Training method of landslide displacement prediction model based on denoising diffusion probability network

Abstract

The application relates to the technical field of artificial intelligence, in particular to a landslide displacement prediction model training method based on a denoising diffusion probability network. The method comprises the following steps: determining a landslide acceleration probability according to a training sample; respectively coding deformation feature data and environmental stress data to obtain displacement features and pressure features, and fusing the displacement features and the pressure features to obtain cross-modal features; superimposing a disturbance signal on the training sample to generate a to-be-recovered sample, estimating a disturbance component in the to-be-recovered sample, and obtaining an initial landslide displacement prediction value; determining a landslide displacement prediction value according to the initial landslide displacement prediction value and the landslide acceleration probability; and optimizing parameters of a landslide displacement prediction model based on the denoising diffusion probability network. The application solves the technical problems that related landslide displacement prediction methods have poor prediction capability for complex landslide systems and sudden acceleration events and low interpretability of model prediction results.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a training method for a landslide displacement prediction model based on a denoised diffusion probability network. Background Technology

[0002] Landslide displacement prediction is a crucial component of geological disaster early warning systems. With the development of surface deformation monitoring technology, high-precision, continuous time-series data on landslide displacement and environmental stress (such as pore water pressure) obtained through methods such as extensometers, interferometric synthetic aperture radar, and global navigation satellite systems have provided a solid foundation for data-driven prediction methods.

[0003] The relevant methods mainly rely on deep learning models to model historical displacement sequences, typically including Long Short-Term Memory networks, gated recurrent units, temporal convolutional networks, and Transformer architectures. Although the above methods have achieved high prediction accuracy in the conventional landslide evolution stage, they are essentially "black box" data-driven models and cannot establish a direct causal relationship between the predicted output and the intrinsic physical mechanism of the landslide.

[0004] To improve interpretability, some studies have introduced signal decomposition and attention mechanisms. However, the interpretability of these methods remains at the level of post-hoc correlation analysis; the decomposed components or attention weights themselves lack clear physical meaning and cannot reflect the threshold triggering mechanism of landslide instability. Especially when facing rare, nonlinear accelerated instability events, due to the scarcity of such event samples in the training data, existing models are prone to prediction lag, misjudgment, or complete failure.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This application provides a training method for a landslide displacement prediction model based on a denoised diffusion probability network, in order to at least solve the technical problems that related landslide displacement prediction methods have poor predictive ability for complex landslide systems and sudden acceleration events, and low interpretability of model prediction results.

[0007] According to one aspect of this application, a training method for a landslide displacement prediction model based on a denoised diffusion probability network is provided, comprising: identifying the deviation between signals, constructing a loss function, and optimizing the parameters of the landslide displacement prediction model based on the loss function until a preset stopping condition is met to obtain training samples, wherein the training samples include at least: deformation feature data for characterizing the geometric evolution of the landslide, and environmental stress data for characterizing the internal state of the landslide; determining the landslide acceleration probability based on the training samples; encoding the deformation feature data and environmental stress data respectively to obtain displacement features and pressure features, and fusing the displacement features and pressure features to obtain cross-modal features; superimposing the disturbance signal with the training samples to generate a sample to be restored, and using the landslide displacement prediction model based on the denoised diffusion probability network, under the guidance of the cross-modal features, estimating the disturbance component in the sample to be restored to obtain an initial landslide displacement prediction value; determining the landslide displacement prediction value based on the initial landslide displacement prediction value and the landslide acceleration probability; and determining the landslide displacement prediction value based on the landslide displacement prediction value and the disturbance.

[0008] Optionally, determining the landslide acceleration probability based on training samples includes: using a value-excited filtering layer to perform feature enhancement processing on the environmental stress data, increasing the weight of environmental stress data above a preset threshold in the feature space and decreasing the weight of environmental stress data below a preset threshold in the feature space to obtain first environmental stress data, wherein the feature space is used to characterize the evolution law of preset feature components in the environmental stress data; using a multi-scale feature extraction module to process the first environmental stress data to obtain temporal dynamic features at different scales, wherein the multi-scale feature extraction module includes at least two parallel convolutional branches, and different convolutional branches use different convolutional kernel sizes; concatenating the temporal dynamic features at different scales to obtain multi-scale fusion features; using a convolutional layer with a channel attention mechanism to process the multi-scale fusion features to adaptively allocate the weights of features at different scales to obtain attention-weighted features; and processing the attention-weighted features through an activation function to obtain the landslide acceleration probability at each sampling time.

[0009] Optionally, displacement features and pressure features are fused to obtain cross-modal features, including: interactively fusing displacement features and pressure features through two symmetrical and parallel encoding paths; wherein the two symmetrical and parallel encoding paths include a first encoding path and a second encoding path; in the first encoding path, using displacement features as the main query vector and pressure features as the key value vector, the dependency relationship between displacement mode and pressure mode is extracted through a cross-modal attention mechanism to obtain a first fused feature; feature enhancement processing is performed on the first fused feature to obtain a first target fused feature; in the second encoding path, using pressure features as the main query vector and displacement features as the key value vector, the dependency relationship between pressure mode and displacement mode is extracted through a cross-modal attention mechanism to obtain a second fused feature; feature enhancement processing is performed on the second fused feature to obtain a second target fused feature; the first target fused feature and the second target fused feature are concatenated to obtain an initial cross-modal feature; the initial cross-modal feature is mapped to a pre-set representation space to obtain the cross-modal feature.

[0010] Optionally, a landslide displacement prediction model based on a denoising diffusion probability network is used to estimate the disturbance component in the sample to be restored under the guidance of cross-modal features, thereby obtaining an initial landslide displacement prediction value. This includes: in the iterative denoising step of the landslide displacement prediction model based on a denoising diffusion probability network, the following sub-steps are executed: modeling the temporal dependency within the sample to be restored in the current round using a self-attention mechanism; extracting and predicting the trend component features and residual component features corresponding to the disturbance component in the sample to be restored through parallel sub-branches; injecting the cross-modal features as physical constraint signals into the sub-branches using a cross-attention mechanism to guide the denoising direction for the trend component features and residual component features; dynamically adjusting the fusion ratio of the trend component features and residual component features using a gating unit to generate an intermediate prediction sample corresponding to the current iteration round; and repeatedly executing the iterative denoising step to gradually restore the real data in the sample to be restored through multiple iterations until a preset stopping condition is met, and determining the intermediate prediction sample obtained in the last iteration as the initial landslide displacement prediction value.

[0011] Optionally, the trend component features and residual component features corresponding to the perturbation components in the sample to be restored are extracted and predicted respectively through parallel sub-branches, including: extracting and predicting the trend component features and residual component features corresponding to the perturbation components in the sample to be restored respectively through parallel first sub-branch and second sub-branch; in the first sub-branch, the trend component features in the landslide movement are characterized by polynomial fitting; in the second sub-branch, the residual component features between historical data and predicted data are captured by neural network.

