Search

CN-122020223-A - PatchTST model landslide risk assessment method based on multi-period rainfall events

CN122020223ACN 122020223 ACN122020223 ACN 122020223ACN-122020223-A

Abstract

The invention discloses a PatchTST model landslide risk assessment method based on a multi-period rainfall event, which comprises data preprocessing and space-time alignment, multi-period rainfall event segmentation, effective rainfall event type segmentation, double-branch PatchTST model construction, displacement prediction and risk assessment, and relates to the technical field of geological disaster monitoring and early warning. According to the PatchTST model landslide risk assessment method based on the multi-period rainfall events, the causal relationship of rainfall event-displacement response is established, effective rainfall is screened through the association coefficient, invalid data is removed, the signal to noise ratio of model training is remarkably improved, a real geological mechanism is learned by the model, a prediction result is more credible, and a PatchTST model with a storm branch and a overcast and rainy branch is designed, so that the model can capture displacement surge caused by the storm and can characterize displacement accumulation caused by continuous and overcast rain, and the distinguishing and modeling capacity of rainfall types is improved.

Inventors

  • YANG TAO
  • TAN RUI
  • Quan Jiatao
  • XIANG XIANG

Assignees

  • 江苏省地质局第四地质大队

Dates

Publication Date
20260512
Application Date
20260128

Claims (8)

