Real-time monitoring data-driven landslide instability time dynamic prediction method and system

Abstract

The present application relates to the technical field of geological early warning, and discloses a real-time monitoring data driven landslide instability time dynamic prediction method and system, real-time collection of target landslide daily cumulative displacement monitoring data, calculation of daily deformation rate sequence; if true acceleration is determined based on daily deformation rate, locking the acceleration starting point and starting prediction; based on the monitoring data from the acceleration starting point to the current time, calculation of a preset index and real-time parameter solving of the Fukuzono model without instability time constraint through linearization fitting, obtaining the real-time creep characteristics of the day; training of an instability time prediction model based on the monitoring data of historical landslides of the same type; input of the real-time creep characteristics of the target landslide into the instability time prediction model, obtaining the remaining instability days predicted by machine learning, fusion of the remaining instability days predicted by the Fukuzono model of the day, real-time obtaining of the final remaining instability days and the predicted instability date. The present application realizes landslide instability time prediction which can be continuously iterated and updated with monitoring data.

Description

Technical Field

[0001] This invention relates to the field of geological early warning technology, specifically to a method and system for dynamic prediction of landslide instability time driven by real-time monitoring data. Background Technology

[0002] Landslides, as a typical progressive geological hazard, generally exhibit decelerating creep, isochronous creep, and accelerating creep evolution patterns under the coupled influence of multiple factors such as long-term creep, reservoir water level fluctuations, rainfall infiltration, and the self-weight of the soil and rock mass. Accurate prediction of their instability time is a core technical challenge for reducing geological hazard losses and ensuring the safety of people's lives and property. With the widespread adoption of automated monitoring technologies such as GNSS, landslide surface displacement can now be collected in high-frequency, continuous, and real-time data. How to fully explore the creep evolution patterns contained in real-time monitoring data and construct data-driven dynamic prediction models has become a core direction for breaking through the bottlenecks of traditional early warning technologies.

[0003] Current landslide instability time prediction methods still rely primarily on traditional creep theory models such as the Fukuzono model. These methods only perform nonlinear fitting based on the priority displacement time series of a single landslide, rather than being driven by real-time monitoring data throughout the entire cycle. This results in significant technical limitations: First, in the early stages of accelerated creep, the effective real-time data sequence is short, making it prone to drastic jumps in creep parameters and instability time, leading to extremely poor prediction stability. Second, they only follow the assumptions of ideal creep theory, without incorporating massive amounts of historical landslide real-time monitoring data for systematic calibration, ignoring individual differences in landslide geological conditions, resulting in significant deviations between theoretical calculations and actual instability behavior. Third, they cannot distinguish between false acceleration caused by environmental disturbances and true acceleration driven by slip zone damage based on real-time data, exhibiting weak anti-interference capabilities and a high false alarm rate. Fourth, they lack the dynamic prediction capability to be continuously updated with new real-time monitoring data, making it difficult to adapt to the actual needs of continuous early warning at engineering sites. Summary of the Invention

[0004] This invention aims to provide a method and system for dynamic prediction of landslide instability time driven by real-time monitoring data, in order to solve the technical problems of traditional Fukuzono models, which only perform theoretical fitting based on monitoring time series, and have shortcomings such as early prediction jumps, sensitivity to environmental disturbances, failure to consider individual geological differences of landslides, and difficulty in direct engineering application.

[0005] The basic solution provided by this invention is: a dynamic prediction method for landslide instability time driven by real-time monitoring data, comprising: S1: Real-time acquisition of daily cumulative displacement monitoring data of the target landslide, followed by preprocessing to calculate the daily deformation rate sequence; preprocessing includes multi-point synchronization verification and overall deformation time series synthesis; S2: Calculate the rate ratio and daily acceleration based on the daily deformation rate, and use the accelerated creep stage and damage-driven true acceleration fusion judgment mechanism to determine true acceleration. If it is determined to be true acceleration, lock the acceleration start point and start the prediction to enter S3; otherwise, it is false acceleration and terminate the prediction and return to S1. S3, based on the monitoring data from the acceleration start point to the current time, calculate the preset index and solve the real-time parameters of the Fukuzono model without instability time constraints through linear fitting to obtain the real-time creep characteristics of the day; the real-time creep characteristics include the baseline rate, the real-time amplitude of the Fukuzono model, the real-time nonlinear exponent, the real-time rate reciprocal cumulative amount and the rate-acceleration linearized characteristic quantity. S4. Based on the monitoring data of the accelerated creep stage of similar landslides that have been unstable in the past, calculate the corresponding real-time creep characteristics and the actual number of days remaining instability. Use the globally standardized real-time creep characteristics as input and the number of days remaining instability as output to train a machine learning model and obtain an instability time prediction model. S5: Input the real-time creep characteristics of the target landslide on the same day into the trained instability time prediction model to obtain the remaining instability days predicted by machine learning. Combine the remaining instability days predicted by the Fukuzono model on the same day with the model fusion to obtain the final remaining instability days and the predicted instability date in real time.

