A method and system for predicting the residual life of a slope

Abstract

The application provides a kind of prediction method and system of slope residual life, and the technical scheme adopted by the application is: converting real-time displacement of slope into velocity sequence, constructing triple joint criterion based on time-frequency domain characteristics to automatically lock displacement acceleration starting point;Introduce incremental trend checking mechanism to filter non-monotonic fluctuation interference, and construct incremental observation data set conforming to the law of monotonic acceleration;Subsequently, input the data set into the phenomenological model under the layered Bayesian framework, and use Markov chain Monte Carlo algorithm to update model parameters on-line, so as to output residual life probability distribution in real time and trigger graded warning signal.The application can solve the technical problems that the identification of slope acceleration starting point is easy to be disturbed by noise, the initial prediction of gradually accumulated monitoring data is easy to diverge, and there is lack of dynamic updating mechanism conforming to physical monotonic characteristics.

Description

Technical Field

[0001] This invention relates to the field of slope safety monitoring technology, specifically to a method and system for predicting the remaining life of a slope. Background Technology

[0002] Slope instability is a major challenge in disaster prevention and mitigation. Currently, prediction methods for slope instability mainly include physics-based numerical simulation, data-driven machine learning, and phenomenological model prediction based on displacement monitoring. Among these, phenomenological models are widely used in early warning of potential landslides in on-site slope engineering due to their advantages such as simple computation and easy online real-time parameter updates.

[0003] However, existing phenomenological models tend to focus on "post-hoc static verification," that is, using corrupted complete data sequences for back-calculation and accuracy assessment, which cannot meet the needs of forward-looking, on-site, dynamic early warning in real-time monitoring scenarios. In actual online monitoring data processing, achieving forward-looking prediction faces the following technical bottlenecks: First, the accuracy of identifying the onset of accelerated slope deformation directly determines the timing of subsequent prediction initiation. Traditional moving average methods or single threshold criteria are easily affected by outliers in measurements and local environmental noise, and their generalization ability is insufficient when faced with complex disturbances, which can easily lead to false alarms or missed alarms in the early warning system.

[0004] Secondly, existing prediction mechanisms fail to simulate the dynamic physical process of data accumulation during actual monitoring. In the early stages of accelerated deformation, there are very few available monitoring data points, and the calibration of model parameters is highly uncertain. Traditional deterministic fitting methods are prone to computational divergence when data is insufficient, making it impossible to reasonably define the range of uncertainty. Furthermore, when the velocity data collected on-site experiences non-monotonic, transient fluctuations or declines due to signal disturbances, it easily disrupts the mathematical boundaries of the phenomenological model, leading to a sharp drop in extrapolation accuracy.

[0005] Finally, existing prediction methods mostly output absolute destruction time and lack a mechanism to intuitively map uncertain prediction results to the on-site response window for disaster prevention. Decision-makers still need to calculate the emergency response time themselves, which reduces the operability of early warning plans.

[0006] Therefore, how to achieve robust online identification of slope acceleration initiation points, overcome the instability of predictions caused by uncertainties in the early stages of data accumulation, and establish a dynamic remaining lifetime probability prediction mechanism that conforms to the physical monotonic acceleration law is an urgent problem to be solved in this field. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a method and system for predicting the remaining life of slopes, thereby solving the technical problems in existing technologies, such as the susceptibility of slope acceleration initiation point identification to noise interference, the tendency for early predictions to diverge due to the gradual accumulation of monitoring data, and the lack of a dynamic update mechanism that conforms to the monotonic characteristics of physics.

[0008] The technical solution adopted in this invention is as follows: Real-time displacement sequences of the slope are acquired and converted into instantaneous velocity sequences; a triple joint criterion is constructed by fusing the velocity statistical features of the time-domain cumulative window and the detail coefficient features of the frequency-domain multi-scale discrete wavelet to achieve robust identification of the acceleration start point; after identifying the acceleration start point, an incremental trend checking algorithm based on a counting mechanism is introduced to skip perturbation data that does not satisfy the monotonic acceleration feature, constructing an incremental observation dataset that conforms to physical laws; the dataset is input into a phenomenological model configured in a hierarchical Bayesian framework, and online posterior sampling is performed using a Markov chain Monte Carlo algorithm to update model parameters in real time, outputting the posterior probability distribution of the remaining lifespan of the slope; based on the comparison result between the posterior probability distribution of the remaining lifespan and a preset risk threshold, a multi-level landslide warning signal is triggered.

[0009] In the first feasible method, a method for predicting the remaining life of a slope is provided, comprising the following steps: Real-time displacement sequences of the slope are obtained using displacement sensors; The real-time displacement sequence is differentially processed to convert it into an instantaneous velocity sequence; The instantaneous velocity sequence is analyzed to identify the acceleration start point and determine whether the slope deformation has entered the acceleration stage. In response to the slope deformation entering the acceleration stage, the corresponding acceleration start time is determined as the acceleration start point, and the starting velocity corresponding to the acceleration start point is output, as well as the real-time velocity sequence from the acceleration start point to the current time in the instantaneous velocity sequence. The real-time velocity sequence is filtered for monotonically increasing properties to construct and output an increasing observation dataset that conforms to the monotonically accelerating law. The incremental observation dataset is used as the observation sample input to the hierarchical Bayesian framework. Based on the phenomenological model configured in the hierarchical Bayesian framework, the Markov chain Monte Carlo algorithm is used to perform online posterior sampling to update the model parameters in real time and output the posterior probability distribution of the slope's remaining life. Based on the posterior probability distribution of the remaining lifespan and the preset risk threshold, a corresponding early warning trigger signal is output.

