A landslide displacement prediction method based on time series decomposition and shape constraint

Abstract

The application discloses a landslide displacement prediction method based on time sequence decomposition and shape constraint, and relates to the cross technical field of early monitoring and early warning of geological disasters and artificial intelligence time sequence data processing. The method comprises the following steps: firstly, decoupling a landslide displacement sequence into a trend component, a seasonal component and a residual component; secondly, inputting the trend component and the seasonal component into corresponding TimesNet models for prediction to obtain a trend prediction value and a seasonal prediction value, and inputting the residual component into an iTransformer model for prediction to obtain a residual prediction value; in the model training stage, a MixLoss mixed loss function is introduced, which fuses mean square error and cosine similarity, and the geometric shape and evolution direction of the displacement sequence are forced to be learned by the network model through trend similarity constraint; finally, the trend prediction value, the seasonal prediction value and the residual prediction value are added to obtain a landslide displacement prediction value sequence of continuous time steps from the current time. The method improves the accuracy of landslide displacement prediction.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of early monitoring and warning of geological disasters and artificial intelligence time-series data processing, and in particular to a landslide displacement prediction method based on time-series decomposition and morphological constraints. Background Technology

[0002] Landslides, as a sudden geological disaster with wide spatial distribution and extremely strong destructive potential, often pose a devastating threat to mountainous transportation infrastructure, water conservancy projects, and the safety of residents' lives and property. Therefore, conducting high-precision landslide displacement trend prediction research before a disaster occurs, and identifying the instability acceleration characteristics of the slope in advance, has immeasurable engineering practical significance and socio-economic value for achieving early warning of disasters, early evacuation of people, and greatly reducing disaster risks.

[0003] With the popularization of modern space geodesy technologies (such as BeiDou / GNSS networks) and the development of artificial intelligence deep learning algorithms, data-driven time series prediction methods, represented by Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and the early Transformer architecture, have been widely introduced and applied to the field of landslide displacement monitoring and early warning due to their powerful implicit mapping capabilities of nonlinear functions.

[0004] Existing deep prediction models generally use basic loss functions such as mean squared error as the evaluation criterion for parameter updates during the training phase. These basic loss functions only focus on the point-to-point absolute numerical approximation between the predicted value and the true label. When faced with abrupt changes in landslide displacement or accelerated deformation stages, the model often exhibits statistical conservatism in order to quickly minimize the absolute error at the current time step, highly relying on and tending to output the observation value from the previous moment. This deficiency in the underlying optimization mechanism causes the model's prediction curve to lag behind the actual evolution trend of the landslide in time phase on the macroscopic time axis, resulting in prediction lag, which further leads to low accuracy in landslide displacement prediction. Summary of the Invention

[0005] Therefore, it is necessary to provide a landslide displacement prediction method based on temporal decomposition and morphological constraints to address the aforementioned technical problems. This method improves the accuracy of landslide displacement prediction.

[0006] The following technical solution is adopted in this specification: This specification provides a landslide displacement prediction method based on temporal decomposition and morphological constraints, including: Obtain historical cumulative displacement time series data of the target landslide area using Global Navigation Satellite System (GNSS), and preprocess the GNSS historical cumulative displacement time series data to obtain a standardized landslide displacement sequence; By using a time-series decomposition algorithm based on local weighted polynomial regression, the standardized landslide displacement sequence is decoupled into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area. Obtain the landslide displacement prediction model; the landslide displacement prediction model includes two parallel TimesNet models and one iTransformer model; the landslide displacement prediction model applies morphological constraints on the direction of displacement sequence evolution during backpropagation to correct the prediction phase deviation during the accelerated deformation stage; the loss function of the landslide displacement prediction model is a MixLoss hybrid loss function that integrates mean square error and cosine similarity constraints. The trend component and seasonal component are respectively input into the corresponding TimesNet model for prediction, and the predicted values ​​of the trend component and seasonal component are obtained for multiple consecutive time steps starting from the current time. The residual component is input into the iTransformer model for prediction, and the predicted values ​​of the residual component are obtained for multiple consecutive time steps starting from the current time. The predicted values ​​of the trend term, seasonal term, and residual term are added together to obtain a sequence of predicted landslide displacement values ​​for multiple consecutive time steps starting from the current moment.

[0007] Optionally, the loss function of the landslide displacement prediction model is: in, The loss value is for the landslide displacement prediction model. This is the mean square error term used to control the overall accuracy of statistical forecasting. The trend similarity error term is used to correct the predicted phase deviation. The output of the landslide displacement prediction model includes The predicted sequence vector at each time step, for The actual landslide displacement sequence vector corresponding to each time step. Predicting time indices in sequence vectors The value at that location, Time index in the actual landslide displacement sequence vector The value at that location, This is a hyperparameter constant used to adjust the weight balance between trend similarity error and mean square error.

[0008] Optionally, using a time-series decomposition algorithm based on locally weighted multinomial regression, the standardized landslide displacement sequence is decoupled into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area. Specifically, these components include: Initialize the local weighted regression parameters and the robust weights of all data points in the standardized landslide displacement sequence; the local weighted regression parameters include window size, weight function type, number of inner loop iterations, and number of outer loop iterations; In the inner loop, robust weights are combined with a low-pass filter window to perform smoothing and denoising, extract the initial trend term of the standardized landslide displacement sequence, and perform subsequence smoothing on the standardized landslide displacement sequence after removing the initial trend term to extract the initial seasonal term. In the outer loop, the difference between the current standardized landslide displacement sequence and the initial trend term and the initial seasonal term is used as the residual. The updated robust weights of each data point are dynamically updated based on the residuals, and the updated robust weights are fed back into the next inner loop. The larger the residual, the smaller the corresponding robust weight. The inner and outer loops are executed alternately until the number of iterations of the inner loop and the number of iterations of the outer loop are reached. The initial trend term and the initial seasonal term corresponding to the last iteration are then determined as the trend term component and the seasonal term component. Subtracting the trend and seasonal components from the standardized landslide displacement sequence yields the residual components.

[0009] Optionally, the TimesNet model includes k A parallel multi-scale two-dimensional convolutional module; the trend component and seasonal component are respectively input into the corresponding TimesNet model for prediction, to obtain the predicted values ​​of the trend component and seasonal component for multiple consecutive time steps from the current time, specifically including: The trend component and the seasonal component are treated as independent one-dimensional time series. For any one-dimensional time series, after performing a Fast Fourier Transform on the one-dimensional time series, the frequency domain features of the one-dimensional time series are extracted. k Each dominant frequency was obtained separately. k The dominant frequency corresponds to k One cycle step size; based on k Each period step is used to reconstruct the one-dimensional time series, resulting in... k A two-dimensional tensor matrix; k Each of the two-dimensional tensor matrices corresponds to k A different cycle step size dimension; Will k Each of the two-dimensional tensor matrices is input to... kIn a parallel multi-scale two-dimensional convolution module, the two-dimensional spatial evolution features of each two-dimensional tensor matrix within and between periods are extracted by jointly using the multi-scale two-dimensional convolution module. Each two-dimensional spatial evolution feature is sequentially subjected to inverse dimensionality reduction mapping to restore its corresponding one-dimensional feature sequence. Then, Softmax weights are used to weight and fuse these one-dimensional feature sequences to obtain the predicted trend or seasonal values ​​for multiple consecutive time steps starting from the current moment in the one-dimensional time series. The Softmax weights are based on... k The amplitude distribution of the dominant frequencies was calculated.