[0006] This invention also provides a real-time monitoring data-driven dynamic prediction system for landslide instability time, to execute a real-time monitoring data-driven method for dynamic prediction of landslide instability time; the system includes: The data processing module is used to collect the daily cumulative displacement monitoring data of the target landslide in real time, and to preprocess and calculate the daily deformation rate sequence. The true acceleration determination module is used to calculate the rate ratio and daily acceleration based on the daily deformation rate, and adopts a true acceleration determination mechanism that combines the accelerated creep stage and damage-driven true acceleration. If it is determined to be true acceleration, the acceleration start point is locked and prediction is started; otherwise, it is a false acceleration and the prediction is terminated. The creep characteristic calculation module is used for response start prediction. Based on the monitoring data from the acceleration start point to the current time, it calculates the preset index and solves the real-time parameters of the Fukuzono model without instability time constraints through linear fitting to obtain the real-time creep characteristics of the day. The prediction model building module is used to calculate the corresponding real-time creep characteristics and the actual remaining number of days of instability based on the monitoring data of the accelerated creep stage of similar landslides that have been unstable in the past. It uses the globally standardized real-time creep characteristics as input and the remaining number of days of instability as output to train a machine learning model and obtain an instability time prediction model. The model fusion prediction module is used to input the real-time creep characteristics of the target landslide into the trained instability time prediction model to obtain the remaining instability days predicted by machine learning. It then combines the remaining instability days predicted by the Fukuzono model on the same day with the model fusion to obtain the final remaining instability days and the predicted instability date in real time.

[0007] The working principle and advantages of this invention are as follows: This invention fully leverages the value of real-time landslide monitoring data, balancing the rationality of creep theory with the accuracy of data-driven predictions. It achieves landslide instability time prediction that can be continuously updated with monitoring data, significantly improving the engineering practicality of geological disaster early warning. Compared with existing technologies, the core advantages of this invention are: First, in the early warning triggering stage, existing technologies typically use displacement time-series monitoring data to directly perform theoretical fitting using mainstream models such as Fukuzono to complete the prediction. This skips the validity verification of the data layer and relies solely on model fitting, which easily leads to the failure of the early warning. Even if some existing technologies use the presence of acceleration as the activation condition, this mechanism cannot distinguish between the landslide truly entering the irreversible accelerated evolution stage and the temporary acceleration phenomenon caused by external disturbances (such as rainfall, construction, seasonal freeze-thaw, etc.), resulting in a large number of false alarms and misreports, which seriously weakens the engineering credibility of the early warning system. There is a fundamental difference between the presence of acceleration and whether it is true acceleration. This invention introduces a true acceleration discrimination mechanism, which accurately identifies the true starting point of the landslide's change from isotropic deformation to accelerated deformation by performing multi-dimensional joint judgment on the persistence, trend, and consistency of deformation stages of acceleration. This effectively eliminates false acceleration signals caused by environmental noise and temporary disturbances, ensuring that the subsequent prediction process is only activated when the landslide truly enters the irreversible evolution stage, thus guaranteeing the reliability of the early warning from the source.

[0008] Second, at the feature extraction level, existing technologies usually use the original displacement monitoring data or the output of the basic physical model as the input of the prediction model. The former lacks the ability to physically represent the evolution law of landslide creep, while the latter produces drastic prediction jumps because the basic model is sensitive to monitoring noise and the early parameters have not converged. This invention constructs a creep feature system with clear physical meaning. It calculates multi-dimensional indicators such as the baseline rate, real-time amplitude of the Fukuzono model, real-time nonlinear exponent, real-time inverse rate cumulative value, and rate-acceleration linearized feature value from real-time monitoring data after true acceleration judgment. This comprehensively characterizes the accelerated evolution of landslides, providing sufficient and effective information for subsequent predictions. Simultaneously, the inverse rate cumulative value quantifies the degree of accelerated evolution, and the rate-acceleration ratio linearly expresses the creep law. Neither of these relies on the unknown instability time, enabling real-time calculation throughout the entire process. Through linearization transformation, the constraint of the unknown instability time in the Fukuzono model is eliminated, achieving real-time stable fitting of creep parameters for undisturbed landslides. This overcomes the technical bottleneck of traditional creep models, which can only be verified post-hoc, and provides an input foundation for subsequent data-driven models that combines physical rationality and individual adaptability. Furthermore, training a machine learning model with these features to output the instability time fully leverages the value of real-time monitoring data, making data-driven predictions physically interpretable.

[0009] Third, regarding model architecture, existing machine learning ensemble methods generally adopt a cascade structure of basic model output and secondary fitting by machine learning. That is, the learning objective of the machine learning model is to correct the prediction results of the basic physical model. This architecture has a fundamental limitation: the prediction accuracy of the machine learning model is highly dependent on the accuracy of the basic model. Once the basic model fails due to poor monitoring data quality, mismatch in evolution stages, or invalid model assumptions, the machine learning model also fails, lacking truly independent prediction capabilities. This invention innovatively constructs a dual-path parallel prediction architecture. One path takes creep characteristics as input and directly outputs the instability time through a machine learning model pre-trained based on multiple sets of historical data of unstable landslides. This model learns the intrinsic mapping relationship between creep evolution laws and instability time, completely independent of any assumptions and calculations of the basic physical model. The other path also independently runs the basic physical model to output the instability time. The two models are completely decoupled in terms of input information, modeling principles, and error sources, independent of each other, and mutually verify each other, achieving a unification of the rationality of geotechnical theory and the reliability of historical data. By setting up a pre-trained model independent of the base model, this model directly outputs predicted values ​​based on the statistical regularity of historical landslides. Even if the base model fails early, it can still provide a stable baseline prediction, effectively suppressing prediction oscillations.

