PatchTST model landslide risk assessment method based on multiple rainfall events

Abstract

The invention discloses a PatchTST model landslide risk assessment method based on multi-period rainfall events, and relates to the technical field of geological disaster monitoring and early warning, and the method comprises the steps of data preprocessing and space-time alignment, multi-period rainfall event segmentation, effective rainfall event type segmentation, double-branch PatchTST model construction and displacement prediction and risk assessment. 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 correlation coefficient, invalid data is eliminated, the signal-to-noise ratio of model training is remarkably improved, the model learns a real geological mechanism, the prediction result is more credible, and the method is suitable for the landslide risk assessment of the PatchTST model. By designing the PatchTST model of the rainstorm branch and the overcast rain branch, the model can not only capture displacement sudden increase caused by rainstorm, but also describe displacement accumulation caused by continuous overcast rain, and the rainfall type distinguishing and modeling capacity is improved.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, specifically to a landslide risk assessment method based on the PatchTST model of multi-phase rainfall events. Background Technology

[0002] Landslides, a common geological hazard, are characterized by their suddenness, destructive power, and wide-ranging impact, often causing numerous casualties and enormous property losses. Statistics show that landslides cause over ten billion US dollars in economic losses globally each year, directly threatening people's lives and property. Research indicates that rainfall is one of the most significant factors triggering landslides, exhibiting a combined mechanism of "rainstorm triggering" and "long-term infiltration accumulation": heavy rainfall rapidly increases pore water pressure within the landslide mass, disrupting the stress balance of the soil and rock, reducing shear strength, and inducing sudden deformation of the landslide.

[0003] Prolonged rainfall, such as continuous rain, will continuously soak and lubricate the landslide, leading to the deterioration of the physical and mechanical properties of the soil and rock, such as reduced cohesion and decreased internal friction angle. This causes the landslide to slowly deform and accumulate, which may eventually lead to instability.

[0004] Secondly, different types of rainfall have significantly different driving mechanisms for landslide displacement. If they are mixed and modeled indiscriminately, it will be difficult to accurately characterize the intrinsic relationship between rainfall characteristics and landslide deformation response, resulting in limited generalization ability of the model in practical applications and possible deviations between the prediction results and the actual displacement evolution law.

[0005] Landslide displacement is a core indicator reflecting landslide stability, and accurate prediction of landslide displacement trends can provide crucial technical support for early warning and emergency response to landslide disasters. Therefore, existing methods in landslide prediction often focus on landslide displacement prediction, mainly including three categories: physical mechanism models, statistical models, and machine learning models. Physical mechanism models are highly dependent on geological parameters, which are difficult to obtain accurately, and cannot fully characterize the nonlinear relationship between "sudden changes in rainfall intensity and sudden displacement," resulting in large prediction errors. Statistical models can only capture linear or simple nonlinear relationships and cannot cope with the complex dynamic changes in displacement driven by rainfall, lacking the ability to predict sudden deformation. Machine learning methods, with their powerful nonlinear fitting capabilities, are widely used in landslide displacement prediction, but they all have the following problems:

[0006] Current rainfall-induced landslide displacement prediction does not differentiate between rainfall types and includes data from no rainfall and invalid rainfall events in the model for training. This approach not only increases the computational load but may also introduce noise interference, affecting prediction accuracy.

[0007] The current method does not construct a dedicated "rainfall type-displacement response" coupling module, which fails to fully explore the dynamic correlation between "heavy rain-sudden increase in displacement" and "continuous rain-gradual increase in displacement", resulting in poor robustness of the model under different rainfall scenarios;

[0008] The inability to convert rainfall changes into landslide risk prediction: Existing methods often only make accurate predictions of landslide displacement, but cannot convert landslide displacement into risk prediction. Professionals are required to conduct risk assessments of landslides based on their professional knowledge, which increases the learning cost for non-professionals.

[0009] To address the aforementioned shortcomings, there is an urgent need for a landslide displacement prediction method that can distinguish between heavy rain and continuous rain, extract rainfall characteristics in a targeted manner, and convert rainfall changes into risk prediction changes to meet the actual needs of geological disaster early warning. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention provides a landslide risk assessment method based on the PatchTST model for multi-period rainfall events. This method solves the problems of poor adaptability of landslide displacement prediction to intermittent rainfall, inaccurate rainfall-displacement coupling modeling, and lack of risk assessment.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a landslide risk assessment method based on the PatchTST model of multi-period rainfall events, specifically including the following steps:

[0012] Step 1: Data Preprocessing and Spatiotemporal Alignment: Collect GNSS displacement time series data and rainfall time series data for the target landslide area. The data is denoised and missing values ​​are filled in. The two types of data are aligned in time and space by timestamp matching to ensure consistent data resolution.

[0013] Step 2: Segmentation of Multiple Rainfall Events: Based on Rainfall Intensity Thresholds And the no-rain interval threshold T, for the preprocessed rainfall sequence Multi-phase rainfall event identification and segmentation were performed, and the displacement rate of the landslide at each moment during the rainfall period was calculated. ,speed With acceleration And solve for the optimal rainfall impact lag time, based on The system identifies invalid and valid rainfall events by size, backtracks and correlates them to obtain valid rainfall events, and constructs an event set.

[0014] Step 3: Effective Rainfall Event Type Segmentation: Based on the rainstorm intensity threshold provided by the local government. Distinguish between heavy rain event sequences and continuous rainy event sequences, resulting in m independent rainfall event sequences (including heavy rain sequences { , … } and continuous rain sequence { , … }, + =m) and the corresponding displacement response sequence { , … Finally, the clustering model for rainfall events is trained.

[0015] Step 4: Construction of the two-branch PatchTST model: Taking the independent rainfall event sequence as input, the complete two-branch PatchTST model is constructed through steps such as two-branch feature sequence construction, sequence segmentation and projection, position encoding, independent feature extraction by Transformer encoder, cross-branch attention fusion, and model training.

[0016] Step 5, Displacement Prediction and Risk Assessment: Input the rainfall data characteristics of the period to be predicted into the clustering model to identify rainstorm events, cloudy and rainy events and invalid rainfall events. For valid rainfall events, the landslide displacement prediction results are output by the two-branch PatchTST model. Combined with the fuzzy comprehensive evaluation method, the risk value is calculated and the warning level is divided. The warning results are output in real time, and corresponding measures are provided.

[0017] Preferably, the specific process of data preprocessing in step one includes:

[0018] a1. Data Acquisition Phase: Displacement data is acquired through a GNSS receiver, rainfall data is acquired through a small weather station, and environmental data is acquired using temperature and humidity sensors and soil moisture sensors. The acquisition frequency is once per hour.

[0019] a2. Data Standardization: Data is processed using Z-score standardization, with the formula as follows: ,in, This is the original data. For standardized data, The mean of the data. The standard deviation of the data;

[0020] a3. Missing Value Imputation: When a single data point is missing but the preceding and following data are valid, linear interpolation is used to impute it. The formula is as follows: If more than three consecutive sampling points are missing, the missing data will be supplemented by combining the average displacement data under similar weather conditions in the same period of history to avoid bias in subsequent event analysis due to missing data.

[0021] a4. Spatiotemporal alignment: Based on the timestamp in the format of “YYYY-MM-DD HH:MM:SS”, the processed GNSS displacement and rainfall data are matched point by point to ensure that each time node corresponds to a unique displacement value and rainfall value, and the data resolution is uniformly 1 hour.

[0022] Preferably, step two, which involves segmenting and transforming multi-stage rainfall events based on displacement response, specifically includes:

[0023] b1. Rainfall Time Recording and Displacement Parameter Calculation: Based on rainfall and non-rainfall times recorded by rain gauges, and combined with historical landslide displacement and rainfall data, the maximum rainfall delay time is determined through statistical analysis. The rainfall event time is the rainfall start time plus the lag time. Based on the preprocessed displacement data, the calculation is performed according to the formula... Calculate the displacement-velocity sequence (Unit: mm / h), according to the formula Calculate displacement acceleration sequence (Unit: mm / h²), where Hour;

[0024] b2. Risk Event Judgment: When Three consecutive sampling points (i.e., 3 hours) exceeding And the displacement increment during that period If the event occurs, it is considered a displacement acceleration event, and the start time of the event is recorded. With end time ( for First time below (time)

[0025] b3. Calculation of Rainfall Lag Time and Event Correlation: The impact of rainfall on landslide displacement is not instantaneous but exhibits a significant lag effect. This lag primarily stems from the time required for rainwater to infiltrate the landslide body, the time needed for soil and rock moisture content to accumulate, and the gradual nature of pore water pressure changes and the deterioration of soil and rock mechanical properties. Ignoring this lag effect and directly modeling the correlation between rainfall data and concurrent displacement data will lead to a misalignment of the causal relationship between rainfall and displacement acceleration events, thereby introducing invalid rainfall interference and reducing the model's prediction accuracy. Therefore, optimizing the rainfall lag time is a crucial step in accurately constructing the "rainfall event-displacement response" correlation and selecting effective rainfall events. The specific steps are as follows:

[0026] b4. Definition of Lag Time Set

[0027] Based on the rapid infiltration effect of short-duration torrential rain, which typically triggers displacement response within hours, and the slow infiltration effect of long-duration continuous rain, which requires several days to accumulate before displacement changes become apparent, a lag time set is defined. However, by using discrete values ​​at 6-hour intervals, computational complexity is reduced and optimization efficiency is improved while ensuring coverage integrity.

[0028] b5. Correlation of time node calibration

[0029] For each identified displacement acceleration event i (whose start time is...) The end time is Based on each candidate value in the lag time set Calculate the corresponding rainfall association start time. It needs to be made clear that This is not the actual meteorological observation start time of the rainfall event, but rather the "effective rainfall start time" calibrated to establish a causal relationship between rainfall and displacement acceleration. The core logic is that the occurrence of a displacement acceleration event is essentially triggered by the cumulative effect of rainfall over a preceding period, through a lag time... By retrospectively matching displacement acceleration events with the rainfall periods that actually caused the events, invalid rainfall interference that did not trigger a displacement response can be eliminated.

