An embedded landslide real-time early warning method based on space-time feature decoupling

Abstract

The application discloses a kind of embedded landslide real-time early warning method based on space-time feature decoupling, belong to geological disaster intelligent monitoring and early warning technical field.It includes: obtaining the historical time series monitoring data of landslide body and constructing training data set;Space-time feature decoupling is carried out to data, and global component and detail component are decomposed;A double-branch lightweight time series convolution network containing global branch and detail branch is constructed, and joint training and adaptive weight fusion are completed in the cloud;The trained network is differentially lightweighted and deployed in embedded terminal;In embedded end, feature decoupling and prediction are carried out using real-time data stream, and the predicted value and uncertainty estimation are obtained by combining lightweight monte carlo dropout method, finally dynamic threshold calculation and graded early warning are realized.The application effectively solves the problem that existing method is difficult to balance prediction accuracy and calculation efficiency in embedded terminal, realizes real-time, accurate and adaptive early warning of landslide disaster.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and early warning technology for geological disasters, and more specifically, to an embedded real-time landslide early warning method based on spatiotemporal feature decoupling. Background Technology

[0002] Landslides are a serious geological disaster, and accurate real-time early warning is crucial to protecting people's lives and property. With the development of deep learning technology, displacement prediction models based on neural networks have become a research hotspot. However, when deploying high-precision models on embedded terminals with limited computing, storage, and power resources, there are several challenges: (1) The contradiction between model complexity and resource constraints: High-precision models (such as standard TCN and LSTM) have a large number of parameters and are computationally complex, making it difficult to run in real time on embedded terminals; (2) The contradiction between early warning sensitivity and model robustness: While the model can be reduced by performing rough uniform lightweighting (such as overall pruning and quantization), it will also impair the model's ability to grasp long-term evolution trends and its sensitivity to short-term sudden displacements (key signals of landslide precursors), leading to missed or false alarms; (3) The contradiction between deterministic prediction and risk decision-making: Traditional point prediction cannot quantify the uncertainty of prediction results and is difficult to support probability-based risk classification early warning decisions.

[0003] Therefore, there is an urgent need in this field for a landslide real-time early warning method that can run efficiently on embedded terminals while maintaining high prediction accuracy and sensitive capture of sudden change signals. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an embedded real-time landslide early warning method based on spatiotemporal feature decoupling. This method trains and performs deep lightweight processing of a high-precision model in the cloud and deploys it on an embedded terminal, aiming to solve the problem that complex deep learning models cannot run in real time under resource-constrained environments. Simultaneously, by introducing a dynamic early warning mechanism based on model uncertainty, the early warning threshold is adaptively adjusted to address the insufficient sensitivity and reliability of traditional static threshold methods when dealing with the uncertainty of landslide evolution. This provides reliable technical support for achieving embedded, high-precision, adaptive real-time early warning of landslide disasters.

[0005] This invention is implemented as follows:

[0006] The technical solution to achieve the purpose of this invention is: an embedded real-time landslide early warning method based on spatiotemporal feature decoupling, comprising the following steps:

[0007] Step 1: Obtain historical time-series monitoring data of the landslide body and construct a training dataset based on the sliding window method. Each training sample is used to predict the monitoring data of one or more time points after its corresponding window.

[0008] Step 2: Decouple the historical time series monitoring data by spatiotemporal features, and decompose the original sequence into global components that characterize long-term evolution trends and local detail components that characterize short-term fluctuations and mutations.

[0009] Step 3: Construct a two-branch lightweight temporal convolutional network. The input of this network is the length. L The monitoring data sequence is output as the predicted values ​​for one or more time points as described in step one; the network includes a global branch and a detail branch.

[0010] The total number of depthwise separable convolutional layers in the global branch is greater than the total number of depthwise separable convolutional layers in the detail branch.

[0011] Step 4: Jointly train the dual-branch lightweight temporal convolutional network using the training dataset in the cloud, and use an adaptive weight fusion module to fuse the outputs of the two branches to obtain the final prediction value;

[0012] Step 5: Perform differentiated lightweighting processing on the trained dual-branch lightweight temporal convolutional network:

[0013] The global branch is subjected to high-intensity pruning and quantization; the detailed branch is subjected to low-intensity pruning and quantization or full precision is preserved.