[0010] Fourth, regarding the result fusion strategy, existing technologies employ an uninterpretable implicit integration method, fusing the prediction results of multiple models through a black-box model. This fusion logic lacks physical interpretability and is difficult to verify and debug in engineering practice. This invention employs an interpretable and controllable weighted fusion mechanism to generate the final instability time. A machine learning model trained using creep characteristics learns the historical evolution patterns of multiple instable landslides. During weighted fusion, it can perform data-driven calibration of the theoretical prediction results of the Fukuzono model, effectively correcting theoretical model biases, smoothing prediction fluctuations, and reducing the probability of false alarms. This solves the technical problems of insufficient prediction accuracy and poor robustness of a single theoretical model. The core weight allocation method is equal weighting, balancing the data-driven advantages of machine learning with the rationality of the mechanical theory of the Fukuzono model, effectively reducing the prediction error of a single model and making the results more robust and with clearer physical meaning. Furthermore, this method allows for continuous model iteration and updates as monitoring data accumulates, with the prediction results gradually converging and stabilizing as the landslide evolves, making the landslide instability time prediction truly valuable for engineering early warning. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the real-time monitoring data-driven dynamic prediction method for landslide instability time provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the real-time monitoring data-driven dynamic prediction system for landslide instability time provided in an embodiment of the present invention. Detailed Implementation

[0012] The following detailed explanation illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 As shown: A real-time monitoring data-driven method for dynamic prediction of landslide instability time, including: S1: Real-time acquisition of daily cumulative displacement monitoring data of the target landslide, followed by preprocessing to calculate the daily deformation rate sequence; preprocessing includes multi-point synchronization verification and overall deformation time series synthesis; details are as follows: Preprocessing also includes data cleaning and denoising. Data cleaning uses the 3σ criterion to remove outliers, and denoising uses the 3-day moving average method for smoothing.

[0013] The preprocessing multi-point synchronization verification includes calculating the Pearson correlation coefficient of the deformation rate time series of each measuring point, using the main control measuring point in the middle of the landslide as a benchmark, and screening valid measuring points with synchronous deformation whose correlation coefficients meet the threshold requirements. The screening method includes retaining measuring points with correlation coefficients greater than the first threshold and removing measuring points with correlation coefficients less than the second threshold (determined as local single-point disturbance measuring points and directly removed), where the first threshold is greater than the second threshold (e.g., the first threshold is 0.8 and the second threshold is 0.5). Measuring points with correlation coefficients between the second threshold and the first threshold (e.g., 0.5 to 0.8) are temporarily retained; if their deformation trend deviates from the main control measuring point for N consecutive periods (e.g., 3 periods), they are subsequently removed, and the remaining points continue to participate in the mean synthesis.

[0014] The preprocessing of the overall deformation time series synthesis includes taking the arithmetic mean of the cumulative displacement of the effective measurement points of synchronous deformation, which is used as the overall cumulative displacement time series of the landslide.

[0015] The formula for calculating the daily deformation rate is:

[0016] In the formula, Let t be the daily deformation rate. The cumulative displacement at time t for The cumulative displacement at time -1 is obtained by verifying the synchronization of multiple measurement points in the preprocessing and synthesizing the overall deformation time sequence.

[0017] The benefits of preprocessing for data cleaning, noise reduction, and multi-point synchronization verification are that it can effectively remove abnormal monitoring data and locally disturbed monitoring points, ensure that the deformation time series reflects the overall true motion state of the landslide, and avoid deviations in subsequent creep characteristic calculation and prediction results due to data distortion.

[0018] S2 calculates the rate ratio and daily acceleration based on the daily deformation rate, and uses a fusion judgment mechanism of accelerated creep stage and damage-driven true acceleration to determine true acceleration. If it is determined to be true acceleration, the acceleration starting point is locked and prediction is started, entering S3; otherwise, it is considered false acceleration, and the current prediction is terminated, returning to S1; the details are as follows: rate ratio The calculation formula is:

[0019] daily acceleration The calculation formula is:

[0020] In the formula, Let t be the daily deformation rate at time t; Let t be the daily deformation rate at time t-1.

[0021] The true acceleration determination condition is: the rate ratio over a continuous preset time (e.g., 5 days) is higher than the target rate. Rate than threshold and daily acceleration And it continues to increase. Among them, a large amount of measured data (historical statistical results of creep landslides) shows that R(t)>1.5 for several consecutive days is the statistical critical index for the slip zone shear damage to enter accelerated creep; therefore, 1.5 is the preferred empirical rate ratio threshold, and in actual engineering, it can be adaptively adjusted in the range of 1.3 to 1.8 according to the landslide lithology.