[0030] b6. Calculation of Event Relevance Coefficient

[0031] To quantify the causal relationship between rainfall and displacement acceleration events with different lag times, a correlation coefficient is defined. The calculation formula is as follows:

[0032] ;

[0033] Among them: molecule Displacement acceleration during the associated time period With rainfall The integral reflects the degree of coordinated change between rainfall intensity and displacement acceleration effect—when the rainfall intensity is large and the displacement acceleration is significant during the same period, the integral value increases, indicating that the two are closely related;

[0034] denominator The total rainfall within the associated period is integrated for normalization, eliminating the impact of differences in total rainfall on the correlation assessment and making the correlation of events with different rainfall intensities and durations comparable.

[0035] b7. Optimal Lag Time Screening

[0036] For the set of lag times For each candidate value, iterate through all identified displacement acceleration events and calculate the correlation coefficient for each event at that lag time. And find all events mean This mean reflects the suitability of the candidate lag time to the overall event, by comparing the values ​​corresponding to all candidate lag times. ,choose Maximum As the optimal lag time;

[0037] b8. Construction of a valid event set: This will satisfy... The events were included in the effective rainfall event set. Each event Corresponding to a unique rainfall sequence With displacement sequence This enables structured storage and management of multi-period events;

[0038] b9. Invalid Rainfall Marker: Rainfall periods not associated with any displacement acceleration events. Rainfall data marked as invalid is not included in model training and is directly assigned a low risk score.

[0039] Preferably, the effective rainfall event feature extraction and type classification in step three specifically includes:

[0040] c1. Characteristic Differentiation: For each identified valid rainfall event, a rainfall intensity threshold is set. The threshold can be obtained from local meteorological files or set manually from the local average rainfall intensity; if the event has rainfall intensity ≥ within a set time period... If so, it is determined to be a rainstorm event, corresponding to a surface landslide caused by rapid saturation of shallow soil;

[0041] If the rainfall intensity of the event is < The event was determined to be a continuous rainy season, which corresponds to a deep landslide caused by the slow infiltration of rainwater raising the groundwater level.

[0042] Extract the corresponding rainfall sequence as the rainstorm sequence Or a continuous rainy sequence Simultaneously, the displacement sequence within the corresponding time interval is extracted as the displacement response sequence. or ,in =1,2,… , =1,2,… ;

[0043] c2. Multi-dimensional feature extraction: This involves extracting features for each heavy rain event, continuous rain event, and invalid rainfall event. Construct a multidimensional feature vector for each event. The calculation methods and physical meanings of each event are as follows:

[0044] Total rainfall According to the formula The calculation, expressed in mm, represents the total amount of rainfall and directly affects the accumulation of soil moisture content; among which, Indicates the start time of rainfall. Indicate the target time;

[0045] Maximum rainfall intensity According to the formula The calculation, in mm / h, reflects the extreme intensity of rainfall and determines the likelihood of rapid soil saturation.

[0046] Average rainfall intensity : Characterizes the average intensity of rainfall and is related to the soil infiltration rate;

[0047] Rainfall duration This reflects the duration of rainfall and affects the depth of rainfall infiltration.

[0048] Rainfall intensity variance Unit is This reflects the degree of fluctuation in rainfall intensity and is related to the frequency of changes in soil stress.

[0049] c3. Feature Standardization: Z-Score standardization is used to eliminate the influence of dimensions. The standardization formula is: ;

[0050] in For the first The mean of each feature, For the first The standard deviation of each feature The total number of effective rainfall events;

[0051] c4. Feature Clustering: Input the multi-dimensional features of the identified rainstorm events, the multi-dimensional features of the continuous rain events, the multi-dimensional features of the invalid rainfall events, and the corresponding rainfall into the semi-supervised K-means model for feature clustering training.

[0052] Preferably, the construction of the two-branch PatchTST model based on rainfall events in step four includes:

[0053] d1, Rainstorm Feature Processing Channel: Processing rainstorm sequences High-frequency sampling (sampling interval) is used =1-3 hours, preferably 1 hour), construct a convolutional layer with a small 3×1 kernel, the formula is:

[0054] ;

[0055] in Convolution operation specifically for heavy rain. , The convolution kernel weights and biases are used respectively, and a 2×1 max pooling operation is applied to enhance the information of the rainstorm peak, resulting in rainstorm sub-features. ;

[0056] d2. Long-term infiltration and cumulative continuous rainy event channel: for continuous rainy sequences Low-frequency sampling (sampling interval) is used =7-24 hours), construct convolutional layers with large-size convolutional kernels of 7×1-9×1, the formula is:

[0057] ;

[0058] in Convolution operation specifically for continuous rainy weather. , The convolutional kernel weights and biases are used respectively. A 4×1 average pooling operation is applied to smooth local rainfall fluctuations while preserving the cumulative trend, thus obtaining the sub-features of continuous overcast and rainy weather. ;

[0059] d3. Sequence Segmentation and Projection: For the rainstorm event sequence { , , …, } and the sequence of continuous rainy events { , , …, Each independent sequence in} is processed. For a one-dimensional sequence of length L, it is divided into N=L / P sequence blocks by a non-overlapping sliding window of size P. Each sequence block is mapped to a D-dimensional block embedding vector through a learnable linear projection layer, and then position encoding is completed.

[0060] d4. Dual-branch input construction: The blocks of all rainstorm events are embedded and stacked in batches to form the rainstorm branch feature tensor. All blocks of continuous rainy events are embedded and stacked in batches to form a continuous rainy branch feature tensor. ;

[0061] d5. Independent feature extraction from the Transformer encoder: and The features are concatenated along the feature dimension and fed into the shared Transformer encoder backbone network, outputting a rainstorm context feature sub-matrix Fr1 and a continuous rainy weather context feature sub-matrix Fr2, respectively. It's important to note that the rainstorm context feature sub-matrix Fr1 is no longer the original one. It contains high-level features that have been deeply mined by Transformer and take into account the complex temporal relationships between all blocks within the rainstorm event. The corresponding continuous rain context feature submatrix Fr2 contains high-level features that have been modeled and show long-term, cumulative temporal dependencies within the continuous rain event.

[0062] d6. Cross-branch attention fusion: Construct a cross-branch attention fusion module, taking Fr1 and Fr2 as inputs, and calculate the cross-attention weights. and ( =1- ), and according to the formula Fr= ·Fr1+ • Fr2 is dynamically weighted and fused to obtain a unified rainfall fusion feature matrix Fr;

[0063] d7. Model Training: The fused feature matrix Fr is used as the model input. The standard PatchTST prediction head is used, which includes a normalization layer, a flattening layer and a linear projection layer. The landslide displacement sequence X corresponding to the time window is used as the prediction target. The training of the two-branch PatchTST model is completed using a hybrid loss function and the AdamW optimizer.

[0064] To meet the specific needs of predicting landslide displacement risks in the following 6, 12, and 24 hours, the prediction tasks at different scales differ significantly. Short-term predictions focus on immediate displacement responses, with relatively smooth data fluctuations but high sensitivity to accuracy. Long-term predictions need to cover the complete lag effect of rainfall-displacement, with drastic data fluctuations and easily accumulating errors. If the model is trained by directly summing the original loss values, the large absolute error in long-term predictions will lead to training bias, weakening the optimization of accuracy in short- and medium-term predictions. Therefore, a multi-scale adaptive hybrid loss function is designed, the specific formula of which is as follows:

[0065] ;

[0066] The definitions of each item and parameter are as follows:

[0067] , ;

[0068] This represents the multi-scale normalized loss value based on the mean squared error (MSE), which is suitable for data with high quality and no significant outliers.

[0069] Furthermore, it is crucial for scenarios with stringent requirements for prediction accuracy, such as short-term displacement monitoring during the stabilization phase of a landslide. This represents the multi-scale normalized loss value based on the mean absolute error (MAE), which is suitable for scenarios where rainfall is concentrated and displacement is prone to sudden outliers, such as predicting the active period of landslides triggered by rainstorms.

[0070] The value ranges from 0 to 1 and is used to adjust the weight ratio of the two core loss terms to adapt to different scenario requirements. k is the scale identifier, with values ​​of 1, 2, and 3, corresponding to three prediction scales of 6 hours, 12 hours, and 24 hours, respectively.

[0071] and The first Mean squared error loss and mean absolute error loss at each scale, among which Predicted values ​​at various scales The actual value;

[0072] Data distribution normalization factor, representing the first The standard deviation of the true displacement sequence at each scale is used to eliminate differences in dimensionality and distribution dispersion of displacement data at different scales. This is the duration balancing factor, with corresponding values ​​of 6, 12, and 24. This is a scale-adaptive adjustment coefficient, an adjustable parameter used to adapt to the priority requirements of different early warning scenarios.

[0073] Preferably, the real-time landslide risk forecasting in step five specifically includes:

[0074] e1. Real-time data acquisition and transmission: The GNSS receiver and rain gauge acquire data once per hour, and trigger an alarm when the data is abnormal;

[0075] e2. Real-time event detection and classification: Real-time rainfall The rainfall event analysis process is initiated when the rainfall is greater than 1 mm / h at 3 consecutive sampling points. If the rainfall is less than this threshold and the duration exceeds 3 hours, it is determined to be a risk-free rainfall.

[0076] Real-time calculation of five-dimensional features, including real-time total rainfall, real-time maximum rainfall intensity, real-time average rainfall intensity, duration, previous 24-hour rainfall, real-time rainfall intensity variance, displacement increment, and real-time peak acceleration.

[0077] The standardized features extracted in real time are input into the clustering model trained in step three, and the event type is output. If the classification result is If the data is not triggered, an alert will not be issued, and the data will be stored in the historical database for subsequent model updates and optimizations.

[0078] e3. Model Branching and Multi-Scale Prediction: Based on Real-Time Event Classification Results Automatically select the corresponding model branch and retrieve the start time of the current rainfall event. up to the current time displacement sequence ;

[0079] If the sequence length is insufficient Then, the average initial displacement data of similar events in the same historical period is used for completion, and the weight of the completed data is set to 0.5. The model outputs the next 6 hours ( ), 12 hours ), 24 hours The displacement prediction value is obtained, and the prediction result is stored in the early warning system database in real time. At the same time, a prediction curve is generated to show the displacement change trend.