[0010] Optionally, the iTransformer model includes an embedding layer, a Transformer encoding module, and a projection layer connected in series; the Transformer encoding module includes a multi-head attention mechanism and a feedforward neural network; the residual components are input into the iTransformer model based on the inverted attention mechanism for prediction, obtaining the predicted values ​​of the residuals for multiple consecutive time steps starting from the current time, specifically including: The entire historical time series of the residual component is treated as an independent variable feature token, its dimensions are inverted, and it is input into the embedding layer for mapping to obtain the initial embedding vector representation of the residual component, so as to break away from the traditional time step sequence dependency. The initial embedding vector representation is input into the Transformer encoding module, and high-frequency nonlinear fluctuation features across time segments are interactively extracted on the global feature dimension of the initial embedding vector representation using a multi-head self-attention mechanism and a feedforward neural network. The high-frequency nonlinear fluctuation characteristics are input into the projection layer for feature dimensionality reduction and nonlinear mapping to obtain the predicted values ​​of the residual terms for multiple consecutive time steps starting from the current moment.

[0011] Optionally, the method further includes: Calculate the difference between the predicted landslide displacement values ​​at adjacent time steps in the landslide displacement prediction value sequence, and divide it by the preset time step length to obtain the predicted landslide displacement change rate for multiple consecutive time steps starting from the current moment. Obtain a pre-set absolute displacement threshold, a first velocity threshold, and a second velocity threshold; the second velocity threshold is greater than the first velocity threshold. Traverse the landslide displacement prediction value sequence and the corresponding landslide displacement prediction change rate. When the landslide displacement prediction change rate is less than the first rate threshold, determine that the target landslide area is in the iso-velocity deformation stage and generate the first-level early warning signal. When the predicted rate of change of landslide displacement is greater than or equal to the first rate threshold and less than the second rate threshold, the target landslide area is determined to have entered the primary accelerated deformation stage, and a second-level early warning signal is generated. When the predicted rate of change of landslide displacement is greater than or equal to the second rate threshold, and the corresponding initial predicted landslide displacement value is less than the absolute displacement threshold, the target landslide area is determined to have entered the medium-accelerated deformation stage, a second-level early warning signal is generated, and a high-frequency monitoring command is triggered simultaneously. When the predicted rate of change of landslide displacement is greater than or equal to the second rate threshold and the corresponding predicted value of initial landslide displacement is greater than or equal to the absolute displacement threshold, the target landslide area is determined to have entered the imminent landslide transition stage, a third-level early warning signal is generated, and an emergency evacuation command is triggered.

[0012] Optionally, the historical cumulative displacement time series data of GNSS is preprocessed to obtain a standardized landslide displacement sequence, specifically including: Gross error removal is performed on the historical cumulative displacement time series data of GNSS to obtain gross error removed data; The outlier data was downsampled according to a preset time scale to obtain a standardized landslide displacement sequence.

[0013] This specification provides a landslide displacement prediction device based on temporal decomposition and morphological constraints, including: The preprocessing module is used to acquire historical cumulative displacement time series data of the Global Navigation Satellite System (GNSS) for the target landslide area, and to preprocess the historical cumulative displacement time series data of GNSS to obtain a standardized landslide displacement sequence. The decoupling module is used to decouple the standardized landslide displacement sequence into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area through a time-series decomposition algorithm based on local weighted multinomial regression. The acquisition module is used to acquire the landslide displacement prediction model. The landslide displacement prediction model includes two parallel TimesNet models and one iTransformer model. The landslide displacement prediction model applies morphological constraints on the direction of displacement sequence evolution during backpropagation to correct the prediction phase deviation during the accelerated deformation stage. The loss function of the landslide displacement prediction model is a MixLoss hybrid loss function that integrates mean square error and cosine similarity constraints. The prediction module is used to input the trend component and the seasonal component into the corresponding TimesNet model for prediction, and obtain the predicted values ​​of the trend component and the seasonal component for multiple consecutive time steps starting from the current time. The residual component is input into the iTransformer model for prediction, and obtain the predicted values ​​of the residual component for multiple consecutive time steps starting from the current time. The determination module is used to add the predicted values ​​of the trend term, the seasonal term, and the residual term to obtain a sequence of predicted landslide displacement values ​​for multiple consecutive time steps starting from the current moment.

[0014] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described landslide displacement prediction method based on temporal decomposition and morphological constraints.

[0015] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described landslide displacement prediction method based on temporal decomposition and morphological constraints.

[0016] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: In the landslide displacement prediction method based on time-series decomposition and morphological constraints provided in this specification, the GNSS historical cumulative displacement time series data is preprocessed by removing gross errors and downsampled according to a preset time scale to obtain a standardized landslide displacement sequence. Using a time-series decomposition algorithm based on locally weighted multinomial regression, the standardized landslide displacement sequence is decoupled into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area. A landslide displacement prediction model is then obtained; the landslide displacement prediction model includes two parallel TimesNet models and one iTran model. The landslide displacement prediction model uses the iTransformer model; the loss function is a MixLoss hybrid loss function that integrates numerical error constraints and geometric constraints. The trend component and seasonal component are input into the corresponding TimesNet model for prediction, obtaining the predicted values ​​of the trend component and seasonal component for multiple consecutive time steps starting from the current time. The residual component is input into the iTransformer model based on the inverted attention mechanism for prediction, obtaining the predicted values ​​of the residual component for multiple consecutive time steps starting from the current time. The predicted values ​​of the trend component, seasonal component, and residual component are added together to obtain the landslide displacement prediction sequence for multiple consecutive time steps starting from the current time.

[0017] This method utilizes a time-series decomposition algorithm based on locally weighted multinomial regression to decouple the cumulative displacement sequence into trend, seasonal, and residual terms, accurately decoupling the cumulative landslide displacement into these terms. This enables differentiated modeling of different displacement components and improves the targeting of feature extraction. Secondly, it employs the TimesNet model, using two-dimensional time variation modeling to capture the regularity of trend and seasonal terms. It also utilizes the inverted attention mechanism of the iTransformer model to extract high-frequency nonlinear fluctuations in the residual term. Combining TimesNet and iTransformer models to capture displacement trends, seasonal patterns, and high-frequency nonlinear fluctuations in the residual term allows for the full exploitation of multi-scale variation characteristics of landslide displacement time series, enhancing the model's ability to express and predict complex time-series changes. Finally, by introducing cosine similarity constraints to force the model to learn the sequence's geometric shape, it overcomes the prediction lag problem caused by traditional loss functions that only pursue numerical approximation and lack morphological constraints during model training, thereby improving the accuracy of landslide displacement prediction. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This document presents a schematic diagram of a landslide displacement prediction method based on temporal decomposition and morphological constraints. Figure 2 This invention provides a flowchart for combined landslide displacement prediction that integrates STL decoupling, heterogeneous model divide-and-conquer, and MixLoss morphological constraints. Figure 3 A comparison chart of the original landslide displacement sequence prediction results based on the traditional mean square error loss function provided by this invention; Figure 4 A comparison chart of the prediction results of the original landslide displacement sequence based on the hybrid loss function of this invention; Figure 5 Comparison of landslide displacement residual term sequence prediction results based on traditional mean square error loss function provided by the present invention; Figure 6 A comparison chart of landslide displacement residual term sequence prediction results based on the hybrid loss function of this invention is provided for this invention. Figure 7 A schematic diagram of a landslide displacement prediction device based on temporal decomposition and morphological constraints provided in this specification; Figure 8 This is a schematic diagram of a computer device used to implement a landslide displacement prediction method based on temporal decomposition and morphological constraints, as provided in this specification. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0021] Devices such as desktop computers, servers, and laptops are capable of executing the solutions described in this manual. For ease of explanation, the following description will focus on servers as the primary execution method.