[0022] It should be noted that meeting the rate ratio condition determines that the target landslide has entered the accelerated creep stage. Further meeting the daily acceleration condition determines it to be true acceleration driven by slip zone damage, thus identifying the acceleration starting point. (i.e., the earliest time point that satisfies the true acceleration criteria); if daily acceleration However, the continuous decrease indicates that the rate of deformation is gradually slowing down. The temporary deformation and uplift were only induced by external instantaneous loads such as rainfall and short-term fluctuations in reservoir water levels. The slip zone did not suffer irreversible and continuous damage. This is a false acceleration caused by environmental disturbance. Therefore, this prediction is terminated, and we will continue to wait for the new real-time data of the target landslide to be connected. The calculation will be restarted after the deformation meets the true acceleration conditions again.

[0023] The beneficial effect of true acceleration determination is that it can accurately distinguish between false acceleration caused by external disturbances such as reservoir water and rainfall and true acceleration driven by slip zone damage, avoiding invalid predictions and false warnings, and improving the reliability and pertinence of the prediction system.

[0024] S3, based on monitoring data from the acceleration start point to the current time, calculates preset indices and solves for the real-time parameters of the Fukuzono model without instability time constraints through linear fitting, obtaining the real-time creep characteristics for the day; the real-time creep characteristics include the baseline rate, the real-time amplitude of the Fukuzono model, the real-time nonlinear exponent, the cumulative inverse of the real-time rate, and the linearized rate-acceleration characteristic quantity; as detailed below: Reference rate The arithmetic mean of the deformation rate during the constant-rate creep stage before the acceleration point is used, which is the average rate during the constant-rate creep stage before the landslide acceleration. The deformation rate during the constant-rate creep stage before the acceleration point is determined after removing abnormal fluctuation values.

[0025] The formula for calculating the real-time rate reciprocal cumulative amount (reflecting the total cumulative degree of accelerated landslide evolution) is as follows:

[0026] In the formula, To accelerate the start, For the current moment, for Daily deformation rate.

[0027] Based on the classic Fukuzono model: By taking the derivative of both sides with respect to time, we can obtain:

[0028] The result of the transformation is: ,Right now: .

[0029] Therefore, the linearized characteristic quantity of the velocity-acceleration ratio is determined, i.e., the formula for calculating the velocity-acceleration ratio is:

[0030] The process of solving the real-time parameters of the Fukuzono model without instability time constraints through linear fitting includes: For the linearized characteristic of velocity-acceleration (i.e., the velocity-acceleration ratio) With time The linear equation is obtained by performing least squares linear fitting. , , These are the fitting parameters; Real-time nonlinear exponent ; The time of instability predicted by the fitting on that day ; Substituting the aforementioned calculation results into the Fukuzono rate formula, the real-time amplitude is obtained. :

[0031] The beneficial effect is that the calculation using the classic Fukuzono model is affected by the predicted instability time. Constraints; this embodiment uses the classical Fukuzono model to derive and deform, making With time The relationship exhibits a strictly linear one-dimensional relationship; then, through linearization transformation, the fitting parameters are obtained, and the results are calculated. , and No unknown instability time needs to be given in advance throughout the process. Thus, the constraint of unknown instability time in the Fukuzono model is eliminated through the above calculation process, realizing real-time stable fitting of creep parameters of unstable landslides, and breaking through the technical bottleneck that traditional creep models can only be verified after the fact.

[0032] Real-time creep characteristics include: baseline rate Real-time amplitude of Fukuzono model Real-time nonlinear exponent , rate reciprocal cumulative quantity and speed-acceleration ratio The final output of the real-time creep characteristics for the day is as follows: .

[0033] The beneficial effects are that multi-dimensional real-time creep characteristics can comprehensively characterize the accelerated evolution of landslides, providing sufficient and effective information for subsequent predictions; at the same time, the cumulative inverse of the rate can quantify the degree of accelerated evolution, and the rate-acceleration ratio can linearly express the creep law. Neither of these depends on the unknown instability time, and the entire process can be calculated in real time.

[0034] S4. Based on monitoring data of the accelerated creep phase of similar landslides that have historically become unstable, calculate the corresponding real-time creep characteristics and the actual remaining number of days until instability. Using globally standardized real-time creep characteristics as input and the remaining number of days until instability as output, train a machine learning model to obtain an instability time prediction model; as detailed below: Landslides of this type are categorized into two types: rock landslides and soil landslides. Historical samples are classified according to landslide lithology (rock / soil), and dedicated prediction models are constructed for each type. For rock landslides, training samples are taken from similar rock creep landslides; for soil landslides, training samples are taken from similar soil creep landslides, to improve the model's prediction accuracy for landslides of different lithologies. In practical use, the appropriate prediction model is selected based on the type of the target landslide.