[0080] e4. Calculation of Multi-Dimensional Risk Indicators: Based on multi-scale displacement prediction values, the predicted velocity for different time periods is calculated using the formula: the average velocity from the present to the next 6 hours. Average speed over the next 6-12 hours Average speed over the next 12-24 hours All units are mm / h;

[0081] Predicted acceleration calculation: Average acceleration from now until the next 6 hours Average acceleration over the next 6-12 hours All units are mm / h², where The current real-time displacement velocity is calculated using the current displacement data and the displacement data from the previous hour.

[0082] e5. Fuzzy comprehensive evaluation and risk level classification: The risk index is quantified by the triangular membership function. Velocity boundary point and acceleration boundary point are set. The specific thresholds are determined based on the historical disaster data and geological conditions of the target landslide area to ensure that they conform to the local landslide instability law.

[0083] The displacement fuzzy membership function is set as follows:

[0084] ;

[0085] in The maximum value of the displacement. This represents the minimum displacement. , ;

[0086] The velocity membership function is set as follows:

[0087] ;

[0088] in This represents the average value of the velocity change during the steady-state period. This represents the maximum value of the velocity change during the steady-state period;

[0089] The acceleration membership function is set as follows:

[0090] ;

[0091] in This represents the average value of the acceleration change during the steady-state period. This represents the maximum value of acceleration change during the stabilization period. The specific threshold is determined based on historical disaster data and geological conditions of the target landslide area to ensure that it conforms to the local landslide instability pattern. Furthermore, for each additional month of data, the displacement, velocity, and acceleration values ​​during the stabilization period are recalculated and dynamically updated to ensure that the risk assessment is consistent with reality.

[0092] The weights of the three risk indicators were determined using the analytic hierarchy process (AHP). A judgment matrix was constructed and a consistency check was performed to finally determine the weight vector. The risk value is calculated based on three risk weight indicators, with displacement having the highest weight (0.4), followed by velocity (0.3) and acceleration (0.3).

[0093] e6. Early warning result output and feedback: Push early warning results through multiple channels, record early warning information to form a log, and regularly evaluate the accuracy of early warning and update model parameters.

[0094] Preferably, the risk value calculation formula is as follows:

[0095] ;

[0096] For the comprehensive risk value, To predict the membership degree of displacement in the next 24 hours, The membership degree for the predicted speed over the next 6-12 hours. The membership degree of the predicted acceleration from now to the next 6 hours is determined, and finally, based on the comprehensive risk value... Determine the scope of the warning.

[0097] Preferably, the one based on comprehensive risk value The specific categories of the warning scope are as follows:

[0098] Blue Alert ( );

[0099] Yellow alert ( );

[0100] Orange alert ( );

[0101] Red Alert ( ).

[0102] Beneficial effects

[0103] This invention provides a landslide risk assessment method based on the PatchTST model using multiple rainfall events. Compared with existing technologies, it has the following advantages:

[0104] (1) The landslide risk assessment method based on the PatchTST model of multiple rainfall events establishes a causal relationship between "rainfall event-displacement response" by inversely correlating displacement acceleration with multiple effective rainfall events, rather than the statistical correlation of traditional models. Effective rainfall is screened by correlation coefficient, and invalid data is eliminated. The signal-to-noise ratio of model training is significantly improved, the model learns the real geological mechanism, and the prediction results are more credible.

[0105] (2) The landslide risk assessment method based on the PatchTST model of multiple rainfall events introduces the threshold of rainstorm intensity and the threshold of continuous rain duration to achieve accurate differentiation between rainstorm and continuous rain. A PatchTST model based on rainstorm branch and rain dual branch is designed, which effectively solves the problem of "indiscriminate modeling" in the existing methods, avoids feature confusion caused by single structure, and enables the model to capture both the sudden increase in displacement caused by rainstorm and the accumulation of displacement caused by continuous rain, further improving the ability to distinguish and model rainfall types.

[0106] (3) The landslide risk assessment method based on the PatchTST model of multiple rainfall events adopts multiple small blocks for batch processing, which reduces memory usage twice. Furthermore, invalid rainfall and no rainfall data are excluded, further reducing resource consumption. By utilizing the independent characteristics of PatchTST channels, the influence of noise in each channel can be effectively resisted, further improving the prediction accuracy.

[0107] (4) The landslide risk assessment method based on the PatchTST model of multiple rainfall events provides a complete end-to-end processing flow from data preprocessing, event identification, model training to real-time early warning. The output result is an intuitive risk level, which does not require professional interpretation and is highly practical. Attached Figure Description

[0108] Figure 1 This is an overall flowchart of the landslide displacement prediction method of the two-branch PatchTST model for multi-phase rainfall events in this invention;

[0109] Figure 2 This is a schematic diagram illustrating the effect of the rainfall event segmentation and type differentiation of the present invention;

[0110] Figure 3 This is a comparison chart of the prediction effects of the present invention and existing methods;

[0111] Figure 4 This is a landslide prediction risk assessment diagram for the present invention. Detailed Implementation

[0112] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0113] Please see Figure 1-4 This invention provides a technical solution: a landslide risk assessment method based on the PatchTST model using multiple rainfall events, with the specific operation steps as follows:

[0114] Step 1: Data Preprocessing and Spatiotemporal Alignment

[0115] A1. Data Preprocessing: Collecting two types of core data from the target landslide area:

[0116] GNSS landslide displacement data were acquired using GNSS monitoring stations, with a sampling frequency of 1 time / hour.

[0117] Rainfall data were obtained from regional meteorological stations, with a sampling frequency of once per hour.

[0118] A2. Outlier Removal: A two-step method combining the "3σ criterion + Isolation Forest algorithm" is used to remove outliers, preventing them from interfering with model training.

[0119] A3. Missing Value Imputation: When a single data point is missing but the preceding and following data are valid, linear interpolation is used to imput it. The formula is:

[0120] ;

[0121] If more than 3 consecutive sampling points are missing, the data should be supplemented by the average displacement data under similar weather conditions in the same period of history to avoid bias in subsequent event analysis due to missing data.

[0122] A4. Standardization Processing: To eliminate the impact of differences in data units on model training, Z-score standardization is used to process displacement data and rainfall data separately, as shown in the following formula:

[0123] ;

[0124] Let the preprocessed displacement sequence be X=[x1,x2,…, (L is the total length of the data), the rainfall sequence is R=[r1,r2,…, Subsequent model training is based on the standardized sequence, and the prediction results are then reverse-normalized to the actual physical quantities before being output.

[0125] A5. Spatiotemporal alignment: Based on the timestamp in the format of "YYYY-MM-DD HH:MM:SS", the processed GNSS displacement data and rainfall data are matched point by point to ensure that each time node corresponds to a unique displacement value and rainfall value. Finally, the data resolution is unified to 1 hour, which lays the data foundation for subsequent multi-stage rainfall event segmentation.

[0126] Step 2: Segmentation of multiple rainfall events:

[0127] B1. Determination of Rainfall Intensity Threshold I: The calculation formula is as follows ,in K represents the average historical rainfall intensity in the target area, and K is the multiplier.

[0128] B2. Determination of the no-rain interval threshold T: Based on the hydrological response characteristics of the landslide body—when the no-rain time exceeds T, the pore water pressure of the landslide body begins to decrease steadily, and the displacement response tends to be gentle. At this time, it is determined to be two independent rainfall events. Different T is set for cohesive soil landslides and sandy soil landslides respectively.

[0129] B3. Displacement Acceleration Event Recognition: Based on the preprocessed displacement data, according to the formula... Calculate the displacement-velocity sequence Unit: mm / h, according to formula Calculate displacement acceleration sequence Unit: mm / h², where Hourly data sampling interval;

[0130] The risk event determination rule is as follows: when Three consecutive sampling points (i.e., 3 hours) exceeding And the displacement increment during that period When this occurs, it is determined to be a displacement acceleration event, and the start time of the event is recorded. With end time ( for First time below (Time period), to ensure that all identified events are "risk period" events that have a real impact on landslide stability;

[0131] B4. Optimization of Rainfall Lag Time: Considering the lag effect of rainfall on landslide displacement, a set of lag times is defined. Calculate the mean of the event correlation coefficient for each quantity. Correlation coefficient According to the formula Calculate, and ,in, This indicates the start time of the acceleration event in the i-th rainfall event. Lag time, what needs to be understood is This does not imply that the rainfall start time, lag time, and landslide location movement of the i-th event are closely related; this is done to avoid the influence of ineffective rainfall. The larger the value, the stronger the causal relationship between rainfall and accelerated displacement. Maximum As the optimal lag time;

[0132] B5. Construction of a valid event set: This will satisfy... The events were included in the effective rainfall event set. Each event Corresponding to a unique rainfall sequence With displacement sequence This enables structured storage and management of multi-period events;

[0133] B6. Invalid Rainfall Marker and Weighting: Rainfall periods not associated with any displacement acceleration events. Rainfall that is marked as invalid only wets the surface soil without causing a significant displacement response is not used in model training and is directly assigned a low risk.

[0134] Step 3: Effective Rainfall Event Type Segmentation: Rainfall Intensity Threshold Threshold and duration threshold of continuous rain Settings: Based on the rainstorm intensity standard set for the landslide area, this distinguishes between rainstorms and ordinary rainfall; based on the permeability characteristics of the soil and rock mass, a threshold for the duration of continuous rain is set. This is used to distinguish between prolonged rain and short-term rainfall.