  1. 1. A PatchTST model landslide risk assessment method based on a multi-period rainfall event is characterized by comprising the following steps: step one, data preprocessing and space-time alignment, namely collecting GNSS displacement time series data and rainfall time series data of a target landslide area Denoising and missing value complement processing is carried out on the data, and time-space alignment of two types of data is achieved through time stamp matching, so that the data resolution is ensured to be consistent; Step two, multi-period rainfall event segmentation, namely based on rainfall intensity threshold value And a threshold T of no rain interval, for the rainfall sequence after pretreatment Carrying out multi-period rainfall event identification and segmentation, and calculating displacement rate of landslide in rainfall period at each moment Speed and velocity of And acceleration Solving the optimal rainfall influence lag time according to Identifying ineffective rainfall events and effective rainfall events according to the sizes, and performing backtracking association to obtain the effective rainfall events and constructing an event set; dividing the effective rainfall event types according to the storm intensity threshold value given by the local government Distinguishing a heavy rain event sequence from a continuous overcast event sequence to obtain m independent rainfall event sequences (including heavy rain sequences {) 、 ... And the sequence of continuous overcast and rain { 、 ... }, + =M) and corresponding displacement response sequence { 、 ... Finally, completing the cluster model training of the rainfall event; Constructing a double-branch PatchTST model, namely taking an independent rainfall event sequence as input, and constructing a complete double-branch PatchTST model through the steps of double-branch feature sequence construction, sequence blocking and projection, position coding, independent feature extraction of a transducer encoder, cross-branch attention fusion, model training and the like; inputting rainfall data characteristics of a period to be predicted into a clustering model, identifying a heavy rainfall event, a overcast rainfall event and an ineffective rainfall event, outputting a landslide displacement prediction result by the effective rainfall event through a double-branch PatchTST model, calculating a risk value by combining a fuzzy comprehensive evaluation method, dividing early warning grades, outputting an early warning result in real time, and providing corresponding measure opinion.
  2. 2. The method for landslide risk assessment based on PatchTST model of multi-period rainfall event according to claim 1, wherein the specific process of data preprocessing in the first step comprises: a1, acquiring displacement data through a GNSS receiver, acquiring rainfall data through a small weather station, and acquiring environmental data by adopting a temperature and humidity sensor and a soil humidity sensor, wherein the acquisition frequency is 1 time per hour; a2, data standardization, namely adopting Z-score standardization to process data, wherein the formula is as follows: , wherein, As the raw data is to be processed, In order to make the data after the normalization, As a mean value of the data, Data standard deviation; a3, supplementing the missing value, namely supplementing by using a linear interpolation method when a single data point is missing and the front and rear data are effective, and supplementing by using the linear interpolation method, wherein the formula is If more than 3 sampling points are continuously deleted, the displacement data average value under the similar weather condition of the same period of history is combined for complementation, so that analysis deviation of subsequent events caused by data deletion is avoided; and a4, space-time alignment, namely taking a timestamp in a 'YYYY-MM-DD HH: MM: SS' format as a reference, and matching the processed GNSS displacement and rainfall data point by point, so as to ensure that each time node corresponds to a unique displacement value and rainfall value, and the unified data resolution is 1 hour.
  3. 3. The method for evaluating landslide risk based on PatchTST model of multi-period rainfall event according to claim 1, wherein the multi-period rainfall event segmentation and risk transformation based on displacement response in the second step specifically comprises: B1, recording rainfall time and displacement parameter calculation, namely recording rainfall and non-rainfall time according to a rainfall gauge, combining historical landslide displacement and rainfall data, determining maximum rainfall delay time through statistical analysis, wherein the rainfall event time is rainfall start time plus delay time, and according to the preprocessed displacement data, according to a formula Calculating a sequence of displacement velocities (Unit: mm/h), according to the formula Calculating a displacement acceleration sequence (Unit: mm/h 2), wherein Hours; b2, judging risk event when With 3 consecutive sampling points (i.e. 3 hours) exceeding And the displacement is increased in the period If the displacement acceleration event is determined to be a displacement acceleration event, the starting time of the event is recorded And end time ( Is that For the first time lower than Is a time of (2); b3, calculating rainfall lag time and event association degree, namely, the influence of rainfall on landslide displacement is not immediate, but has obvious lag effect, wherein the lag is mainly caused by time consumption of a process of rainwater infiltration into a landslide body, accumulation of water content of a rock-soil body, and progressive deterioration of pore water pressure change and mechanical properties of the rock-soil body, if the lag effect is ignored, direct association modeling of rainfall data and contemporaneous displacement data can lead to causal relation dislocation of rainfall and displacement acceleration events, and ineffective rainfall interference is introduced, and model prediction precision is reduced, so that rainfall lag time optimization is a key link for accurately constructing 'rainfall event-displacement