[0035] Construction of a historical landslide sample database and training of a prediction model, including: S41, The time-series training sample library is constructed as follows: For historically unstable creep landslides of the same type (20 in this example), the full process calculation of S1-S3 is repeated for each landslide. Based on the daily monitoring data from the landslide acceleration point to the actual instability time, the real-time creep characteristics of the day are extracted. Simultaneously, the number of days remaining until the actual instability of the landslide is determined by the current date is used as the label value. The label calculation formula is as follows:

[0036] In the formula, The actual date of instability in historical landslides. This allows for arbitrary calculation of dates during the acceleration phase.

[0037] The daily real-time creep characteristics of all landslides are integrated with the corresponding remaining instability days labels to construct a time-series training sample library for creep landslides; the sample library contains complete evolution data of all historical landslide acceleration stages.

[0038] S42, the global feature standardization process is as follows: The Min-Max normalization method is used to globally standardize the real-time creep features of the training sample library, eliminating differences in the magnitude of different features and biases in different landslide evolution scales. The standardization formula is as follows:

[0039] In the formula, These are the original eigenvalues. , These are the global minimum and maximum values ​​of the features corresponding to all samples of the historical landslide; it can be understood that each feature in the real-time creep feature is standardized accordingly.

[0040] After standardization, all feature values ​​in the real-time creep feature are mapped to the [0,1] interval. Standardization is performed only on the input features, and the remaining unstable days label retains the original natural days value without any transformation. The global standardization mapping parameters are saved simultaneously for feature normalization in subsequent target landslide prediction.

[0041] The beneficial effect of global Min-Max normalization of real-time creep features is to eliminate the differences in the magnitude of different features and the bias of different landslide evolution scales, and to ensure that the model learns the contributions of various features fairly.

[0042] S43, the training and test sets are divided as follows: The training and test sets are constructed by independently partitioning the data according to landslides to avoid overfitting caused by leakage of time-series data. The ratio of training to test set samples is 8:2, which significantly improves the model's generalization ability and prediction reliability. In this embodiment, all time-series samples of 16 landslides are randomly selected from 20 unstable landslides as the training set, and all time-series samples of the remaining 4 landslides are used as the test set.

[0043] S44, The machine learning model uses a random forest regression model, and the training and validation are as follows: A random forest regression prediction model is constructed using standardized real-time creep characteristics as model input and the remaining number of days of instability as model output. The core parameter of the model is set as: number of decision trees. =100, maximum depth of a single tree =6, minimum number of sample splits is 2, minimum number of sample leaf nodes is 1; the model is trained using training set samples, and the model accuracy is verified using test set samples, with the determination coefficient as the determining factor. As the core validation metric, the model validation decision coefficient is required. After meeting the engineering prediction accuracy requirements, the final instability time prediction model is obtained.

[0044] In this embodiment, training is performed using a common model architecture in the field, and the standardized real-time creep characteristics are assumed to be... = The total number of decision trees is K, and each decision tree... The output is:

[0045] in, For this decision tree The feature space region corresponding to the mth leaf node; This represents the average number of remaining days for the training samples at that leaf node. M is the indicator function (1 if the sample falls into the region, 0 otherwise); M is the decision tree. Total number of leaves.

[0046] The feature space region is automatically determined by each decision tree during training, based on randomly selected features and the optimal split threshold, through recursive axis-parallel bisection. Ultimately, each leaf node corresponds to a non-overlapping superrectangular region, and the average remaining days of the samples within that region is the predicted output for that region.

[0047] In reality, the region is a set of nested if-then conditions, specifically learned automatically from the samples during model training. For any input sample... = Each decision tree determines its region by: starting from the root node, sequentially comparing the features and thresholds of the current node (e.g., "... (≤ 2.3? , 2.3 is for illustration only). Based on the comparison result, proceed to the left or right child node, repeating this process until a leaf node is reached. All conditions corresponding to this leaf node (overlapping feature inequalities) define a hyperrectangular region in the feature space. A sample falls into this region if and only if it satisfies all conditions on the path. Since the tree structure is predetermined through training, the entire judgment process only involves the comparison of feature values, without the need to explicitly calculate the region boundaries.

[0048] The final number of predicted days for the random forest is: .

[0049] By training the existing model architecture based on the above process, the mapping relationship between input features and output days can be determined.

[0050] S5: Input the real-time creep characteristics of the target landslide on the current day into the trained instability time prediction model to obtain the remaining instability days predicted by machine learning. Combine this with the remaining instability days predicted by the Fukuzono model on the same day for dual-model fusion to obtain the final remaining instability days and predicted instability date in real time; as detailed below: S51, the real-time creep characteristics of the target landslide obtained in S3 for the current day. Substitute the historical landslide global standardized mapping parameters obtained from S42, perform Min-Max normalization processing that is completely consistent with the training samples, and obtain the standardized real-time features of the target landslide on the same day, ensuring that the distribution of the model input data is consistent with the distribution of the training data.

[0051] S52 inputs the globally standardized real-time creep characteristics of the target landslide into the trained random forest prediction model. Through model ensemble inference calculation, it outputs a machine learning-driven prediction of the remaining instability days. The results are based on data-driven fitting of historical landslide evolution patterns and have strong generalization ability.