[0022] From the perspective of geotechnical mechanics and kinematic evolution, the temporal evolution of landslide displacement is not a linear movement in a single dimension. Instead, it is influenced by a nonlinear deep coupling of multiple inducing factors, including the internal geological structure of the slope, fluctuations in the external hydrological environment (such as seasonal heavy rainfall cycles and periodic reservoir water level adjustments), and frequent human engineering activities. This results in extremely strong dynamic non-stationarity and frequency complexity. Although traditional physical and mechanical static models based on limit equilibrium theory or numerical simulation have clear physical mechanisms, they are often limited by the extreme difficulty in obtaining physical parameters of deep soil and rock masses and the systematic calculation errors caused by the generalization of multi-field coupled boundary conditions. Consequently, they are difficult to accurately track and predict minute deformations in real monitoring scenarios with complex terrain environments.

[0023] Although deep learning algorithms in artificial intelligence have improved the automation of prediction to some extent, existing architectures still have the following problems when faced with landslide monitoring sequence data with high-frequency sampling and long-term spans: I. Landslide displacement time series in the Global Navigation Satellite System (GNSS) are typically composed of a low-frequency long-term trend component reflecting slow geological evolution, a mid-frequency periodic fluctuation component influenced by rainfall or water storage, and a high-frequency random noise component caused by microseismic events or the propagation of shallow surface cracks. Current mainstream research often employs an end-to-end single deep learning network architecture to attempt to extract features from all frequency bands. This approach not only easily leads to gradient vanishing but also limits the model's ability to characterize multi-frequency coupling features. Even when some studies introduce Ensemble Empirical Mode Decomposition (EEMD) or Variational Mode Decomposition (VMD) for signal preprocessing, these decomposition methods are prone to mode aliasing under extreme environmental noise. More importantly, because the decomposed components have drastically different physical evolution mechanisms, continuing to use the same model for fitting will result in a mismatch between the model architecture and the physical signal characteristics, thus preventing the effective discovery of deeper disaster-causing patterns.

[0024] II. Existing deep prediction models almost universally employ Mean Squared Error (MSE) or Mean Absolute Error (MAE) as the sole criterion for parameter updates in their training and optimization loss function design. The initial intention of MSE is to constrain the overall point-to-point numerical closeness between the model output and the true label, which performs well in stationary time series prediction. However, landslide evolution undergoes a sudden transition from a constant-rate deformation stage to an accelerated deformation stage in the middle and later stages of its life cycle, resulting in a near-vertical, nonlinear, steep step-like displacement curve. Under this extreme condition, since MSE essentially penalizes Euclidean distance anomalies, according to the principle of statistical expectation, when the model encounters rapidly changing trend data, in order to minimize the total variance penalty in the safest and most conservative way at the current time step, the model tends to output the statistical mean of data within the past historical observation window or simply shift the output to the observation value of the previous moment. This internal optimization mechanism, based on statistical conservatism, directly results in the model's predicted curve always lagging behind the actual accelerated deformation curve of the landslide on the macroscopic time axis (i.e., forming a severe phase bias or prediction lag). In actual disaster emergency early warning projects, the lag in early warning response will hinder the implementation of emergency evacuation work.

[0025] To address the limitations of existing technologies, such as the inability of a single model to accurately characterize multi-frequency coupling features and the lag in landslide step prediction caused by the traditional MSE loss function, this invention proposes a combined landslide displacement prediction method that integrates temporal decomposition and morphological constraints. The core idea of ​​this invention is to achieve high-precision, low-phase-delay landslide displacement prediction by locally weighted temporal decoupling, constructing a heterogeneous model combination architecture oriented towards the physical characteristics of different frequency bands, and introducing a geometric morphological loss function.

[0026] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0027] Figure 1 This is a schematic diagram of a landslide displacement prediction method based on temporal decomposition and morphological constraints, as described in this specification. The method specifically includes the following steps: S101: Obtain historical cumulative displacement time series data of the target landslide area using the Global Navigation Satellite System (GNSS), preprocess the GNSS historical cumulative displacement time series data to obtain a standardized landslide displacement sequence.

[0028] Historical cumulative displacement time series data were collected using GNSS positioning devices deployed on and around the landslide surface. After acquiring the data, a Deep Learning Approach for Unsupervised Anomaly Detection in Time Series (DeepAnt) was used to remove spike pseudo-random gross errors caused by satellite communication obstruction or multipath scattering effects. A moving average strategy was then applied to uniformly downsample the data to a 12-hour spatial scale to generate a standardized landslide displacement sequence.

[0029] In an exemplary embodiment, the GNSS historical cumulative displacement time series data is preprocessed to obtain a standardized landslide displacement sequence. Specifically, this includes: performing gross error removal processing on the GNSS historical cumulative displacement time series data to obtain gross error removed data; and downsampling the gross error removed data according to a preset time scale to obtain a standardized landslide displacement sequence.

[0030] S102: By using a time-series decomposition algorithm based on local weighted polynomial regression, the standardized landslide displacement sequence is decoupled into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area.

[0031] In an exemplary embodiment, a time-series decomposition algorithm based on locally weighted multinomial regression is used to decouple the standardized landslide displacement sequence into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area. Specifically, this includes: initializing locally weighted regression parameters and initializing robust weights for all data points in the standardized landslide displacement sequence; the locally weighted regression parameters include window size, weight function type, number of inner loop iterations, and number of outer loop iterations; in the inner loop, combined with the robust weights, a low-pass filter window is run to perform smoothing and denoising, extracting the initial trend of the standardized landslide displacement sequence. The initial trend term is extracted, and the standardized landslide displacement sequence after removing the initial trend term is smoothed by subsequence smoothing to extract the initial seasonal term. In the outer loop, the difference between the current standardized landslide displacement sequence and the initial trend term and initial seasonal term is used as the residual. The updated robust weights of each data point are dynamically updated based on the residuals, and the updated robust weights are fed back to the next inner loop. The larger the residual, the smaller the corresponding robust weight. The inner and outer loops are executed alternately until the number of iterations of the inner loop and the number of iterations of the outer loop are reached. The initial trend term and initial seasonal term are corresponding to the last iteration and are determined as the trend term component and seasonal term component. The trend term component and seasonal term component are subtracted from the standardized landslide displacement sequence to obtain the residual term component.