[0010] Furthermore, the instantaneous velocity sequence is analyzed to identify the acceleration start point and determine whether the slope deformation has entered the acceleration phase. Specific steps include: Real-time calculation of the start time of the constant velocity deformation stage up to the current moment The mean, standard deviation, skewness, and kurtosis of velocity within the cumulative time window are used to construct a dynamic 95% confidence interval. The real-time displacement sequence is subjected to J-level multi-scale discrete wavelet transform to extract detail coefficients at each scale, which are used to characterize the local high-frequency abrupt change features of the displacement signal. Using the constructed dynamic 95% confidence interval, skewness, kurtosis, and detail coefficient, a statistical abrupt change criterion, a distribution evolution criterion, and a wavelet synchronization criterion are respectively constructed to form a triple joint criterion; Based on the aforementioned triple joint criteria, the slope deformation is determined to have entered the acceleration phase according to the preset joint identification logic.

[0011] Furthermore, the aforementioned triple joint criterion specifically includes: Statistical mutation criterion: The current instantaneous velocity remains within the dynamic 95% confidence interval for multiple consecutive time periods. Above the upper limit, and must be maintained for at least 3 consecutive monitoring cycles; Distribution evolution criterion: skewness of instantaneous velocity distribution and kurtosis It exhibits a continuous monotonically increasing trend; Wavelet synchronization criterion: the magnitude of detail coefficients at two adjacent wavelet decomposition scales. It exhibits synchronous and continuous growth; In response to the detection that the statistical mutation criterion and the distribution evolution criterion are satisfied simultaneously, or the statistical mutation criterion and the wavelet synchronization criterion are satisfied simultaneously, the slope deformation is determined to have entered the acceleration stage, and the current moment is declared as the acceleration start point.

[0012] Furthermore, the hierarchical Bayesian framework is configured with the INV phenomenological model and the SLO phenomenological model, wherein: The prediction formula of the INV phenomenological model is as follows:

[0013] in, The predicted velocity at time t, To predict the time of destruction, A and For material-related constants, and Set to 2; The prediction formula of the SLO phenomenological model is as follows:

[0014] Where B is a constant, representing the slope velocity in the unit of time before failure.

[0015] Furthermore, after inputting the incremental observation dataset into the hierarchical Bayesian framework, based on the phenomenological model configured in the hierarchical Bayesian framework, online posterior sampling is performed using the Markov chain Monte Carlo algorithm to update the model parameters in real time, outputting the posterior probability distribution of the slope's remaining lifespan, and outputting corresponding early warning trigger signals according to the posterior probability distribution of the remaining lifespan and the preset risk threshold. Specific steps include: Based on the deterministic function corresponding to the phenomenological model, the velocity observation values ​​in the incremental observation dataset are uniformly transformed to construct a unified observation likelihood function; Based on the observed likelihood function and the preset prior distribution of the model parameters, the Markov chain Monte Carlo algorithm is run to perform online posterior sampling of the model parameters and update them in real time, including the predicted destruction time. The model parameters, including those included, are used to output the predicted destruction time. The posterior probability distribution; The predicted destruction time obtained from sampling Subtract the current time from the posterior distribution sample The posterior probability distribution and confidence interval of the remaining life of the slope at the current moment are calculated. The calculated posterior probability distribution of the remaining lifespan is compared with a preset risk threshold. When the remaining lifespan is lower than the corresponding risk threshold, it is converted into a unified early warning trigger signal and output.

[0016] Furthermore, the real-time velocity sequence is filtered for monotonically increasing properties to construct and output an increasing observation dataset that conforms to a monotonically accelerating law. Specific steps include: The initial velocity at the acceleration starting point Set as the initial reference speed and initialize a continuously incrementing counter. ; For each newly acquired monitoring speed Perform the following judgment: In response to Accept the monitoring speed and update. and the counter Increase by 1; In response to Skip this monitoring speed and set the counter. Reset to 1; Response to counter When the value reaches 3 for the first time, it is determined that an acceleration trend has been formed, and it will start from... Up to now The incremental velocity sequence formed by the accumulated velocity points is used as the initial data to construct the incremental observation dataset, which is used to perform the first Bayesian Markov chain Monte Carlo prediction and output an initial estimate of the remaining lifetime. After the initial prediction is completed, a continuous dynamic update phase begins, where each newly acquired monitoring velocity is updated. Perform the following judgment: In response to Accept the monitoring speed and update. The velocity point is then added to the incremental observation dataset, and the Bayesian Markov chain Monte Carlo prediction is re-executed to update the remaining lifetime distribution. In response to Skip this monitoring speed, keep the reference speed unchanged, and pause updates.

[0017] Furthermore, the observations of the INV phenomenological model and the SLO phenomenological model are uniformly transformed within the hierarchical Bayesian framework: For the INV phenomenological model, the transformed observations are defined as follows: ; For the SLO phenomenological model, the transformed observations are defined as follows: ; The unified expression for the two models after transformation is:

[0018] in, For deterministic model functions, namely the INV phenomenological model and the SLO phenomenological model, Let be the vector of parameters to be estimated. The error between the observed and predicted values ​​is independently and identically distributed according to a normal distribution. , This represents the standard deviation of the error for the unknown parameter.

[0019] Furthermore, in the hierarchical Bayesian framework, the prior distribution of the model parameters is set as follows: The parameter vector of the model to be estimated This includes parameter A or parameter B, which is related to material properties, and the estimated predicted failure time. ; Assign a log-normal prior distribution to the constant parameter A, which represents the material properties in the INV phenomenological model; Assign a normal prior distribution to the constant parameter B, which represents the baseline slope velocity per unit time before failure, in the SLO phenomenological model; The predicted destruction time The dynamic prior distribution is set as follows:

[0020] in, This indicates that it follows a normal distribution. For the current moment, The average remaining life constant during the acceleration phase is obtained based on statistics from historical landslide cases. The variance of remaining life is obtained from statistics of historical landslide cases.