[0135] C1. Independent event classification: Calculate the time interval Δt between two adjacent active rainfall moments—if Δt≤T, it is determined to be the same rainfall event;

[0136] If Δt > T, they are determined to be different rainfall events, thus obtaining the time intervals [start1, end1], [start2, end2], ..., [start1, end1] for m independent rainfall events. m end m ];

[0137] C2. Rainfall type differentiation: For each independent rainfall event, combined with... and Judgment type - If there is continuous rainfall intensity of ≥ 3 hours in the event It was determined to be a rainstorm event;

[0138] If the duration of the event (end-start+1) ≥ And the rainfall intensity at all times < It was determined to be a continuous rainy event;

[0139] C3. Sequence Extraction: Based on the event time interval, extract the corresponding rainfall sequence as the rainstorm sequence. ( =1,2,… (or continuous rainy sequence) ( =1,2,… Simultaneously, the displacement sequence within the corresponding time interval is extracted as the displacement response sequence. or ,satisfy + =m, and , for length, for length;

[0140] C4. Multi-dimensional feature extraction: This involves extracting features for each heavy rain event, continuous rain event, and invalid rainfall event. Construct feature vectors Specifically, this includes total rainfall. Maximum rainfall intensity average rainfall intensity Rainfall duration Rainfall intensity variance And Z-Score standardization is used to eliminate the influence of dimensions;

[0141] C5. Clustering Model Parameter Training: The multi-dimensional features and rainfall data of the categorized rainfall events (such as heavy rain events, continuous rainy weather events, and invalid rainfall events) are input into the clustering model to train the model parameters and enhance its accuracy.

[0142] Step 4: Building the dual-branch PatchTST model:

[0143] D1, Rainstorm Feature Processing Channel (Rainfall Feature Extraction Branch 1)

[0144] This channel is designed specifically for the characteristics of heavy rainfall, namely "sudden increases and decreases in intensity and high-frequency fluctuations." Through differentiated sampling and dedicated convolution operations, it accurately captures key rainfall features that trigger sudden deformations in landslides. The specific process is as follows: High-frequency sampling uses the same high-frequency sampling interval as the original data. =1 hour, fully preserving triggering characteristics such as hourly rainfall peak, sudden increase, and sudden decrease rates, and obtaining high-frequency subsequences of rainstorms through this sampling. The length of each subsequence is matched with the duration (in hours) of the corresponding rainstorm event. Then, a dedicated convolution operation is performed using a 3×1 small-size convolution kernel. The formula is as follows:

[0145] ;

[0146] in, Convolution operation specifically for heavy rain. , These are the kernel weights and biases, respectively. The receptive field of a small-sized kernel focuses on a 1-3 hour window, which can accurately capture drastic changes in rainfall intensity within the time frame, such as the moment of peak occurrence and the point of sudden drop in intensity.

[0147] Finally, max pooling is applied to the convolution results using a 2×1 max pooling operation. This prioritizes retaining key triggering information such as rainfall peaks and sudden changes in intensity, including maximum rainfall intensity and peak occurrence time, while filtering out redundant data from non-peak periods. The final result is the rainstorm feature sub-matrix. ;

[0148] D2. Long-term infiltration and continuous rainy events (branch two of rainfall feature extraction)

[0149] This channel is designed to address the characteristics of prolonged, continuous rainfall with its "stable intensity and sustained accumulation." It fully integrates the cumulative effect of long-term rainfall through low-frequency sampling and large-size convolution operations. The specific process is as follows: Low-frequency sampling over a long period of time with varying low-frequency sampling intervals. This sampling method weakens hourly fluctuations and emphasizes cumulative characteristics such as daily cumulative rainfall and duration of rainfall, thereby obtaining a low-frequency subsequence of continuous rainy weather. The subsequence length is consistent with the duration of the corresponding continuous rainy event. Then, a dedicated convolution operation is performed using a 7×1 large-size convolution kernel. The formula is as follows:

[0150] ;

[0151] in, Convolution operation specifically for continuous rainy weather. , These are the kernel weights and biases, respectively.

[0152] Finally, average pooling was applied to the convolution result using a 4×1 average pooling operation to smooth out local rainfall fluctuations and highlight the overall trend of rainfall accumulation, ultimately yielding the feature submatrix of continuous rain. ;

[0153] D3. Construction of the dual-branch PatchTST model:

[0154] A PatchTST model is constructed, which includes a shared backbone network and dual-branch parallel input. The backbone network is based on the standard Transformer encoder structure. The core processing flow of the model is as follows:

[0155] Sequence segmentation and projection: For each independent rainfall event sequence (rainstorm sequence) Or a continuous rainy sequence First, the sequence is divided into blocks of length 1. A one-dimensional sequence, through a non-overlapping array of size The sliding window is divided into Each block is mapped through a learnable linear projection layer to... dimensional block embedding vector;

[0156] Learnable positional encoding: Add learnable positional encoding to all block embedding vectors to preserve the temporal order information of the sequence;

[0157] Two-branch parallel processing:

[0158] Heavy Rain Branch: Sequence of all heavy rain events { , , …, The block embeddings are stacked in batches to form the rainstorm feature tensor. ;

[0159] Continuous Rain Branch: Sequence all continuous rain events { , , …, The block embeddings are stacked in batches to form the continuous rain feature tensor. ;

[0160] Channel-independent feature extraction: Following the channel independence principle of PatchTST, T1 and T2 are regarded as two independent "channels". They are concatenated and input into the shared Transformer encoder backbone network. The self-attention mechanism in the encoder only operates on the time dimension of each block, which naturally realizes the isolation of the two branches of information in the deep layer of the model and targeted feature learning, and outputs the rainstorm context feature sub-matrix Fr1 and the continuous rain context feature sub-matrix Fr2 respectively.

[0161] D4. Cross-branch attention fusion:

[0162] Construct a lightweight cross-branch attention module to dynamically fuse features from both branches. The specific formula for cross-branch attention fusion is as follows:

[0163] ;

[0164] ;

[0165] ;

[0166] in , The query and key transformation matrices for heavy rain features and continuous rain features are respectively (both with dimensions (C / 2)×d, where d is the attention feature dimension). , They are respectively , Attention weights This is the final rainfall fusion feature matrix;

[0167] D5. Model Training:

[0168] Input-output construction: using fused feature matrices As input to the model, the landslide displacement sequence corresponding to the time window. As a prediction target;

[0169] Prediction Head: A prediction head follows the shared Transformer encoder. This head typically consists of a series of normalization layers, a Flatten operation, and a linear projection layer. Mapped to prediction step size Displacement prediction value .

[0170] Training details: A multi-scale adaptive hybrid loss function is used as the loss function. The AdamW optimizer is used to complete the shift training for 6h, 12h, and 24h, and an early stopping strategy is adopted to prevent overfitting.

[0171] Step 5: Displacement Prediction and Risk Assessment

[0172] E1. Real-time data acquisition and transmission: Displacement and rainfall data are acquired in real time through GNSS receivers and rain gauges deployed in the landslide area, with an acquisition frequency of 1 time / hour.

[0173] E2. Real-time event detection and classification: When real-time rainfall... If the rainfall is greater than 1 mm / h for 3 consecutive sampling points (i.e. 3 hours), the rainfall event analysis process is initiated. If the rainfall is lower than the threshold and the duration exceeds 6 hours, it is determined to be risk-free rainfall and the analysis is not initiated. For the rainfall event that is currently occurring, its five-dimensional characteristics are calculated in real time, including real-time total rainfall, real-time maximum rainfall intensity, real-time average rainfall intensity, rainfall duration, and real-time rainfall intensity variance.

[0174] E3. Event Classification: Input the standardized features extracted in real time into the clustering model trained in step three, and output the event type. If the classification result is If the data is not triggered, an alert will not be issued, and the data will be stored in the historical database for subsequent model updates and optimizations.

[0175] E4. Model branching and multi-scale prediction:

[0176] Model branch selection: based on real-time event classification results Automatically select the corresponding model branch and retrieve the start time of the current rainfall event. up to the current time displacement sequence If the sequence length is insufficient If the initial displacement data of similar events (same type of rainfall events) in the same historical period are used for completion, the weight of the completed data is set to 0.5 to avoid prediction deviation due to insufficient sequence length;

[0177] Multi-scale prediction execution: The constructed input sequence is input into the corresponding model branch, and the model outputs the next 6 hours ( ), 12 hours ), 24 hours The displacement prediction value is obtained, and the prediction result is stored in the early warning system database in real time. At the same time, a prediction curve is generated to intuitively display the displacement change trend.

[0178] E5. Fuzzy Comprehensive Evaluation and Risk Level Classification:

[0179] Setting a fuzzy membership function for displacement: The displacement risk index is quantified into a membership degree in the range [0,1] using a triangular membership function. The closer the membership degree is to 0, the lower the risk; the closer it is to 1, the higher the risk. The specific expression is as follows:

[0180] ;

[0181] in The maximum value of the displacement. This represents the minimum displacement. , ;

[0182] The velocity membership function is set as follows:

[0183] ;

[0184] in This represents the average value of the velocity change during the steady-state period. This represents the maximum value of the velocity change during the steady-state period;

[0185] The acceleration membership function is set as follows:

[0186] ;

[0187] in This represents the average value of the acceleration change during the steady-state period. This represents the maximum value of acceleration change during the stabilization period. The specific threshold is determined based on historical disaster data and geological conditions of the target landslide area to ensure that it conforms to the local landslide instability pattern. Furthermore, for each additional month of data, the displacement, velocity, and acceleration values ​​during the stabilization period are recalculated and dynamically updated to ensure that the risk assessment is consistent with reality.

[0188] Indicator weight determination: The weights of the three risk indicators were determined using the analytic hierarchy process (AHP). A judgment matrix was constructed and a consistency check was performed to ultimately determine the weight vector. Among them, displacement has the highest weight (0.4), followed by velocity (0.3) and acceleration (0.3).

[0189] Risk value calculation: according to the formula Calculate the overall risk value , To predict the membership degree of displacement in the next 24 hours, The membership degree for the predicted speed over the next 6-12 hours. The membership degree of the predicted acceleration from now to the next 6 hours;

[0190] Risk level classification: based on risk value Landslide risks are classified into four levels, with the specific classification criteria and corresponding countermeasures as follows:

[0191] Blue Alert ( The landslide is in a stable state with no risk of instability; Response measures: Maintain the regular monitoring frequency, no additional intervention is required.

[0192] Yellow alert ( The landslide showed slight deformation and the risk was low. Countermeasures: Increase monitoring frequency to once every 30 minutes, and assign dedicated personnel to inspect surface cracks and seepage at the toe of the landslide.