[0032] Specifically, a time-series decomposition algorithm based on local weighted multinomial regression is invoked to strip the standardized landslide displacement sequence.

[0033] The formula corresponding to the time series decomposition algorithm based on local weighted multinomial regression is formula (1): (1); in, To standardize the landslide displacement sequence at time 1 displacement value, Indicates the trend component. Indicates the seasonal component. For the components of the residual term, , n This represents the number of time steps.

[0034] The time-series decomposition algorithm constructs an inner and outer dual recursive network during execution. The inner loop first runs a low-pass filter window to perform smoothing and denoising, thereby refining the initial trend term. Then, the trend term is subtracted from the original signal to extract the initial seasonal term containing the periodic factor. The outer loop activates a robust threshold weight function on this basis, weakening the interference of isolated data such as extreme rainstorms on the process of stripping away the dominant stationary components.

[0035] The component signals of the decoupled output all obtained clear physical characteristics within the scope of geotechnical mechanics: Trend component ( ): This reflects the deep shear creep of the slope under self-weight stress, the deterioration of the geological background, and long-term tectonic slippage.

[0036] Seasonal component ( ): This characterizes the periodic accumulation and dissipation of pore water pressure caused by the seasonal concentrated infiltration of rainwater, as well as the regular fluctuations in groundwater level.

[0037] Residual component ( This primarily characterizes the high-frequency, unpredictable disturbance noise caused by sudden microseismic waves, disturbances from surrounding artificial engineering excavation and blasting, and instability and collapse of internal soil fissures. This physical separation step lays the foundation for heterogeneous divide-and-conquer prediction of downstream systems.

[0038] This invention introduces a time series decomposition algorithm (Seasonal and Trend decomposition using Loess, STL) with local weighted polynomial regression characteristics, which decouples the landslide cumulative displacement time series signal into three independent subsequences representing different disaster-causing mechanisms: namely, the trend component that maps the continuous evolution of the gravity geological background, the seasonal component that maps the disturbance of meteorological and hydrological cycles, and the residual component that characterizes sudden uncontrollable factors, thereby transforming the non-stationary sequence into modeling elements with a clear physical structure.

[0039] S103: Obtain the landslide displacement prediction model; the landslide displacement prediction model includes two parallel TimesNet models and one iTransformer model; the landslide displacement prediction model applies morphological constraints on the direction of displacement sequence evolution during backpropagation to correct the prediction phase deviation during the accelerated deformation stage; the loss function of the landslide displacement prediction model is a MixLoss hybrid loss function that integrates mean square error and cosine similarity constraints.

[0040] In engineering applications of landslide prediction algorithms, training and testing on data with significant upward trends often reveals a prediction phase lag phenomenon (when the actual displacement jumps significantly at a certain moment, the prediction model often only begins to show an upward trend several time steps later). Current solutions involve increasing the number of layers or attention heads in the neural network, but this can sometimes lead to model overfitting. This invention, based on the principles of mathematical statistics, reveals the mathematical logic behind this phenomenon induced by the MSE loss function.

[0041] In this invention, the final output vector of the neural network is denoted as... The absolute true value of landslide monitoring is recorded as The traditional mean square error (MSE) can be expressed as formula (2): (2); in, Mean square error, Time index in the actual landslide displacement sequence vector The value at that location, Predicting time indices in sequence vectors The value at that location, n The amount of input data.

[0042] In the context of machine learning sampling training, as the amount of input data increases... As the value increases, the sample mean approaches the expected value. At this point, by taking the expected values ​​on both sides of the equation, we can derive formula (3):

[0043] (3); in, For the expectation, For variance, For covariance, f To predict the sequence vector, For the first The actual landslide displacement sequence vector corresponding to each time step. E The expectation operator is Cov, which stands for covariance calculation.

[0044] From formula (3), we can conclude that the MSE loss function is a statistical constraint on the overall global numerical difference of the training batch data. In the accelerated sliding phase with a steep upward trend, there is a large variance within the data sequence. Fluctuations occur. At this point, the MSE optimizer, in order to minimize the current global variance penalty, tends to generate smooth, conservative predictions, driving the model output to directly approximate the statistical mean of the data within the historical observation window. This process ignores the geometric slope trend of local instants in the sequence, ultimately resulting in delayed early warning response in applications.

[0045] To address this mathematical limitation, this invention designs a MixLoss hybrid loss function system. This system, while retaining the MSE control for global distance approximation, superimposes a trend direction similarity error term (SimilarLoss), constructing a dual-constraint framework as shown in formula (4):

[0046] (4); in, The loss value is for the landslide displacement prediction model. This is the MSE loss value. This represents the trend similarity error term.

[0047] This invention uses cosine similarity to measure the consistency of evolution direction between two time series vectors. The advantage of cosine similarity is that it focuses on the angle between vectors in multidimensional space, without being directly affected by the absolute magnitude of the vectors. For time series prediction vectors of the same dimensional length... and real label vector The cosine similarity score is calculated using formula (5):

[0048] (5); in, The cosine similarity score is calculated.

[0049] This indicator ranges from -1 to 1. The closer the value is to 1, the more perfectly synchronized the predicted curve is with the actual landslide curve in terms of geometric rise or fall.

[0050] This invention performs a reverse mapping of the similarity score, transforming it into an error penalty term for gradient descent optimization as shown in formula (6): (6); in, The cosine similarity score is calculated.

[0051] The loss function of the landslide displacement prediction model is given by formula (7): (7); in, The loss value is for the landslide displacement prediction model. The output of the landslide displacement prediction model includes The predicted sequence vector at each time step, for The actual landslide displacement sequence vector corresponding to each time step. Predicting time indices in sequence vectors The value at that location, Time index in the actual landslide displacement sequence vector The value at that location, This is the mean square error term used to control the overall accuracy of statistical forecasting. The trend similarity error term is used to correct the predicted phase deviation. This is a hyperparameter constant used to adjust the weight balance between trend similarity error and mean square error.

[0052] In formula (7), the hyperparameter constants It plays a dynamic adjustment role, mapping the dimensionless decimal of cosine similarity to a numerical range that matches the magnitude of the mean square error, and adjusting the weights between the absolute spatial distance fitting and the alignment with future evolutionary trends during iterative training. This loss function design guides the neural network to not only pursue numerical approximation during non-stationary abrupt changes, but also to maintain a high degree of consistency in the evolutionary direction.

[0053] This invention designs and introduces a novel hybrid loss function, MixLoss, to improve prediction lag. The MixLoss function, based on the traditional mean squared error architecture, incorporates a trend-shape constraint evolution term based on cosine similarity. The physical essence of this mechanism lies in the fact that during the forward propagation and backward gradient update of the error in the neural network, the model is required not only to approximate the true historical displacement in absolute numerical terms, but also to impose a directional consistency penalty on the evolution direction (i.e., the trend of the tangent angle) between the predicted curve and the true displacement curve in the geometric vector space. This optimization strategy, which integrates numerical point constraints and overall curve geometric shape constraints, prompts the model to focus on the future tilt direction of the sequence, thereby effectively correcting the phase deviation prediction lag problem caused by the compromise of numerical fitting in traditional models.

