Landslide displacement prediction method based on time series data mining

Abstract

The application discloses a landslide displacement prediction method based on time sequence data mining and relates to the fields of geological data processing and disaster prediction, and comprises the following steps: obtaining a three-dimensional coordinate time sequence collected by a displacement sensor and an original sound wave signal collected by a sound wave sensor near a vegetation root system, identifying root fracture of the sound wave signal, counting the number of fractures in a unit time, and constructing a fracture time sequence; combining rate and acceleration time sequence information calculated from the displacement sequence, identifying stage characteristics of a slope body changing from being constrained by the root system to losing the constraint, and switching a displacement prediction model when the root system collapse characteristics are detected, so as to predict the displacement of the landslide body after entering the root system collapse stage. The application solves the problem that it is difficult to reflect sudden sliding in a timely manner in a traditional displacement prediction process by ignoring the change of the root system constraint.

Description

Technical Field

[0001] This invention relates to the fields of geological data processing and disaster prediction, and more specifically, to a landslide displacement prediction method based on time-series data mining. Background Technology

[0002] Landslide displacement prediction is typically based on time series data acquired by displacement sensors. Future displacement trends are inferred by fitting historical displacement patterns. In vegetated slopes, plant roots often penetrate the potential sliding surface and anchor to deeper, stable soil and rock, keeping the slope constrained by roots for extended periods. This constraint not only increases the shear strength of the soil but also limits the relative displacement of the sliding surface through the tensile force of the root fibers. Since soil and rock masses are often brittle or elastoplastic, while plant root fibers are highly resilient, as external loads gradually increase, slope displacement may be continuously suppressed within a very small range by the roots. Displacement curves monitored near plant roots tend to appear approximately flat or exhibit slow creep over extended periods.

[0003] Existing displacement sequence-based prediction methods typically assume that slope constraints remain constant throughout the prediction period, focusing on learning stable temporal evolution patterns from historical displacement changes, while rarely considering changes in the internal constraint state of the slope. As a slope gradually transitions from being constrained by plant roots to losing its constraints, the internal root system often undergoes a process of gradual accumulation from a few micro-fractures to concentrated fractures. Once the root constraint experiences a chain reaction of attenuation within a short period, the slope may rapidly enter an accelerated sliding state. Currently, there is no specific prediction model to predict displacement for this phenomenon. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a landslide displacement prediction method based on time series data mining, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A landslide displacement prediction method based on time-series data mining includes:

[0007] Acquire the three-dimensional coordinate time sequence collected by the displacement sensor, and the raw acoustic signal collected by the acoustic sensor installed near the root system of vegetation.

[0008] The original acoustic signal is input into a pre-trained root fracture recognition model to identify signal segments in the original acoustic signal that match the fracture characteristics of root fibers, and outputs the number of fractures per unit time to generate a fracture time sequence.

[0009] The landslide body's velocity and acceleration time series are calculated based on the three-dimensional coordinate time series. If the fracture time series shows an increasing trend and the velocity time series is lower than a preset creep threshold, the landslide body is determined to have entered the transition stage from the root constraint stage to the root collapse stage.

[0010] If it is further detected that the fracture time series drops in value after reaching a local peak, and is accompanied by a positive step in the acceleration time series, then the landslide body is identified as being in the root collapse stage, and the displacement prediction logic is triggered to switch from the baseline prediction model to the accelerated sliding prediction model.

[0011] The three-dimensional coordinate time series and the fracture time series are input into the accelerated slip prediction model, and the displacement prediction of the landslide body after entering the root system failure stage is output.

[0012] Preferably, the root fracture identification model is a one-dimensional time-series classification model based on supervised learning, which includes a sound wave signal preprocessing layer, a one-dimensional convolutional feature extraction layer, and a fracture event discrimination layer connected in sequence.

[0013] The acoustic signal preprocessing layer is used to perform frame division and amplitude normalization on the original acoustic signal.

[0014] The one-dimensional convolutional feature extraction layer is used to extract time-domain or frequency-domain features related to root fiber fracture from the time series of the acoustic signal.

[0015] The fracture event discrimination layer outputs a discrimination result based on the feature extraction results, indicating whether root fracture has occurred within the corresponding time slice.