[0012] Optionally, a cross-attention mechanism is used to inject cross-modal features as physical constraint signals into sub-branches to guide the denoising direction for trend component features and residual component features. This includes: using the cross-attention mechanism, taking the sample to be restored in the current round as the query vector and the cross-modal features as the key vector and value vector, calculating the correlation weight distribution of the sample to be restored relative to the environmental stress data at each sampling time; and performing weighted correction on the representation of the sample to be restored based on the correlation weight distribution to adjust the denoising update direction of the sample to be restored. In this way, through weighted correction, the sample to be restored is constrained to maintain physical evolution consistency with the environmental stress data on the denoising trajectory, and the landslide displacement prediction model based on the denoising diffusion probability network is guided to generate a target evolution path that matches the changing pattern of the environmental stress data.

[0013] Optionally, obtaining training samples includes: obtaining original training samples; removing data segments with consecutive missing values ​​exceeding a preset threshold from the original training samples to obtain a first training sample; filling in missing sampling points in the first training sample to obtain a second training sample; identifying and removing outlier sampling points in the second training sample to obtain a third training sample; denoising the third training sample to obtain a fourth training sample; and resampling the deformation feature data and environmental stress data in the fourth training sample at the same multiple sampling times to obtain training samples.

[0014] According to another aspect of this application, a landslide displacement prediction method is also provided, comprising: acquiring landslide monitoring data; analyzing the landslide monitoring data using a landslide displacement prediction model based on a denoised diffusion probability network to obtain a predicted landslide displacement value, wherein the landslide displacement prediction model based on a denoised diffusion probability network is obtained by training using a training method for a landslide displacement prediction model based on a denoised diffusion probability network with weights above.

[0015] According to another aspect of this application, a training device for a landslide displacement prediction model based on a denoised diffusion probability network is also provided, comprising: an acquisition module for acquiring training samples, wherein the training samples include at least: deformation feature data characterizing the geometric evolution of the landslide and environmental stress data characterizing the internal state of the landslide; a first determination module for determining the landslide acceleration probability based on the training samples; an encoding module for encoding the deformation feature data and the environmental stress data respectively to obtain displacement features and pressure features, and fusing the displacement features and pressure features to obtain cross-modal features; and an estimation module. The first module is used to superimpose the disturbance signal with the training sample to generate the sample to be restored. Then, using the landslide displacement prediction model based on the denoising diffusion probability network, the disturbance component in the sample to be restored is estimated under the guidance of cross-modal features to obtain the initial landslide displacement prediction value. The second determination module is used to determine the landslide displacement prediction value based on the initial landslide displacement prediction value and the landslide acceleration probability. The third module is used to construct a loss function based on the deviation between the landslide displacement prediction value and the disturbance signal, and optimize the parameters of the landslide displacement prediction model based on the denoising diffusion probability network based on the loss function until the preset stopping condition is met.

[0016] According to another aspect of this application, a non-volatile storage medium is also provided, the storage medium including a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above-described training method for the landslide displacement prediction model based on a denoising diffusion probability network.

[0017] According to another aspect of this application, an electronic device is also provided, comprising: a memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the above-described training method for a landslide displacement prediction model based on a denoised diffusion probability network.

[0018] According to another aspect of this application, a computer program is also provided, wherein when the computer program is executed by a processor, it implements the above-described training method for a landslide displacement prediction model based on a denoised diffusion probability network.

[0019] According to another aspect of this application, a computer program product is also provided, the computer program product including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above-described training method for the landslide displacement prediction model based on a denoising diffusion probability network.

[0020] In this application, a loss function is constructed using the deviation between signals, and the parameters of a landslide displacement prediction model based on a denoising diffusion probability network are optimized based on the loss function until a preset stopping condition is met to obtain training samples. The training samples include at least: deformation feature data characterizing the geometric evolution of the landslide, and environmental stress data characterizing the internal state of the landslide. Based on the training samples, the landslide acceleration probability is determined. The deformation feature data and environmental stress data are encoded separately to obtain displacement features and pressure features, which are then fused to obtain cross-modal features. The disturbance signal is superimposed on the training samples to generate samples to be restored. Then, guided by the cross-modal features, the landslide displacement prediction model based on a denoising diffusion probability network is used to analyze the samples to be restored. The perturbation component in the model is estimated to obtain the initial landslide displacement prediction. Based on the initial landslide displacement prediction and the landslide acceleration probability, the landslide displacement prediction is determined. According to the landslide displacement prediction and the perturbation mode, by establishing an explicit coupling between the landslide acceleration probability and the diffusion generation process under multimodal physical constraints, the model output can be traced back to the geological, hydrological and physical mechanisms without relying on ex post facto interpretation. This achieves the technical effect of improving the generalization ability of complex landslide system evolution behavior, the prediction robustness of rare accelerated instability events, and the clear physical interpretability of the prediction results. This solves the technical problem that related landslide displacement prediction methods have poor prediction ability for complex landslide systems and sudden acceleration events and low interpretability of model prediction results. Attached Figure Description

[0021] 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:

[0022] Figure 1 This is a flowchart of a training method for a landslide displacement prediction model based on a denoised diffusion probability network according to an embodiment of this application.

[0023] Figure 2 This is a schematic diagram of a training method for a multimodal gated diffusion model according to an embodiment of this application;

[0024] Figure 3 This is a structural diagram of a multimodal fusion encoder according to an embodiment of this application;

[0025] Figure 4 This is a structural diagram of a denoising diffusion probability model decoder according to an embodiment of this application;

[0026] Figure 5 This is a flowchart of a landslide displacement prediction method according to an embodiment of this application;

[0027] Figure 6 This is a schematic diagram of raw landslide monitoring data according to an embodiment of this application;

[0028] Figure 7 This is a schematic diagram of a test result according to an embodiment of this application;

[0029] Figure 8 An example diagram of a trend component and an acceleration probability component according to an embodiment of this application;

[0030] Figure 9 A structural diagram of a training device for a landslide displacement prediction model based on a denoised diffusion probability network according to an embodiment of this application;

[0031] Figure 10 A hardware block diagram of a computer terminal for a training method of a landslide displacement prediction model based on a denoised diffusion probability network, according to an embodiment of this application. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] According to an embodiment of this application, a method embodiment for training a landslide displacement prediction model based on a denoised diffusion probability network is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] Figure 1 This is a flowchart illustrating a training method for a landslide displacement prediction model based on a denoised diffusion probability network, according to an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0036] Step S101: Obtain training samples, wherein the training samples include at least: deformation feature data for characterizing the geometric evolution of the landslide, and environmental stress data for characterizing the internal state of the landslide.