[0054] S104: Input the trend component and seasonal component into the corresponding TimesNet model for prediction to obtain the predicted values ​​of the trend component and seasonal component for multiple consecutive time steps starting from the current time. Input the residual component into the iTransformer model for prediction to obtain the predicted values ​​of the residual component for multiple consecutive time steps starting from the current time.

[0055] In one exemplary embodiment, the TimesNet model includes k The system employs a parallel multi-scale two-dimensional convolutional module. The trend and seasonal components are input into their respective TimesNet models for prediction, yielding predicted trend and seasonal values ​​for multiple consecutive time steps from the current moment. Specifically, this involves treating the trend and seasonal components as independent one-dimensional time series; for any one-dimensional time series, performing a Fast Fourier Transform, and then extracting the frequency domain features of the one-dimensional time series. k Each dominant frequency was obtained separately. k The dominant frequency corresponds to k Each cycle step size; based on k Each period step is used to reconstruct the one-dimensional time series, resulting in... k A two-dimensional tensor matrix; k Each of the two-dimensional tensor matrices corresponds to k Each different cycle step size dimension; k Each of the two-dimensional tensor matrices is input to... kIn a series of parallel multi-scale two-dimensional convolutional modules, the two-dimensional spatial evolution features of each two-dimensional tensor matrix within and between periods are extracted jointly. These two-dimensional spatial evolution features are then sequentially subjected to inverse dimensionality reduction mapping to restore their corresponding one-dimensional feature sequences. Softmax weights are then used to weight and fuse these one-dimensional feature sequences to obtain the predicted trend or seasonality values ​​for multiple consecutive time steps from the current moment in the one-dimensional time series. The Softmax weights are based on... k The amplitude distribution of the dominant frequencies was calculated.

[0056] Specifically, a TimesNet multidimensional tensor transformation and modeling engine for low- to mid-frequency regular evolution (trend and seasonal terms): The long-term creep evolution of landslide slopes and their cyclical response dominated by the rainy season contain periodic rules. This invention selects the TimesNet model to solve the backbone extraction task. The front end of the TimesNet model first uses Fast Fourier Transform (FFT) to scan the frequency domain features of the sequence and anchor the underlying signals. The highest peak dominant frequency component and the corresponding real physical cycle step size : .

[0057] Then, based on the step size of each cycle... Matrix splitting command, splitting a one-dimensional input sequence Reconstructed A two-dimensional tensor matrix The reconstruction process is shown in formula (8): (8); in, , It is a two-dimensional tensor matrix. For reconstruction, Given a one-dimensional input sequence, The number of dominant frequency components.

[0058] After the dimensional transformation, the vertical extension of the two-dimensional tensor matrix represents the cross-cycle changes (macroscopically corresponding to the cumulative geological evolution across quarters), while the horizontal slice represents the microscopic changes within a single cycle (corresponding to the stress surge response within a single week of concentrated rainfall during the flood season).

[0059] Subsequently, these two-dimensional tensor matrices are fed into the Inception multi-scale composite two-dimensional convolutional kernel cluster to perform cross-dimensional spatial local feature receptive field extraction, the extraction process of which is Equation (9): (9); in, The output is a two-dimensional feature tensor after convolution. This represents a multi-scale two-dimensional convolutional feature extraction module.

[0060] Finally, through inverse dimensionality reduction mapping and Softmax weighted fusion, a smooth one-dimensional future trend time series prediction sequence and seasonal time series prediction sequence are restored and output.

[0061] The inverse dimensionality reduction mapping is given by formula (10): (10); in, This represents the inverse reconstruction operation that flattens and restores a two-dimensional feature tensor. The corresponding number obtained after inverse reconstruction A one-dimensional feature sequence.

[0062] Softmax weighted fusion is expressed as formula (11): (11); in, This represents an activation function that weights the importance of different periodic features. First, take the natural exponent of the amplitude at each frequency; then sum the exponents as the denominator; divide the individual exponents by the sum to obtain the Softmax weights (ensuring that all weights add up to 1); the final weighted fusion involves multiplying the calculated weight scalars by the corresponding feature sequences, and then... k The sequences are summed up.

[0063] In an exemplary embodiment, the iTransformer model includes an embedding layer, a Transformer encoding module, and a projection layer connected in series. The Transformer encoding module includes a multi-head attention mechanism and a feedforward neural network. The residual term components are input into the iTransformer model based on the inverted attention mechanism for prediction, obtaining the predicted residual term values ​​for multiple consecutive time steps starting from the current moment. Specifically, this includes: inverting the dimension of the entire historical time series of the residual term components as an independent variable feature token and inputting it into the embedding layer for mapping to obtain the initial embedding vector representation of the residual term components, thus escaping the traditional time step sequence dependency; inputting the initial embedding vector representation into the Transformer encoding module, and using the multi-head self-attention mechanism and the feedforward neural network, interactively extracting high-frequency nonlinear fluctuation features across time segments on the global feature dimension of the initial embedding vector representation; and inputting the high-frequency nonlinear fluctuation features into the projection layer for feature dimensionality reduction and nonlinear mapping to obtain the predicted residual term values ​​for multiple consecutive time steps starting from the current moment.

[0064] Specifically, the high frequency, nonlinearity, and random noise characteristics of the remaining residual components increase the difficulty of prediction. Traditional Transformer models calculate dependencies along time steps, and when encountering sequences without obvious periodicity, the attention mechanism tends to smooth out abrupt changes and anomalies by averaging weight allocation.

[0065] This invention introduces the iTransformer architecture, which implements variable dimension inversion embedding. Instead of decomposing time, iTransformer embeds the entire historical residual evolution sequence of each independent monitoring variable as a whole and inputs it into the network mapping layer at once.

[0066] iTransformer is a novel time series modeling framework that represents a structural rethinking and reconstruction of the traditional Transformer model for time series forecasting tasks. Traditional Transformer models embed all variables at the same time step as a single token, with their attention mechanism focusing on the correlations between time steps. However, in multivariate time series, different variables often exhibit heterogeneity and temporal inconsistency, making this modeling approach prone to obscuring the essential connections between variables. iTransformer takes the opposite approach, embedding the entire time series of each variable as an independent token into the network, thus focusing on modeling the multidimensional correlations between variables. This inverted structure effectively preserves the independent dynamic processes of each variable, enabling the model to more accurately capture the collaborative relationships and potential dependencies between variables through a multi-head attention mechanism.

[0067] The iTransformer model adopts a simplified encoder-only architecture, mainly consisting of an embedding layer, a multi-head attention mechanism, a feedforward network, and a projection layer. Its prediction process can be formally represented by equations (12)-(14):

[0068] (12); (13); (14); in, Indicates the first Historical time series of variables, For the corresponding initial embedding representation, Indicates the first The set of hidden states of all variable tokens in the Transformer layer. This is a Transformer encoding module that includes a multi-head self-attention mechanism and a feedforward neural network; This refers to the multilayer perceptron (MLP) used to generate the prediction results.

[0069] All variables in the model are input in parallel as variable tokens. The self-attention mechanism is used to model the interdependence structure between variables, rather than the temporal dependency in the traditional model.