[0014] Step Six: Deploy the lightweighted model on an embedded terminal;

[0015] Step 7: In the embedded terminal, maintain a length of [length missing] for the real-time acquired monitoring data stream. L The latest data sliding window; the data in this window is preprocessed using the same spatiotemporal feature decoupling method as in step two;

[0016] Step 8: Input the decoupled global component and detail component into the deployed dual-branch network respectively to obtain the predicted values. T b ;

[0017] Step 9: Obtain the uncertainty estimate of the model predictions based on the lightweight Monte Carlo Dropout method. U t And based on the predicted value T b and uncertainty estimation U t Calculate the dynamic early warning threshold;

[0018] Step 10: Compare real-time monitoring data with dynamic early warning thresholds to achieve tiered early warning.

[0019] Furthermore, in step two, the spatiotemporal feature decoupling is achieved through adaptive frequency threshold filtering, specifically including:

[0020] S21: Perform a fast Fourier transform on the original time-series data and analyze its power spectral density;

[0021] S22: Reconstruct the low-frequency components in the power spectrum that have a cumulative contribution rate of 80%~90% into the global components;

[0022] S23: Reconstruct the remaining high-frequency components into the detailed components.

[0023] Furthermore, in step three, each branch of the dual-branch lightweight temporal convolutional network is composed of sequentially connected lightweight temporal convolutional modules; the lightweight temporal convolutional module sequentially includes: a depthwise separable convolutional layer, a linear rectified unit activation function layer, and a random deactivation layer.

[0024] Furthermore, in step four, the joint training refers to using the global component and detail component obtained by decoupling from the same original sequence as the inputs to the global branch and detail branch, respectively, and simultaneously optimizing the parameters of the two branches end-to-end under the supervision of a unified loss function.

[0025] Furthermore, in step four, the adaptive weight fusion module generates its weights based on data, specifically as follows:

[0026] S41: Concatenate the feature maps of the final layers of the two branches;

[0027] S42: Generate a two-dimensional weight vector through a lightweight fully connected layer or a 1x1 convolutional layer. α ,1- α ];

[0028] S43: Final predicted value = α ×Global branch output+(1- α )×Detailed branch output

[0029] The weight α reflects the relative importance of global trends and local details in the current input data.

[0030] Furthermore, in step five, the differentiated lightweight processing specifically refers to: for global branches, adopting an iterative pruning strategy to achieve a target sparsity of 70% to 90%, and performing 8-bit integer quantization; for detail branches, adopting one-time pruning to achieve a target sparsity of less than 30%, and performing 16-bit floating-point quantization or maintaining 32-bit full precision.

[0031] The iterative pruning strategy specifically includes:

[0032] Step S51: Perform sparsification training on the converged training model;

[0033] Step S52: Perform pruning, removing the 10%-20% of connections with the smallest absolute weight;

[0034] Step S53: Retrain the pruned model to recover the lost performance;

[0035] Step S54: Repeat the pruning and retraining steps 3-5 times until the target sparsity is achieved.

[0036] Furthermore, the lightweight Monte Carlo Dropout method is used to obtain an uncertainty estimate of the model predictions. U t The method is as follows:

[0037] When performing inference on an embedded terminal, the same input data is executed. K Each forward propagation randomly activates the Dropout layer, resulting in... K Different prediction outputs ;

[0038] By calculating this K The standard deviation σ of the predicted output yields an uncertainty estimate. U t ,Right now:

[0039]

[0040] in, K It is an integer between 2 and 4.

[0041] Furthermore, in step nine, the dynamic early warning threshold is dynamically calculated based on historical data volatility and model uncertainty. The calculation method is as follows:

[0042] T dmax = T b ×(1+ α × U t / U max )

[0043] T dmin = T b ×(1- α × Ut / U max )

[0044] in: T dmax This is the upper limit of the dynamic early warning threshold; T dmin This is the lower limit of the dynamic early warning threshold; T b This is a predicted value; U max This represents the historical maximum value of the uncertainty estimate during the training phase; α The adjustment coefficient has a range of 0.1 to 0.3.