[0037] The deformation characteristic data includes at least landslide displacement data, and the environmental stress data includes at least pore pressure data. Landslide displacement data refers to the change in the geometric position of a landslide body over time, obtained through surface monitoring equipment. It can be recorded with millimeter to centimeter precision, characterizing the cumulative deformation process of the landslide body as a whole or locally along the slope or vertical direction. Landslide displacement data reflects the macroscopic response results of landslide movement, including long-term slow creep trends, seasonal periodic fluctuations, and sudden accelerated sliding, among other multi-scale movement characteristics. Pore water pressure data refers to the hydrostatic pressure exerted on the water in the pores of the soil, measured by pore water pressure gauges embedded within the landslide body. Its unit can be kilopascals or meters of water column. Pore water pressure data reflects the dynamic changes in the hydrodynamic state within the landslide body. When rainfall or snowmelt causes water infiltration, the increased pore water pressure reduces the effective stress of the soil and weakens its shear strength, which is a key internal factor triggering landslide instability and accelerated movement.

[0038] Step S102: Determine the landslide acceleration probability based on the training samples.

[0039] In step S102, the landslide acceleration probability is estimated based on the acceleration probability estimation module. The landslide acceleration probability refers to the conditional probability that, at a certain moment, the landslide body will change from a stable or slowly creeping state to a significantly accelerated unstable state. Its value is between 0 and 1 and is dynamically estimated by the model based on the real-time changes in environmental stress (such as pore water pressure). It is used to quantify the critical triggering risk of landslide instability.

[0040] The landslide acceleration probability is a dynamic risk assessment quantification obtained by modeling the physical driving mechanism of landslides through a deep learning module, combining historical landslide movement characteristics with the threshold response relationship of pore water pressure. The core physical basis of the landslide acceleration probability stems from the effective stress principle and Coulomb's law of friction: when pore water pressure rises to a certain critical threshold, the effective stress of the soil decreases, the shear strength weakens, and the shear force on the sliding surface exceeds the resistance, thus triggering a nonlinear jump in displacement rate. Therefore, the landslide acceleration probability is essentially a mathematical representation of the aforementioned physical critical process.

[0041] Step S103: Encode the deformation feature data and environmental stress data respectively to obtain displacement features and pressure features, and fuse the displacement features and pressure features to obtain cross-modal features.

[0042] In step S103, a cross-modal encoder is used to independently encode the landslide displacement data and pore water pressure data to extract the deep temporal dynamic features contained in each mode. Specifically, a coding network based on convolution and self-attention mechanisms is used to process the deformation feature data, capturing the long-term trend evolution, periodic fluctuations, and local abrupt change patterns in the landslide displacement sequence, and outputting a high-dimensional, semantically rich displacement feature vector. Simultaneously, another set of structurally symmetrical coding modules is used to process the environmental stress data. This module employs multi-scale convolutional branching and channel attention mechanisms to enhance the sensitivity to sudden increases in pore water pressure, cumulative effects, and gradually changing backgrounds, thereby generating a pressure feature vector that reflects the hydrological driving intensity and evolutionary stage.

[0043] The two types of features are input into a dual-path interactive cross-modal fusion architecture, with displacement as the primary query and pressure as the reference, and pressure as the primary query and displacement as the reference, respectively. By alternately executing cross-modal attention operations, the displacement features can adaptively focus on the water pressure change periods that are strongly correlated with its motion response, while the pressure features can identify the deformation response patterns most sensitive to landslide acceleration. The fused features are then refined by a self-attention module to enhance the temporal dependency consistency within each modality. Finally, the outputs of the two paths are concatenated along the feature dimension and mapped to a unified semantic space through linear projection, forming a high-dimensional cross-modal feature representing the coupled dynamics of landslide deformation and hydrology.

[0044] Step S104: The disturbance signal is superimposed with the training sample to generate the sample to be restored. Then, the disturbance component in the sample to be restored is estimated by using the landslide displacement prediction model based on the denoising diffusion probability network under the guidance of cross-modal features, so as to obtain the initial landslide displacement prediction value.

[0045] The denoising diffusion probability network can also be defined as the denoising diffusion probability model (DDPM). The landslide displacement prediction model based on the denoising diffusion probability network can also be defined as the multi-modal gated diffusion model (MG-DS).

[0046] In step S104, Gaussian noise is added to the landslide displacement sequence in the training samples as a perturbation signal step by step to form the corrupted samples to be restored. It is worth noting that the above process simulates the forward diffusion process from the real displacement distribution to a completely random noise distribution, providing a training basis for subsequent inverse denoising and reconstruction.

[0047] The sample to be restored, along with the previously generated cross-modal features, is input into the landslide displacement prediction model based on a denoising diffusion probability network. In each denoising time step, the landslide displacement prediction model based on the denoising diffusion probability network models the long-range dependencies within the displacement sequence through a self-attention mechanism, and uses a cross-attention mechanism to introduce the deformation evolution law and pore water pressure driving signal included in the cross-modal features into the decoding process, thereby achieving conditionally guided recovery of the displacement sequence under noise interference.

[0048] In the above process, the model iteratively learns the denoising function, successively estimates the noise residual at the current moment, and gradually corrects the prediction results accordingly. This effectively suppresses the interference of observation noise and abnormal fluctuations while preserving the original displacement structure. After several rounds of denoising iterations, the model outputs a preliminarily recovered landslide displacement sequence with temporal continuity, serving as the initial landslide displacement prediction value. It can be understood that the initial landslide displacement prediction value is structurally constrained by the physical driving characteristics, possessing stronger stability and generalization ability than traditional end-to-end regression models.

[0049] Step S105: Determine the predicted landslide displacement value based on the initial predicted landslide displacement value and the landslide acceleration probability.

[0050] In step S105, the initial landslide displacement prediction value obtained in step S104 is multiplied element-wise with the landslide acceleration probability obtained in step S102, thereby applying a physical mechanism-based dynamic modulation to the initial prediction result. It should be noted that the multiplication operation uses the landslide acceleration probability as a gating switch. During periods when the pore water pressure has not yet reached the instability threshold, the predicted displacement value is suppressed within a low amplitude range to inhibit non-physical fluctuations. During periods when the pore water pressure significantly increases and the acceleration probability approaches or exceeds the critical value, the initial prediction value is allowed to be fully released, accurately reproducing the nonlinear accelerated displacement of the landslide body due to strength weakening.

[0051] Step S105 is based on the core physical law of landslide dynamics: the change in displacement rate is not a linear response to environmental driving forces, but rather exhibits a significant threshold triggering characteristic. Therefore, through step S105, the landslide displacement prediction model based on a denoised diffusion probability network can explicitly model the "stable-accelerated" state switching of landslides during the prediction stage. The output landslide displacement prediction value is the result after physical constraint correction, which not only is closer to the actual observation in terms of numerical value, but also accurately reproduces the unique intermittent acceleration and slow recovery evolution characteristics of landslides in terms of temporal morphology.

[0052] Step S106: Based on the deviation between the predicted landslide displacement value and the disturbance signal, a loss function is constructed, and based on the loss function, the parameters of the landslide displacement prediction model based on the denoising diffusion probability network are optimized until the preset stopping condition is met.

[0053] In step S106, based on the difference between the landslide displacement prediction value output by the model and the real displacement sequence in the training samples, a reconstruction loss function based on the L1 norm or L2 norm is constructed. The loss function is used to quantify the degree of deviation between the final recovery result of the denoising process and the real physical state, and serves as a supervision signal for model learning.