[0052] S53 uses a dual-model fusion formula:

[0053] In the formula, The final remaining number of instability days obtained from the weighted fusion calculation of the two models; The remaining number of days of instability predicted by machine learning; The remaining number of days of instability predicted based on the Fukuzono model is calculated using the following formula:

[0054] In the formula, This represents the predicted instability time obtained from the fitting in S3 for the current day. This represents the current monitoring date. The result, derived from creep mechanics theory, indicates the upper limit of the predicted number of days remaining until the target landslide becomes unstable, ensuring the physical validity of the prediction.

[0055] and The weighting coefficients, preferably 0.5 and 0.5, form an equally weighted dual-model fusion. This equal-weighted fusion balances the data-driven advantages of machine learning with the theoretical rationale of the Fukuzono model, effectively reducing the prediction error of a single model and making the results more robust and physically meaningful. Of course, in practical applications, the weights can be dynamically adjusted based on the prediction confidence of each model in historical stages or the current deformation stage to generate a final instability time that combines data-driven fitting capabilities with the constraints of the physical model mechanism.

[0056] In other embodiments, models can be added and weights adjusted to achieve weighted fusion of multiple models.

[0057] S54, based on the current monitoring date Remaining instability days until final fusion Calculate the final predicted instability date of the target landslide. The calculation formula is:

[0058] The prediction results are rounded to one decimal place, which meets the accuracy requirements for engineering monitoring and early warning.

[0059] Understandably, this method employs a dynamic prediction strategy, which involves repeating the entire S1-S5 process for newly acquired surface displacement monitoring data each day, updating multi-point synchronous data, rate sequences, real-time creep characteristic parameters, model prediction of remaining days, and the final instability date in real time. This enables the prediction results to be dynamically corrected as the landslide evolves in real time, forming a closed-loop dynamic prediction system.

[0060] The beneficial effect is that the present invention adopts a dual-model fusion prediction method. The Fukuzono model provides a basic prediction that conforms to the theory of creep mechanics, ensuring the physical rationality of the results. The machine learning model learns from a large amount of real-time monitoring data of historical landslides to calibrate the deviation of the theoretical model. This solves the shortcomings of the traditional Fukuzono model, such as accelerating early prediction jumps, weak anti-interference ability, and poor adaptability to individual differences. The dual-model fusion takes into account both theoretical rationality and data-driven engineering practicality, effectively improving the stability and accuracy of the prediction results.

[0061] The following describes the implementation process of this method using a specific example of a soil landslide in a certain region: S1. Data Preprocessing: Daily cumulative displacement data from 2020 to 2024 were collected from 8 GNSS surface deformation monitoring points of the landslide. After outlier removal using 3σ and noise reduction using 3-day moving average, synchronization verification was carried out using the central main control monitoring point as the benchmark. The Pearson correlation coefficient of the deformation rate sequence of each monitoring point was calculated, and 6 valid synchronous deformation monitoring points with a correlation coefficient greater than 0.8 were retained. The average cumulative displacement of these 6 points was taken as the overall deformation time series, and the daily deformation rate was calculated. Complete monitoring data of 20 similar landslides that have already become unstable were collected simultaneously and the same preprocessing was performed.

[0062] S2, True Acceleration Detection: Calculation speed ratio With daily acceleration For five consecutive days starting August 15, 2020 ,and The continuous increase indicates true acceleration driven by slip belt damage, thus locking in the acceleration starting point. It is August 15, 2020.

[0063] S3, Real-time creep characteristic calculation: Reference rate : Select the isostatic rate during the 45 days prior to the acceleration start, remove outliers, and calculate the average value to obtain This is a globally fixed value; Reciprocal cumulative rate :

[0064] Linearized features:

[0065] Fukuzono model parameter fitting: obtained by linear fitting , Calculations yielded , Fitted amplitude .

[0066] Thus, the real-time creep characteristics for that day are obtained: [0.1, 0.85, 1.8, 150, 20].

[0067] S4, Random Forest Model Construction A time-series training sample library was constructed based on 20 unstable landslides. After global Min-Max normalization, the training and test sets were divided in an 8:2 ratio. A random forest regression model was trained, and the decision coefficients of the model were validated. This meets the accuracy requirements for engineering predictions.

[0068] S5, Dual-model fusion prediction Standardization of the creep characteristics of the target landslide on the same day: The real-time characteristics were normalized using historical global parameters; The daily creep characteristics of the target landslide, after global standardization, are input into the trained instability time prediction model (the random forest regression model in this embodiment) to obtain the remaining instability days predicted by machine learning: ; Predict the instability time using the fitting method of the day. , Fukuzono model parameter fitting 183.7 is rounded down to 184. The current calculation time t is 139. Using 184-139=45, the prediction result of the Fukuzono model is obtained: =45 days; Equal weight fusion results: ; Predicted instability date: Current monitoring date + 43.5 days.