[0193] Orange alert ( The landslide deformation is obvious and the risk is high; countermeasures: activate the emergency monitoring plan, evacuate temporary personnel and equipment within the landslide-affected area, close surrounding roads, and prepare to evacuate supplies.

[0194] Red Alert ( The landslide is on the verge of instability and poses an extremely high risk. Response measures: Immediately organize the emergency evacuation of all personnel, vehicles, and equipment within the landslide's impact area, activate the geological disaster emergency response, and notify the local government and rescue departments to prepare for rescue operations.

[0195] E6. Early Warning Result Output and Feedback: Early warning results are pushed out in real time through multiple channels, including: sending SMS warnings to managers, which include the warning level, predicted displacement, and recommended measures;

[0196] The monitoring and early warning platform displays early warning information and forecast curves.

[0197] Warning signals are issued through on-site audible and visual alarms. At the same time, an early warning feedback mechanism is established to record information such as early warning time, risk value, predicted data, and actual response status, forming an early warning log. The accuracy of the early warning is evaluated regularly, such as the early warning accuracy rate, false alarm rate, and missed alarm rate. The model parameters, such as the fuzzy membership function threshold and early warning weight, are updated based on the evaluation results to continuously improve the performance of the early warning system.

[0198] Using a typical cohesive soil landslide in Suzhou as the experimental area, this region is a high-incidence area for landslides and has coexisting rainfall characteristics, making the experimental data representative: the plum rain season typically begins in mid-to-late June and ends in mid-July; the city's average annual precipitation is concentrated in the range of 1100-1500 mm; the rainfall process is characterized by "intermittent, mainly light to moderate rain, with concentrated heavy rainfall"; there are as many as 11 potential landslide sites; to conduct professional landslide monitoring, many monitoring stations were set up, including 9 sets of GNSS equipment for comprehensive and high-precision displacement monitoring of the landslide surface, 1 rain gauge, and 1 water level gauge. Data from D4 GNSS monitoring points, rainfall monitoring equipment, and groundwater level monitoring equipment were selected for analysis. The GNSS equipment collected displacement changes at the landslide points; the rainfall monitoring equipment collected meteorological data characteristics including current rainfall, air temperature, and air pressure; and the groundwater level monitoring equipment collected characteristics including soil temperature and soil moisture content. Each device collected soil temperature and soil moisture content data at four different points.

[0199] Experimental data details

[0200] The collected data was compiled into an Excel spreadsheet. The headers represent: Timestamp (time), value_rainfull (current rainfall), value_temperature (temperature), value_air_pressure (air pressure), value_soil_temperature (soil temperature), value_soil_moisture (soil moisture), and displacement (GNSS displacement). The collected rainfall data covers the hourly rainfall intensity from March 7, 2025 to August 19, 2025 (from regional meteorological station sensors, totaling 3951 data points).

[0201] Experimental Environment Details

[0202] Experimental environment: Operating system Windows 11, 13th Gen Intel(R) Core(TM) i7-13700HX 32G RAM, hardware is NVIDIA RTX 5060 graphics card (8GB VRAM), Python environment: Python 3.10, using PyTorch framework, with dependencies including scikit-learn, pandas, numpy, matplotlib, etc.

[0203] Parameter settings and model building

[0204] Data preprocessing parameters

[0205] The first 5 rows of the collected data are in the following format:

[0206] time Current rainfall Total rainfall Temperature air pressure Displacement 2025-03-07 14:00:00 0 0 6.97 99.61 -7.29 2025-03-07 15:00:00 0 0 7.66 99.60 -7.95 2025-03-07 16:00:00 0 0 8.07 99.55 -8.62 2025-03-07 17:00:00 0 0 7.1 99.49 -9.28 2025-03-07 18:00:00 0 0 5.51 99.52 -9.94

[0207] Data preprocessing:

[0208] Outlier removal: The confidence interval for the 3σ criterion is 99.7%, and the outlier proportion threshold for isolated forests is 0.05;

[0209] First, gross errors were removed from the GNSS displacement data using a three-fold mean square error method. Problematic data was then eliminated. Missing acquisition times were filled in using linear interpolation, and Z-score standardization and spatiotemporal alignment were applied. Finally, the processed GNSS displacement data and rainfall data were matched point by point using timestamps in the format "YYYY-MM-DD HH:MM:SS" as a benchmark. This ensured that each time point corresponded to a unique displacement and rainfall value. Ultimately, the data resolution was unified to 1 hour, laying the data foundation for subsequent segmentation of multiple rainfall events.

[0210] Rainfall event segmentation and type differentiation

[0211] The rainfall intensity threshold I is determined using the following calculation formula. ,in The average historical rainfall intensity of the target area in recent years is denoted by K, which is a multiple of 0.2 in this embodiment.

[0212] Determination of the no-rain interval threshold T: Based on the hydrological response characteristics of the landslide body - when the no-rain time exceeds T, the pore water pressure of the landslide body begins to decrease steadily, and the displacement response tends to be gentle. At this time, it is determined to be two independent rainfall events. In this embodiment, T=72 hours.

[0213] Independent event segmentation: When real-time rainfall When three consecutive sampling points (i.e., 3 hours) show rainfall exceeding 1 mm / h, the rainfall event analysis process is initiated. The time interval Δt between two adjacent active rainfall moments is calculated. If Δt ≤ T, it is determined to be the same rainfall event; if Δt > T, it is determined to be different rainfall events. This yields the time intervals [start1, end1], [start2, end2], ..., [start1, end1] for m independent rainfall events. m end m ];

[0214] The standard displacement change during the landslide stabilization period is calculated based on the preprocessed displacement data using the formula. Calculate the displacement-velocity sequence (Unit: mm / h), according to the formula Calculate displacement acceleration sequence (Unit: mm / h²), where The data sampling interval is 1 hour; and the mean μ of the raw data is calculated. In this embodiment, according to historical data, the average historical velocity is 0.3549 mm / h, the maximum velocity is 0.935 mm / h, the average value of the acceleration change is 0.125, and the maximum acceleration is 0.643 mm / h².

[0215] Rainfall lag time optimization: Considering the lag effect of rainfall on landslide displacement, a set of lag times is defined. Calculate the mean of the event correlation coefficient for each quantity. correlation coefficient According to the formula Calculate, and ,in, This represents the start time of the acceleration event in the i-th rainfall event. Calculations show that in this invention… ;

[0216] Construction of a valid event set: will satisfy The events were included in the effective rainfall event set. Each event Corresponding to a unique rainfall sequence With displacement sequence This enables structured storage and management of multi-period events;

[0217] Effective rainfall event type segmentation

[0218] The thresholds for rainstorm intensity (I1) and duration of continuous rain (T1) are set as follows: Rainstorm intensity standards are set according to the municipal government documents of the landslide area to distinguish between rainstorms and ordinary rainfall. The threshold for duration of continuous rain (T1) is set according to the permeability characteristics of the soil and rock mass to distinguish between continuous rain and short-term rainfall.

[0219] According to the design rainfall intensity formula for the urban area of Suzhou, when the duration is less than or equal to 1440 min, the design rainfall intensity formula is:

[0220] ;

[0221] In the formula is the design rainfall intensity (mm / min), is the rainfall duration (min), is the design return period (years). In this invention, the recurrence period is 5 years;

[0222] The distinction between heavy rain and continuous rainy weather can refer to the standards specified by the meteorological bureau or be based on the data collection formula , where is the average of the historical rainfall intensities in the target area in recent years, and K is a multiple, which is taken as 1 in this embodiment;

[0223] Rainfall type distinction: For each independent rainfall event, combine I1 and T1 to determine the type. If there is continuous rainfall with an intensity ≥ I1 for 3 hours or more in the event, it is determined as a heavy rain event. If the duration of the event (end - start + 1) ≥ T1 and the rainfall intensity at all times < I1, it is determined as a continuous rainy event;

[0224] Sequence interception: According to the event time interval, intercept the corresponding rainfall sequence as the heavy rain sequence R1ᵢ (i = 1, 2, …, α) or the continuous rainy sequence R2ⱼ (j = 1, 2, …, β), and at the same time intercept the displacement sequence corresponding to the time interval as the displacement response sequence X1ᵢ or X2ⱼ, satisfying α + β = m, and (L1ᵢ is the length of R1ᵢ, and L2ⱼ is the length of R2ⱼ);

[0225] Multi - dimensional feature extraction: For each heavy rain rainfall event, continuous rainy rainfall event and ineffective rainfall event , construct a feature vector , specifically including the total rainfall , the maximum rainfall intensity , the average rainfall intensity , the rainfall duration , the variance of rainfall intensity , and use Z - Score normalization to eliminate the influence of dimension;

[0226] K-means model parameter training: The multi-dimensional features and rainfall amounts of the classified rainfall events, such as heavy rain, cloudy rain, and invalid rain, are input into the k-means model to train the K-means model parameters and enhance the accuracy of K-means. In this implementation example, a total of 26 rainfall events were identified, including 16 short-term heavy rain events, 4 continuous cloudy rain events, and 6 invalid rain events. The first 5 rainfall events are shown below as examples.

[0227] Rainstorm Event 1: From 03:00:00 on March 12, 2025 to 01:00:00 on March 13, 2025, lasting 22.00 hours;

[0228] Rainstorm Event 2: From 02:00:00 on April 12, 2025 to 01:00:00 on April 13, 2025, lasting 23.00 hours; Rainstorm Event 3: From 14:00:00 on April 21, 2025 to 01:00:00 on April 22, 2025, lasting 11.00 hours;

[0229] Continuous Rainy Weather Event 1: From 07:00:00 on March 27, 2025 to 01:00:00 on March 28, 2025, lasting 18.00 hours;

[0230] Continuous Rainy Weather Event 2: From 10:00:00 on May 5, 2025 to 01:00:00 on May 6, 2025, lasting 15.00 hours;

[0231] Continuous Rainy Weather Event 3: From 06:00:00 on June 10, 2025 to 01:00:00 on June 11, 2025, lasting 19.00 hours;

[0232] Combining the correlation characteristics between rainfall event segmentation and displacement response, from the attached Figure 2 The coupling relationship between rainfall and landslide displacement changes was clearly observed. After segmentation of the original rainfall sequence, the temporal characteristics of short-duration torrential rain events and continuous rainy events were accurately separated. The landslide displacement changed after each rainfall event, exhibiting a lag effect. This effectively demonstrates the crucial influence of rainfall on landslide displacement changes. Figure 2 It can also be observed that after rainy events, landslide displacement changes relatively slowly, while after heavy rain events, landslide displacement often changes abruptly. This confirms that the two types of rainfall have different impact mechanisms on landslides and produce different hazards. Therefore, it is necessary to classify and calculate the impact of rainfall events on landslides.