[0045] Furthermore, in step ten, the tiered early warning is determined according to the following rules:

[0046] (1) Landslide displacement anomaly early warning: When real-time monitoring data > T dmax At that time, calculate the ratio of real-time monitoring data exceeding the upper limit of the dynamic early warning threshold:

[0047] Exceeding the limit ratio = (real-time monitoring data - ...) T dmax ) / U t

[0048] The warning level is determined based on the excess ratio:

[0049] When the excess ratio is greater than 1.0, it is determined to be a Level 3 red alert;

[0050] When the ratio of 1.0 to the above exceeds 0.5, it is determined to be a Level 2 orange alert;

[0051] When 0.5 ≥ the above excess ratio > 0, it is determined to be a Level 1 Yellow Alert;

[0052] (2) Monitoring system anomaly alert: When real-time monitoring data < T dmin At that time, calculate the ratio of the real-time monitoring data to the lower limit of the dynamic early warning threshold:

[0053] Excess ratio = ( T dmin -Real-time monitoring data) / U t

[0054] When the current exceedance ratio is greater than 0.5, a system anomaly warning is triggered;

[0055] (3) Other situations are classified as Level 0 normal.

[0056] The beneficial effects of this invention are as follows: This invention provides an embedded real-time landslide early warning method based on spatiotemporal feature decoupling. First, through the collaborative design of "spatiotemporal feature decoupling" and "differentiated lightweighting," the prediction accuracy for landslide displacement, especially short-term abrupt signals, is preserved to the maximum extent while achieving model volume compression. Second, dynamic thresholds are output based on the uncertainty of the model itself, and multi-level hierarchical early warning is achieved, making the early warning strategy more intelligent and refined. Third, it is embedded-friendly: the entire method, from model structure (depth-separable convolution) to back-end processing (differentiated compression, lightweight MC-Dropout), fully considers the resource constraints of embedded terminals, enabling the efficient and real-time operation of complex algorithms on low-power devices. In summary, the method of this invention can provide more reliable technical support for real-time prediction and early warning of slope and dam displacement. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0058] Figure 1 This is a flowchart of an embedded landslide real-time early warning method based on spatiotemporal feature decoupling provided by an embodiment of the present invention;

[0059] Figure 2 This is a diagram of the dual-branch lightweight temporal convolutional network structure provided in the embodiments of the present invention;

[0060] Figure 3 This is a schematic diagram of the slope displacement prediction and graded early warning results provided by the embodiments of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the implementation examples and accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The method of this invention will be illustrated below using a specific case of real-time landslide displacement prediction.

[0063] like Figure 1 An embedded real-time landslide early warning method based on spatiotemporal feature decoupling includes the following steps:

[0064] Step 1: Obtain historical time-series monitoring data of the landslide body and construct a training dataset based on the sliding window method. Each training sample is used to predict the monitoring data of one or more time points after its corresponding window.

[0065] Step 2: Decouple the historical time series monitoring data by spatiotemporal features, and decompose the original sequence into global components that characterize long-term evolution trends and local detail components that characterize short-term fluctuations and mutations.

[0066] Step 3: Construct a two-branch lightweight temporal convolutional network. The input of this network is the length. L The monitoring data sequence is output as the predicted values ​​for one or more time points as described in step one; the network includes a global branch and a detail branch.

[0067] The total number of depthwise separable convolutional layers in the global branch is greater than the total number of depthwise separable convolutional layers in the detail branch.

[0068] Step 4: Jointly train the dual-branch lightweight temporal convolutional network using the training dataset in the cloud, and use an adaptive weight fusion module to fuse the outputs of the two branches to obtain the final prediction value;

[0069] Step 5: Perform differentiated lightweighting processing on the trained dual-branch lightweight temporal convolutional network:

[0070] The global branch is subjected to high-intensity pruning and quantization; the detailed branch is subjected to low-intensity pruning and quantization or full precision is preserved.

[0071] Step Six: Deploy the lightweighted model on an embedded terminal;

[0072] Step 7: In the embedded terminal, maintain a length of [length missing] for the real-time acquired monitoring data stream. L The latest data sliding window; the data in this window is preprocessed using the same spatiotemporal feature decoupling method as in step two;

[0073] Step 8: Input the decoupled global component and detail component into the deployed dual-branch network respectively to obtain the predicted values. T b ;

[0074] Step 9: Obtain the uncertainty estimate of the model predictions based on the lightweight Monte Carlo Dropout method. U t And based on the predicted value T b and uncertainty estimation U t Calculate the dynamic early warning threshold;

[0075] Step 10: Compare real-time monitoring data with dynamic early warning thresholds to achieve tiered early warning.