[0021] Furthermore, the specific process of real-time updating using the MCMC sampling algorithm includes: The determination coefficient is found within the extreme range of parameters in historical landslide cases using a grid search. The maximum parameter combination is used as the initial parameter set for MCMC sampling. ; The Metropolis-Hastings algorithm is used to run multiple independent Markov chains for posterior sampling, discarding the pre-warmed samples and performing sparsification. Using the Gelman-Rubin statistic and effective sample size Verify sampling convergence; Based on the destruction time after convergence The posterior sample set of the remaining lifetime RL is calculated from the posterior sample, and its posterior mode is taken as the point estimate of the remaining lifetime. The uncertainty is characterized by the 95% highest posterior density interval.

[0022] In conjunction with the first feasible approach, a second feasible approach provides a system for predicting the remaining life of a slope, comprising: The data acquisition and preprocessing module is used to acquire the real-time displacement sequence of the slope through the displacement sensor; and to perform differential processing on the real-time displacement sequence to convert it into an instantaneous velocity sequence. An acceleration start point identification module is used to perform acceleration start point identification analysis on the instantaneous velocity sequence to determine whether the slope deformation has entered the acceleration stage; and when it is determined that the slope deformation has entered the acceleration stage, the corresponding acceleration start time is determined as the acceleration start point, and the starting velocity corresponding to the acceleration start point is output, as well as the real-time velocity sequence from the acceleration start point to the current time in the instantaneous velocity sequence. The monotonic filtering module is used to perform monotonically increasing filtering on the real-time velocity sequence, and to construct and output an increasing observation dataset that conforms to the monotonically accelerating law. The remaining life prediction module is used to input the incremental observation dataset as observation samples into the hierarchical Bayesian framework, and based on the phenomenological model configured in the hierarchical Bayesian framework, to perform online posterior sampling using the Markov chain Monte Carlo algorithm to update the model parameters in real time, and output the posterior probability distribution of the remaining life of the slope. The early warning triggering module is used to output a corresponding early warning triggering signal based on the posterior probability distribution of the remaining lifespan and a preset risk threshold.

[0023] As can be seen from the above technical solution, the beneficial technical effects of the present invention are as follows: 1. By extracting high-order statistical moment indices such as mean, skewness, and kurtosis from the velocity sequence, and combining them with the ability of multi-scale wavelet transform to represent local abrupt changes in different frequency bands, a triple joint judgment logic was constructed. This effectively eliminates false alarms caused by single-point outliers or local sporadic oscillations, ensuring accurate and sensitive capture of the transformation point of the deformation mechanism in the very early stages of deformation acceleration.

[0024] 2. To address the non-monotonic fluctuations caused by sensor disturbances in actual monitoring, an incremental trend detection algorithm with a counter mechanism was designed. This mechanism only activates prediction and includes data in the sample when data that meets the incremental pattern is continuously observed. It automatically filters out non-monotonic interference noise, effectively avoiding the problem of phenomenological model parameter distortion or prediction accuracy divergence caused by data decline, and enhancing the robustness of online rolling prediction.

[0025] 3. The hierarchical Bayesian framework enables online adaptive correction of parameters as monitoring data accumulates point by point. In the early stages when there are few monitoring data points, the uncertainty of the prediction can be reasonably quantified through posterior confidence intervals, and the range of uncertainty gradually converges as the data increases. At the same time, the abstract moment of destruction is directly transformed into a specific remaining response time, and graded early warnings are triggered based on time thresholds, directly serving the decision-making of emergency rescue. Attached Figure Description

[0026] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0027] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a statistical prior distribution diagram of parameters A and B and remaining lifetime RL in Embodiment 1 of the present invention; Figure 3 This is a panoramic view of the landslide case in Embodiment 1 of the present invention, showing the layout of measuring points at various time periods. Figure 4 The figure shows the displacement monitoring curves of each GNSS measuring point in the reservoir landslide case of Embodiment 1 of the present invention. Figure 5 This is an OOA identification result diagram based on GNSS05 measurement point displacement data in Embodiment 1 of the present invention. Figure 5 (a) shows the true velocity curve, the upper bound of the dynamic 95% confidence interval, and the identified inflection point. Figure 5 (b) represents the dynamic monotonic evolution of the skewness and kurtosis of the velocity distribution. Figure 5 (c) shows the dynamic time-frequency variation of wavelet detail coefficients 1 and 2; Figure 6 This is a graph showing the prediction results of the INV model in Embodiment 1 of the present invention. Figure 6 (a) is the parameter A and The joint posterior distribution, Figure 6 (b) is the time of destruction The posterior probability distribution, Figure 6 (c) is the convergence diagnosis of MCMC. Figure 6 (d) shows the predicted fit of the INV model; Figure 7 This is a graph showing the prediction results of the SLO model in Embodiment 1 of the present invention; Figure 8 This is a comprehensive comparison chart of the prediction results of Embodiment 1 of the present invention; Figure 9 This is a system structure diagram of Embodiment 2 of the present invention; Figure label: A slope remaining life prediction system, 21-data acquisition module, 22-acceleration starting point identification module, 23-monotonic filtering module, 24-remaining life prediction module, 25-early warning triggering module. Detailed Implementation

[0028] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0029] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0030] Example 1

[0031] This embodiment provides a method for predicting the remaining life of a slope. The working principle of Embodiment 1 is explained in detail below: This embodiment provides a method for predicting the remaining life of a slope, such as... Figure 1 The diagram shown is a flowchart of the method in this embodiment. This method integrates two key components: automatic identification of acceleration start point (OOA) and dynamic probability prediction of remaining lifetime (RL), and includes the following steps: Real-time displacement sequences of the slope are obtained using displacement sensors; The real-time displacement sequence is differentially processed to convert it into an instantaneous velocity sequence; The instantaneous velocity sequence is analyzed to identify the acceleration start point and determine whether the slope deformation has entered the acceleration stage. In response to the slope deformation entering the acceleration stage, the corresponding acceleration start time is determined as the acceleration start point, and the starting velocity corresponding to the acceleration start point is output, as well as the real-time velocity sequence from the acceleration start point to the current time in the instantaneous velocity sequence. The real-time velocity sequence is filtered for monotonically increasing properties to construct and output an increasing observation dataset that conforms to the monotonically accelerating law. The incremental observation dataset is used as the observation sample input to the hierarchical Bayesian framework. Based on the phenomenological model configured in the hierarchical Bayesian framework, the Markov Chain Monte Carlo (MCMC) algorithm is used to perform online posterior sampling to update the model parameters in real time and output the posterior probability distribution of the slope's remaining life. Based on the posterior probability distribution of the remaining lifespan and the preset risk threshold, a corresponding early warning trigger signal is output.

