract the time sequence node x3 and substitute it into the second fitting formula to calculate y3=211.2623, determine the correction point (x3,y3) and replace the measured displacement corresponding to the original x3; ④ Continue sliding the window backward, fitting the scattered points within window 3-32 to obtain the third analytical expression y=0.0049xb,..., repeating the cycle of "window sliding - removing 10% extreme values ​​- least squares fitting - single-point replacement correction", iteratively processing all 30 sets of time series data point by point. Each iteration of the fitting uses the corrected points from the previous step as the benchmark to ensure the continuity and consistency of the time series.

[0035] Step 5: Deformation Risk Classification and Processing Result Output; Based on the deformation rate and displacement increment, combined with the duration and amplitude of spurious fluctuations, a comprehensive judgment is made; the deformation rate and displacement increment indices of the corrected data are extracted to determine that the monitoring point is in a stable creep state; noise-reduced and smoothed displacement time series and velocity time series curves are output to obtain the original velocity. Figure 11 Corrected rate Figure 12 The output results are directly fed into the landslide disaster early warning model.

[0036] contrast Figure 11 , Figure 12 As can be seen, after the above-mentioned sliding window point-by-point iterative correction, the original data obtained is compared with the data graph processed by this method and the data under other processing methods. It can be seen that the original monitoring data contains a large amount of high-frequency noise and sharp peaks. After the adaptive fitting processing of this invention, abnormal peaks are filtered out, the curve is continuous and smooth, and the slight deformation trend of the slope is preserved, reducing the risk of false alarms.

[0037] After processing using this method, the sum of squared residuals between the original data and the processed results is 8904.70, with 90 sampling points. According to the root mean square error formula RMSE = The calculated noise residual is (8904.70 / 90) = 9.95 mm. After processing with the moving average method, the sum of squared residuals is 80086.14. Substituting this into the formula, the noise residual is (80086.14 / 90) = 29.83 mm. The standard deviation of the original data is 44.10 mm. According to the noise suppression rate formula, the noise suppression rate of this method is (44.10-9.95) / 44.10×100% = 77.44%, and the moving average method is (44.10-29.83) / 44.10×100% = 32.36%. Regarding the integrity of deformation retention, the accelerated segments (numbers 51 to 90) were selected for calculation. The calculation was performed using the formula (cumulative increment of the accelerated segment after processing / cumulative increment of the original accelerated segment) × 100%. The original cumulative increment was 366.23 - 272.67 = 93.56 mm. After processing using this method, the cumulative increment was 371.44 - 281.59 = 89.85 mm, resulting in an integrity of 89.85 / 93.56 × 100% = 96.04%. After processing using the moving average method, the cumulative increment was 421.01 - 300.19 = 120.82 mm, resulting in an integrity of 120.82 / 93.56 × 100% = 129.14%. Based on the above calculation results, the noise residual of this method is lower than that of the moving average method, the noise suppression rate is about 45 percentage points higher, and the deformation preservation in the acceleration phase is 96.04%, which is within a reasonable range. In contrast, the moving average method shows an amplification of deformation amplitude in the acceleration phase (129.14%), indicating that this method achieves a better balance between noise reduction and deformation preservation than the traditional moving average method.

[0038] Example 2 In this embodiment, another landslide monitoring point in the Heifangtai area was selected to obtain 30 sets of GNSS measured displacement-time data and execute the complete process of this invention.