[0016] Preferably, the training process of the root fracture identification model includes: using acoustic signal samples labeled with the root fracture time as training data to perform supervised training on the model parameters.

[0017] Preferably, the identification of the number of fractures includes the following steps:

[0018] The root fracture identification results output by the root fracture identification model are arranged in chronological order, and the continuous acoustic signal is divided into multiple time statistical windows based on a preset time resolution.

[0019] Within each time statistical window, the acoustic signal segment that is determined to have caused root fracture is counted as a fracture event.

[0020] When the time interval between adjacent acoustic signal segments is less than the preset minimum event interval, the adjacent acoustic signal segments are combined and counted as the same breakage event.

[0021] The number of break events within each time window is used as the break count for the corresponding time period.

[0022] Preferably, the accelerated slip prediction model is a time-series-based supervised learning model, which is constructed using a neural network structure or a tree-based ensemble learning structure. The neural network structure is any one of a recurrent neural network, a gated recurrent unit network, or a long short-term memory network; the tree-based ensemble learning structure is a temporal random forest model.

[0023] Preferably, the training process of the accelerated slip prediction model includes: extracting multi-source time-series data segments from historical landslide samples after the landslide body enters the root system failure stage, and using the actual displacement changes within the corresponding time period as supervision labels to train the model parameters.

[0024] Preferably, the determination of the numerical drop includes the following steps: within a preset time window after the fracture time sequence reaches a local peak, the average number of fractures within the preset time window is calculated; when the decrease of the average value relative to the local peak exceeds a preset proportional threshold, a numerical drop is determined to have occurred.

[0025] Preferably, the positive step refers to the situation where, within a preset time window, the value of the acceleration sequence abruptly increases from a first value range less than a first preset threshold to a second value range greater than a second preset threshold, and the difference in acceleration values ​​before and after the abrupt change is greater than a preset step amplitude threshold.

[0026] Preferably, the benchmark prediction model is used to characterize the displacement evolution law of the landslide body under the root constraint stage. The benchmark prediction model constructs the displacement-time relationship based on the three-dimensional coordinate time series. The benchmark prediction model is any one of a linear regression model, an exponential creep model, or a low-order time series model.

[0027] Preferably, it also includes a landslide early warning step, wherein the landslide early warning includes a first-stage early warning and a second-stage early warning;

[0028] The first stage warning is triggered when the fracture time sequence is detected to change from a zero value state to a continuous non-zero fracture count, which is used to indicate that fracture behavior has begun to occur in the vegetation root system.

[0029] The second-stage warning is triggered when the value of the fracture time series drops after reaching a local peak, and the positive step of the acceleration time series is detected simultaneously. This is used to indicate that the vegetation root fracture process has been basically completed and the landslide body has entered the accelerated movement stage.

[0030] The advantage of this invention over existing technologies lies in its ability to address the difficulty of traditional displacement prediction methods in depicting the transition of a slope from root-constrained to unconstrained states. This invention incorporates acoustic information from near vegetation roots into the displacement time-series data. By identifying and statistically analyzing acoustic signals related to root fracture, a fracture time-series sequence reflecting changes in root constraint is constructed. When the slope is in the root-constrained stage, even with small displacement changes, the fracture time-series sequence still exhibits a gradual, cumulative trend, providing additional information for identifying potential constraint weakening processes. When the fracture time-series sequence reaches a local peak and shows a significant decrease, while the acceleration time-series sequence undergoes a sudden change, it indicates that the slope has transitioned from a root-constrained dominant state to a dynamically controlled state after unconstraint. Based on the above stage identification results, this invention introduces a prediction model switching mechanism during the prediction process. During the root-constrained stage, a baseline prediction model based on historical displacement evolution is used; after identifying the root collapse stage, the model switches to a prediction model trained for accelerated slip characteristics. By defining the statistical rules for the number of fractures, the determination methods for numerical drops and positive step jumps, and setting the early warning trigger conditions for different stages, this invention can provide clear and distinct early warning information at key nodes such as the initial appearance of root fractures and the completion of collapse, reducing the risk of prediction lag caused by the sluggishness of displacement sequences, and improving the applicability and reliability of landslide displacement prediction results in the abrupt change stage. In particular, the detection and early warning of displacement sensors near plant roots will be more accurate. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the present invention;