[0054] Meanwhile, to enhance the model's sensitivity to acceleration events, a weighting mechanism can be introduced into the loss function to assign higher weights to time points where the displacement rate of change increases significantly, so that the model focuses more on accurately reconstructing the key acceleration phase before landslide instability during the optimization process.

[0055] Building upon this, the loss gradient is propagated layer by layer along the entire landslide displacement prediction model based on a denoising diffusion probability network using the backpropagation algorithm. This updates the parameters of each layer in the accelerated probability estimation module, cross-modal encoder, and DDPM decoder, enabling the model to gradually enhance its learning ability regarding the mapping relationship between noise input, physical guidance, and true output in each training iteration. When the loss value converges to a preset threshold, or when the prediction performance on the validation set no longer significantly improves, it indicates that the model has fully captured the nonlinear coupling between landslide displacement and pore water pressure and possesses good generalization ability. At this point, optimization is stopped, the optimal parameters are retained, and the model training process is complete.

[0056] The above steps, by establishing an explicit coupling between the landslide acceleration probability and the diffusion generation process under multimodal physical constraints, achieve the goal of making the model output traceable to geological, hydrological, and physical mechanisms without relying on ex post facto interpretation. This achieves the technical effect of improving the generalization ability of the evolution behavior of complex landslide systems, the predictive robustness of rare accelerated instability events, and giving the prediction results clear physical interpretability.

[0057] The core advantage of the above steps lies in achieving a balance between high accuracy and physical interpretability in the prediction results. Compared to traditional black-box deep learning models, this method reflects the long-term geological evolution of landslides through trend components and accelerates the probabilistic quantification of the regulatory effect of pore water pressure on instability, thus giving the prediction output a clear geomechanical meaning. This characteristic gives it unique application value in the field of landslide disaster monitoring and early warning, and it is particularly suitable for public safety scenarios where transparency in the model decision-making process is required.

[0058] In the daily risk assessment work of emergency management departments, relevant systems often rely on manual experience combined with threshold alarms (such as alarming when the daily displacement increase exceeds 5mm). This is prone to false alarms due to a brief acceleration after rainfall, or missed alarms due to slow creep instability. The above steps can be used as an auxiliary decision-making module, embedded in the geological disaster early warning platform, to output two independent but related indicators: the trend accumulation rate and the probability of instability. Emergency personnel can then determine whether to initiate patrols, increase monitoring, or prepare personnel evacuation plans.

[0059] In the geological hazard investigation and dynamic assessment system of the natural resources and water conservancy departments, the above steps can support the construction of risk profiles for hazard points. Traditional methods rely on periodic field surveys and point monitoring, which are difficult to capture the nonlinear characteristics of landslide evolution. This application can access continuous data streams from pore water pressure and displacement sensors for a long period of time, automatically identifying which landslides exhibit a low-speed continuous accumulation + water pressure sensitive response mode (belonging to the high-risk slow-moving type) and which exhibit intermittent acceleration + water pressure threshold triggering characteristics (belonging to the sudden instability type). This application provides a data-driven classification basis for risk grading, rather than relying solely on morphology or historical event experience.

[0060] In long-term observation projects conducted by the Geological Survey and geological environment monitoring agencies, this application can serve as an auxiliary analytical tool for retrospective modeling of historical landslide events. For example, for landslides that have already experienced instability, historical monitoring data can be used to train models to infer their pre-instability trend evolution path and accelerating triggering conditions, thereby extracting the behavioral characteristics of typical landslides in the area. Furthermore, undeformed potential hazard points with similar structures, lithology, and hydrological conditions can be identified as priority targets for deploying monitoring equipment.

[0061] In slope health monitoring for engineering design and infrastructure operation and maintenance units (such as railways, highways, hydropower stations, and mining enterprises), this application can be integrated into automated monitoring systems to replace or enhance traditional displacement over-limit alarm logic. Many engineering slopes are in a state of slow deformation for a long time. Traditional systems suffer from alarm fatigue due to frequent alarms, and maintenance personnel may easily overlook real risks. This application identifies intrinsic creep through trend components and abnormal disturbances through acceleration probability, which can effectively distinguish between normal deformation and early signs of danger. For example, when pore water pressure increases due to heavy rain but the acceleration probability does not increase significantly, it can be determined as a water pressure response fluctuation and no alarm can be triggered temporarily; conversely, if the acceleration probability continues to rise, even if the displacement increment is still small, it can indicate potential instability risk and provide an earlier intervention window for engineering reinforcement.

[0062] Furthermore, the model structure obtained in this application supports lightweight deployment and edge computing. Its inference process is unidirectionally generated, does not rely on complex online training, and can be deployed on low-power edge devices (such as industrial-grade IoT gateways and embedded AI modules) to achieve localized prediction in remote mountainous areas with no network or limited communication. At the same time, its input only requires data from two common types of sensors: displacement and pore water pressure. It does not require high-cost interferometric synthetic aperture radar, lidar, or high-frequency acquisition by drones, and has high compatibility with existing monitoring networks, facilitating rapid upgrades at existing sites.

[0063] Figure 2 This is a schematic diagram of a training method for a multimodal gated diffusion model according to an embodiment of this application, as shown below. Figure 2 As shown, the multimodal gated diffusion model (MG-DS) includes: an Acceleration Probability Estimation Module (DDPM), a Multimodal Fusion Encoder (MFE), and a Denoising Diffusion Probability Model Decoder (DDPM). The DDPM estimates the landslide acceleration probability at each time point. The MFE jointly encodes landslide displacement and pore pressure data to extract cross-modal features. The DDPM decoder iteratively optimizes the predicted displacement values ​​using these cross-modal features. The output of the DDPM decoder is then multiplied by the estimated acceleration probability to obtain the final output of the MG-DS.

[0064] It is worth explaining that the design of the multimodal gated diffusion model follows the guidelines below and is further enhanced through multimodal fusion to improve its ability to characterize complex geomechanical processes.

[0065] Landslide displacement prediction methods are mainly based on the effective stress principle and Coulomb's friction law. It is deduced that pore pressure exceeding a critical threshold can cause landslide instability to transition from stable sliding to an accelerated state. Indoor and field observations further demonstrate that pore pressure exceeding this threshold can significantly increase the displacement rate within the shear zone. However, changes in pore pressure may also induce other feedback mechanisms, such as shear dilatation strengthening effects, making the identification of the threshold for predicting landslide acceleration more complex. Directly studying these mechanisms remains challenging. Therefore, this application proposes a landslide prediction method that combines signal decomposition with a simplified characterization of the pore pressure-driven process (Equation 1). Trend Term and residuals As a mathematical abstraction of the driving force under different characteristic frequencies: Capture stable, continuous low-frequency motion. This indicates relatively unstable, high-frequency fluctuations. Gating function Characterizing the control of pore pressure on displacement rate: displacement accelerates when pore pressure exceeds a threshold, otherwise it remains relatively stable.