[0039] GNSS field measurement data from the JJ7 monitoring point at Heifangtai Station, Yongjing County, Linxia Prefecture, Gansu Province, with a sampling interval of 1 day; the first 8 representative raw time series (time, displacement mm) are extracted: (2015 / 8 / 10, 2.585455), (2015 / 8 / 11, 2.910952), (2015 / 8 / 12, 2.49771), (2015 / 8 / 1, 2.342723), (2015 / 8 / 14, 2.96962), (2015 / 8 / 15, 3.397842), (2015 / 8 / 16, 2.908897), (2015 / 8 / 17, 3.2904778); the complete dataset contains a total of 30 sets. The reason for selecting a window size of 30 is as follows: Both excessively short and excessively long window lengths negatively impact processing effectiveness. With a window size of 10, fluctuations such as the 30th point (40.0), 35th point (40.5), and 40th point (41.0) in the original data still show significant residuals in the fitted curve, indicating large curve fluctuations. This suggests that when the window is too short, individual outliers have an excessively large impact on the fitting result, failing to effectively filter out environmental interference. With a window size of 20, the fluctuation amplitude decreases, but some peaks remain visible. With a window size of 30, the curve smoothness improves significantly, most abnormal fluctuations are effectively suppressed, and the overall trend remains consistent with the original data. With a window size of 50, the curve becomes overly smoothed, flattening the normal upward trend in the original data and masking some true deformation information. Taking this set of data as an example, with a window size of 10, the noise residual is approximately 0.8 mm; with a window size of 30, it decreases to approximately 0.3 mm; and with a window size of 50, although it further decreases to approximately 0.2 mm, the deformation retention rate drops to approximately 70%. Therefore, the window length needs to strike a balance between noise reduction and shape preservation. If the window is too small, it cannot effectively filter out noise; if the window is too large, it will mask the actual deformation. After multiple sets of comparative experiments, the 30-point window showed the best overall performance in this embodiment.

[0040] First, a fixed sliding window of 30 points is set, and points are sequentially truncated in the order of 1-30, 2-31, 3-32, etc., to eliminate the interference from environmental factors and equipment status changes at different time periods through a segmented isolation strategy. Then, the extreme values ​​are removed first, followed by the least squares method to fit a straight line to all scattered points within the first fixed time period, and the analytical expression of the fitted line is obtained: y = 1.6204x - 68422.

[0041] Iterative correction steps: ① Take the x-coordinate x2 of the second time-series node in the window and substitute it into the initial analytical expression y=1.6204x-68422 to calculate the theoretical correction displacement y2 and replace the original measured point; ② Slide the window back one point to get window 2-31. Starting from the corrected (x2,y2), remove 10% of the extreme values ​​and refit to get the second fitting analytical expression: y=3.7019x-156314; ③ Extract the time sequence node x3 and substitute it into the second formula to solve y3. Determine the correction point (x3, y3) and replace the corresponding measured value in the original data. Continue to slide backward. Remove the extreme values ​​of the scattered points in window 3-32 and then fit the third analytical expression with (x3, y3) as the starting point: y=2.9271x-123599. ④ Continuously cycle through the iterative process of "removing extrema - fitting - calculating - replacing - sliding" to complete the correction of all time-series data until the entire data sequence is covered. Each iteration is based on the data corrected in the previous iteration to fit the next segment, ensuring the continuity and consistency of the correction process.

[0042] The measured GNSS data were corrected using a sliding window iterative correction method, as shown in the example below. Figure 13 , Figure 14 It is easy to see from the correction of the rate curves before and after that the original field environmental data has large negative abrupt changes. After processing by the method of this invention, the high-frequency noise and data abrupt changes caused by these changes are effectively filtered out, resulting in a smoother fitting curve with smaller fluctuations. This processing method can completely preserve the true deformation development characteristics of loess landslides, greatly reduce the risk of false alarms caused by environmental interference, and significantly improve the accuracy and stability of deformation early warning in the pre-disaster stage.

[0043] After processing using this method, the sum of squares of the residuals at each point in the original data and the processed results is 249.74, with 90 sampling points. The noise residual calculated using the root mean square error formula is 1.67 mm. After processing using the translation-fill method, the sum of squares of the residuals is 526.36, and the calculated noise residual is 2.42 mm. The noise residual of this method is 0.75 mm lower than that of the translation-fill method, a reduction of approximately 31%. The standard deviation of the original data is 2.31 mm. The noise suppression rate of this method is 27.71%, while the noise suppression rate of the translation-fill method is negative 4.76%, indicating that the translation-fill method failed to effectively suppress noise and instead amplified the data deviation. Regarding deformation preservation, 30 sampling points were selected during the acceleration phase for evaluation. The cumulative displacement increment of the original data was 3.55 mm. After processing by this method, the cumulative displacement increment was 2.76 mm, resulting in a deformation preservation rate of 77.75%. The cumulative displacement increment after processing by the translation-filling method was 4.59 mm, with a deformation preservation rate of 129.25%, indicating that the translation-filling method amplified the actual deformation amplitude during the acceleration phase. In summary, this method achieves a better balance between noise suppression and deformation preservation than the translation-filling method.