response' association and screening effective rainfall events, and the specific steps are as follows: b4, lag time set definition Based on the rapid infiltration effect of short duration heavy rain, the slow infiltration effect of displacement response and long duration continuous overcast and overcast rain is usually induced in a few hours, displacement change is shown after accumulation for a few days, and a lag time set is defined On the premise of ensuring the coverage integrity, the discrete value taking 6 hours as an interval is adopted, so that the calculation complexity is reduced, and the optimization efficiency is improved; b5, associated time node calibration For each identified displacement acceleration event i (the start time of which is The end time is ) Based on each candidate value in the set of lag times Calculating corresponding rainfall association start time It should be clear that The central logic of the method is that the occurrence of the displacement acceleration event is essentially the accumulated effect trigger of rainfall in a certain period in the earlier stage, and the delay time is used for the generation of the displacement acceleration event Backtracking, namely accurately matching a displacement acceleration event with a rainfall period actually causing the event, so as to eliminate ineffective rainfall interference which does not cause displacement response; b6, calculating the event association coefficient To quantify causal correlation strengths of rainfall and displacement acceleration events at different lag times, a correlation coefficient is defined The calculation formula is as follows: ; Wherein the molecule is For displacement acceleration during the associated period And rainfall amount Reflecting the cooperative variation degree of the rainfall intensity and the displacement acceleration effect, when the rainfall intensity is large and the contemporaneous displacement acceleration is obvious, the integral value is increased, which indicates that the rainfall intensity and the displacement acceleration are closely related; Denominator of denominator The total rainfall integral in the association period is used for normalization processing, so that the influence of the rainfall total quantity difference on the association degree evaluation is eliminated, and the event association degrees of different rainfall intensities and different duration lengths are comparable; b7, optimal lag time screening For sets of lag times Traversing all identified displacement acceleration events, calculating the relevance coefficient of each event under the lag time And find all events Mean of (2) The mean value reflects the suitability of the candidate lag time for the overall event by comparing all candidate lag times Selecting At maximum As an optimal lag time; b8, constructing an effective event set, namely, meeting the following conditions Event inclusion effective rainfall event collection of (2) Each event Corresponding to a unique rainfall sequence And a shift sequence Realizing the structured storage and management of the multi-period event; b9, ineffective rainfall marking, namely rainfall period not associated by any displacement acceleration event The rainfall is marked as ineffective rainfall, does not participate in the training of the model, and is directly given low risk.
  4. 4. The PatchTST model landslide risk assessment method based on the multi-period rainfall events, which is characterized in that the effective rainfall event feature extraction and type division in the third step specifically comprises the following steps: c1, distinguishing characteristic features, namely setting a storm intensity threshold value for each identified effective rainfall event The threshold value is obtained from local meteorological files or is set automatically from local average rainfall intensity, if the event exists, the rainfall intensity is not less than within the continuous set time Judging a storm event, and corresponding to a surface landslide caused by rapid saturation of the shallow soil body; If the rainfall intensity of the event is equal Judging that the rain event is continuous, and corresponding to deep landslide caused by slowly penetrating and lifting the underground water level by rainwater; intercepting a corresponding rainfall sequence as a storm sequence Or continuous rainy and overcast sequences Simultaneously intercepting a displacement sequence of a corresponding time interval as a displacement response sequence Or (b) Wherein =1,2,... , =1,2,... ; C2, extracting multidimensional features, namely aiming at each heavy rain rainfall event, continuous overcast rain rainfall event and ineffective rainfall event Constructing multidimensional feature vectors in each event The calculation mode and physical meaning of each event are as follows: Total rainfall According to the formula Calculating, wherein the unit is mm, representing the total amount scale of rainfall, directly influencing the accumulation of water content of soil mass, and the method comprises the steps of, The start time of the rainfall is indicated, Representing a target time; Maximum rainfall intensity According to the formula Calculating, wherein the unit is mm/h, reflecting the extreme intensity of rainfall, and determining the possibility of rapid saturation of soil; Average rainfall intensity Representing the average intensity of rainfall, and correlating with soil infiltration rate; duration of rainfall Reflecting the duration of rainfall and influencing the rainfall infiltration depth; variance of rainfall intensity The unit is Reflecting the fluctuation degree of rainfall intensity, and relating to the stress change frequency of the soil body; And c3, performing feature standardization treatment, namely adopting Z-Score standardization to eliminate dimension influence, wherein a standardization formula is as follows: ; Wherein the method comprises the steps of Is the first The average value of the individual features is used, Is the first The standard deviation of the individual features is used, Is the total number of effective rainfall events; And c4, feature clustering, namely inputting the identified multi-dimensional features of the storm event, the multi-dimensional features of the continuous overcast event, the multi-dimensional features of the ineffective rainfall event and the corresponding rainfall into a semi-supervised K-means model, and performing feature clustering training.