[0066]

[0067]

[0068] Furthermore, the trend-residual decomposition architecture in the multimodal gated diffusion model separates the low-frequency trend component and the high-frequency residual component in landslide displacement, providing an explanatory basis for the landslide displacement prediction process.

[0069] Figure 2 The specific training process for the multimodal gated diffusion model shown is as follows.

[0070] In step S201, the model receives multimodal observation data (Data), which includes landslide displacement time series and pore water pressure time series. The two are aligned in the time dimension and constitute the joint input of the model.

[0071] In step S202, the accelerated probability estimation module outputs a probability vector using pore water pressure data.

[0072] By employing multi-scale convolution and attention mechanisms, the conditional probability value (i.e., landslide acceleration probability) of landslide acceleration at each time point is dynamically calculated. The landslide acceleration probability, ranging from 0 to 1, represents the likelihood of a nonlinear jump in landslide displacement rate triggered by current hydrological conditions and is a core intermediate variable for achieving physical interpretability of the model.

[0073] Specifically, the accelerated probability estimation module first performs initial feature enhancement processing on the input pore water pressure time series. By introducing an incentive mechanism, it nonlinearly amplifies the high-amplitude features in the signal while suppressing low-amplitude background noise, thereby significantly enhancing the response weight of the key warning signal generated when the pore water pressure exceeds the critical threshold in the feature space, and generating the first enhanced feature sequence after preliminary screening.

[0074] The first enhanced feature sequence is input into a multi-scale convolutional layer (CONV) for deep feature extraction. This convolutional layer consists of multiple parallel convolutional branches, each using convolutional kernels of different sizes to capture dynamic change patterns at different time scales. Large convolutional kernels are used to extract long-term trends and cumulative effects, while small convolutional kernels are used to capture short-term mutations and high-frequency fluctuations. The multi-scale local features output by each branch are concatenated along the channel dimension to form a fusion feature representation containing rich spatiotemporal context information.

[0075] After completing the multi-scale feature fusion, the accelerated probability estimation module further performs secondary adaptive weighting on the fused features. Through the channel incentive mechanism, it performs global statistical analysis on the spliced ​​multi-scale features, calculates the importance score of each feature channel, and generates adaptive weight coefficients accordingly. The original features are then subjected to channel-by-channel weighted modulation to strengthen the key feature channels that are highly correlated with the landslide instability mechanism, while weakening the interference of irrelevant or redundant information, resulting in a finely modulated second enhanced feature sequence.

[0076] The second enhanced feature sequence is input into the probability estimation layer (Estimate), which can be composed of a fully connected layer or a linear projection layer combined with a sigmoid activation function. This layer maps the high-dimensional feature representation to a low-dimensional probability space and outputs the landslide acceleration probability value corresponding to each time sampling point through nonlinear transformation. The probability values ​​of all time points together constitute the landslide acceleration probability vector (Probability Vector).

[0077] Step S203: The multimodal encoder fuses pore water pressure and slip displacement data to generate cross-modal features.

[0078] Through a dual-path cross-attention architecture, bidirectional feature interaction is achieved, with displacement as the primary driver and pressure as the secondary driver, and pressure as the primary driver and displacement as the secondary driver, respectively. The goal of the multimodal encoder is to extract the nonlinear coupling relationship between the two types of data, and generate a unified high-dimensional semantic representation that fuses the deformation response mechanism and hydrological driving signals, i.e., cross-modal features.

[0079] Specifically, the multimodal encoder comprises two symmetrical and complementary processing paths, each primarily driven by a different modality for feature interaction. In the first processing path, the landslide displacement time series is used as the main query vector, and the pore water pressure time series is used as the key and value vectors, respectively, and input into the fusion self-attention module (Fusion, Self-Attn). The model calculates the correlation weights between displacement and pressure features, enabling the displacement mode to adaptively focus on feature segments in the pore water pressure that are closely related to its deformation response, thereby generating cross-modal guided displacement enhancement features. These enhanced features are then subjected to dimensionality mapping and normalization through a linear mapping layer (Projection) to obtain the first fused feature representation.

[0080] In the second processing path, the processing logic is symmetrical to the first path, but the dominant modes are reversed. The pore water pressure time series is used as the main query vector, and the landslide displacement time series is used as the key and value vectors, both of which are input into the fusion self-attention module. Here, the model models the regulatory effect of pore water pressure on displacement evolution through the attention mechanism, enabling the pressure mode to capture the dynamic changes in the displacement history sequence that indicate potential instability modes, generating pressure enhancement features guided by cross-modal characteristics. These pressure enhancement features are also subjected to dimensionality mapping and normalization through a linear mapping layer (projection) to obtain the second fused feature representation.

[0081] The design of the two paths mentioned above can ensure that the displacement and pressure modes maintain their respective temporal dependencies during feature extraction, while also achieving bidirectional information complementarity and deep coupling.

[0082] Step S204: In the prediction phase, the prediction window is initialized with Gaussian noise.

[0083] In the forward diffusion during the training phase, within the time window to be predicted, the initial displacement sequence is replaced with random noise that follows a standard normal distribution as the starting point of the diffusion model. This simulates the process of reconstructing the real signal from a state without information and is the core design of the denoising diffusion probability model. This ensures that the model does not depend on the initial value but gradually recovers the real displacement pattern through iteration.

[0084] In step S205, the DDPM decoder obtains a clear prediction result through an iterative denoising process.

[0085] Using Gaussian noise as the initial input and cross-modal features from a multimodal encoder as the guiding condition, a multi-round iterative neural network denoising process gradually eliminates noise and recovers a clear displacement sequence. This process includes two parallel branches: a trend branch for modeling long-term smooth displacement evolution and a residual branch for capturing high-frequency fluctuations and unmodeled nonlinear responses. The sum of these two branches forms a preliminary prediction.

[0086] In other words, starting from the initial time step t=T, the input is a completely random Gaussian noise sequence. The decoder uses the conditional features extracted by the multimodal fusion encoder as a guiding signal. Through self-attention and cross-attention mechanisms, it predicts the noise components in the current noise sample at each step t, and uses a predefined scheduling strategy to gradually reduce the noise variance and update the sample state to t-1.

[0087] Specifically, in the early stages (e.g., t=T→t=T-1), the model primarily relies on macroscopic conditions to guide the reconstruction of the global trend structure of the displacement sequence, ensuring that the initially generated sequence possesses a reasonable physical evolutionary form (e.g., monotonically increasing or periodically fluctuating). In the later stages (e.g., t=1→t=0), the model focuses on fine-tuning high-frequency details, eliminating residual random noise, and ensuring the numerical accuracy and smoothness of the prediction results. After T iterations of denoising, the model finally outputs the clean displacement prediction sequence at time t=0, which is the initial displacement prediction value of the MG-DS model.