[0076] Furthermore, in step two, the spatiotemporal feature decoupling is achieved through adaptive frequency threshold filtering, specifically including:

[0077] S21: Perform a fast Fourier transform on the original time-series data and analyze its power spectral density;

[0078] S22: Reconstruct the low-frequency components in the power spectrum that have a cumulative contribution rate of 80%~90% into the global components;

[0079] S23: Reconstruct the remaining high-frequency components into the detailed components.

[0080] Furthermore, in step three, each branch of the dual-branch lightweight temporal convolutional network is composed of sequentially connected lightweight temporal convolutional modules; the lightweight temporal convolutional module sequentially includes: a depthwise separable convolutional layer, a linear rectified unit activation function layer, and a random deactivation layer.

[0081] Furthermore, in step four, the joint training refers to using the global component and detail component obtained by decoupling from the same original sequence as the inputs to the global branch and detail branch, respectively, and simultaneously optimizing the parameters of the two branches end-to-end under the supervision of a unified loss function.

[0082] Furthermore, in step four, the adaptive weight fusion module generates its weights based on data, specifically as follows:

[0083] S41: Concatenate the feature maps of the final layers of the two branches;

[0084] S42: Generate a two-dimensional weight vector through a lightweight fully connected layer or a 1x1 convolutional layer. α ,1- α ];

[0085] S43: Final predicted value = α ×Global branch output+(1- α )×Detailed branch output

[0086] The weight α reflects the relative importance of global trends and local details in the current input data.

[0087] Furthermore, in step five, the differentiated lightweight processing specifically refers to: for global branches, adopting an iterative pruning strategy to achieve a target sparsity of 70% to 90%, and performing 8-bit integer quantization; for detail branches, adopting one-time pruning to achieve a target sparsity of less than 30%, and performing 16-bit floating-point quantization or maintaining 32-bit full precision.

[0088] The iterative pruning strategy specifically includes:

[0089] Step S51: Perform sparsification training on the converged training model;

[0090] Step S52: Perform pruning, removing the 10%-20% of connections with the smallest absolute weight;

[0091] Step S53: Retrain the pruned model to recover the lost performance;

[0092] Step S54: Repeat the pruning and retraining steps 3-5 times until the target sparsity is achieved.

[0093] Furthermore, the lightweight Monte Carlo Dropout method is used to obtain an uncertainty estimate of the model predictions. U t The method is as follows:

[0094] When performing inference on an embedded terminal, the same input data is executed. K Each forward propagation randomly activates the Dropout layer, resulting in... K Different prediction outputs ;

[0095] By calculating this K The standard deviation σ of the predicted output yields an uncertainty estimate. U t ,Right now:

[0096]

[0097] in, K It is an integer between 2 and 4.

[0098] Furthermore, in step nine, the dynamic early warning threshold is dynamically calculated based on historical data volatility and model uncertainty. The calculation method is as follows:

[0099] T dmax = T b ×(1+α × U t / U max )

[0100] T dmin = T b ×(1- α × U t / U max )

[0101] in: T dmax This is the upper limit of the dynamic early warning threshold; T dmin This is the lower limit of the dynamic early warning threshold; T b This is a predicted value; U max This represents the historical maximum value of the uncertainty estimate during the training phase; α The adjustment coefficient has a range of 0.1 to 0.3.

[0102] Furthermore, in step ten, the tiered early warning is determined according to the following rules:

[0103] (1) Landslide displacement anomaly early warning: When real-time monitoring data > T dmax At that time, calculate the ratio of real-time monitoring data exceeding the upper limit of the dynamic early warning threshold:

[0104] Exceeding the limit ratio = (real-time monitoring data - ...) T dmax ) / U t

[0105] The warning level is determined based on the excess ratio:

[0106] When the excess ratio is greater than 1.0, it is determined to be a Level 3 red alert;

[0107] When the ratio of 1.0 to the above exceeds 0.5, it is determined to be a Level 2 orange alert;

[0108] When 0.5 ≥ the above excess ratio > 0, it is determined to be a Level 1 Yellow Alert;

[0109] (2) Monitoring system anomaly alert: When real-time monitoring data < T dmin At that time, calculate the ratio of the real-time monitoring data to the lower limit of the dynamic early warning threshold:

[0110] Excess ratio = (T dmin -Real-time monitoring data) / U t

[0111] When the current exceedance ratio is greater than 0.5, a system anomaly warning is triggered;

[0112] (3) Other situations are classified as Level 0 normal.