[0070] In landslide displacement sequences, residual terms typically represent unstructured disturbances and unpredictable signals, such as the impact of short-term or non-periodic factors like sudden rainstorms, earthquakes, and construction disturbances. This type of information often exhibits highly nonlinear, non-stationary, and difficult-to-model characteristics. iTransformer, with its inverted modeling mechanism at its core, expands the residual sequence along the feature dimension, emphasizing its internal representational ability across time segments. It can uncover deep patterns and potential dependencies in the residual sequence even without explicit periodic or trend structures. This model demonstrates good generalization and disturbance resistance in handling landslide residual terms, and is particularly suitable for capturing small-amplitude, random non-stationary fluctuations in GNSS displacement monitoring, improving the overall prediction model's sensitivity and robustness to anomalous changes.

[0071] This global inversion review mechanism better preserves the independent dynamic process of the evolution of small burst signals and broadens the model's perspective on capturing the characteristics of sudden step jitter.

[0072] This invention constructs a heterogeneous divide-and-conquer joint prediction architecture to adapt the model to the frequency band characteristics of the signal. For the decoupled trend and seasonal components, which exhibit significant long-term inertia and periodic oscillations, this invention deploys the TimesNet model. This model leverages its two-dimensional temporal tensor transformation and multi-scale convolutional modeling capabilities to overcome the limitations of a one-dimensional sequence perspective, simultaneously capturing multi-scale overlapping temporal periodic features on a two-dimensional plane. For the high-frequency nonlinearity and irregular jump characteristics inherent in the residual components, this invention employs the iTransformer model. This model uses an inverted attention structure, expanding each time variable slice along the feature mapping dimension and enhancing interactive extraction, thereby improving the model's ability to perceive and capture short-wave random disturbances and transient anomalous fluctuations.

[0073] S105: Add the predicted values ​​of the trend term, seasonal term, and residual term to obtain a sequence of predicted landslide displacement values ​​for multiple consecutive time steps starting from the current moment.

[0074] In an exemplary embodiment, the method further includes: calculating the difference between landslide displacement prediction values ​​at adjacent time steps in the landslide displacement prediction value sequence, and dividing it by a preset time step length to obtain the landslide displacement prediction change rate corresponding to multiple consecutive time steps from the current moment; obtaining a preset absolute displacement threshold, a first rate threshold, and a second rate threshold; the second rate threshold being greater than the first rate threshold; traversing the landslide displacement prediction value sequence and the corresponding landslide displacement prediction change rate; when the landslide displacement prediction change rate is less than the first rate threshold, determining that the target landslide area is in a constant-velocity deformation stage, and generating a first-level early warning signal; when the landslide displacement prediction change rate is greater than or equal to ... When the landslide area reaches the first rate threshold and is less than the second rate threshold, it is determined that the target landslide area has entered the initial accelerated deformation stage, and a second-level early warning signal is generated. When the predicted landslide displacement rate is greater than or equal to the second rate threshold, and the corresponding initial predicted landslide displacement value is less than the absolute displacement threshold, it is determined that the target landslide area has entered the intermediate accelerated deformation stage, and a second-level early warning signal is generated and a high-frequency monitoring command is triggered simultaneously. When the predicted landslide displacement rate is greater than or equal to the second rate threshold, and the corresponding initial predicted landslide displacement value is greater than or equal to the absolute displacement threshold, it is determined that the target landslide area has entered the imminent landslide abrupt change stage, and a third-level early warning signal is generated and an emergency evacuation command is triggered.

[0075] To verify the technical advantages of the algorithm in this invention, a high-precision GNSS monitoring station (number HF05) on a typical landslide (No. 5) was selected for engineering verification. This area is affected by agricultural irrigation and rainwater infiltration during the rainy season, exhibiting a step-like progressive failure characteristic.

[0076] The experimental data was segmented as follows: 745 high-quality long-term monitoring sequences were extracted after cleaning. The first 565 sequences were used as the model training set, the middle 60 sequences as the hyperparameter tuning and validation set, and the last 120 sequences as the test set.

[0077] The evaluation index system includes: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (COD) for regression fit. The models included in the comparison are: LSTM, BPNN, DLinear, TimesNet, iTransformer, Autoformer, and PatchTST. Table 1 shows the benchmark evaluation indicators for the overall prediction performance of each model.

[0078] Table 1 Comparison of benchmark evaluation indicators for the overall prediction performance of each model As shown in Table 1, the model represented by LSTM exhibits a relatively large error in long sequence extrapolation, with an RMSE of 29.83 mm. In contrast, the model constructed in this invention demonstrates a significant performance advantage, with a MAE of 3.02 mm and an RMSE of 4.00 mm. Compared to the single-model PatchTST, the integrated system of this invention achieves a reduction in RMSE error by approximately 56.0%. This confirms the effectiveness of the STL decoupling and divide-and-conquer approach in improving prediction accuracy.

[0079] To verify the contribution of the MixLoss hybrid constraint loss function mechanism in this patent, the residual term component was selected separately for comparative testing. Table 2 shows the comparison of prediction performance of the residual term for high-frequency extreme mutations.

[0080] Table 2 Comparison of prediction performance of high-frequency extreme mutation residuals Table 2 shows that when the model is evaluated using traditional MSE, it performs poorly in the face of frequent residual mutations. After introducing the MixLoss constraint mechanism, the error index of the comparison model is significantly reduced. Especially when iTransformer is combined with MixLoss, the prediction fit is greatly improved. The value was increased to 0.8022. This indicates that the geometric angle penalty imposed on the gradient descent channel can effectively constrain the neural network, enabling it to better track the dramatic fluctuations in the actual displacement.

[0081] After adopting the MixLoss constraint mechanism, the comprehensive prediction system architecture of this invention achieved good prediction results in the final landslide displacement prediction. Its RMSE was reduced to 3.22 mm. The accuracy was improved to 0.9991. Furthermore, during the critical acceleration deformation stage, the prediction curve closely matched the displacement changes and exhibited highly synchronous following, effectively improving the lag problem.

[0082] In one exemplary embodiment, to further explore whether the present invention possesses generalization ability across different geological dynamic mechanisms, the algorithm was directly applied to a landslide monitoring project without modifying the underlying logic code. The landslide in this monitoring project is affected by the periodic large-scale rise and fall of the water level and alternating buoyancy forces of the Three Gorges Dam, and its surface monitoring data exhibits a step-type nonlinear periodic fluctuation characteristic, which differs from the ascent trajectory of a landslide in the previous embodiment. Table 3 compares the prediction performance of different prediction schemes for the Bazimen landslide.

[0083] Table 3 Comparison of the prediction performance of different prediction schemes for the Bazimen landslide In cross-regional verification scenarios, the combined model of this invention demonstrated good environmental adaptability. The RMSE was 6.81 mm. The accuracy of 0.98 fully demonstrates that the combined prediction model overcomes the limitations of a single geological background and possesses strong cross-regional generalization value. Through this precise tracking, once the front-end data shows anomalies and is identified as about to cross the boundary and trigger an alert, the early warning central system can accurately issue early warning instructions in advance using the prediction time window gained by this algorithm, providing valuable emergency response time for the organization and evacuation of people in dangerous areas.