[0088] The initial displacement prediction output from the DDPM decoder is multiplied element-wise with the probability vector output from the acceleration probability estimation module. When the acceleration probability is close to 0, the prediction is suppressed, simulating a steady state; when the probability is close to 1, the prediction is fully preserved or enhanced, simulating acceleration instability.

[0089] After gating modulation, the final prediction result is output. This final prediction result not only approximates the actual observed displacement numerically, but also possesses physical rationality in its temporal pattern. Specifically, it exhibits real landslide evolution characteristics such as low-amplitude fluctuations during stable periods, sudden increases in water pressure after rises, and slow decay after water pressure falls, achieving a balance between high accuracy and high interpretability.

[0090] According to some optional embodiments of this application, step S102, determining the landslide acceleration probability based on training samples, can be achieved through the following steps: Using a value-excited filtering layer, feature enhancement processing is performed on the environmental stress data to increase the weight of environmental stress data above a preset threshold in the feature space and decrease the weight of environmental stress data below a preset threshold in the feature space, resulting in first environmental stress data. The feature space is used to characterize the evolution of preset feature components in the environmental stress data. A multi-scale feature extraction module is used to process the first environmental stress data to obtain temporal dynamic features at different scales. The multi-scale feature extraction module includes at least two parallel convolutional branches, each using a different kernel size. The temporal dynamic features at different scales are concatenated to obtain multi-scale fusion features. A convolutional layer with channel attention is used to process the multi-scale fusion features to adaptively allocate weights for features at different scales, resulting in attention-weighted features. An activation function is used to process the attention-weighted features to obtain the landslide acceleration probability at each sampling time.

[0091] Step S102 can be achieved through Figure 2 The acceleration probability estimation module is implemented in this embodiment. It is understood that landslide acceleration is typically related to pore pressure exceeding a certain threshold. These thresholds vary significantly among different landslides. For a specific landslide, its triggering threshold may also change slightly over time. Therefore, this embodiment employs a deep learning-based method to dynamically estimate the landslide acceleration probability, specifically including the following steps.

[0092] First, a value-excited filtering layer is used to nonlinearly enhance the environmental stress data. This layer employs an activation function based on intensity response, significantly amplifying environmental stress values ​​above a preset threshold to give them higher representation weight in the feature space. Simultaneously, it effectively suppresses low-amplitude fluctuations below the threshold, thus highlighting key pressure surge events that may trigger landslide instability and weakening the interference of background noise and normal fluctuations. The resulting enhanced first environmental stress data is output. This process, through continuous nonlinear mapping, simulates the sensitivity transition mechanism of geological bodies when water pressure accumulation reaches a critical state, making the feature space more focused on the evolution path of high-energy signals directly related to instability.

[0093] Secondly, the first environmental stress data is input into the multi-scale feature extraction module, which consists of at least two parallel one-dimensional convolutional branches. Each branch uses a convolutional kernel of different lengths, such as 3, 7, and 15 time steps, to simultaneously capture short-term sudden pressure pulses, medium-term cumulative upward trends, and long-term slowly increasing hydrological background, thereby comprehensively depicting the dynamic evolution pattern of pore pressure at multiple time scales and generating feature maps that reflect the response characteristics at different time granularities.

[0094] Furthermore, temporal features from different convolutional branches are concatenated along the channel dimension to form a multi-scale fusion feature containing rich multi-scale information, thus fully preserving the complete hydraulic response spectrum from instantaneous perturbations to long-term accumulation. In addition, to avoid uneven contributions from features at different scales in subsequent processing, a convolutional layer with a channel attention mechanism is introduced. This layer adaptively weights global features, dynamically identifies the scale channel most discriminative for accelerating prediction in the current landslide environment, and assigns it a higher attention weight, suppressing redundant or interfering features. This achieves context-aware optimization of feature selection, outputting a highly discriminative feature representation after attention weighting.

[0095] Finally, the attention-weighted features are mapped to a continuous probability value sequence between 0 and 1 through a fully connected layer and a Sigmoid activation function, with each sampling time corresponding to an independently estimated landslide acceleration probability.

[0096] This embodiment uses a collaborative mechanism of value-excited filtering, multi-scale convolutional feature extraction, and channel attention weighting to adaptively identify and enhance high-amplitude environmental stress signals closely related to landslide instability, effectively suppressing background noise and non-critical fluctuations, thereby accurately estimating the landslide acceleration probability at each moment.

[0097] According to some optional embodiments of this application, the fusion of displacement features and pressure features in step S103 to obtain cross-modal features can be achieved through the following steps: interactive fusion of displacement features and pressure features through two symmetrical and parallel encoding paths; wherein, the two symmetrical and parallel encoding paths include a first encoding path and a second encoding path; in the first encoding path, using displacement features as the main query vector and pressure features as the key value vector, the dependency relationship between displacement mode and pressure mode is extracted through a cross-modal attention mechanism to obtain a first fused feature; feature enhancement processing is performed on the first fused feature to obtain a first target fused feature; in the second encoding path, using pressure features as the main query vector and displacement features as the key value vector, the dependency relationship between pressure mode and displacement mode is extracted through a cross-modal attention mechanism to obtain a second fused feature; feature enhancement processing is performed on the second fused feature to obtain a second target fused feature; the first target fused feature and the second target fused feature are concatenated to obtain an initial cross-modal feature; the initial cross-modal feature is mapped to a preset representation space to obtain the cross-modal feature.

[0098] It is understandable that the fusion of displacement and pressure features in step S103 to obtain cross-modal features can be achieved through... Figure 2 The multimodal fusion encoder is implemented in [the code / platform]. Figure 3 This is a structural diagram of a multimodal fusion encoder according to an embodiment of this application, such as... Figure 3As shown, the multimodal fusion encoder first performs embedding processing on the landslide deformation time series and pore pressure time series, mapping them to a high-dimensional vector space, and introduces positional encoding to add unique positional identification information to the embedding vector of each time step.

[0099] Then, the encoder uses two symmetrical and complementary paths to perform cross-modal feature interaction and fusion.

[0100] The first path uses the embedded representation of landslide deformation as the main query vector and the embedded representation of pore pressure as the key and value vectors, which are then input into the first cross-modal fusion module. Figure 3 The Fusion Block above. This module first executes a cross-modal attention mechanism (Cross-Attn), which calculates the correlation weights between displacement and pressure features, enabling the displacement mode to adaptively focus on the feature segments in pore water pressure that are most closely related to its deformation response, thereby extracting the landslide response signal driven by hydrology. The feature sequence after cross-modal interaction is fed into the feedforward network within the module for nonlinear transformation to extract higher-order semantic features and generate a preliminary displacement-dominant feature sequence.

[0101] The initial displacement-dominant feature sequence is input into the first self-attention refining module. Figure 3 The Self-attnBlock module (above) includes a self-attention mechanism (Self-Attn) and a feedforward network: the self-attention mechanism captures the temporal dependencies within the displacement sequence, reinforcing the learning of long-term evolutionary patterns; the feedforward network further performs nonlinear mapping and dimensional transformation on the temporal features. After processing by this module, a first intermediate feature with rich feature representation is obtained.