[0032] The working principle and specific execution steps of this embodiment are explained in detail below: In this embodiment, the system acquires the real-time displacement sequence of the slope using displacement sensors (such as GNSS receivers, multi-point displacement meters, etc.) deployed at the slope site. .

[0033] The input real-time displacement sequence is converted into an instantaneous velocity sequence using the differential method. By calculating the difference between the displacement changes of adjacent sampling points, some long-period drift errors can be quickly filtered out, resulting in an instantaneous velocity sequence that is highly sensitive to changes in slope deformation acceleration.

[0034] In this embodiment, the system further performs OOA identification analysis on the instantaneous velocity sequence to determine whether the slope deformation has entered the acceleration phase.

[0035] During the isochronous deformation stage, slope deformation is mainly controlled by elastic deformation or steady creep processes, and the velocity sequence typically fluctuates around a certain mean. Due to the randomness of measurement errors, its data distribution can be approximated as a normal distribution. When the slope enters the accelerated deformation stage, the accumulation of internal damage leads to a continuous increase in the deformation driving force. The velocity sequence not only exhibits a systematic shift in the mean, but its higher-order statistical moments (skewness and kurtosis) also change significantly, reflecting the enhanced nonlinearity and nonstationarity of the deformation process.

[0036] To quantify this evolution process, this embodiment employs a dynamic statistics module, setting the starting time from the constant velocity deformation stage. up to the current moment The cumulative time window for velocity sequences Real-time analysis is performed. The specific formulas for calculating the four core statistics are as follows: average speed Characterizing the overall level of velocity:

[0037] in, From up to the current moment The cumulative number of samples, These are the velocity observation values ​​at each moment within this time period.

[0038] Standard deviation : Measures the degree of dispersion of velocity distribution:

[0039] Skewness : Measuring the asymmetry of velocity distribution:

[0040] An increase in positive skewness indicates that the velocity distribution has a longer tail in the direction of higher velocity, reflecting the emergence of an acceleration trend.

[0041] Kudo Characterizing the kurtosis of the distribution:

[0042] An increase in positive kurtosis indicates that the distribution is more angular and has thicker tails than a normal distribution, which means that the probability of extreme velocity values ​​increases.

[0043] Based on average speed and standard deviation The system constructs a dynamic 95% confidence interval. :

[0044] During the constant velocity phase, the observed velocity should remain stable within this range. When the velocity continuously exceeds the upper limit of this range for multiple consecutive time periods, it indicates that the slope may be transitioning to an acceleration phase.

[0045] To independently verify the physical activation of accelerated deformation, this embodiment simultaneously introduces a wavelet analysis module. Although the displacement curves of some landslides appear visually smooth, their deformation process is essentially the result of the coupling of multi-scale mechanical processes such as internal micro-fractures, particle rearrangement, and frictional sliding along structural surfaces. The suddenness and collective nature of these micromechanical events can excite specific high-frequency components or enhance signal singularities in macroscopic deformation signals. Traditional time-domain analysis methods (such as moving average methods) cannot effectively extract such hidden transient features.

[0046] Wavelet transform, as an excellent time-frequency localization tool, utilizes the mother wavelet... Scaling and translating the basis functions can decompose the signal into different scales (frequency) and locations (time). For discrete monitoring data, displacement signals... conduct Multi-scale discrete wavelet transform:

[0047] in, For the specified number of decomposition levels, For the first The approximate signal of the layer represents the lowest frequency component or overall trend of the displacement signal. For the first Layer detail signals, including signals at scale Local high-frequency abrupt changes (corresponding to a specific frequency band). Detail coefficients of the j-th layer. The result is obtained by calculating the inner product of the displacement signal and the mother wavelet function after scaling and translation:

[0048] In this embodiment, the db5 wavelet function is used as the wavelet basis. When the slope enters the acceleration phase, the frequency and energy of internal micro-fracture events increase significantly, which is reflected in the amplitude of detail coefficients at multiple adjacent decomposition scales. Simultaneously increasing. The system makes a joint judgment using the following three joint criteria: Statistical mutation criterion: The currently observed instantaneous velocity must remain within the dynamic 95% confidence interval. The slope deformation must be above the upper limit and maintained for at least three consecutive monitoring periods. This criterion aims to ensure that a continuous accelerating trend in slope deformation is detected.

[0049] Distribution evolution criterion: skewness of velocity distribution and kurtosis It must simultaneously exhibit continuous monotonically increasing behavior. This criterion captures the characteristics of the deformation process evolving from an approximately symmetrical, mild random fluctuation (normal distribution) to an asymmetric, sharp, and heavy-tailed distribution, and is a strong statistical indicator characterizing an impending acceleration.

[0050] Wavelet synchronization criterion: the magnitude of detail coefficients at two adjacent wavelet decomposition scales. The growth must be synchronous, significant, and continuous. This criterion verifies the physical activation of accelerated slope deformation from the perspective of the physical changes in the signal's intrinsic structure.