  5. 5. The method for landslide risk assessment based on PatchTST models of multi-stage rainfall events according to claim 1, wherein the construction of the dual-branch PatchTST model based on rainfall events in the fourth step comprises the following steps: d1, storm feature processing channel, for storm sequence Using high-frequency sampling (sampling interval =1-3 Hours, preferably 1 hour), a convolution layer of a 3×1 small-sized convolution kernel is constructed, with the formula: ; Wherein the method comprises the steps of For the exclusive convolution operation of a storm, 、 Respectively convolution kernel weight and bias, and strengthening storm peak value information through 2×1 maximum pooling operation to obtain storm sub-characteristics ; D2, long-term infiltration accumulated continuous overcast and rainy event channel, namely continuous overcast and rainy sequence Using low frequency sampling (sampling interval =7-24 Hours), a convolution layer of 7×1-9×1 large-size convolution kernels is constructed, with the formula: ; Wherein the method comprises the steps of For the special convolution operation of continuous overcast and rainy, 、 The convolution kernel weight and the bias are respectively adopted, the local rainfall fluctuation is smoothed through 4 multiplied by 1 average pooling operation, and the accumulated trend is reserved, so that the continuous overcast and rainy sub-characteristics are obtained ; D3, sequence blocking and projection, namely { the sequence of the storm event respectively , , ..., Sequence of events of } and consecutive overcast and rains { , , ..., Processing each independent sequence in the sequence, dividing the sequence into N=L/P sequence blocks through a non-overlapping sliding window with the size of P for a one-dimensional sequence with the length of L, mapping each sequence block into a D-dimensional block embedded vector through a learnable linear projection layer, and then finishing position coding; d4, constructing double-branch input, namely embedding all blocks of the storm event into blocks which are stacked in batches to form a storm branch characteristic tensor Embedding all blocks of the continuous rainy events into blocks stacked in batches to form continuous rainy branch characteristic tensors ; D5, independent feature extraction of the transducer encoder, namely And Splicing in characteristic dimension, respectively inputting the backbone network of the shared transducer encoder, and respectively outputting a storm context characteristic sub-matrix Fr1 and a continuous overcast and rainy context characteristic sub-matrix Fr2, wherein the storm context characteristic sub-matrix Fr1 is not original any more The method comprises advanced features which are deeply excavated by a transducer and consider complex time sequence relations among all blocks in a heavy rain event, and a corresponding soft and overcast context feature submatrix Fr2 comprises advanced features which are modeled and connected with long-term and accumulated time sequence dependence in the heavy rain event; d6, cross-branch attention fusion, namely constructing a cross-branch attention fusion module, taking Fr1 and Fr2 as input, and calculating cross attention weight And ( =1- ) And is expressed by the formula Fr = ·Fr1+ Performing dynamic weighted fusion on Fr2 to obtain a unified rainfall fusion feature matrix Fr; d7, model training, namely taking a fusion feature matrix Fr as model input, adopting a standard PatchTST pre-measuring head comprising a normalization layer, a flattening layer and a linear projection layer, taking a landslide displacement sequence X corresponding to a time window as a prediction target, and adopting a mixed loss function and a AdamW optimizer to complete training of a dual-branch PatchTST model; In order to predict the special requirements of the following landslide displacement risks of 6 hours, 12 hours and 24 hours, the prediction tasks of different scales have obvious differences, the short-term prediction focuses on the instant displacement response, the data fluctuation is gentle but the sensitivity to precision is high, the long-term prediction needs to cover the complete hysteresis effect of rainfall-displacement, the data fluctuation is severe and the errors are easy to overlap, if the model training is directly carried out by summing up the original loss values, the training bias is caused by the large absolute error of the long-term prediction, the precision optimization of the short-term and medium-term prediction is weakened, and a multi-scale self-adaptive mixed loss function is designed for the purpose, and the specific formula is as follows: ; Wherein, each item and parameter are defined as follows: , ; representing a multiscale normalized loss value based on a mean square error MSE, and being applicable to high data quality and no significant abnormal value; and in a scenario with strict requirements on prediction accuracy, such as short-term displacement monitoring during landslide stabilization The method is used for representing the multiscale normalized loss value based on the mean absolute error MAE, and is suitable for scenes in which sudden variation constant values are easy to occur in the middle period of rainfall collection and displacement, such as landslide active period prediction caused by heavy rain; The value range is 0-1, the weight ratio of two core loss items is used for adjusting, so that adaptation of different scene requirements is realized, k is a scale mark, and values 1,2 and 3 correspond to three prediction scales of 6 hours, 12 hours and 24 hours respectively; And Respectively the first Mean square error loss and mean absolute error loss at individual scales, where The predicted value at the individual scale is used to determine, Is a true value; Data distribution normalization factor, representing the first Standard deviation of a real displacement sequence under each scale is used for eliminating dimension difference and distribution dispersion difference of displacement data of different scales, The corresponding values are 6, 12 and 24 for the time length balance factors, The system is a scale self-adaptive adjusting coefficient, belongs to adjustable parameters and is used for adapting to priority demands of different early warning scenes.