[0233] Dual-branch PatchTST model construction

[0234] The heavy rain branch is set as follows: sampling interval τ1 = 1 hour, 3×1 convolution kernel (number 32), 2×1 max pooling kernel. The continuous rain branch is set as follows: sampling interval τ2 = 7 hours, 7×1 convolution kernel (number 32), 4×1 average pooling kernel.

[0235] Time Series Patch Branching: Adaptive Patch Partitioning: Patch Length (L is the sequence length);

[0236] Linear embedding and positional encoding: embedding dimension d e =64, and uses sine-cosine position encoding;

[0237] Transformer encoder: 3 layers, 8 multi-head attention heads h, 256 dimensions of FFN hidden layers (GELU activation function).

[0238] Cross-branch attention fusion: Attention feature dimension d=64, weight calculation uses Softmax normalization;

[0239] Dataset Construction: Sliding window size W=24, prediction step size S=[6, 12, 24], training set:validation set=8:2, using the AdamW optimizer, initial learning rate 1e-4, weight decay coefficient 1e-5; training strategy: training epochs=50, batch size=32, early stopping threshold 10 epochs, learning rate using cosine annealing, loss function using multi-scale adaptive hybrid loss function. Experimental verification showed that when... When the time is right, the effect is optimal. All model parameters are obtained based on practical experience and can be used as a reference.

[0240] Landslide displacement prediction and risk assessment

[0241] GNSS receivers and rain gauges deployed in the landslide area are used to collect displacement and rainfall data in real time, at a frequency of once per hour, along with meteorological forecast data for the next 24 hours or more, and real-time rainfall data. If the rainfall is greater than 1 mm / h for 3 consecutive sampling points (i.e. 3 hours), the rainfall event analysis process is initiated. If the rainfall is lower than the threshold and the duration exceeds 6 hours, it is determined to be risk-free rainfall and the analysis is not initiated. For the rainfall event that is currently occurring, its 5-dimensional characteristics are calculated in real time, including real-time total rainfall, real-time maximum rainfall intensity, real-time average rainfall intensity, rainfall duration, and real-time rainfall intensity variance.

[0242] Event Classification

[0243] The standardized features extracted in real time are input into the trained improved K-Means clustering model, which outputs the event type. If the classification result is If the data is not triggered, an alert will not be issued, and the data will only be stored in the historical database for subsequent model updates and optimizations; if (Rainstorm type), select the completed rainstorm branch (Branch1), if (Rainy type), select the continuous rainy branch (Branch2) for multi-scale prediction:

[0244] Multi-scale prediction

[0245] Input sequence construction: Take the start time of the current rainfall event. up to the current time displacement sequence If the sequence length is insufficient If the initial displacement data of similar events (same type of rainfall events) in the same historical period are used for completion, the weight of the completed data is set to 0.5 to avoid prediction deviation due to insufficient sequence length;

[0246] Multi-scale prediction execution: The constructed input sequence is input into the corresponding model branch, and the model outputs the next 6 hours ( ), 12 hours ), 24 hours The displacement prediction value is obtained, and the prediction result is stored in the early warning system database in real time. At the same time, a prediction curve is generated to intuitively display the displacement change trend.

[0247] Multi-dimensional risk indicator calculation

[0248] Predicted velocity calculation: Based on the multi-scale displacement prediction values, according to the formula... Calculate the average speed from now until the next 6 hours. (Average speed over the next 6-12 hours) (Average velocity over the next 12-24 hours), all in mm / h; and calculate the predicted acceleration based on the predicted velocity using the formula: (Average acceleration from now until the next 6 hours) (Average acceleration over the next 6-12 hours), all in mm / h²; where The current real-time displacement velocity is calculated using the current displacement data and the displacement data from the previous hour.

[0249] Fuzzy comprehensive evaluation and risk level classification:

[0250] The formula for the displacement fuzzy membership function is:

[0251] ;

[0252] in The maximum value of the displacement. This represents the minimum displacement. , In this embodiment, the minimum displacement is -42.679 mm, and the maximum displacement is 0.762 mm. The calculation method is as follows:

[0253] Core area length: ;

[0254] Left saturation point: ;

[0255] Right-side saturation point: ;

[0256] Based on the above calculations, the specific formula for the displacement fuzzy membership function in this embodiment is as follows:

[0257] ;

[0258] The velocity membership function can be set as follows:

[0259] ;

[0260] in This represents the average velocity change during the steady-state period. In this embodiment, the average historical velocity is 0.3549 mm / h. The maximum velocity change during the steady-state period is 0.935 mm / h, and the acceleration membership function is set as follows:

[0261] ;

[0262] in The average value of the acceleration change during the steady-state period is 0.125. The maximum value of the acceleration change during the steady-state period is 0.643;

[0263] The weights of the three risk indicators were determined using the Analytic Hierarchy Process (AHP). A judgment matrix was constructed and a consistency check was performed to finally determine the weight vector. Among them, displacement has the highest weight (0.4), because displacement is a direct manifestation of landslide instability. Velocity (0.3) and acceleration (0.3) serve as auxiliary indicators to reflect the trend of displacement changes; according to the formula Calculate the overall risk value , To predict the membership degree of displacement in the next 24 hours, The membership degree for the predicted speed over the next 6-12 hours. The membership degree of the predicted acceleration from now to the next 6 hours;

[0264] Calculations show that the displacement risk for the next 6 hours is 0.000, the velocity risk is 0.57, and the acceleration risk is 0.47. Based on the risk assessment method, the change in the R-risk value over the next 6 hours is 0.313, therefore a yellow alert is issued. Monitoring frequency is increased to once every 30 minutes, and personnel are assigned to inspect the landslide surface for cracks and seepage at the toe of the slope. The change in the R-risk value over the next 24 hours is as follows: Figure 4 As shown in the figure, the risk gradually decreases.

[0265] Early warning result output and feedback:

[0266] Based on comprehensive analysis, the following measures are recommended: 1. Continuously monitor displacement changes; 2. Pay attention to rainfall, especially the effect of a 6-hour delay; 3. Adjust the monitoring frequency according to the risk level.

[0267] Experimental Results and Analysis

[0268] To fully verify the effectiveness of the method of this invention, existing mainstream models (multivariate linear, SVM, random forest, PatchTST) were selected as comparison models. The root mean square error (RMSE) and mean absolute percentage error (MAPE) were used as core evaluation indicators to test the method under heavy rain, continuous rain, and overall scenarios. The experimental results are shown in the table below:

[0269] Model RMSE MAE Linear Regression 7.9498 6.1091 SVM 5.0744 4.013 Original PatchTST 4.3575 3.6675 Rainfall event PatchTST 3.6007 2.9571 Method of the present invention 3.1855 2.635

[0270] Results Analysis

[0271] In this embodiment, to fully verify the superiority of the proposed method, five types of models were selected: multiple linear regression, SVM (Support Vector Machine), original PatchTST, rainfall event PatchTST, and the method of this invention. Under the same experimental data and environment, comparative tests were conducted using two core indicators: root mean square error (RMSE) and mean absolute error (MAE). The test results showed that the linear regression model had the worst prediction accuracy, with an RMSE of 7.9498 and an MAE of 6.1091. This result fully demonstrates that linear regression, as a traditional statistical model, can only capture the linear correlation between data. However, landslide displacement is affected by multiple nonlinear factors such as rainfall type, lag effect, and cumulative effect, and its coupling relationship with rainfall is extremely complex. Therefore, the linear regression method cannot accurately characterize the nonlinear correlation mechanism in landslide prediction, resulting in a significantly higher prediction error.

[0272] Secondly, the SVM model, with an RMSE of 5.0744 and a MAE of 4.013, showed a significant improvement in prediction accuracy compared to linear regression. This improvement is attributed to the kernel function mapping capability of SVM as an intelligent algorithm, which can transform the complex relationship in the original feature space into a linearly separable problem in a high-dimensional space through nonlinear mapping. This allows it to capture the nonlinear correlation between rainfall and landslide displacement to a certain extent, demonstrating that intelligent algorithms have better adaptability in landslide displacement prediction scenarios and can effectively improve prediction accuracy.

[0273] The prediction performance of the original PatchTST model has been further improved, with RMSE reduced to 4.3575 and MAE to 3.6675. This model relies on the Patch partitioning strategy to divide long time series data into several local feature blocks. The Transformer encoder captures the temporal dependencies between feature blocks, giving full play to the core advantages of Patch processing in time series data modeling. It can more accurately explore the intrinsic evolution law of landslide displacement time series. Therefore, compared with traditional linear models and basic intelligent algorithms, it shows better prediction results.

[0274] The PatchTST model for rainfall events has been specifically optimized based on the original PatchTST. Its core improvement lies in using rainfall events as the data processing unit and filtering valid rainfall events through displacement acceleration correlation analysis. Invalid rainfall data that did not trigger a significant landslide response were eliminated. This design reduces the interference of invalid information from the data source and reduces noise in model training, enabling the model to focus on the key rainfall events that actually affect landslide stability. Therefore, the robustness is significantly enhanced and the prediction accuracy is further improved, with RMSE reaching 3.6007 and MAE reaching 2.9571.