[0032] Figure 2 This is a schematic diagram of the acoustic wave detection of the present invention;

[0033] Figure 3 This is a schematic diagram of the forward transition of the present invention;

[0034] Figure 4 This is a schematic diagram of the software interface for the monitoring process of this invention. Detailed Implementation

[0035] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0036] The purpose of this invention is to address the issue that, in the early stages, the soil near the root system is held in place by the roots like reinforcing bars, and the displacement curve may remain nearly flat for a long time. However, fiber breakage within the root system occurs first and gradually accumulates. Only after the breakage is complete does the landslide enter a significant accelerated sliding phase. To transform this process of breakage followed by acceleration into a predictive process that is calculable, predictable, and allows for model switching, this invention combines displacement time series with root breakage time series analysis. This allows for early risk identification before displacement becomes significant and switching to a more suitable predictive model once the acceleration phase begins. Figure 1As shown.

[0037] The monitoring system includes displacement sensors and acoustic sensors. Displacement sensors acquire three-dimensional coordinate time-series sequences of monitoring points on the landslide, while acoustic sensors are deployed near vegetation roots to acquire raw acoustic signals, such as... Figure 2 As shown. The computing side is used to perform acoustic wave recognition, fracture counting, velocity and acceleration calculation, stage determination, prediction model switching, displacement prediction output, and early warning output. It can also display key curves and alarm status through a software interface, such as... Figure 4 As shown.

[0038] The three-dimensional coordinate time series is recorded in chronological order as P(t) = [x(t), y(t), z(t)], where P(t) is the position vector at time t, and x, y, and z are the coordinate values ​​in the three directions. To make the subsequent velocity and acceleration more stable, the acquired coordinate series can be smoothed first. Smoothing methods can include moving average, Savitzky-Golay smoothing, or Kalman filtering. The sliding window length can be 3 to 31 sampling points. The displacement can be calculated using one of the following methods according to engineering requirements:

[0039] The first method calculates the displacement modulus D(t) = |P(t) - P(t0)| based on the initial moment, which is suitable for overall displacement evaluation that is not sensitive to direction. t0 is the reference moment used to define the starting point for displacement calculation.

[0040] The second approach is to calculate the projected displacement D(t) = [P(t) - P(t0)]·u along the sliding direction if the main sliding direction vector u is known, so that the velocity and acceleration are closer to the actual sliding direction.

[0041] After obtaining the displacement sequence, calculate the velocity and acceleration time sequences. Differential calculations can be used.

[0042] t k The velocity V(t) at time t k ) is: V(t) k )=[D(t k )-D(t k-1 )] / Δt;

[0043] t k acceleration A(t) at time t k ) is: A(t) k )=[V(t k )-V(t k-1 )] / Δt;

[0044] Where, Δt = t k -t k-1 . t k For the k-th time, t k-1This is the (k-1)th time step.

[0045] To avoid differential amplification noise, low-pass or median filtering can be applied to D, V, or A before or after differential amplification. The low-pass cutoff frequency can be from 0.001 Hz to 0.1 Hz. Subsequent specific embodiments also include setting a preset creep threshold. The purpose is to distinguish between the low-speed creep during the root constraint stage and the accelerated slip after root failure, thereby avoiding misjudgment of safety when the displacement remains relatively flat for a long period. The rate threshold can be determined by field calibration or historical sample statistics, and can be from 0.01 mm / h to 5 mm / h, or from 0.1 mm / d to 50 mm / d on a daily scale.

[0046] As shown in Figure 2, the acoustic signal is acquired by an acoustic sensor. The sampling rate can range from 1 kHz to 200 kHz, with 20 kHz to 100 kHz being commonly used to cover the transient signals generated by root fiber breakage. To suppress environmental noise, a bandpass filter can be applied, with a bandpass range of 50 Hz to 20 kHz, or it can be adaptively adjusted according to the background noise spectrum. Subsequently, the acoustic wave is framed and its amplitude is normalized. The frame length can range from 5 ms to 200 ms, and the frame shift can range from 1 ms to 100 ms. The purpose of framing is to transform the continuous signal into a localizable time slice, providing a basis for subsequently outputting the time location and counting of breakage occurrences.