[0102] The processing logic for the second path is symmetrical to that of the first path, but it uses pore water pressure embedding as the main query vector and landslide displacement embedding as the key and value vectors. This path sequentially passes through the second cross-modal fusion module ( Figure 3 The Fusion Block below), the second self-attention refining module ( Figure 3 The second intermediate feature is obtained from the Self-attn Block below.

[0103] Finally, the first and second intermediate features output from the two paths are concatenated along the channel dimension and compressed into a unified semantic representation space via a global projection layer to generate high-dimensional, bidirectionally aligned, and semantically rich cross-modal features.

[0104] Specifically, cross-attention operations are defined as follows:

[0105]

[0106]

[0107] in, Key matrix Dimensions This represents the pre-trained weight matrix across attention layers. Query AND value Originating from the dominant mode ,key Derived from auxiliary mode This allows the dominant mode to selectively focus on relevant features in the auxiliary modes.

[0108] In some optional embodiments of this application, step S104, which utilizes a landslide displacement prediction model based on a denoising diffusion probability network to estimate the disturbance component in the sample to be restored under the guidance of cross-modal features, and obtains the initial landslide displacement prediction value, can be achieved through the following steps: In the iterative denoising step of the landslide displacement prediction model based on a denoising diffusion probability network, the following sub-steps are executed: Modeling the temporal dependency within the sample to be restored in the current round using a self-attention mechanism; Extracting and predicting the trend component features and residual component features corresponding to the disturbance component in the sample to be restored through parallel sub-branches; Injecting cross-modal features as physical constraint signals into the sub-branches using a cross-attention mechanism to guide the denoising direction for the trend component features and residual component features; Dynamically adjusting the fusion ratio of the trend component features and residual component features using a gating unit to generate the intermediate prediction sample corresponding to the current iteration round; Repeating the iterative denoising step to gradually restore the real data in the sample to be restored through multiple iterations until the preset stopping condition is met, and determining the intermediate prediction sample obtained in the last iteration as the initial landslide displacement prediction value.

[0109] Specifically, through parallel sub-branches, the trend component features and residual component features corresponding to the perturbation components in the sample to be restored are extracted and predicted respectively, including the following sub-steps: through parallel first and second sub-branches, the trend component features and residual component features corresponding to the perturbation components in the sample to be restored are extracted and predicted respectively; in the first sub-branch, polynomial fitting is used to characterize the trend component features in the landslide movement; in the second sub-branch, a neural network is used to capture the residual component features between historical data and predicted data.

[0110] In this embodiment, it can be achieved through Figure 2 The denoising diffusion probability model decoder in the middle generates the initial landslide displacement prediction value. Figure 4 This is a structural diagram of a denoising diffusion probability model decoder according to an embodiment of this application. The following is in conjunction with... Figure 4 This section details how to generate initial landslide displacement predictions.

[0111] The decoder takes the cross-modal semantic features output by the multimodal fusion encoder as conditional input, and the landslide displacement time series initialized with Gaussian noise within the time period to be predicted as the initial noise input, forming the starting point of the denoising process. In each denoising iteration, it first performs stepwise optimization through a series of stacked denoising residual blocks. Each residual block includes a self-attention mechanism (self-attn) to model the long-term temporal dependencies within the displacement sequence, and uses a cross-attention mechanism (cross-attn) to use the cross-modal features as conditional queries to dynamically guide the attention weight of the current denoising step to key historical patterns, thereby achieving semantically driven noise correction.

[0112] The output of cross-attention undergoes a nonlinear transformation via a feed-forward network and is mapped to a unified feature space through a linear projection layer to ensure stable information transmission across the multi-layered structure. Building upon this, the decoder performs trend-residual decoupling through two parallel and structurally symmetrical sub-modules: the Trend Block employs a parameterized polynomial fitter, projecting the input features using low-order polynomial basis functions (e.g., 3rd order), outputting a smooth, monotonic, low-frequency trend component. Its parameters are constrained during training to only express irreversible cumulative deformation dominated by gravitational stress accumulation or soil creep, prohibiting it from capturing any high-frequency disturbances. Simultaneously, the Residual Block, through a deep feed-forward neural network, calculates the deviation from historical evolution patterns point-by-point and outputs a residual component specifically designed to characterize sudden instability behavior through nonlinear activation and multi-layer projection. This component is significantly enhanced only when triggered by preset mechanisms such as a sudden increase in pore water pressure, slip surface shear zone instability, or local rock mass slippage.

[0113] At the end of each iteration, the trend component and the residual component are added point by point to form the preliminary displacement prediction for the current iteration. This preliminary prediction result is then modulated element by element with the dynamic probability vector (ProbabilityVector) output by the accelerated probability estimation module to achieve physical constraints on the displacement amplitude. That is, when the pore water pressure is below the threshold, the prediction is suppressed; when the pressure reaches the critical state, the prediction is amplified, thus directly embedding the instability risk control mechanism during the generation process.

[0114] Finally, after several rounds of iterative denoising and semantic guidance, a clear displacement prediction sequence with both long-term trend stability and burst response accuracy is stably converged from the initial Gaussian noise, completing the complete generation process from noise to physically interpretable displacement.

[0115] Specifically, the features are projected onto two branches, each dedicated to predicting a single component. The trend branch captures low-frequency, smoothly varying signals in the landslide displacement and is parameterized using a multinomial regressor. ,in, This represents the index of the corresponding decoder module at diffusion step t. It is the decoder in the diffusion step Input for the time trend section, It is the first The average output of each decoder module. Slowly varying polynomial space. It is a vector / The power matrix, It is a smaller order (e.g.) =3), used to simulate low-frequency behavior.

[0116] The residual branch simulates high-frequency fluctuations in landslide displacement caused by other factors by capturing the nonlinear correlation between historical and predicted data. This component can be equivalently viewed as the output of a multilayer perceptron, expressed as: ,in It is the DDPM decoder in the diffusion step The input of the residual part at time, and These are branch parameters learned through deep learning pre-training.

[0117] Optionally, the trend component features and residual component features corresponding to the perturbation components in the sample to be restored are extracted and predicted respectively through the first and second sub-branches in parallel, including the following sub-steps:

[0118] In each iteration, the first sub-branch receives a global semantic guidance signal and adjusts the fitting strategy based on the global semantic guidance signal, so that the trend component features output by the first sub-branch change along the path of the landslide body's evolution on the geological time scale. The global semantic guidance signal is generated by semantically encoding the joint time-series features composed of deformation feature data and environmental stress data. The global semantic guidance signal is used to constrain the trend component features to only express irreversible cumulative deformation.

[0119] In each iteration, the second sub-branch calculates the deviation signal between the current prediction environment and the historical evolution pattern point by point, based on the noise residual of the current denoising step, the trend feature output by the first sub-branch in the previous iteration, and the contextual information of the historical observed displacement sequence in the training samples. The deviation signal is used to characterize the abnormal displacement changes that the trend component cannot explain.