[0113] The experimental data came from online monitoring data of a tailings dam in Sichuan Province. Real-time prediction and early warning were implemented for the tailings dam slope. A total of 2500 time-series monitoring samples, including rainfall, water level, and displacement, were acquired. The data were divided into training, validation, and test sets in a 70%:15%:15% ratio. To optimize memory usage and clearly demonstrate the prediction results, 200 samples were selected from the test set for model prediction and visualization analysis.

[0114] The specific implementation process is as follows:

[0115] like Figure 1 An embedded real-time landslide early warning method based on spatiotemporal feature decoupling includes the following steps:

[0116] Step 1: Obtain historical time-series monitoring data of the landslide body and construct a training dataset based on the sliding window method. Each training sample is used to predict the monitoring data of one or more time points after its corresponding window.

[0117] Preferably, the sliding window length is set to L=120, predicting the monitoring data after the window;

[0118] Step 2: Decouple the historical time series monitoring data by spatiotemporal features, and decompose the original sequence into global components that characterize long-term evolution trends and local detail components that characterize short-term fluctuations and mutations.

[0119] Step 3: Construct a two-branch lightweight temporal convolutional network. The input of this network is the length. L The monitoring data sequence is output as the predicted values ​​for one or more time points as described in step one; the network includes a global branch and a detail branch.

[0120] The total number of depthwise separable convolutional layers in the global branch is greater than the total number of depthwise separable convolutional layers in the detail branch.

[0121] Preferably, the global branch consists of four sequentially connected lightweight temporal convolutional modules, a global average pooling layer, and a fully connected layer. This deeper hierarchical structure allows it to obtain a large receptive field, thereby effectively capturing long-term dependencies in the sequence. The detail branch consists of two sequentially connected lightweight temporal convolutional modules, a global average pooling layer, and a fully connected layer. This shallower hierarchical structure helps maintain sensitivity to short-term dynamics, thus focusing on the rapid extraction of local features. The structure diagram of the dual-branch lightweight temporal convolutional network is shown below. Figure 2 As shown;

[0122] Step 4: Jointly train the dual-branch lightweight temporal convolutional network using the training dataset in the cloud, and use an adaptive weight fusion module to fuse the outputs of the two branches to obtain the final prediction value;

[0123] Step 5: Perform differentiated lightweighting processing on the trained dual-branch lightweight temporal convolutional network:

[0124] The global branch is subjected to high-intensity pruning and quantization; the detailed branch is subjected to low-intensity pruning and quantization or full precision is preserved.

[0125] Preferably, global branches are iteratively pruned with a target sparsity of 85% and converted to INT8 format; detail branches are pruned only once with a sparsity of 20% and FP16 format is preserved.

[0126] Step Six: Deploy the lightweighted model on an embedded terminal;

[0127] Step 7: In the embedded terminal, maintain a length of [length missing] for the real-time acquired monitoring data stream. L The latest data sliding window; the data in this window is preprocessed using the same spatiotemporal feature decoupling method as in step two;

[0128] Step 8: Input the decoupled global component and detail component into the deployed dual-branch network respectively to obtain the predicted values. T b ;

[0129] Step 9: Obtain the uncertainty estimate of the model predictions based on the lightweight Monte Carlo Dropout method. U t And based on the predicted value T b and uncertainty estimation U t Calculate the dynamic early warning threshold;

[0130] Preferably, in the lightweight Monte Carlo Dropout method, the number of forward propagations K=3, in order to achieve the best balance between estimation uncertainty and computational cost;

[0131] Step 10: Compare real-time monitoring data with dynamic early warning thresholds to achieve tiered early warning.

[0132] The slope monitoring data of this embodiment are modeled and predicted using the specific implementation methods described above. Figure 3 The results of displacement prediction and graded early warning using the method of this invention on a test set are shown. The top figure shows the changes of monitored values, predicted values, and dynamic early warning thresholds over time, while the bottom figure shows the corresponding distribution of early warning levels.