[0051] To balance the sensitivity and reliability of identification, the system establishes the following OOA point confirmation logic: Logic 1: When the system detects that both the statistical mutation criterion and the distribution evolution criterion are satisfied, it declares the current moment as the acceleration starting point; Logic 2: When the system detects that both the statistical mutation criterion and the wavelet synchronization criterion are satisfied, it also declares the current moment as the acceleration starting point.

[0052] This rule indicates that the statistical mutation criterion is a necessary condition and must be supported by the distribution evolution criterion or the wavelet synchronization criterion. This design ensures that the basis for triggering accelerated detection is continuous velocity growth, while incorporating deeper statistical distribution evolution or signal structure mutations for verification.

[0053] In this embodiment, further, in response to the slope deformation entering the acceleration phase, the corresponding acceleration start time is determined as the acceleration start point, and the system outputs the starting velocity corresponding to that point. .

[0054] While phenomenological models are widely used in acceleration phases, their mathematical form dictates that parameter standardization and extrapolation accuracy will significantly decrease when input data declines due to local fluctuations. Therefore, this embodiment introduces an increasing trend checking mechanism to filter the real-time velocity sequence for monotonically increasing properties. The specific steps are as follows: The initial velocity at the acceleration starting point Set as the initial reference speed and initialize a continuously incrementing counter. ; For each newly acquired monitoring speed Perform the following judgment: In response to Accept the monitoring speed and update. and the counter Increase by 1; In response to Skip this monitoring speed and set the counter. Reset to 1; When the counter Upon first reaching level 3, a clear trend of physical acceleration has been established. At this point, the system immediately utilizes... Up to now The accumulated increasing velocity sequence is used as initial data to construct an increasing observation dataset and trigger the first Bayesian MCMC prediction, outputting an initial probability estimate of the remaining lifetime RL.

[0055] After the initial prediction is completed, the system enters a continuous dynamic update phase. In this phase, it no longer relies on... Whether to include the count data is determined solely by comparing the new speed with the reference speed: If the newly acquired speed If so, accept the data point and update. The data point is then added to the incremental observation dataset, immediately triggering a Bayesian update and outputting a newly updated RL distribution.

[0056] If the newly acquired speed If so, skip that data point and maintain the reference speed. Remain unchanged and pause updates until a new event occurs that exceeds the current one. Until the speed reaches its limit.

[0057] In this embodiment, further, existing research has evaluated various classical phenomenological models (such as INV, MINV, SLO, VOA, etc.). Among them, the VOA model is extremely sensitive to monitoring noise and computational errors, and its stability is insufficient; while the MINV model introduces a finite destruction rate threshold, which needs to be subjectively set, introducing additional uncertainty. Therefore, this embodiment only configures and uses two classical phenomenological models, INV and SLO, and their prediction formulas are as described above.

[0058] To achieve unified computation in Bayesian inference, the system first performs a unified transformation on the observation equations of the two models: For the INV model (in this embodiment, it will be...) Setting it to 2 enables high-precision deformation and failure prediction. The transformed observation is defined as follows:

[0059] For the SLO model, the transformed observations are defined as follows:

[0060] Therefore, both models can be uniformly represented as:

[0061] in, For the deterministic function of the corresponding model, Let be the vector of parameters to be estimated. The error between the observed and predicted values ​​is independently and identically distributed according to a normal distribution. . Let the unknown standard deviation represent the noise level measured by the sensor. According to Bayes' principle, the likelihood function is defined as:

[0062] This embodiment utilizes statistical data from 19 historical landslide cases to determine the prior distribution of model parameters. For example... Figure 2 The figure shown is a statistical prior distribution diagram of parameters A and B and remaining lifetime RL in this embodiment.

[0063] Parameter A is strictly positive, therefore it is assigned a log-normal prior distribution. Based on historical case fitting, the parameters are determined to be: log-normal mean. Standard deviation .

[0064] Parameter B is assigned to the normal prior distribution, and its parameters are: mean Standard deviation .

[0065] Error standard deviation Assign no informational prior – Half-Cauchy distribution.

[0066] Predicted time of destruction It is a dynamic process whose prior is determined using the historical distribution of RL. Because... The historical RL statistical mean and standard deviation are respectively and Therefore, the prior distribution is set as follows:

[0067] To improve the robustness of MCMC sampling, the system first performs a grid search within the historical landslide parameter range (e.g., ... Figure 2 Find the coefficient of determination within the extreme value range shown. Maximize the combination of parameters This was used as the starting point for MCMC sampling. Subsequently, the Metropolis-Hastings algorithm was used to run four independent Markov chains for posterior sampling. Each chain extracted 40,000 samples, discarding the first 8,000 as a burn-in period, and sparsification was performed at intervals of 8. The Gelman-Rubin statistic was used. and effective sample size Verify the sampling convergence.

[0068] Based on the destruction time after convergence Based on the posterior samples, the system directly calculates the remaining life RL distribution of the slope at the current moment using the following formula:

[0069] The mode of the posterior sample of the RL is taken as the point estimate, and the 95% highest posterior density interval (HPD) is used to intuitively characterize the uncertainty of its prediction. Finally, the calculated RL distribution is compared with the preset risk threshold (i.e., the safety preparation response window, which includes "3 days" and "24 hours" in this embodiment). If the remaining lifespan is lower than the corresponding threshold, the system will automatically output a unified early warning trigger signal to directly guide the emergency response at the engineering site.

[0070] To verify the scientific validity and engineering practical value of the treatment method in this embodiment, this embodiment uses a case of landslide on the slope of a large hydropower station reservoir for application and testing.