[0275] Finally, the method of this invention exhibits the best prediction performance, with an RMSE of only 3.1855 and a MAE as low as 2.635. This method not only continues the event-driven data processing logic of PatchTST for rainfall events, but also innovatively introduces a dual-branch model architecture. It achieves accurate differentiation of rainfall types by using the threshold of rainstorm intensity and the threshold of continuous rain duration. It specifically designs short-term rainstorm channels and continuous rain channels. The rainstorm channel uses high-frequency sampling and small-sized convolutional kernels to accurately capture triggering features such as sudden changes in rainfall intensity and peak fluctuations. The continuous rain channel uses low-frequency sampling and large-sized convolutional kernels to fully explore the cumulative infiltration effect of long-term rainfall. At the same time, the cross-branch attention fusion module dynamically adjusts the weight ratio of the two types of features to achieve full-process optimization of "type differentiation - dedicated extraction - dynamic fusion", which effectively solves the problem of feature confusion under different rainfall scenarios. Therefore, it has the highest processing accuracy in landslide displacement prediction.

[0276] The above comparative results fully demonstrate that the method of this invention, through multi-dimensional technological innovation, has successfully overcome the limitations of existing models in nonlinear modeling, rainfall type adaptation, and invalid information filtering. It has achieved accurate modeling of the combined disaster-causing mechanism of "rainstorm triggering + long-term infiltration accumulation," maintaining high prediction accuracy under different rainfall scenarios and exhibiting significantly better robustness than various comparative models. Furthermore, this invention does not merely stop at the displacement prediction level but uses fuzzy comprehensive evaluation to transform predicted displacement into intuitive risk levels (blue, yellow, orange, and red warnings), directly outputting landslide risk values ​​for a future period without requiring professional interpretation. Simultaneously, it pushes warning information in real time through multiple channels such as SMS, monitoring platforms, and on-site audible and visual alarms, allowing managers sufficient emergency response time. This effectively solves the practical problems of existing methods such as "disconnect between prediction and risk assessment," "high professional threshold," and "delayed response," providing comprehensive and reliable technical support for landslide disaster monitoring and early warning.

Claims

1. A landslide risk assessment method based on the PatchTST model of multi-period rainfall events, characterized in that: Specifically, the steps include the following: Step 1: Data Preprocessing and Spatiotemporal Alignment: Collect GNSS displacement time series data and rainfall time series data for the target landslide area. The data is denoised and missing values ​​are filled in. The two types of data are aligned in time and space by timestamp matching to ensure consistent data resolution. Step 2: Segmentation of Multiple Rainfall Events: Based on Rainfall Intensity Thresholds And the no-rain interval threshold T, for the preprocessed rainfall sequence Multi-phase rainfall events were identified and segmented, and the displacement rate of landslides at each moment during the rainfall period was calculated. ,speed With acceleration And solve for the optimal rainfall impact lag time, based on The system identifies invalid and valid rainfall events by size, backtracks and correlates them to obtain valid rainfall events, and constructs an event set. Step 3: Effective Rainfall Event Type Segmentation: Based on the rainstorm intensity threshold provided by the local government. Distinguish between heavy rain event sequences and continuous rainy event sequences, resulting in m independent rainfall event sequences (including heavy rain sequences { , … } and continuous rain sequence { , … }, + =m) and the corresponding displacement response sequence { , … Finally, the clustering model for rainfall events is trained. Step 4: Construction of the two-branch PatchTST model: Taking the independent rainfall event sequence as input, the complete two-branch PatchTST model is constructed through steps such as two-branch feature sequence construction, sequence segmentation and projection, position encoding, independent feature extraction by Transformer encoder, cross-branch attention fusion, and model training. Step 5, Displacement Prediction and Risk Assessment: Input the rainfall data characteristics of the period to be predicted into the clustering model to identify rainstorm events, cloudy and rainy events and invalid rainfall events. For valid rainfall events, the landslide displacement prediction results are output by the two-branch PatchTST model. Combined with the fuzzy comprehensive evaluation method, the risk value is calculated and the warning level is divided. The warning results are output in real time, and corresponding measures are provided.

2. The landslide risk assessment method based on the PatchTST model using multi-period rainfall events as described in claim 1, characterized in that: The specific process of data preprocessing in step one includes: a1. Data Acquisition Phase: Displacement data is acquired through a GNSS receiver, rainfall data is acquired through a small weather station, and environmental data is acquired using temperature and humidity sensors and soil moisture sensors. The acquisition frequency is once per hour. a2. Data Standardization: Data is processed using Z-score standardization, with the formula as follows: ,in, The original data, For standardized data, The mean of the data. The standard deviation of the data; a3. Missing Value Imputation: When a single data point is missing but the preceding and following data are valid, linear interpolation is used to impute it. The formula is as follows: If more than three consecutive sampling points are missing, the missing data will be supplemented by combining the average displacement data under similar weather conditions in the same period of history to avoid bias in subsequent event analysis due to missing data. a4. Spatiotemporal alignment: Based on the timestamp in the format "YYYY-MM-DD HH:MM:SS", the processed GNSS displacement and rainfall data are matched point by point to ensure that each time node corresponds to a unique displacement value and rainfall value, and the data resolution is unified to 1 hour.

3. The landslide risk assessment method based on the PatchTST model using multi-period rainfall events as described in claim 1, characterized in that: Step two, which involves segmenting and transforming multiple rainfall events based on displacement response, specifically includes: b1. Rainfall Time Recording and Displacement Parameter Calculation: Based on rainfall and non-rainfall times recorded by rain gauges, and combined with historical landslide displacement and rainfall data, the maximum rainfall delay time is determined through statistical analysis. The rainfall event time is the rainfall start time plus the lag time. Based on the preprocessed displacement data, the calculation is performed according to the formula... Calculate the displacement-velocity sequence (Unit: mm / h), according to the formula Calculate displacement acceleration sequence (Unit: mm / h²), where Hour; b2. Risk Event Judgment: When Three consecutive sampling points (i.e., 3 hours) exceeding And the displacement increment during that period If the event occurs, it is considered a displacement acceleration event, and the start time of the event is recorded. With end time ( for First time below (time) b3. Calculation of Rainfall Lag Time and Event Correlation: The impact of rainfall on landslide displacement is not instantaneous but exhibits a significant lag effect. This lag primarily stems from the time required for rainwater to infiltrate the landslide body, the time needed for soil and rock moisture content to accumulate, and the gradual nature of pore water pressure changes and the deterioration of soil and rock mechanical properties. Ignoring this lag effect and directly modeling the correlation between rainfall data and concurrent displacement data will lead to a misalignment of the causal relationship between rainfall and displacement acceleration events, thereby introducing invalid rainfall interference and reducing the model's prediction accuracy. Therefore, optimizing the rainfall lag time is a crucial step in accurately constructing the "rainfall event-displacement response" correlation and selecting effective rainfall events. The specific steps are as follows: b4. Definition of Lag Time Set Based on the rapid infiltration effect of short-duration torrential rain, which typically triggers displacement response within hours, and the slow infiltration effect of long-duration continuous rain, which requires several days to accumulate before displacement changes become apparent, a lag time set is defined. However, by using discrete values ​​with a 6-hour interval, the computational complexity is reduced and the optimization efficiency is improved while ensuring the integrity of the coverage. b5. Correlation of time node calibration For each identified displacement acceleration event i (whose start time is...) The end time is Based on each candidate value in the set of lag times Calculate the corresponding rainfall association start time. It needs to be made clear that This is not the actual meteorological observation start time of the rainfall event, but rather the "effective rainfall start time" calibrated to establish a causal relationship between rainfall and displacement acceleration. The core logic is that the occurrence of a displacement acceleration event is essentially triggered by the cumulative effect of rainfall over a preceding period, through a lag time... By retrospectively matching displacement acceleration events with the rainfall periods that actually caused the events, invalid rainfall interference that did not trigger a displacement response can be eliminated. b6. Calculation of Event Relevance Coefficient To quantify the causal relationship between rainfall and displacement acceleration events with different lag times, a correlation coefficient is defined. The calculation formula is as follows: ； Among them: molecule Displacement acceleration during the associated time period With rainfall The integral reflects the degree of coordinated change between rainfall intensity and displacement acceleration effect—when the rainfall intensity is large and the displacement acceleration is significant during the same period, the integral value increases, indicating that the two are closely related; denominator The total rainfall within the associated period is integrated for normalization, eliminating the impact of differences in total rainfall on the correlation assessment and making the correlation of events with different rainfall intensities and durations comparable. b7. Optimal Lag Time Screening For the set of lag times For each candidate value, iterate through all identified displacement acceleration events and calculate the correlation coefficient for each event at that lag time. And find all events mean This mean reflects the suitability of the candidate lag time to the overall event, by comparing the values ​​corresponding to all candidate lag times. ,choose Maximum As the optimal lag time; b8. Construction of a valid event set: This will satisfy... The events were included in the effective rainfall event set. Each event Corresponding to a unique rainfall sequence With displacement sequence This enables structured storage and management of multi-period events; b9. Invalid Rainfall Marker: Rainfall periods not associated with any displacement acceleration events. Rainfall data marked as invalid is not included in model training and is directly assigned a low risk score.

4. The landslide risk assessment method based on the PatchTST model using multi-period rainfall events as described in claim 1, characterized in that: Step three, the extraction of effective rainfall event features and classification, specifically includes: c1. Characteristic Differentiation: For each identified valid rainfall event, a rainfall intensity threshold is set. The threshold can be obtained from local meteorological files or set manually from the local average rainfall intensity; if the event has rainfall intensity ≥ within a set time period... If so, it is determined to be a rainstorm event, corresponding to a surface landslide caused by rapid saturation of shallow soil; If the rainfall intensity of the event is < The event was determined to be a continuous rainy season, which corresponds to a deep landslide caused by the slow infiltration of rainwater raising the groundwater level. Extract the corresponding rainfall sequence as the rainstorm sequence Or a continuous rainy sequence Simultaneously, the displacement sequence within the corresponding time interval is extracted as the displacement response sequence. or ,in =1,2,… , =1,2,… ; c2. Multi-dimensional feature extraction: This involves extracting features for each heavy rain event, continuous rain event, and invalid rainfall event. Construct a multidimensional feature vector for each event. The calculation methods and physical meanings of each event are as follows: Total rainfall According to the formula The calculation, expressed in mm, represents the total amount of rainfall and directly affects the accumulation of soil moisture content; among which, Indicates the start time of rainfall. Indicate the target time; Maximum rainfall intensity According to the formula The calculation, in mm / h, reflects the extreme intensity of rainfall and determines the likelihood of rapid soil saturation. Average rainfall intensity : Characterizes the average intensity of rainfall and is related to the soil infiltration rate; Rainfall duration This reflects the duration of rainfall and affects the depth of rainfall infiltration. Rainfall intensity variance Unit is This reflects the degree of fluctuation in rainfall intensity and is related to the frequency of changes in soil stress. c3. Feature Standardization: Z-Score standardization is used to eliminate the influence of dimensions. The standardization formula is: ; in For the first The mean of each feature, For the first The standard deviation of each feature The total number of effective rainfall events; c4. Feature Clustering: Input the multi-dimensional features of the identified rainstorm events, the multi-dimensional features of the continuous rain events, the multi-dimensional features of the invalid rainfall events, and the corresponding rainfall into the semi-supervised K-means model for feature clustering training.