[0069] The calculation results are presented with specific dates. t=0 corresponds to August 15, 2020 (the acceleration starting point), representing the time series for the current feature calculation. t=139 corresponds to the actual calculation date of January 1, 2021. The machine learning prediction of the instability date is 42 + 139 days, corresponding to February 12, 2021. The Fukuzono model prediction date is 183.7 rounded down to 184, corresponding to February 15, 2021. The fusion model prediction date is February 13, 2021. It should be noted that the input time series data are daily cumulative displacement monitoring data. t=0 and t=139 represent the time series determined by daily statistical units. The units of measurement for this part of the data are unified with the existing model.

[0070] S6, dynamically updated New monitoring data is entered daily, and the above steps are repeated to correct creep characteristics and prediction results in real time, thereby achieving dynamic tracking and prediction of landslide instability time.

[0071] like Figure 2 As shown, this embodiment also provides a real-time monitoring data-driven dynamic prediction system for landslide instability time, which executes the above-described real-time monitoring data-driven dynamic prediction method for landslide instability time. The system includes: The data processing module is used to collect the daily cumulative displacement monitoring data of the target landslide in real time, and to preprocess and calculate the daily deformation rate sequence. The true acceleration determination module is used to calculate the rate ratio and daily acceleration based on the daily deformation rate, and adopts a true acceleration determination mechanism that combines the accelerated creep stage and damage-driven true acceleration. If it is determined to be true acceleration, the acceleration start point is locked and prediction is started; otherwise, it is a false acceleration and the prediction is terminated. The creep characteristic calculation module is used for response start prediction. Based on the monitoring data from the acceleration start point to the current time, it calculates the preset index and solves the real-time parameters of the Fukuzono model without instability time constraints through linear fitting to obtain the real-time creep characteristics of the day. The prediction model building module is used to calculate the corresponding real-time creep characteristics and the actual remaining number of days of instability based on the monitoring data of the accelerated creep stage of similar landslides that have been unstable in the past. It uses the globally standardized real-time creep characteristics as input and the remaining number of days of instability as output to train a machine learning model and obtain an instability time prediction model. The dual-model fusion prediction module is used to input the real-time creep characteristics of the target landslide into the trained instability time prediction model to obtain the remaining instability days predicted by machine learning. The remaining instability days predicted by the Fukuzono model on the same day are then fused to obtain the final remaining instability days and the predicted instability date in real time.

[0072] It is understandable that the above system can fully execute the above method, with the same process and effect, so it will not be described in detail here.

[0073] The real-time monitoring data-driven dynamic prediction method and system for landslide instability time provided in this embodiment fully explores the value of real-time monitoring data, takes into account the rationality of creep theory and the accuracy of data-driven prediction, and realizes the prediction of landslide instability time that can be continuously updated with monitoring data, significantly improving the engineering practicality of geological disaster early warning.

[0074] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for dynamic prediction of landslide instability time driven by real-time monitoring data, characterized in that, include: S1: Real-time acquisition of daily cumulative displacement monitoring data of the target landslide, and calculation of daily deformation rate sequence after preprocessing; Preprocessing includes multi-point synchronization verification and overall deformation time sequence synthesis; S2: Calculate the rate ratio and daily acceleration based on the daily deformation rate, and use the accelerated creep stage and damage-driven true acceleration fusion judgment mechanism to determine true acceleration. If it is determined to be true acceleration, lock the acceleration start point and start the prediction to enter S3; otherwise, it is false acceleration and terminate the prediction and return to S1. S3, based on the monitoring data from the acceleration start point to the current time, calculate the preset index and solve the real-time parameters of the Fukuzono model without instability time constraints through linear fitting to obtain the real-time creep characteristics of the day; the real-time creep characteristics include the baseline rate, the real-time amplitude of the Fukuzono model, the real-time nonlinear exponent, the real-time rate reciprocal cumulative amount and the rate-acceleration linearized characteristic quantity. S4. Based on the monitoring data of the accelerated creep stage of similar landslides that have been unstable in the past, calculate the corresponding real-time creep characteristics and the actual number of days remaining instability. Use the globally standardized real-time creep characteristics as input and the number of days remaining instability as output to train a machine learning model and obtain an instability time prediction model. S5: Input the real-time creep characteristics of the target landslide on the same day into the trained instability time prediction model to obtain the remaining instability days predicted by machine learning. Combine the remaining instability days predicted by the Fukuzono model on the same day to perform model fusion and obtain the final remaining instability days and predicted instability date in real time.

2. The method for dynamic prediction of landslide instability time driven by real-time monitoring data according to claim 1, characterized in that, In S1, the formula for calculating the daily deformation rate is: In the formula, Let t be the daily deformation rate. The cumulative displacement at time t The cumulative displacement at time t-1 is obtained by combining the multi-point synchronization verification in the preprocessing with the overall deformation time sequence.

3. The method for dynamic prediction of landslide instability time driven by real-time monitoring data according to claim 1 or 2, characterized in that, In S1, the multi-point synchronization verification includes taking the main control point in the middle of the landslide as the benchmark, calculating the Pearson correlation coefficient of the deformation rate time series of each point, and screening out effective synchronous deformation points whose correlation coefficients meet the threshold requirements. The overall deformation time series synthesis includes taking the arithmetic mean of the cumulative displacement of effective measurement points of synchronous deformation, which is used as the overall cumulative displacement time series of the landslide.