[0071] like Figure 3 The image shown is a panoramic view of a reservoir landslide case used for method verification in this embodiment, along with a diagram of the monitoring point layout across multiple time periods. Five GNSS displacement monitoring points (GNSS01~GNSS05) were deployed on the landslide, and their spatial distribution and the deformation and instability process of the landslide at each key time point are clearly presented. Figure 3 The entire landslide deformation evolution process can be divided into the following three stages: Initial Deformation Stage (May 18, 2023 - October 18, 2023): Deformation of the landslide began in May 2023, and frame beam anchor cable reinforcement was implemented on-site in July of the same year. The average deformation rate during this stage was 1.56 mm / day. Under the reinforcement, the deformation rate gradually decreased, and by mid-October, the deformation tended to stabilize. As shown in the dynamic image, the frame lattice beam remained intact.

[0072] Uniform Deformation Stage (October 18, 2023 - July 17, 2024): After the reinforcement project was completed, the deformation entered a quasi-steady-state stage. The displacement monitoring rate decreased and remained at an extremely low level of about 0.3 mm / day for a long period of time, indicating that the slope was in a temporary equilibrium state.

[0073] Accelerated Failure Phase (July 17, 2024 - July 20, 2024): Triggered by heavy rainfall, the slope of the accumulation mass began to collapse rapidly from around 16:00 on July 17 (at an elevation of approximately 2606 meters), initially showing localized damage at the leading edge. By July 19, approximately 1000 cubic meters of the leading edge had slid into the bottom of the gully. At 10:00 on July 20, a large area of ​​the frame beam anchor cables broke, causing approximately 20,000 cubic meters of material behind them to become unstable and slide.

[0074] like Figure 4 The diagram shows the displacement monitoring curves of various GNSS measuring points in a reservoir landslide case of this embodiment, including detailed magnified views of the overall deformation segment and the local acceleration segment. Since the GNSS05 measuring point is located at the leading edge of the landslide and is most sensitive to the early instability response of the landslide body, this embodiment uses the data from the GNSS05 measuring point during the pre-slide deformation period as an example to illustrate the specific method application.

[0075] like Figure 5 The image shown is an OOA identification result based on GNSS05 measurement point displacement data in this embodiment. Figure 5 (a) Shows the actual velocity curve, the upper bound of the dynamic 95% confidence interval, and the identified inflection points. Figure 5 (b) illustrates the dynamic monotonic evolution of the skewness and kurtosis of the velocity distribution. Figure 5 (c) shows the dynamic time-frequency variations of wavelet detail coefficients 1 and 2.

[0076] Analysis shows that on day 423 of the monitoring time series (corresponding to July 12, 2024), the mean displacement rate of the GNSS05 measuring point exceeded the upper limit of the dynamic 95% confidence interval. Furthermore, over the next three consecutive days, the velocity skewness and kurtosis increased monotonically, while the amplitude of the wavelet detail coefficients increased significantly, perfectly triggering the triple joint criterion of this embodiment. The system thus declared July 12 as the acceleration start point (OOA point). This identification time was 5 days earlier than the local instability on July 17 and 8 days earlier than the overall instability on July 20.

[0077] After successfully locking the OOA point, the system immediately initiates an incremental trend check mechanism and uses a hierarchical Bayesian framework to conduct dynamic rolling prediction of the remaining life before slippage.

[0078] like Figure 6 The image shown is a graph illustrating the INV model prediction results based on monitoring data from July 12th to July 14th in this embodiment, including: Figure 6 (a) Parameter A and The joint posterior distribution, Figure 6 (b) Time of destruction The posterior probability distribution (95% HPD). Figure 6 (c) Convergence diagnosis of MCMC Figure 6 (d) Predictive fit of the INV model.

[0079] like Figure 7 The figure shown is the prediction result of the SLO model based on the monitoring data from July 12 to July 14 in this embodiment, which shows its posterior estimation and prediction under the same monitoring window.

[0080] Depend on Figure 6 and Figure 7 The prediction results show that the system output a remaining life prediction of about 3.05 days on July 14, which is highly consistent with the actual time of the local damage that occurred on the afternoon of July 17 (about 3 days later), and the prediction result is extremely accurate.

[0081] like Figure 8 The figure shows a comprehensive comparison of the prediction results obtained at various monitoring times in this embodiment. The figure illustrates the evolution of the remaining lifetime probability distribution predicted by the INV model (left) and the SLO model (right) as monitoring data gradually accumulates over time (July 12th to July 18th) after the OOA point is triggered on July 12th. Figure 8It can be seen that in the early stage of prediction (when there are fewer data points), the prediction uncertainty range (95% HDI gray shaded area) is relatively wide, but its probability mode (pink solid line) has accurately pointed to the actual local damage time (blue dashed line). As time goes on and more monitoring data is injected, the prediction interval gradually converges and narrows, the uncertainty decreases rapidly, and the predicted lifetime almost perfectly matches the actual remaining lifetime (red solid line).

[0082] The bar chart below shows the model's determination coefficients at each rolling prediction step. SLO model The value remained stable around 1.00 for a long time, while the INV model also remained stable above 0.85 in the later stages of deformation.

[0083] The above verification results fully demonstrate that the method proposed in this embodiment can robustly identify the physical initiation time of slope acceleration online, and quantitatively output a high-precision remaining lifetime probability distribution during the dynamic process of data accumulation, providing intuitive, scientific and highly reliable technical support for early warning and risk avoidance decision-making in engineering landslide disasters.