[0044] Simulates pure environmental interference noise conditions: Simulated time series data is constructed: the time series exhibits short-term, large-amplitude jumps, but no sustained displacement increments; the data fluctuation amplitude is large, with a single sudden change lasting 2-4 sampling periods, and the displacement returns to the original baseline after the fluctuation ends. This simulated time series is input into the complete algorithm flow of this invention; after judgment: the fluctuation duration is short, there is no sustained unidirectional displacement increment, and the comprehensive risk level is determined to be environmental interference noise, and a corresponding risk warning is output.

[0045] Simulated accelerated sliding conditions before a disaster: The algorithm constructs a simulated time series of accelerated sliding before a disaster: the deformation rate continuously increases, and the displacement increment increases rapidly and continuously. Using deformation rate and displacement increment as trigger indicators, the algorithm identifies the entry into the accelerated disaster phase and automatically reduces the noise reduction weight from 0.7-0.9 in the creep phase range to 0.2-0.4. The algorithm preserves the characteristics of accelerated deformation without excessive smoothing. Based on comprehensive indicators, the risk level is determined to be accelerated sliding, and a disaster warning level is output.

[0046] For the same set of original monitoring data from Examples 1 and 2, the traditional global least squares method was used for fitting, resulting in... Figures 19-22 .

[0047] The processing result of the technical solution of this invention is as follows: Figure 15 , Figure 17 ,pass Figures 15 to 22 The comparison shows that, for a long time, the least squares method, with its advantages of simple operation and fast calculation speed, has been widely used in the basic fitting analysis of various slope monitoring data. It can integrate and process scattered monitoring data and roughly show the overall trend of soil deformation over time. However, combined with the actual field monitoring conditions in this study area, it can be found that the field operation environment of Heifangtai is harsh, with many uncertain factors such as wind and sand, temperature difference, and interference from field electromagnetic signals. At the same time, equipment observation and manual reading will produce small random errors, resulting in the acquisition of raw GNSS data containing a large amount of random noise and abnormal mutation points. Since the least squares method is a basic linear fitting method, it does not have the data optimization capabilities of noise reduction and error elimination, and cannot handle high-frequency interference data generated by complex field environments. This also makes the displacement-time curve obtained by the final fitting chaotic, with serious numerical jumps, poor overall curve smoothness, and unable to accurately reflect the slow and subtle deformation process in the early stage of landslide formation. It is difficult to capture the weak deformation signals in the early stage of slope, which is not conducive to accurately analyzing the early deformation law of loess landslides.

[0048] To address the shortcomings of the least squares method in processing landslide data, this paper employs a novel, optimized data processing method to fit the same set of original monitoring data. A comparison of the fitting curves from the two different methods clearly demonstrates the significant advantages of the improved method. The displacement time-series curve optimized by the new method effectively filters out random noise and anomalous changes in the monitoring data, greatly reducing data errors caused by harsh field conditions and equipment observation biases. This results in more continuous and stable temporal changes in displacement data, thoroughly improving the problems of large fluctuations and frequent jumps in traditional fitting curves. Simultaneously, this optimized algorithm, while achieving data smoothing and noise reduction, preserves the true deformation trend and deformation characteristics of the slope at different stages to the greatest extent possible. It avoids the shortcomings of traditional methods, such as susceptibility to abnormal data interference, distorted fitting results, and ambiguous deformation patterns, thus realistically reproducing the complete deformation process of the sudden loess landslide in the study area, showing its gradual development and dynamic changes. This data processing method is more suitable for the unique development characteristics of early micro-deformation and rapid instability before landslides in irrigated loess landslides, and can provide more accurate and reliable measured data support for subsequent in-depth analysis of landslide deformation mechanisms, definition of disaster evolution stages, and construction of landslide early warning models.