5. The landslide risk assessment method based on the PatchTST model using multi-period rainfall events as described in claim 1, characterized in that: The construction of the two-branch PatchTST model based on rainfall events in step four includes: d1, Rainstorm Feature Processing Channel: Processing rainstorm sequences High-frequency sampling (sampling interval) is used =1-3 hours, preferably 1 hour), construct a convolutional layer with a small 3×1 kernel, the formula is: ； in Convolution operation specifically for heavy rain. , The convolution kernel weights and biases are used respectively, and a 2×1 max pooling operation is applied to enhance the information of the rainstorm peak, resulting in rainstorm sub-features. ; d2. Long-term infiltration and cumulative continuous rainy event channel: for continuous rainy sequences Low-frequency sampling (sampling interval) is used =7-24 hours), construct convolutional layers with large-size convolutional kernels of 7×1-9×1, the formula is: ； in Convolution operation specifically for continuous rainy weather. , The convolutional kernel weights and biases are used respectively. A 4×1 average pooling operation is applied to smooth local rainfall fluctuations while preserving the cumulative trend, thus obtaining the sub-features of continuous overcast and rainy weather. ; d3. Sequence Segmentation and Projection: For the rainstorm event sequence { , , …, } and the sequence of continuous rainy events { , , …, Each independent sequence in} is processed. For a one-dimensional sequence of length L, it is divided into N=L / P sequence blocks by a non-overlapping sliding window of size P. Each sequence block is mapped to a D-dimensional block embedding vector through a learnable linear projection layer, and then position encoding is completed. d4. Dual-branch input construction: The blocks of all rainstorm events are embedded and stacked in batches to form the rainstorm branch feature tensor. All blocks of continuous rainy events are embedded and stacked in batches to form a continuous rainy branch feature tensor. ; d5. Independent feature extraction from the Transformer encoder: and The features are concatenated along the feature dimension and fed into the shared Transformer encoder backbone network, outputting a rainstorm context feature sub-matrix Fr1 and a continuous rainy weather context feature sub-matrix Fr2, respectively. It's important to note that the rainstorm context feature sub-matrix Fr1 is no longer the original one. It contains high-level features that have been deeply mined by Transformer and take into account the complex temporal relationships between all blocks within the rainstorm event. The corresponding continuous rain context feature submatrix Fr2 contains high-level features that have been modeled and show long-term, cumulative temporal dependencies within the continuous rain event. d6. Cross-branch attention fusion: Construct a cross-branch attention fusion module, taking Fr1 and Fr2 as inputs, and calculate the cross-attention weights. and ( =1- ), and according to the formula Fr= ·Fr1+ • Fr2 is dynamically weighted and fused to obtain a unified rainfall fusion feature matrix Fr; d7. Model Training: The fused feature matrix Fr is used as the model input. The standard PatchTST prediction head is used, which includes a normalization layer, a flattening layer and a linear projection layer. The landslide displacement sequence X corresponding to the time window is used as the prediction target. The training of the two-branch PatchTST model is completed using a hybrid loss function and the AdamW optimizer. To meet the specific needs of predicting landslide displacement risks in the following 6, 12, and 24 hours, the prediction tasks at different scales differ significantly. Short-term predictions focus on immediate displacement responses, with relatively smooth data fluctuations but high sensitivity to accuracy. Long-term predictions need to cover the complete lag effect of rainfall-displacement, with drastic data fluctuations and easily accumulating errors. If the model is trained by directly summing the original loss values, the large absolute error in long-term predictions will lead to training bias, weakening the optimization of accuracy in short- and medium-term predictions. Therefore, a multi-scale adaptive hybrid loss function is designed, the specific formula of which is as follows: ; The definitions of each item and parameter are as follows: ， ; This represents the multi-scale normalized loss value based on the mean squared error (MSE), which is suitable for data with high quality and no significant outliers. Furthermore, it is crucial for scenarios with stringent requirements for prediction accuracy, such as short-term displacement monitoring during the stabilization phase of a landslide. This represents the multi-scale normalized loss value based on the mean absolute error (MAE), which is suitable for scenarios where rainfall is concentrated and displacement is prone to sudden outliers, such as predicting the active period of landslides triggered by rainstorms. The value ranges from 0 to 1 and is used to adjust the weight ratio of the two core loss terms to adapt to different scenario requirements. k is the scale identifier, with values ​​of 1, 2, and 3, corresponding to three prediction scales of 6 hours, 12 hours, and 24 hours, respectively. and The first Mean squared error loss and mean absolute error loss at each scale, among which Predicted values ​​at various scales The actual value; Data distribution normalization factor, representing the first The standard deviation of the true displacement sequence at each scale is used to eliminate differences in dimensionality and distribution dispersion of displacement data at different scales. This is the duration balancing factor, with corresponding values ​​of 6, 12, and 24. This is a scale-adaptive adjustment coefficient, an adjustable parameter used to adapt to the priority requirements of different early warning scenarios.

6. The landslide risk assessment method based on the PatchTST model using multi-period rainfall events as described in claim 1, characterized in that: The real-time landslide risk forecasting in step five specifically includes: e1. Real-time data acquisition and transmission: The GNSS receiver and rain gauge acquire data once per hour, and trigger an alarm when the data is abnormal; e2. Real-time event detection and classification: Real-time rainfall The rainfall event analysis process is initiated when the rainfall is greater than 1 mm / h at 3 consecutive sampling points. If the rainfall is less than this threshold and the duration exceeds 3 hours, it is determined to be a risk-free rainfall. Real-time calculation of five-dimensional features, including real-time total rainfall, real-time maximum rainfall intensity, real-time average rainfall intensity, duration, previous 24-hour rainfall, real-time rainfall intensity variance, displacement increment, and real-time peak acceleration. The standardized features extracted in real time are input into the clustering model trained in step three, and the event type is output. If the classification result is If the data is not triggered, an alert will not be issued, and the data will be stored in the historical database for subsequent model updates and optimizations. e3. Model Branching and Multi-Scale Prediction: Based on Real-Time Event Classification Results Automatically select the corresponding model branch and retrieve the start time of the current rainfall event. up to the current time displacement sequence ; If the sequence length is insufficient Then, the average initial displacement data of similar events in the same historical period is used for completion, and the weight of the completed data is set to 0.

5. The model outputs the next 6 hours ( ), 12 hours ), 24 hours The displacement prediction value is obtained, and the prediction result is stored in the early warning system database in real time. At the same time, a prediction curve is generated to show the displacement change trend. e4. Calculation of Multi-Dimensional Risk Indicators: Based on multi-scale displacement prediction values, the predicted velocity for different time periods is calculated using the formula: the average velocity from the present to the next 6 hours. Average speed over the next 6-12 hours Average speed over the next 12-24 hours All units are mm / h; Predicted acceleration calculation: Average acceleration from now until the next 6 hours Average acceleration over the next 6-12 hours All units are mm / h², where The current real-time displacement velocity is calculated using the current displacement data and the displacement data from the previous hour. e5. Fuzzy comprehensive evaluation and risk level classification: The risk index is quantified by the triangular membership function. Velocity boundary point and acceleration boundary point are set. The specific thresholds are determined based on the historical disaster data and geological conditions of the target landslide area to ensure that they conform to the local landslide instability law. The displacement fuzzy membership function is set as follows: ； in The maximum value of the displacement. This represents the minimum displacement. , ; The velocity membership function is set as follows: ； in This represents the average value of the velocity change during the steady-state period. This represents the maximum value of the velocity change during the steady-state period; The acceleration membership function is set as follows: ； in This represents the average value of the acceleration change during the steady-state period. This represents the maximum value of acceleration change during the stabilization period. The specific threshold is determined based on historical disaster data and geological conditions of the target landslide area to ensure that it conforms to the local landslide instability pattern. Furthermore, for each additional month of data, the displacement, velocity, and acceleration values ​​during the stabilization period are recalculated and dynamically updated to ensure that the risk assessment is consistent with reality. The weights of the three risk indicators were determined using the analytic hierarchy process (AHP). A judgment matrix was constructed and a consistency check was performed to finally determine the weight vector. The risk value is calculated based on three risk weight indicators, with displacement having the highest weight (0.4), followed by velocity (0.3) and acceleration (0.3). e6. Early warning result output and feedback: Push early warning results through multiple channels, record early warning information to form a log, and regularly evaluate the accuracy of early warning and update model parameters.

7. The landslide risk assessment method based on the PatchTST model using multi-period rainfall events as described in claim 6, characterized in that: The formula for calculating the risk value is: ； For the comprehensive risk value, To predict the membership degree of displacement in the next 24 hours, The membership degree for the predicted speed over the next 6-12 hours. The membership degree of the predicted acceleration from now to the next 6 hours is determined, and finally, based on the comprehensive risk value... Determine the scope of the warning.

8. The landslide risk assessment method based on the PatchTST model using multi-period rainfall events as described in claim 7, characterized in that: The basis of comprehensive risk value The specific categories of the warning scope are as follows: Blue Alert ( ); Yellow alert ( ); Orange alert ( ); Red Alert ( ).

Images