[0047] The root fracture identification model is a supervised learning one-dimensional temporal classification model, which structurally includes an acoustic signal preprocessing layer, a one-dimensional convolutional feature extraction layer, and a fracture event discrimination layer.

[0048] A one-dimensional convolutional feature extraction layer is used to extract time-domain or frequency-domain features related to root fiber fracture, and can be implemented as follows:

[0049] You can directly input time-domain waveforms and extract pulse-type transient features through multi-layer 1D convolution and pooling;

[0050] Alternatively, feature sequences such as short-time energy, zero-crossing rate, spectral centroid, spectral entropy, or Mel spectrum can be calculated first, and then higher-order temporal textures can be extracted using 1D convolution.

[0051] A multi-scale convolutional kernel parallel structure can also be adopted, with the kernel size taking several options from 3, 5, 9, and 15, to take into account both short breaks and longer frictional acoustic emissions.

[0052] Furthermore, the number of convolutional layers can be 2 to 8, and the number of channels per layer can be 16 to 256. To improve generalization ability, batch normalization and random deactivation can be added between convolutional layers, with the random deactivation ratio being 0.1 to 0.5. The fracture event discrimination layer outputs whether root fracture has occurred within the corresponding time slice. It can use a fully connected layer with a sigmoid function to output binary classification probabilities, or a softmax function to output the probability distribution of fracture and non-fracture. The loss function can be cross-entropy loss. If fracture samples are significantly scarce, cross-entropy with class weights can be used to alleviate class imbalance. The optimization algorithm can be Adam or SGD, with a learning rate of 0.00001 to 0.01 and a batch size of 16 to 512. The intention of adopting the above structure is to capture the abrupt changes in fracture sound using relatively stable one-dimensional convolutional features and output fracture discrimination results that can be used for counting and trend analysis.

[0053] The model is trained under supervised training using acoustic samples labeled with the root fracture time. Labeling sources can include manual listening combined with time-frequency plots, synchronous recordings from root tensile failure tests, or event times obtained through multi-sensor consistency verification. To adapt the model to different soil types and root systems, data augmentation can be performed, such as amplitude scaling, temporal perturbations, additive noise, and frequency band perturbations, thereby improving robustness in complex field environments.

[0054] Furthermore, the fracture time sequence is obtained by counting the fracture discrimination results. First, the fracture discrimination results are arranged in chronological order, and then the continuous sound wave is divided into multiple time statistical windows according to a preset time resolution. The window length can be from 0.5 s to 120 s, and is commonly used from 1 s to 30 s. Within each statistical window, the sound wave signal segment determined to have fractured is counted as one fracture event. To avoid the same fracture being divided into multiple segments and counted repeatedly, when the time interval between adjacent sound wave signal segments is less than a preset minimum event interval, the adjacent segments are merged and counted as the same fracture event. The minimum event interval can be from 0.001 s to 2 s, and is commonly used from 0.01 s to 0.5 s. The number of fracture events within each statistical window is the number of fractures in that time period, arranged chronologically to form the fracture time sequence. The purpose of this processing is to transform the high-frequency sound wave identification results into a low-frequency event counting curve that can be aligned with displacement, velocity, and acceleration, facilitating stage determination.

[0055] Furthermore, the stage determination is carried out according to the root system constraint logic. First, it is detected whether the fracture time series shows an increasing trend, and at the same time, it is determined whether the rate time series is lower than the above-mentioned preset creep threshold. If both conditions are met, the landslide body is considered to have entered the transition stage from the root system constraint stage to the root system collapse stage.

[0056] An increasing trend can be achieved in several ways: a linear fit can be performed on the number of fractures within a sliding window, and an increase in the slope is considered sufficient; a Mann-Kendall trend test can also be performed; or an exponential moving average can be used to compare the current value with the historical mean and combine it with the standard deviation threshold. The purpose of introducing the parallel judgment of increasing trend and low-velocity constraints is to identify the fact that internal root damage is accumulating rapidly in advance, even when the displacement is still suppressed by the root system and the rate is still low.