  6. 6. The method for evaluating landslide risk based on PatchTST models of multi-period rainfall events according to claim 1, wherein the real-time landslide risk prediction in the fifth step specifically comprises the following steps: e1, acquiring and transmitting real-time data, namely acquiring data 1 time/hour by a GNSS receiver and a rain gauge, and triggering an alarm when the data is abnormal; e2, real-time event detection and classification, namely real-time rainfall Starting a rainfall event analysis flow when 3 continuous sampling points are more than 1mm/h, wherein if the rainfall is lower than the threshold value and the duration exceeds 3 hours, the rainfall is judged to be risk-free; Calculating five-dimensional characteristics in real time, wherein the five-dimensional characteristics comprise real-time total rainfall, real-time maximum rainfall intensity, real-time average rainfall intensity, lasting time, early 24-hour rainfall, real-time rainfall intensity variance, generated displacement increment and real-time acceleration peak value; inputting the standardized features extracted in real time into the clustering model trained in the third step, and outputting event types If the classification result is The early warning is not triggered temporarily, and only the data is stored in the historical database for updating and optimizing the subsequent model; e3, model branch selection and multi-scale prediction, namely classifying results according to real-time events Automatically selecting corresponding model branches, and taking the starting time of the current rainfall event To the current time Is a shift sequence of (2) ; If the sequence length is insufficient The initial displacement data mean value of the history synchronous similar event is adopted for complementation, the weight of the complete data is set to 0.5, and the model outputs the future 6 hours # ) The time is 12 hours ) 24 Hours% ) The displacement prediction value of the (2) is stored in real time in an early warning system database, and a prediction curve is generated at the same time to display the displacement variation trend; e4, calculating multidimensional risk indexes, namely calculating the predicted speeds of different time periods according to a formula and the average speed from the current time to the future time for 6 hours according to the multiscale displacement predicted value Average speed of 6-12 hours in future Average speed of 12-24 hours in future The units are mm/h; calculation of the predicted acceleration the average acceleration from the present to the future 6 hours Average acceleration of 6-12 hours in future In mm/h2, where The current real-time displacement speed is obtained by calculating displacement data of the current hour and the previous hour; e5, fuzzy comprehensive evaluation and risk grade division, namely quantifying risk indexes by adopting a triangle membership function, setting a speed demarcation point and an acceleration demarcation point, and calibrating a specific threshold value according to historical disaster data and geological conditions of a target landslide area to ensure that the local landslide instability rule is met; Wherein, the displacement fuzzy membership function is set as: ; Wherein the method comprises the steps of The maximum value of the displacement of the table, Representing the minimum value of the displacement and, , ; The speed membership function is set as: ; Wherein the method comprises the steps of Represents the average value of the change in the velocity during the stationary phase, A maximum value representing a change in speed during the stationary phase; The acceleration membership function is set as: ; Wherein the method comprises the steps of Represents the average value of the acceleration change during the stationary phase, The maximum value of acceleration change in the stabilization period is represented, a specific threshold value is calibrated according to historical disaster data and geological conditions of a target landslide area, the local landslide instability rule is guaranteed to be met, and displacement, speed and acceleration values in the stabilization period are recalculated and updated dynamically every 1 month of data is added subsequently, so that risk assessment is promoted to be fit with reality; adopting an analytic hierarchy process to determine weights of three risk indexes, constructing a judgment matrix and carrying out consistency test to finally determine weight vectors Wherein the displacement weight is the highest (0.4), the speed (0.3) and the acceleration (0.3) are the next highest, and finally the risk value is calculated according to three risk weight indexes; And e6, outputting and feeding back the early warning result, namely pushing the early warning result in multiple channels, recording early warning information to form a log, periodically evaluating early warning accuracy and updating model parameters.
  7. 7. The method for evaluating landslide risk based on PatchTST models of multi-period rainfall events according to claim 6, wherein the risk value calculation formula is: ; in order to integrate the risk value(s), For predicting the membership of the displacement for the next 24 hours, For a membership of the predicted speed for 6-12 hours in the future, Predicting the membership degree of acceleration for 6 hours from the current time to the future time, and finally according to the comprehensive risk value And determining an early warning range.
  8. 8. The method for landslide risk assessment based on PatchTST model of multi-period rainfall event according to claim 7, wherein the comprehensive risk value is based on The specific classification of the early warning range is as follows: blue early warning ); Yellow early warning @ ); Orange early warning @ ); Red early warning [ ] )。