[0084] Example 2

[0085] In conjunction with the method provided in Embodiment 1, this embodiment provides a slope remaining life prediction system 2, the system structure diagram of which is shown below. Figure 9 As shown, it includes: Data acquisition module 21 is used to acquire the real-time displacement sequence of the slope through displacement sensors; and to perform differential processing on the real-time displacement sequence to convert it into an instantaneous velocity sequence. The acceleration start point identification module 22 is used to perform acceleration start point identification analysis on the instantaneous velocity sequence to determine whether the slope deformation has entered the acceleration stage; and when it is determined that the slope deformation has entered the acceleration stage, the corresponding acceleration start time is determined as the acceleration start point, and the starting velocity corresponding to the acceleration start point is output, as well as the real-time velocity sequence from the acceleration start point to the current time in the instantaneous velocity sequence. The monotonic filtering module 23 is used to perform monotonically increasing filtering on the real-time velocity sequence, and to construct and output an increasing observation dataset that conforms to the monotonically accelerating law. The remaining life prediction module 24 is used to input the incremental observation dataset as observation samples into the hierarchical Bayesian framework, and based on the phenomenological model configured in the hierarchical Bayesian framework, use the Markov chain Monte Carlo algorithm to perform online posterior sampling to update the model parameters in real time, and output the posterior probability distribution of the remaining life of the slope. The early warning triggering module 25 is used to output a corresponding early warning triggering signal based on the posterior probability distribution of the remaining lifespan and a preset risk threshold.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for predicting the remaining life of a slope, characterized in that, Includes the following steps: Real-time displacement sequences of the slope are obtained using displacement sensors; The real-time displacement sequence is differentially processed to convert it into an instantaneous velocity sequence; The instantaneous velocity sequence is analyzed to identify the acceleration start point and determine whether the slope deformation has entered the acceleration phase. Specific steps include: Real-time calculation of the start time of the constant velocity deformation stage up to the current moment The mean, standard deviation, skewness, and kurtosis of velocity within the cumulative time window are used to construct a dynamic 95% confidence interval. The real-time displacement sequence is subjected to J-level multi-scale discrete wavelet transform to extract detail coefficients at each scale, which are used to characterize the local high-frequency abrupt change features of the displacement signal. Using the constructed dynamic 95% confidence interval, skewness, kurtosis, and detail coefficient, a statistical abrupt change criterion, a distribution evolution criterion, and a wavelet synchronization criterion are respectively constructed to form a triple joint criterion; Based on the aforementioned triple joint criteria, the slope deformation is determined to have entered the acceleration phase according to the preset joint identification logic. The aforementioned triple joint criterion specifically includes: Statistical mutation criterion: The current instantaneous velocity remains within the dynamic 95% confidence interval for multiple consecutive time periods. Above the upper limit, and must be maintained for at least 3 consecutive monitoring cycles; Distribution evolution criterion: skewness of instantaneous velocity distribution and kurtosis It exhibits a continuous monotonically increasing trend; Wavelet synchronization criterion: the magnitude of detail coefficients at two adjacent wavelet decomposition scales. It exhibits synchronous and continuous growth; In response to the detection that the statistical mutation criterion and the distribution evolution criterion are satisfied simultaneously, or the statistical mutation criterion and the wavelet synchronization criterion are satisfied simultaneously, the slope deformation is determined to have entered the acceleration stage, and the current moment is declared as the acceleration start point. In response to the slope deformation entering the acceleration stage, the corresponding acceleration start time is determined as the acceleration start point, and the starting velocity corresponding to the acceleration start point is output, as well as the real-time velocity sequence from the acceleration start point to the current time in the instantaneous velocity sequence. The real-time velocity sequence is filtered for monotonically increasing properties to construct and output an increasing observation dataset that conforms to the monotonically accelerating law. The incremental observation dataset is used as the observation sample input to the hierarchical Bayesian framework. Based on the phenomenological model configured in the hierarchical Bayesian framework, the Markov chain Monte Carlo algorithm is used to perform online posterior sampling to update the model parameters in real time and output the posterior probability distribution of the slope's remaining life. Based on the posterior probability distribution of the remaining lifespan and the preset risk threshold, a corresponding early warning trigger signal is output.

2. The method for predicting the remaining life of a slope according to claim 1, characterized in that, The hierarchical Bayesian framework includes the INV phenomenological model and the SLO phenomenological model, wherein: The prediction formula of the INV phenomenological model is as follows: in, The predicted velocity at time t, To predict the time of destruction, A and For material-related constants, and Set to 2; The prediction formula of the SLO phenomenological model is as follows: Where B is a constant, representing the slope velocity in the unit of time before failure.

3. The method for predicting the remaining life of a slope according to claim 2, characterized in that, After inputting the incremental observation dataset into the hierarchical Bayesian framework, based on the phenomenological model configured in the hierarchical Bayesian framework, online posterior sampling is performed using the Markov chain Monte Carlo algorithm to update the model parameters in real time, outputting the posterior probability distribution of the slope's remaining life, and outputting corresponding early warning trigger signals according to the posterior probability distribution of the remaining life and the preset risk threshold. The specific steps include: Based on the deterministic function corresponding to the phenomenological model, the velocity observation values ​​in the incremental observation dataset are uniformly transformed to construct a unified observation likelihood function; Based on the observed likelihood function and the preset prior distribution of the model parameters, the Markov chain Monte Carlo algorithm is run to perform online posterior sampling of the model parameters and update them in real time, including the predicted destruction time. The model parameters, including those included, are used to output the predicted destruction time. The posterior probability distribution; The predicted destruction time obtained from sampling Subtract the current time from the posterior distribution sample The posterior probability distribution and confidence interval of the remaining life of the slope at the current moment are calculated. The calculated posterior probability distribution of the remaining lifespan is compared with a preset risk threshold. When the remaining lifespan is lower than the corresponding risk threshold, it is converted into a unified early warning trigger signal and output.