[0057] When the root system enters the collapse stage, the fracture activity typically exhibits an initial strengthening followed by attenuation, while the landslide movement begins to accelerate significantly. Therefore, the synchronicity of two conditions is further examined: first, the fracture time series experiences a numerical drop after reaching a local peak; second, the acceleration time series experiences a positive step. When both conditions are simultaneously met, the landslide is identified as being in the root collapse stage, and the displacement prediction logic is triggered to switch from the baseline prediction model to the accelerated sliding prediction model, such as... Figure 3 As shown. The purpose of setting the switch is to avoid misapplying the accelerated model to approximately stationary data during the root constraint phase, which would cause false alarms. At the same time, after entering the accelerated glide phase, a more suitable predictor is used in a timely manner to improve the ability to track short-term rapid changes.

[0058] In a specific embodiment, the determination of numerical drop can be implemented as follows: After the fracture time sequence reaches a local peak, a preset time window is set, and the average number of fractures is calculated within the window. When the decrease of this average value relative to the local peak exceeds a preset proportional threshold, a numerical drop is determined to have occurred. The preset time window length can be from 5 s to 600 s, and the proportional threshold can be from 0.2 to 0.9, with 0.3 to 0.7 being commonly used. The purpose of adopting the logic of drop after peak is to express the process of root fracture from concentrated occurrence to decay completion with a quantifiable index.

[0059] Furthermore, a positive step is used to characterize the abrupt change in acceleration from near zero or a low value to a high value range. For example... Figure 3 As shown. The determination method is as follows: within a preset time window, the acceleration value abruptly increases from a first value range less than a first preset threshold to a second value range greater than a second preset threshold, and the difference in acceleration before and after the abrupt change is greater than a preset step amplitude threshold. The preset time window can be 1 to 30 sampling points. The first preset threshold can be 0.000001 m / s² to 0.001 m / s², the second preset threshold can be 0.0001 m / s² to 0.1 m / s², and the step amplitude threshold can be 0.0001 m / s² to 0.5 m / s². The above thresholds can be calibrated according to the station noise level and monitoring scale, or automatically given by the quantiles of historical samples. The purpose of defining the step in this way is to exclude slow changes and noise fluctuations, and only respond to true acceleration abrupt changes.

[0060] In a further embodiment, the baseline prediction model is used to characterize the displacement evolution under root constraint conditions. Essentially, it constructs the displacement-time relationship using a three-dimensional coordinate time series. The model type can be any of a linear regression model, an exponential creep model, or a low-order time series model. Linear regression is suitable for short-term, approximately linear, slow changes; exponential creep models are suitable for scenarios with a clear creep curve; and low-order time series models can be AR models, ARMA models, or exponential smoothing models. The significance of the baseline model lies in providing a stable prediction baseline during periods of small and approximately stable displacement changes, and maintaining a conservative, low-false-prone prediction logic when collapse conditions are not triggered.

[0061] After the model switch is triggered, an accelerated slip prediction model is used to predict the displacement after entering the root failure stage. This model is a time-series-based supervised learning model, which can be structured as a neural network or a tree-based ensemble learning structure. The neural network structure can be any of the following: recurrent neural network, gated recurrent unit network, or long short-term memory network. The number of recurrent layers can be 1 to 4, the number of hidden units can be 16 to 512, and the sequence input length can be 10 to 200 time steps. The input features must include at least the three-dimensional coordinate time series and the failure time series, and can also include derived features such as the moving average and slope of the rate, acceleration, and number of failures to enhance the fitting of the acceleration stage. The tree-based ensemble learning structure can be a temporal random forest model. A common implementation is to construct a feature vector with lag terms from the sequence. The lag order can be 3 to 200, the number of trees can be 50 to 1000, and the maximum depth can be 3 to 30. It allows for the simultaneous training of two types of models in engineering, with the option to choose one based on validation performance or computational resources during the online phase, or to output a weighted fusion result, thus balancing nonlinear fitting capability and robustness to small samples.

[0062] The accelerated landslide prediction model is trained using historical landslide sample data. The training data extracts multi-source time-series data segments from historical samples after the landslide body enters the root collapse stage, using the actual displacement changes within the corresponding time period as supervision labels. The starting point of the extraction can be manually labeled or obtained by replaying historical data using the stage identification logic of this invention. The supervision label can be defined as a sequence of displacement increments in the next H steps or a displacement value at a future moment, where H can range from 1 to 200. The loss function can employ mean squared error, Huber loss, or quantile loss to account for both anomalous acceleration points and normal segments. When dividing the training and validation sets, it is advisable to divide them by event to avoid adjacent segments of the same landslide event appearing simultaneously in both training and validation, thereby reducing inflated assessments due to information leakage.