Description

PatchTST model landslide risk assessment method based on multi-period rainfall events Technical Field The invention relates to the technical field of geological disaster monitoring and early warning, in particular to a PatchTST model landslide risk assessment method based on a multi-period rainfall event. Background Landslide is a common geological disaster, and has the characteristics of strong burst property, large destructive power, wide influence range and the like, and a large amount of casualties and huge property loss are often caused. It is counted that the global economic loss caused by landslide disasters exceeds billions of dollars and directly threatens the life and property safety of people. Researches show that rainfall is one of the most important factors for inducing landslide, and the influence of the rainfall on the landslide is represented as a composite mechanism of 'storm triggering' and 'long-term infiltration accumulation', wherein the storm can rapidly increase the pore water pressure of a landslide body, break the stress balance of a rock-soil body, reduce the shear strength and induce the sudden deformation of the landslide; long-term rainfall, such as continuous overcast and overcast rain, can continuously infiltrate the landslide body, and the physical and mechanical properties of the rock-soil body are deteriorated, such as reduced cohesive force and reduced internal friction angle, so that the landslide slowly deforms and accumulates, and finally instability can be possibly caused. And secondly, the driving mechanisms of the landslide displacement caused by different types of rainfall have obvious differences, if the driving mechanisms are subjected to mixed modeling without distinction, the inherent relation between the rainfall characteristics and the landslide deformation response is difficult to accurately describe, the generalization capability of the model in practical application is limited, and the prediction result may deviate from the practical displacement evolution law. Landslide displacement is a core index reflecting landslide stability, and accurate prediction of landslide displacement change trend can provide key technical support for early warning and emergency treatment of landslide disasters. Therefore, the existing method is often concentrated in landslide displacement prediction research, and mainly comprises three types of a physical mechanism model, a statistical model and a machine learning model, wherein the physical mechanism model has extremely strong dependence on geological parameters, the parameters are difficult to accurately acquire, and the nonlinear association of 'rainfall intensity abrupt change-displacement burst' cannot be fully described, so that a prediction error is large, the statistical model can only capture linear or simple nonlinear relation, cannot cope with complex dynamic changes of displacement under rainfall drive, and has insufficient prediction capability on burst deformation, and the machine learning method is widely applied to landslide displacement prediction by virtue of strong nonlinear fitting capability, but has the following problems: the rainfall type is not distinguished in the current rainfall landslide displacement prediction, and rainfall-free data and ineffective rainfall event data are brought into a model to carry out data training, so that the processing mode not only increases the calculated amount, but also can introduce noise interference to influence the prediction accuracy; The current method does not construct a special rainfall type-displacement response coupling module, and cannot fully mine the dynamic association relation of 'heavy rain-displacement surge', 'continuous overcast and rainy-displacement buffer increase', so that the robustness of the model under different rainfall scenes is poor; The rainfall change cannot be converted into landslide risk prediction, namely the existing method is usually used for accurately predicting the landslide displacement, the landslide displacement cannot be converted into risk prediction, and a professional is required to realize risk assessment on the landslide according to professional knowledge, so that the learning cost of non-professional staff is increased. In view of the above drawbacks, there is a need for a landslide displacement prediction method capable of distinguishing heavy rain from continuous overcast rain and extracting rainfall features specifically, and converting rainfall variation into risk prediction variation, so as to meet the actual requirements of geological disaster early warning. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a PatchTST model landslide risk assessment method based on a multi-period rainfall event, which solves the problems of poor adaptability of device landslide displacement prediction to intermittent rainfall, inaccurate rainfall-displacement coupling modeling and lack of ris