[0133] The results analysis shows that:

[0134] 1. By implementing 85% sparsity pruning and 8-bit quantization on the global branches, and light pruning on the detail branches while preserving FP16 accuracy, the final model size is reduced by 15.2 times compared to the original uncompressed model. Simultaneously, the model's ability to fit short-term abrupt changes in displacement sequences is significantly preserved. For example... Figure 2 As shown, at multiple abrupt change points such as time steps 150 and 175, the predicted values ​​can closely follow the rapid changes in the monitored values, which proves that the 'differentiated lightweight' strategy effectively solves the problem of decreased early warning sensitivity caused by uniform compression.

[0135] 2. Based on the uncertainty estimate Ut generated by the lightweight Monte Carlo Dropout method, the system constructs a dynamic early warning threshold. This dynamic early warning threshold range (gray filled area) can reasonably encompass most of the monitoring data, and its upper and lower limits (red and green lines) are dynamically adjusted according to the uncertainty estimate of the model, reflecting the adaptive capability of the early warning system and enabling the early warning system to intelligently cope with the inherent uncertainties in the landslide evolution process.

[0136] 3. The system successfully identified multiple risk periods and triggered tiered warnings. For example, around time steps 150 and 175, the monitored values ​​significantly exceeded the upper limit of the dynamic threshold, and the system accurately issued "orange warnings" and "red alerts"; around time steps 50 and 125, a "yellow alert" was triggered. This demonstrates that the method of this invention can not only capture significant displacement abrupt changes, but also has a keen perception of potential, early-stage deformation anomalies, achieving a leap from "coarse alarms" to "refined warnings."

[0137] To quantitatively evaluate model performance, mean absolute error (MAE), root mean square error (RMSE), and goodness of fit (R²) were used as evaluation metrics. The calculated model performance on the test set was: MAE 0.7105 mm, RMSE 0.8957 mm, and R² 0.8758.

[0138] In summary, the method provided by this invention can not only achieve high-precision prediction of slope displacement, but also provide reliable graded early warning information by combining uncertainty estimation, thus providing effective technical support for slope safety monitoring and risk prevention.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. An embedded real-time landslide early warning method based on spatiotemporal feature decoupling, characterized in that, Includes the following steps: Step 1: Obtain historical time-series monitoring data of the landslide body and construct a training dataset based on the sliding window method. Each training sample is used to predict the monitoring data of one or more time points after its corresponding window. Step 2: Decouple the historical time series monitoring data by spatiotemporal features, and decompose the original sequence into global components that characterize long-term evolution trends and local detail components that characterize short-term fluctuations and mutations. Step 3: Construct a two-branch lightweight temporal convolutional network. The input of this network is the length. L The monitoring data sequence is output as the predicted values ​​for one or more time points as described in step one; the network includes a global branch and a detail branch. The total number of depthwise separable convolutional layers in the global branch is greater than the total number of depthwise separable convolutional layers in the detail branch. Step 4: Jointly train the dual-branch lightweight temporal convolutional network using the training dataset in the cloud, and use an adaptive weight fusion module to fuse the outputs of the two branches to obtain the final prediction value; Step 5: Perform differentiated lightweighting processing on the trained dual-branch lightweight temporal convolutional network: The global branch is subjected to high-intensity pruning and quantization; the detailed branch is subjected to low-intensity pruning and quantization or full precision is preserved. Step Six: Deploy the lightweighted model on an embedded terminal; Step 7: In the embedded terminal, maintain a length of [length missing] for the real-time acquired monitoring data stream. L The latest data sliding window; the data in this window is preprocessed using the same spatiotemporal feature decoupling method as in step two; Step 8: Input the decoupled global component and detail component into the deployed dual-branch network respectively to obtain the predicted values. T b ; Step 9: Obtain the uncertainty estimate of the model predictions based on the lightweight Monte Carlo Dropout method. U t And based on the predicted value T b and uncertainty estimation U t Calculate the dynamic early warning threshold; The lightweight Monte Carlo Dropout method is used to obtain an uncertainty estimate of the model's predicted values. U t The method is as follows: When performing inference on an embedded terminal, the same input data is executed. K Each forward propagation randomly activates the Dropout layer, resulting in... K Different prediction outputs { y 1, y 2, ..., y k }; By calculating this K The standard deviation σ of the predicted output yields an uncertainty estimate. U t ,Right now: U t =σ({ y 1, y 2, ..., y k }) in, K It is an integer between 2 and 4; The dynamic early warning threshold is dynamically calculated based on historical data volatility and model uncertainty. The calculation method is as follows: T dmax = T b ×(1+ α × U t / U max ) T dmin = T b ×(1- α × U t / U max ) in: T dmax This is the upper limit of the dynamic early warning threshold; T dmin This is the lower limit of the dynamic early warning threshold; T b This is a predicted value; U max This represents the historical maximum value of the uncertainty estimate during the training phase; α The adjustment factor has a range of 0.1 to 0.