[0063] In another embodiment, the accelerated slip prediction model can employ a temporal network structure based on a multi-scale attention mechanism, such as the MSA-TimesNet model, as an extension of the neural network structure. This model captures multi-periodic and time-varying features by converting one-dimensional temporal data into a two-dimensional tensor representation, thereby better adapting to the accelerated slip dynamics after the root breakage stage. Specifically, the MSA-TimesNet model includes a temporal embedding layer, a multi-scale transformation layer, an attention fusion layer, and a prediction output layer. The temporal embedding layer initially embeds the input three-dimensional coordinate temporal sequence and the breakage temporal sequence to generate a feature representation suitable for multi-scale analysis. The multi-scale transformation layer performs two-dimensional transformation and convolution operations on the input data by processing subsequences of different periods in parallel to extract the periodic variation features hidden at different time scales. The attention fusion layer, based on a multi-head self-attention mechanism, weights and fuses the features extracted from multiple scales to highlight key time-varying patterns and suppress noise interference. The prediction output layer generates predicted values ​​for future displacements through a fully connected structure. The training process for this model is similar to that described above. It extracts multi-source time-series data segments from historical landslide samples after the root system collapse stage and uses actual displacement changes as supervision labels for parameter optimization. By introducing multi-scale attention, this embodiment can improve prediction accuracy under complex time-series patterns, and is particularly suitable for scenarios with dense fracture events and abrupt accelerations in displacement, while maintaining compatibility with the baseline prediction model.

[0064] After the switch is completed, the three-dimensional coordinate time series and the fracture time series are input into the accelerated slip prediction model, and the displacement prediction results of the landslide body after entering the root system collapse stage are output. The output can be a displacement curve over a future period of time, or the displacement values ​​and growth rates at several key moments. The prediction time range can be from 10 minutes to 72 hours. To facilitate emergency decision-making, the prediction uncertainty can be output simultaneously. The uncertainty can be obtained from the model integration variance, Monte Carlo random inactivation, or quantile regression output, so that the alarm strategy automatically becomes more conservative under high uncertainty.

[0065] In terms of early warning, a first-stage warning and a second-stage warning are set up. The first-stage warning is triggered when the fracture time series changes from a state of 0 to a state of continuously showing non-zero fracture counts, indicating that fracture behavior has begun to appear in the root system. The continuous determination can adopt the rule that the number of fractures is greater than 0 for N consecutive statistical windows, where N can be 2 to 20. The second-stage warning is triggered when the fracture time series reaches a local peak and then the value drops, and a positive step is simultaneously detected in the acceleration time series, indicating that the root fracture process has been basically completed and the landslide has entered the accelerated movement stage. To reduce false alarms, consistency constraints can be added, such as requiring the acceleration to be continuously greater than a second preset threshold within M sampling points after the step, where M can be 2 to 30, or requiring the velocity to exceed the creep threshold by a multiple factor after the step, where the multiple factor can be 2 to 50.

[0066] The above implementation transforms the displacement insensitivity problem caused by root constraints into an observable fracture event time series. It also constructs stage identification and model switching logic using fracture increment, peak drop and positive acceleration step, so that the system can provide early signals when the displacement is not obvious, and can also output displacement prediction and graded early warning by using a more suitable predictor after entering the acceleration stage.

[0067] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A landslide displacement prediction method based on time-series data mining, characterized in that, include: Acquire the three-dimensional coordinate time sequence collected by the displacement sensor, and the raw acoustic signal collected by the acoustic sensor installed near the root system of vegetation. The original acoustic signal is input into a pre-trained root fracture recognition model to identify signal segments in the original acoustic signal that match the fracture characteristics of root fibers, and outputs the number of fractures per unit time to generate a fracture time sequence. The landslide body's velocity and acceleration time series are calculated based on the three-dimensional coordinate time series. If the fracture time series shows an increasing trend and the velocity time series is lower than a preset creep threshold, the landslide body is determined to have entered the transition stage from the root constraint stage to the root collapse stage. If it is further detected that the fracture time series drops in value after reaching a local peak, and is accompanied by a positive step in the acceleration time series, then the landslide body is identified as being in the root collapse stage, and the displacement prediction logic is triggered to switch from the baseline prediction model to the accelerated sliding prediction model. The three-dimensional coordinate time series and the fracture time series are input into the accelerated slip prediction model, and the displacement prediction of the landslide body after entering the root system failure stage is output.