[0049] Therefore, the method of the present invention not only filters out environmental noise but also completely preserves the deformation characteristics of landslides at different evolution stages, proving that the present invention has the following significant technical advancements: (1) Adaptive dynamic adjustment, balancing noise reduction and feature preservation. Abandoning the traditional fixed parameter fitting mode, the deformation rate and displacement increment are used as trigger indicators. The noise reduction intensity is dynamically adjusted according to the deformation characteristics of different stages of landslide creep and accelerated instability. This can effectively filter out random noise and data jumps caused by wind and sand, temperature difference, electromagnetic interference, and equipment drift, while fully preserving the severe deformation characteristics of the disaster stage. This avoids excessive smoothing that masks the signs of instability and greatly reduces the probability of false alarms and missed alarms.

[0050] (2) Real-time online processing eliminates early warning lag. The sliding window point-by-point iterative buffer operation mode is adopted, which does not require the accumulation of complete historical data. It can realize the synchronous completion of monitoring data collection, correction and output. It is compatible with encrypted sampling data streams of field equipment. The algorithm is lightweight and has high computing efficiency. It can be directly deployed on field intelligent monitoring terminals to meet the high timeliness requirements of early warning of sudden landslide disasters.

[0051] (3) Quantitative judgment criteria and accurate risk assessment. Based on massive field measurement data, a quantitative distinction criterion for real deformation and interference noise is established by comprehensively considering deformation rate, displacement increment, duration and amplitude of false fluctuations. The threshold can be adapted to different landslide site configurations, getting rid of the drawbacks of subjective fitting of traditional algorithms. The data correction accuracy is high and the stability is strong, which can accurately restore the entire evolution process of landslide from weak creep to instantaneous instability.

[0052] (4) It has strong engineering adaptability and high promotion value. It is specifically adapted to the development characteristics of loess landslides, such as strong collapsibility, high water sensitivity and sudden instability. After verification by measured datasets, the continuity and stability of the corrected curves are significantly improved. It can provide standardized and high-precision data support for landslide mechanism analysis, stage division and early warning model construction. It is suitable for various slope geological disaster monitoring and early warning scenarios.

[0053] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive fitting and early warning method for geological disaster time-series monitoring data, characterized in that: include: Step 1: Sorting out the temporal deformation and fluctuation characteristics of landslides. Integrating deformation data and original displacement time series data obtained from indoor tests and field measurements, the deformation characteristics of the slope are divided into three stages: slow creep, uniform sliding, and accelerated instability before disaster. The real deformation and false fluctuations are distinguished to provide experimental basis for error judgment and window construction. A 30-point fixed sliding window and a 10% extreme value removal preprocessing method are adopted to eliminate data deviation caused by environmental differences in time periods. The deformation trend of each time period is characterized by piecewise least squares fitting. Step 2: Quantitatively define the characteristics of errors and false fluctuations, statistically analyze the data caused by various interferences in field operations, including abnormal manifestations of instrument zero drift, temperature fluctuations, electromagnetic interference, and engineering vibrations, record the duration, amplitude, and cycle of false fluctuations, clarify the duration, amplitude, and cycle characteristics of false fluctuations, delineate the distinction between the actual accelerated deformation of landslides and environmental noise, and establish quantitative discrimination criteria for effective deformation signals and interference data; Step 3: Comparative experiments and adaptive parameter optimization of multiple filtering algorithms. Comparative experiments were conducted using multiple filtering algorithms, including moving average, piecewise least squares, Kalman filtering, and spline interpolation, to compare the noise reduction performance of different algorithms and select a dynamic parameter adjustment scheme for window length, smoothing weight, and noise reduction threshold that is suitable for the entire evolution cycle of landslides. The moving average, Kalman filtering, and spline interpolation algorithms were only used as comparative experimental algorithms for parameter selection and were not used in the actual monitoring data correction calculation. Step 4: Real-time buffering and filtering of the sliding segmented window, relying on GNSS equipment for encrypted sampling, setting a fixed-length, point-by-point shifting data recognition interval, using a fixed sliding window that shifts point by point, implementing segmented iterative correction of the data stream, realizing data acquisition, processing, and synchronous output, and achieving real-time online processing of the data stream through iterative correction mode; Step 5: Deformation risk classification and result output. Based on the corrected high-precision time series data, extract the core indicators of deformation rate and displacement increment, compare with the quantitative discrimination standard, complete the slope state classification, and simultaneously output the noise-reduced and smoothed displacement time series curve and real-time risk level. The relevant results can be directly substituted into the landslide disaster early warning model.