3. Step 10: Compare real-time monitoring data with dynamic early warning thresholds to achieve tiered early warning.

2. The embedded landslide real-time early warning method based on spatiotemporal feature decoupling according to claim 1, characterized in that, In step two, the spatiotemporal feature decoupling is achieved through adaptive frequency threshold filtering, specifically including: S21: Perform a fast Fourier transform on the original time-series data and analyze its power spectral density; S22: Reconstruct the low-frequency components in the power spectrum that have a cumulative contribution rate of 80%~90% into the global components; S23: Reconstruct the remaining high-frequency components into the detailed components.

3. The embedded landslide real-time early warning method based on spatiotemporal feature decoupling according to claim 1, characterized in that, In step three, each branch of the dual-branch lightweight temporal convolutional network is composed of sequentially connected lightweight temporal convolutional modules; the lightweight temporal convolutional module sequentially includes: a depthwise separable convolutional layer, a linear rectified unit activation function layer, and a random deactivation layer.

4. The embedded landslide real-time early warning method based on spatiotemporal feature decoupling according to claim 1, characterized in that, In step four, the joint training refers to using the global component and detail component obtained by decoupling from the same original sequence as the inputs to the global branch and detail branch, respectively, and simultaneously optimizing the parameters of the two branches end-to-end under the supervision of a unified loss function.

5. The embedded landslide real-time early warning method based on spatiotemporal feature decoupling according to claim 1, characterized in that, In step four, the adaptive weight fusion module generates weights based on data, specifically as follows: S41: Concatenate the feature maps of the final layers of the two branches; S42: Generate a two-dimensional weight vector through a lightweight fully connected layer or a 1x1 convolutional layer. α ,1- α ]; S43: Final predicted value = α ×Global branch output+(1- α )×Detailed branch output; The weight α reflects the relative importance of global trends and local details in the current input data.

6. The embedded landslide real-time early warning method based on spatiotemporal feature decoupling according to claim 1, characterized in that, In step five, the differentiated lightweight processing specifically refers to: for global branches, adopting an iterative pruning strategy to achieve a target sparsity of 70% to 90%, and performing 8-bit integer quantization; for detail branches, adopting one-time pruning to achieve a target sparsity of less than 30%, and performing 16-bit floating-point quantization or maintaining 32-bit full precision. The iterative pruning strategy specifically includes: Step S51: Perform sparsification training on the converged training model; Step S52: Perform pruning, removing the 10%-20% of connections with the smallest absolute weight; Step S53: Retrain the pruned model to recover the lost performance; Step S54: Repeat the pruning and retraining steps 3-5 times until the target sparsity is achieved.

7. The embedded landslide real-time early warning method based on spatiotemporal feature decoupling according to claim 1, characterized in that, In step ten, the tiered early warning is determined according to the following rules: (1) Landslide displacement anomaly early warning: When real-time monitoring data > T dmax At that time, calculate the ratio of real-time monitoring data exceeding the upper limit of the dynamic early warning threshold: Exceeding the limit ratio = (real-time monitoring data - ...) T dmax ) / U t The warning level is determined based on the excess ratio: When the excess ratio is greater than 1.0, it is determined to be a Level 3 red alert; When the ratio of 1.0 to the above exceeds 0.5, it is determined to be a Level 2 orange alert; When 0.5 ≥ the above excess ratio > 0, it is determined to be a Level 1 Yellow Alert; (2) Monitoring system anomaly alert: When real-time monitoring data < T dmin At that time, calculate the ratio of the real-time monitoring data to the lower limit of the dynamic early warning threshold: Excess ratio = ( T dmin -Real-time monitoring data) / U t When the current exceedance ratio is greater than 0.5, a system anomaly warning is triggered; (3) Other situations are classified as Level 0 normal.

Images