[0084] Figure 3 This is a comparison chart of the prediction results of the original landslide displacement sequence based on the traditional mean square error loss function provided by this invention, as shown in the figure. Figure 3 As shown, the model of this invention exhibits high consistency with the measured displacement curve throughout the entire prediction interval, especially in the final accelerated deformation stage, where it can still track the displacement change process well. In contrast, while TimesNet and PatchTST are relatively close to the measured results in terms of overall trend, they still show some deviation in the section where the deformation rate increases significantly; LSTM, iTransformer, and DLinear show varying degrees of response lag or amplitude deviation in the later stages, and their ability to characterize the nonlinear rapid growth process of landslide displacement is relatively insufficient.

[0085] Figure 4 A comparison chart of the prediction results of the original landslide displacement sequence based on the hybrid loss function of this invention is provided for this invention, as shown in the figure. Figure 4 As shown, although single deep learning models exhibit varying degrees of enhanced ability to follow macroscopic displacement trends after incorporating MixLoss morphological constraints, they still inevitably show some amplitude deviation when dealing with local nonlinear and drastic fluctuations. In contrast, the model proposed in this invention effectively overcomes the aforementioned bottlenecks, demonstrating the highest dynamic consistency with the real displacement sequence in both macroscopic long-term evolution trends and microscopic details of local high-frequency fluctuations.

[0086] Figure 5 This is a comparison chart of the prediction results of the landslide displacement residual term sequence based on the traditional mean square error loss function provided by this invention, as shown in the figure. Figure 5 As shown, each model could barely keep up in the early stable fluctuation range, but in the sudden and large jump region at the end of the sequence, all models showed response lag and could not effectively capture sudden displacement signals.

[0087] Figure 6 A comparison chart of landslide displacement residual term sequence prediction results based on the hybrid loss function of this invention is provided in the present invention, as shown in the figure. Figure 6As shown, the prediction curves of each model exhibit significantly improved sensitivity to abrupt changes. Particularly in extreme displacement jump regions, MixLoss, by introducing trend direction constraints, forces the model to more actively fit the true abrupt change pattern, resulting in predicted amplitudes that closely approximate the true extreme values, effectively mitigating the smoothing effect. Among them, the iTransformer model's prediction curve shows the highest degree of fit with the true residual sequence in terms of amplitude and abrupt change pattern, fully demonstrating its advantages in handling extreme non-stationary time series data.

[0088] When applying the landslide displacement prediction method based on temporal decomposition and morphological constraints provided in this manual, it is not necessary to consider... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this manual does not impose any restrictions on it.

[0089] The above describes one or more embodiments of a landslide displacement prediction method based on temporal decomposition and morphological constraints. Based on the same approach, this specification also provides a corresponding landslide displacement prediction device based on temporal decomposition and morphological constraints, such as... Figure 7 As shown.

[0090] Figure 7 A schematic diagram of a landslide displacement prediction device based on temporal decomposition and morphological constraints provided in this specification includes: The preprocessing module 701 is used to acquire the historical cumulative displacement time series data of the Global Navigation Satellite System (GNSS) of the target landslide area, and to preprocess the historical cumulative displacement time series data of the GNSS to obtain a standardized landslide displacement sequence.

[0091] The decoupling module 702 is used to decouple the standardized landslide displacement sequence into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area through a time-series decomposition algorithm based on local weighted polynomial regression.

[0092] The acquisition module 703 is used to acquire the landslide displacement prediction model. The landslide displacement prediction model includes two parallel TimesNet models and one iTransformer model. The landslide displacement prediction model applies morphological constraints on the direction of displacement sequence evolution during backpropagation to correct the prediction phase deviation during the accelerated deformation stage. The loss function of the landslide displacement prediction model is a MixLoss hybrid loss function that integrates mean square error and cosine similarity constraints.

[0093] The prediction module 704 is used to input the trend component and the seasonal component into the corresponding TimesNet model for prediction, and obtain the predicted values ​​of the trend component and the seasonal component for multiple consecutive time steps starting from the current time. The residual component is input into the iTransformer model for prediction, and obtain the predicted values ​​of the residual component for multiple consecutive time steps starting from the current time.

[0094] The determination module 705 is used to add the predicted values ​​of the trend term, the seasonal term, and the residual term to obtain a sequence of predicted landslide displacement values ​​for multiple consecutive time steps starting from the current moment.

[0095] Specific limitations regarding the landslide displacement prediction device based on temporal decomposition and morphological constraints can be found in the limitations of the landslide displacement prediction method based on temporal decomposition and morphological constraints described above, and will not be repeated here. Each module in the aforementioned landslide displacement prediction device based on temporal decomposition and morphological constraints can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0096] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A landslide displacement prediction method based on temporal decomposition and morphological constraints is provided.

[0097] This instruction manual also provides Figure 8 The schematic diagram of the computer device shown is as follows: Figure 8 At the hardware level, the computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 A landslide displacement prediction method based on temporal decomposition and morphological constraints is provided.

[0098] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A landslide displacement prediction method based on temporal decomposition and morphological constraints, characterized in that, include: Obtain historical cumulative displacement time series data of the target landslide area using Global Navigation Satellite System (GNSS), and preprocess the GNSS historical cumulative displacement time series data to obtain a standardized landslide displacement sequence; The standardized landslide displacement sequence is decoupled into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area by using a time-series decomposition algorithm based on local weighted polynomial regression. A landslide displacement prediction model is obtained; the landslide displacement prediction model includes two parallel TimesNet models and one iTransformer model; the landslide displacement prediction model applies morphological constraints on the direction of displacement sequence evolution during backpropagation to correct the prediction phase deviation during the accelerated deformation stage; the loss function of the landslide displacement prediction model is a MixLoss hybrid loss function that integrates mean square error and cosine similarity constraints. The trend component and the seasonal component are respectively input into the corresponding TimesNet model for prediction to obtain the predicted values ​​of the trend component and the seasonal component for multiple consecutive time steps starting from the current time. The residual component is input into the iTransformer model for prediction to obtain the predicted values ​​of the residual component for multiple consecutive time steps starting from the current time. The predicted values ​​of the trend term, the seasonal term, and the residual term are added together to obtain a sequence of predicted landslide displacement values ​​for multiple consecutive time steps starting from the current moment.

2. The landslide displacement prediction method based on temporal decomposition and morphological constraints as described in claim 1, characterized in that, The loss function of the landslide displacement prediction model is: in, The loss value is for the landslide displacement prediction model. This is the mean square error term used to control the overall accuracy of statistical forecasting. The trend similarity error term is used to correct the predicted phase deviation. The output of the landslide displacement prediction model includes The predicted sequence vector at each time step, for The actual landslide displacement sequence vector corresponding to each time step. Predicting time indices in sequence vectors The value at that location, Time index in the actual landslide displacement sequence vector The value at that location, This is a hyperparameter constant used to adjust the weight balance between trend similarity error and mean square error.