2. The landslide displacement prediction method based on time-series data mining according to claim 1, characterized in that: The root fracture identification model is a one-dimensional temporal classification model based on supervised learning, which includes a sound wave signal preprocessing layer, a one-dimensional convolutional feature extraction layer and a fracture event discrimination layer connected in sequence. The acoustic signal preprocessing layer is used to perform frame division and amplitude normalization on the original acoustic signal. The one-dimensional convolutional feature extraction layer is used to extract time-domain or frequency-domain features related to root fiber fracture from the time series of the acoustic signal. The fracture event discrimination layer outputs a discrimination result based on the feature extraction results, indicating whether root fracture has occurred within the corresponding time slice.

3. The landslide displacement prediction method based on time-series data mining according to claim 2, characterized in that: The training process of the root fracture identification model includes: using acoustic signal samples labeled with the root fracture time as training data to supervise the training of model parameters.

4. The landslide displacement prediction method based on time-series data mining according to claim 2, characterized in that: The identification of the number of fractures includes the following steps: The root fracture identification results output by the root fracture identification model are arranged in chronological order, and the continuous acoustic signal is divided into multiple time statistical windows based on a preset time resolution. Within each time statistical window, the acoustic signal segment that is determined to have caused root fracture is counted as a fracture event. When the time interval between adjacent acoustic signal segments is less than the preset minimum event interval, the adjacent acoustic signal segments are combined and counted as the same breakage event. The number of break events within each time window is used as the break count for the corresponding time period.

5. The landslide displacement prediction method based on time-series data mining according to claim 1, characterized in that: The accelerated slip prediction model is a time-series-based supervised learning model, which is constructed using a neural network structure or a tree-based ensemble learning structure. The neural network structure is any one of a recurrent neural network, a gated recurrent unit network, or a long short-term memory network; the tree-based ensemble learning structure is a temporal random forest model.

6. The landslide displacement prediction method based on time-series data mining according to claim 5, characterized in that: The training process of the accelerated slip prediction model includes: extracting multi-source time-series data segments from historical landslide samples after the landslide body enters the root system failure stage, and using the actual displacement changes within the corresponding time period as supervision labels to train the model parameters.

7. The landslide displacement prediction method based on time-series data mining according to claim 1, characterized in that, The determination of the numerical drop includes the following steps: within a preset time window after the fracture time sequence reaches a local peak, the average number of fractures within the preset time window is calculated. When the decrease of the average value relative to the local peak exceeds a preset proportional threshold, a numerical drop is determined to have occurred.

8. The landslide displacement prediction method based on time-series data mining according to claim 1, characterized in that: The positive step refers to the situation where, within a preset time window, the value of the acceleration time sequence changes abruptly from a first value range less than a first preset threshold to a second value range greater than a second preset threshold, and the difference in acceleration values ​​before and after the abrupt change is greater than a preset step amplitude threshold.

9. The landslide displacement prediction method based on time-series data mining according to claim 1, characterized in that: The benchmark prediction model is used to characterize the displacement evolution law of the landslide body under the root constraint stage. The benchmark prediction model constructs the displacement-time relationship based on the three-dimensional coordinate time series. The benchmark prediction model is any one of the linear regression model, exponential creep model or low-order time series model.

10. The landslide displacement prediction method based on time-series data mining according to claim 1, characterized in that: It also includes a landslide early warning procedure, which includes a first-stage early warning and a second-stage early warning. The first stage warning is triggered when the fracture time sequence is detected to change from a zero value state to a continuous non-zero fracture count, which is used to indicate that fracture behavior has begun to occur in the vegetation root system. The second-stage warning is triggered when the value of the fracture time series drops after reaching a local peak, and the positive step of the acceleration time series is detected simultaneously. This is used to indicate that the vegetation root fracture process has been basically completed and the landslide body has entered the accelerated movement stage.

Images