4. The method for predicting the remaining life of a slope according to claim 3, characterized in that, The real-time velocity sequence is filtered for monotonically increasing properties to construct and output an increasing observation dataset that conforms to a monotonically accelerating law. Specific steps include: The initial velocity at the acceleration starting point Set as the initial reference speed and initialize a continuously incrementing counter. ; For each newly acquired monitoring speed Perform the following judgment: In response to Accept the aforementioned monitoring speed and update. and the counter Increase by 1; In response to Skip the aforementioned monitoring speed and set the counter Reset to 1; Response to counter When the value reaches 3 for the first time, it is determined that an acceleration trend has been formed, and it will start from... Up to now The incremental velocity sequence formed by the accumulated velocity points is used as the initial data to construct the incremental observation dataset, which is used to perform the first Bayesian Markov chain Monte Carlo prediction and output an initial estimate of the remaining lifetime. After the initial prediction is completed, a continuous dynamic update phase begins, where each newly acquired monitoring velocity is updated. Perform the following judgment: In response to Accept the current monitoring speed and update. The velocity point is then added to the incremental observation dataset, and the Bayesian Markov chain Monte Carlo prediction is re-executed to update the remaining lifetime distribution. In response to Skip the aforementioned monitoring speed, keep the reference speed unchanged, and pause updates.

5. The method for predicting the remaining life of a slope according to claim 2, characterized in that, The hierarchical Bayesian framework performs a unified transformation on the observations of the INV phenomenological model and the SLO phenomenological model: For the INV phenomenological model, the transformed observations are defined as follows: ; For the SLO phenomenological model, the transformed observations are defined as follows: ; The unified expression for the two models after transformation is: in, For deterministic model functions, namely the INV phenomenological model and the SLO phenomenological model, Let be the vector of parameters of the model to be estimated. The error between the observed and predicted values ​​is independently and identically distributed according to a normal distribution. , This represents the standard deviation of the error for the unknown parameter.

6. The method for predicting the remaining life of a slope according to claim 5, characterized in that, In the hierarchical Bayesian framework, the prior distribution of model parameters is set as follows: The parameter vector of the model to be estimated This includes parameter A or parameter B, which is related to material properties, and the estimated predicted failure time. ; Assign a log-normal prior distribution to the constant parameter A, which represents the material properties in the INV phenomenological model; Assign a normal prior distribution to the constant parameter B, which represents the baseline slope velocity per unit time before failure, in the SLO phenomenological model; The predicted destruction time The dynamic prior distribution is set as follows: in, This indicates that it follows a normal distribution. For the current moment, The average remaining life constant during the acceleration phase is obtained based on statistics from historical landslide cases. The variance of remaining life is obtained from statistics of historical landslide cases.

7. The method for predicting the remaining life of a slope according to claim 6, characterized in that, The specific process of real-time updating using the MCMC sampling algorithm includes: The determination coefficient is found within the extreme range of parameters in historical landslide cases using a grid search. The maximum parameter combination is used as the initial parameter set for MCMC sampling. ; The Metropolis-Hastings algorithm is used to run multiple independent Markov chains for posterior sampling, discarding the pre-warmed samples and performing sparsification. Using the Gelman-Rubin statistic and effective sample size Verify sampling convergence; Based on the destruction time after convergence The posterior sample set of the remaining lifetime RL is calculated from the posterior sample, and its posterior mode is taken as the point estimate of the remaining lifetime. The uncertainty is characterized by the 95% highest posterior density interval.

8. A system for predicting the remaining life of a slope, characterized in that, include: The data acquisition and preprocessing module is used to acquire the real-time displacement sequence of the slope through displacement sensors; The real-time displacement sequence is differentially processed to convert it into an instantaneous velocity sequence; An acceleration start point identification module is used to perform acceleration start point identification analysis on the instantaneous velocity sequence to determine whether the slope deformation has entered the acceleration stage. Specific steps include: Real-time calculation of the start time of the constant velocity deformation stage up to the current moment The mean, standard deviation, skewness, and kurtosis of velocity within the cumulative time window are used to construct a dynamic 95% confidence interval. The real-time displacement sequence is subjected to J-level multi-scale discrete wavelet transform to extract detail coefficients at each scale, which are used to characterize the local high-frequency abrupt change features of the displacement signal. Using the constructed dynamic 95% confidence interval, skewness, kurtosis, and detail coefficient, a statistical abrupt change criterion, a distribution evolution criterion, and a wavelet synchronization criterion are respectively constructed to form a triple joint criterion; Based on the aforementioned triple joint criteria, the slope deformation is determined to have entered the acceleration phase according to the preset joint identification logic. The aforementioned triple joint criterion specifically includes: Statistical mutation criterion: The current instantaneous velocity remains within the dynamic 95% confidence interval for multiple consecutive time periods. Above the upper limit, and must be maintained for at least 3 consecutive monitoring cycles; Distribution evolution criterion: skewness of instantaneous velocity distribution and kurtosis It exhibits a continuous monotonically increasing trend; Wavelet synchronization criterion: the magnitude of detail coefficients at two adjacent wavelet decomposition scales. It exhibits synchronous and continuous growth; In response to the detection that the statistical mutation criterion and the distribution evolution criterion are satisfied simultaneously, or the statistical mutation criterion and the wavelet synchronization criterion are satisfied simultaneously, the slope deformation is determined to have entered the acceleration stage, and the current moment is declared as the acceleration start point. In response to the slope deformation entering the acceleration stage, the corresponding acceleration start time is determined as the acceleration start point, and the starting velocity corresponding to the acceleration start point is output, as well as the real-time velocity sequence from the acceleration start point to the current time in the instantaneous velocity sequence. The monotonic filtering module is used to perform monotonically increasing filtering on the real-time velocity sequence, and to construct and output an increasing observation dataset that conforms to the monotonically accelerating law. The remaining life prediction module is used to input the incremental observation dataset as observation samples into the hierarchical Bayesian framework, and based on the phenomenological model configured in the hierarchical Bayesian framework, to perform online posterior sampling using the Markov chain Monte Carlo algorithm to update the model parameters in real time, and output the posterior probability distribution of the remaining life of the slope. The early warning triggering module is used to output a corresponding early warning triggering signal based on the posterior probability distribution of the remaining lifespan and a preset risk threshold.

Images