3. The landslide displacement prediction method based on temporal decomposition and morphological constraints as described in claim 1, characterized in that, The time-series decomposition algorithm based on local weighted multinomial regression decouples the standardized landslide displacement sequence into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area. Specifically, this includes: Initialize the local weighted regression parameters and the robust weights of all data points in the standardized landslide displacement sequence; the local weighted regression parameters include window size, weight function type, number of inner loop iterations, and number of outer loop iterations; In the inner loop, the robust weights are combined with a low-pass filter window to perform smoothing and denoising, extract the initial trend term of the standardized landslide displacement sequence, and perform subsequence smoothing on the standardized landslide displacement sequence after removing the initial trend term to extract the initial seasonal term. In the outer loop, the difference between the current standardized landslide displacement sequence and the initial trend term and the initial seasonal term is used as the residual. The updated robust weight of each data point is dynamically updated based on the residual, and the updated robust weight is fed back to the next inner loop; the larger the residual, the smaller the corresponding robust weight. The inner loop and the outer loop are executed alternately until the number of iterations of the inner loop and the number of iterations of the outer loop are reached. The initial trend item and the initial seasonal item corresponding to the last iteration are determined as the trend item component and the seasonal item component. Subtracting the trend component and the seasonal component from the standardized landslide displacement sequence yields the residual component.

4. The landslide displacement prediction method based on temporal decomposition and morphological constraints as described in claim 1, characterized in that, The TimesNet model includes k A parallel multi-scale two-dimensional convolutional module; the step of inputting the trend component and the seasonal component into the corresponding TimesNet model for prediction to obtain the predicted values ​​of the trend component and the seasonal component for multiple consecutive time steps starting from the current time specifically includes: The trend component and the seasonal component are treated as independent one-dimensional time series; For any one-dimensional time series, after performing a Fast Fourier Transform on the one-dimensional time series, the frequency domain features of the one-dimensional time series are extracted. k Each dominant frequency was obtained separately. k The dominant frequency corresponds to k One cycle step size; Based on the above k Each period step is used to reconstruct the one-dimensional time series, resulting in... k A two-dimensional tensor matrix; k Each of the two-dimensional tensor matrices corresponds to k A different cycle step size dimension; The k Each of the two-dimensional tensor matrices is input to... k In a parallel multi-scale two-dimensional convolution module, the two-dimensional spatial evolution features of each two-dimensional tensor matrix within and between periods are extracted by jointly using the multi-scale two-dimensional convolution module. Each two-dimensional spatial evolution feature is sequentially subjected to inverse dimensionality reduction mapping to restore its corresponding one-dimensional feature sequence. Then, Softmax weights are used to weight and fuse these one-dimensional feature sequences to obtain the predicted trend or seasonal values ​​for multiple consecutive time steps starting from the current moment in the one-dimensional time series. The Softmax weights are based on... k The amplitude distribution of the dominant frequencies was calculated.

5. The landslide displacement prediction method based on temporal decomposition and morphological constraints as described in claim 1, characterized in that, The iTransformer model comprises an embedding layer, a Transformer encoding module, and a projection layer connected in series. The Transformer encoding module includes a multi-head attention mechanism and a feedforward neural network. The process of inputting the residual term components into the iTransformer model based on the inverted attention mechanism for prediction, to obtain the predicted residual term values ​​for multiple consecutive time steps starting from the current time, specifically includes: The entire historical time series of the residual component is treated as an independent variable feature token, its dimensions are inverted, and it is input into the embedding layer for mapping to obtain the initial embedding vector representation of the residual component, so as to break away from the traditional time step sequence dependency. The initial embedding vector representation is input into the Transformer encoding module, and the high-frequency nonlinear fluctuation features across time segments are interactively extracted on the global feature dimension of the initial embedding vector representation using the multi-head self-attention mechanism and the feedforward neural network. The high-frequency nonlinear fluctuation features are input into the projection layer for feature dimensionality reduction and nonlinear mapping to obtain the predicted residual terms for multiple consecutive time steps starting from the current moment.

6. The landslide displacement prediction method based on temporal decomposition and morphological constraints as described in claim 1, characterized in that, The method further includes: Calculate the difference between the landslide displacement prediction values ​​at adjacent time steps in the landslide displacement prediction value sequence, and divide it by the preset time step length to obtain the landslide displacement prediction change rate corresponding to multiple consecutive time steps from the current moment. Obtain a pre-set absolute displacement threshold, a first velocity threshold, and a second velocity threshold; the second velocity threshold is greater than the first velocity threshold. Traverse the landslide displacement prediction value sequence and the corresponding landslide displacement prediction change rate. When the landslide displacement prediction change rate is less than the first rate threshold, determine that the target landslide area is in the iso-velocity deformation stage and generate a first-level early warning signal. When the predicted rate of change of landslide displacement is greater than or equal to the first rate threshold and less than the second rate threshold, the target landslide area is determined to have entered the primary accelerated deformation stage, and a second-level early warning signal is generated. When the predicted rate of change of landslide displacement is greater than or equal to the second rate threshold, and the corresponding initial predicted landslide displacement value is less than the absolute displacement threshold, the target landslide area is determined to have entered the medium-accelerated deformation stage, a second-level early warning signal is generated, and a high-frequency monitoring command is triggered simultaneously. When the predicted rate of change of landslide displacement is greater than or equal to the second rate threshold and the corresponding initial predicted landslide displacement value is greater than or equal to the absolute displacement threshold, the target landslide area is determined to have entered the imminent landslide transition stage, a third-level early warning signal is generated, and an emergency evacuation command is triggered.

7. The landslide displacement prediction method based on temporal decomposition and morphological constraints as described in claim 1, characterized in that, The preprocessing of the GNSS historical cumulative displacement time series data to obtain a standardized landslide displacement sequence specifically includes: Gross error removal is performed on the GNSS historical cumulative displacement time series data to obtain gross error removed data; The gross error removal data is downsampled according to a preset time scale to obtain a standardized landslide displacement sequence.

8. A landslide displacement prediction device based on temporal decomposition and morphological constraints, characterized in that, include: The preprocessing module is used to acquire historical cumulative displacement time series data of the Global Navigation Satellite System (GNSS) for the target landslide area, and to preprocess the GNSS historical cumulative displacement time series data to obtain a standardized landslide displacement sequence. The decoupling module is used to decouple the standardized landslide displacement sequence into a trend component reflecting the long-term geological evolution of the target landslide area, a seasonal component reflecting the periodic environmental impact of the target landslide area, and a residual component reflecting the high-frequency random disturbance of the target landslide area through a time-series decomposition algorithm based on local weighted polynomial regression. The acquisition module is used to acquire the landslide displacement prediction model; the landslide displacement prediction model includes two parallel TimesNet models and one iTransformer model; the landslide displacement prediction model applies morphological constraints on the direction of displacement sequence evolution during backpropagation to correct the prediction phase deviation during the accelerated deformation stage; the loss function of the landslide displacement prediction model is a MixLoss hybrid loss function that integrates mean square error and cosine similarity constraints. The prediction module is used to input the trend component and the seasonal component into the corresponding TimesNet model for prediction, and obtain the predicted values ​​of the trend component and the seasonal component for multiple consecutive time steps starting from the current time. The module also inputs the residual component into the iTransformer model for prediction, and obtains the predicted values ​​of the residual component for multiple consecutive time steps starting from the current time. The determination module is used to add the predicted values ​​of the trend term, the seasonal term, and the residual term to obtain a sequence of predicted landslide displacement values ​​for multiple consecutive time steps starting from the current moment.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 7.

Images