4. The method for dynamic prediction of landslide instability time driven by real-time monitoring data according to claim 1, characterized in that, In S2, the rate is higher than The calculation formula is: daily acceleration The calculation formula is: In the formula, Let t be the daily deformation rate. The daily deformation rate at time t-1; The true acceleration determination condition is: the ratio of the continuous preset time rate. Rate than threshold and daily acceleration And it continues to increase; among them, if the rate ratio condition is met, the target landslide is determined to have entered the accelerated creep stage, and if the daily acceleration condition is further met, it is determined to be true acceleration driven by slip zone damage.

5. The method for dynamic prediction of landslide instability time driven by real-time monitoring data according to claim 1, characterized in that, In S3, the reference rate The arithmetic mean of the deformation rate during the constant-rate creep stage before the acceleration start; Real-time rate reciprocal cumulative amount The calculation formula is: In the formula, To accelerate the start, For the current moment, for Daily deformation rate; The velocity-acceleration linearization feature is the velocity-acceleration ratio. The calculation formula is: In the formula, Let t be the daily acceleration. Let t be the daily deformation rate at time t.

6. The method for dynamic prediction of landslide instability time driven by real-time monitoring data according to claim 1, characterized in that, In S3, the process of solving for the real-time parameters of the Fukuzono model without instability time constraints through linear fitting includes: linearized velocity-acceleration characteristic With time Least squares linear fitting was performed to obtain ,in , These are the fitting parameters; Real-time nonlinear exponent ; The time of instability predicted by the fitting on that day ; Substituting the aforementioned calculation results into the Fukuzono rate formula, the real-time amplitude is obtained. : In the formula, Let t be the daily deformation rate at time t.

7. The method for dynamic prediction of landslide instability time driven by real-time monitoring data according to claim 1, characterized in that, In S4, S1-S3 are repeated for each historically unstable landslide of the same type. Based on the daily cumulative displacement monitoring data from the landslide acceleration point to the actual instability time, the real-time creep characteristics of the day are extracted, and the remaining number of instability days from the day to the actual instability time of the landslide is used as the label value. The label calculation formula is as follows: In the formula, The actual date of instability in historical landslides. For any date to be calculated during the acceleration phase; The daily real-time creep characteristics of all landslides are integrated with the corresponding remaining instability days labels to construct a time-series training sample library for creep landslides; training sets and test sets are independently divided according to landslides for model training.

8. The method for dynamic prediction of landslide instability time driven by real-time monitoring data according to claim 1, characterized in that, In S4, the Min-Max normalization method is used to globally standardize the acquired real-time creep features: In the formula, These are the original eigenvalues. , These are the global minimum and maximum values ​​of the corresponding features for all samples of historical landslides, respectively; after standardization, all feature values ​​are mapped to the interval [0,1].

9. The method for dynamic prediction of landslide instability time driven by real-time monitoring data according to claim 1, characterized in that, In S5, the dual-model fusion formula is: In the formula, The final remaining number of instability days obtained from the weighted fusion calculation of the two models; The remaining instability days are predicted using machine learning. The remaining number of days of instability predicted based on the Fukuzono model is calculated using the following formula: In the formula, The time of instability predicted for the current day is t; t is the current monitoring date. Final predicted instability date of the target landslide The calculation formula is: .

10. A real-time monitoring data-driven dynamic prediction system for landslide instability time, characterized in that, The system comprises: a real-time monitoring data-driven dynamic prediction method for landslide instability time according to any one of claims 1-9; the system includes: The data processing module is used to collect the daily cumulative displacement monitoring data of the target landslide in real time, and to preprocess and calculate the daily deformation rate sequence. The true acceleration determination module is used to calculate the rate ratio and daily acceleration based on the daily deformation rate, and adopts a true acceleration determination mechanism that combines the accelerated creep stage and damage-driven true acceleration. If it is determined to be true acceleration, the acceleration start point is locked and prediction is started; otherwise, it is a false acceleration and the prediction is terminated. The creep characteristic calculation module is used for response start prediction. Based on the monitoring data from the acceleration start point to the current time, it calculates the preset index and solves the real-time parameters of the Fukuzono model without instability time constraints through linear fitting to obtain the real-time creep characteristics of the day. The prediction model building module is used to calculate the corresponding real-time creep characteristics and the actual remaining number of days of instability based on the monitoring data of the accelerated creep stage of similar landslides that have been unstable in the past. It uses the globally standardized real-time creep characteristics as input and the remaining number of days of instability as output to train a machine learning model and obtain an instability time prediction model. The model fusion prediction module is used to input the real-time creep characteristics of the target landslide into the trained instability time prediction model to obtain the remaining instability days predicted by machine learning. It then combines the remaining instability days predicted by the Fukuzono model on the same day with the model fusion to obtain the final remaining instability days and the predicted instability date in real time.

Images