2. The adaptive fitting and early warning method for geological disaster time-series monitoring data according to claim 1, characterized in that: The specific operation method of step one is as follows: 30 consecutive sets of monitoring data are used as the basic calculation window. The window is shifted backward by one monitoring point each time to extract segmented data in sequence, thereby reducing the data deviation caused by changes in environment and equipment conditions at different times. For all time-displacement measurement points in each segment, the extreme values ​​of 10% are removed first, and then the least squares method is used for linear fitting to obtain the segmented trend formula y=ax+b. This straight line reflects the overall deformation trend of the slope in the corresponding time period. Here, a and b are the coefficients obtained by dynamic solution through least squares fitting after removing the extreme values ​​of 10% in the corresponding sliding window. Different sliding windows correspond to a set of independent a and b coefficients.

3. The adaptive fitting and early warning method for geological disaster time-series monitoring data according to claim 1, characterized in that: The quantitative judgment criteria in step two are constructed as follows: the deformation rate and displacement increment are combined with the duration, amplitude, and cycle period of false fluctuations for judgment; if the data fluctuation amplitude is large and the duration is short, and there is no continuous displacement increment, it is judged to be environmental interference noise; if the displacement continues to increase and the rate is stable, it is judged to be stable creep; if the rate continues to rise significantly and the displacement increment increases rapidly, it is judged to be accelerated sliding. Each judgment threshold can be pre-configured according to the landslide type and on-site conditions.

4. The adaptive fitting and early warning method for geological disaster time-series monitoring data according to claim 1, characterized in that: The fourth step of iterative correction is as follows: after removing 10% of the extreme values ​​from the data within the window, the data is fitted piecewise by least squares. The theoretical correction displacement is calculated to replace the original measured displacement points. The corrected data is used as the input for the next round of window fitting. The real-time data stream is processed iteratively point by point. The iterative correction operation process is as follows: take the x-coordinate x2 corresponding to the second time-series node of the current window, substitute x=x2 into the analytical expression y=ax+b of the current piecewise fitting line, calculate the theoretical correction displacement value y2, and replace the measured displacement value corresponding to the x2 node in the original monitoring data with this theoretical calculation value. The replaced data will be used as the new starting point for the next set of data. After completing the single-point correction, slide one time-series node backward, extract all scattered points in the new window, remove extreme values, replace the (x2, y2) value, and repeat the process of fitting, value extraction, replacement, and refitting to generate the second fitting analytical expression y=a2x+b2. Then, extract the time-series node x3, substitute it into the second fitting formula to solve for the correction displacement y3, replace the original measured data of x3, and then use (x3, y3) as the starting point to remove extreme values ​​and fit the next set of data to finally obtain y=a3x+b3. Continuously loop the entire set of operations of sliding window extraction, extreme value removal, piecewise least squares fitting, single-point value replacement, and refitting correction, iterating along the time sequence point by point to process all real-time monitoring data, and completing the adaptive smoothing filtering of the entire deformed time-series sequence.

5. The adaptive fitting and early warning method for geological disaster time-series monitoring data according to claim 4, characterized in that: The switching logic of the dynamic parameter adjustment scheme is as follows: the deformation rate and displacement increment are used as the parameter switching trigger indicators; when the monitoring data determines that the slope is in the creep stage, the noise reduction weight is increased. When the deformation rate increases and the slope is determined to be entering the accelerated disaster-prone stage, the smoothing and noise reduction efforts should be reduced. The thresholds for determining creep and acceleration stages can be pre-configured according to different landslide engineering sites.

Images