[0120] The second sub-branch, based on the deviation signal, drives the neural network to perform dynamic weight updates in order to identify and enhance residual patterns related to the preset sudden mechanism, thereby obtaining residual component features that reflect sudden instability behavior. The neural network is used to capture the residual between historical data and predicted data. The preset sudden mechanism includes at least one of the following: a sudden increase in pore water pressure, instability of the shear zone of the sliding surface, or local rock mass slippage.

[0121] In this implementation, during each round of denoising iteration, the first sub-branch receives a global semantic guidance signal. This signal is an abstract expression learned by the encoder during long-term training, characterizing the geological evolution inertia of the landslide body. It includes the irreversible displacement trend of the landslide body over time, dominated by gravity stress accumulation, soil creep, or slow slippage of the basement. Instead of relying on end-to-end fitting in traditional neural networks, the first sub-branch uses this guidance signal as a static constraint, forcing its output trend component to increase slowly only within a very low rate of change. This physically restricts the displacement increment between adjacent time points from exceeding the natural upper limit of geological creep and suppresses any high-frequency oscillations or non-monotonic fluctuations. This ensures that the trend component truly reflects the intrinsic geological evolution behavior of the landslide body, rather than an overfitted response to short-term noise or hydrological disturbances.

[0122] Simultaneously, in each iteration, the second sub-branch synchronously receives the Gaussian noise residual introduced by the current denoising step, the trend component features output by the first sub-branch in the previous iteration, and the local context code extracted from the historical observed displacement sequence. These three elements together constitute its input. Based on these three sets of information, the second sub-branch calculates the difference between the current predicted environment and the historical evolution pattern point by point, generating a deviation signal. This deviation signal is used to capture anomalous displacement increments that the trend component fails to explain. Notably, anomalous displacement increments are non-stationary signals that are highly temporally aligned with physical events such as local structural instability of the landslide body, shear zone slippage, or sudden increases in pore water pressure.

[0123] Based on this, the neural network of the second sub-branch triggers an adaptive weight update mechanism according to the deviation signal. Through backpropagation, the network is guided to focus on residual patterns that are highly correlated with preset sudden mechanisms in the time series (such as pore water pressure rising above a threshold within three consecutive steps, displacement acceleration exceeding three times the geological background average, or synchronous with sudden shear strain changes on the sliding surface). The response intensity of these patterns is dynamically enhanced, so that the network no longer generalizes to fit the overall signal, but learns and reproduces the instantaneous acceleration behavior triggered by geological or hydrological changes. The final output only includes residual component features of sudden instability characteristics.

[0124] At the end of each iteration, the residual component features of the second sub-branch output are superimposed point by point onto the trend component features of the first sub-branch output to form the intermediate prediction sample of the current round. The intermediate prediction sample retains the smooth cumulative trend dominated by geological evolution and embeds the nonlinear perturbation component triggered by the sudden mechanism.

[0125] Intermediate prediction samples are used as input for the next round of denoising, participating in iterative optimization together with noise residuals. After completing all denoising steps, they are modulated element-wise with acceleration probabilities to generate a high-fidelity landslide displacement prediction sequence that conforms to long-term geological laws and can accurately respond to sudden instability events. This enables a physically interpretable and mechanism-driven dynamic reconstruction of the landslide evolution process without the need for manual event labeling.

[0126] In some optional embodiments of this application, a cross-attention mechanism is used to inject cross-modal features as physical constraint signals into sub-branches to guide the denoising direction for trend component features and residual component features. Specifically, this includes the following sub-steps: using the cross-attention mechanism, with the sample to be restored in the current round as the query vector and the cross-modal features as the key vector and value vector, the correlation weight distribution of the sample to be restored relative to the environmental stress data at each sampling time is calculated; based on the correlation weight distribution, the representation of the sample to be restored is weighted and corrected to adjust the denoising update direction of the sample to be restored; wherein, through weighted correction, the sample to be restored is constrained to maintain physical evolution consistency with the environmental stress data on the denoising trajectory, and the landslide displacement prediction model based on the denoising diffusion probability network is guided to generate a target evolution path that matches the changing pattern of the environmental stress data.

[0127] Understandably, the goal of the inverse denoising process is to learn how to gradually recover a noise-free data distribution from a Gaussian noise distribution. DDPM prediction can be viewed as, guided by condition y, recovering the data from a probability distribution... The process of sequential sampling. In each diffusion step t of the denoising process, This can be represented as a probability distribution of multiple independent denoising steps: The introduction of conditional guidance requires modification to the gradient update process during model training. This is achieved by applying Bayes' theorem to... The scoring function for gradient updates is given by equation (3). Here, the gradient originates from the forward process. Provided directly by DDPM; and This is provided by a predefined condition encoder, specifically the multimodal encoder in MG-DS in this invention.

[0128]

[0129] In summary, the training process can be interpreted as the model, guided by the encoder, gradually shifting the learned data distribution towards a higher likelihood region. During prediction, a complete sample is drawn from this distribution. This process can be formalized as follows:

[0130]

[0131] in, This represents the prediction result obtained after the model has completely removed noise. The hyperparameter γ is used to balance conditional consistency and generation smoothness. The gradient term can be interpreted as a guiding signal based on reconstruction, the strength of which is controlled by η. During the generation process, multiple gradient updates are performed at each diffusion step to improve control quality. The prediction is refined through multiple loops, represented as... from get After all denoising steps are completed, the final prediction is generated. .

[0132] As some alternative embodiments of this application, step S101, obtaining training samples, can be achieved through the following steps: obtaining original training samples; removing data segments with consecutive missing values ​​exceeding a preset threshold from the original training samples to obtain a first training sample; filling in the missing sampling points in the first training sample to obtain a second training sample; identifying and removing outlier sampling points in the second training sample to obtain a third training sample; performing denoising processing on the third training sample to obtain a fourth training sample; and resampling the deformation feature data and environmental stress data in the fourth training sample at the same multiple sampling times to obtain training samples.

[0133] In this embodiment, the original long-term training samples collected by the original landslide monitoring system are first obtained. These samples include continuously recorded surface displacement (deformation characteristic data) of the landslide body and pore water pressure (environmental stress data) buried inside the landslide body. The sampling frequencies of the two may differ, and due to factors such as sensor failure, communication interruption or environmental interference, the data generally contains missing, drift and noise pollution.

[0134] To ensure the integrity and reliability of the training data, the original training samples are first screened for missing values. All data segments with more than a preset threshold of consecutive missing sampling points (e.g., missing for 5 or more consecutive time steps) are removed. Such segments are removed because they have lost a lot of information and cannot provide effective temporal dependencies. Only segments with a sufficient length of continuous observations are retained to form the first training sample.

[0135] For non-continuous and sporadic missing sampling points in the first training sample, linear interpolation is used to fill them in. This method estimates the missing values ​​based on the linear relationship between adjacent valid observations on the time axis, ensuring the continuity of the time series, while avoiding the introduction of non-physical abrupt changes or high-frequency artifacts, thereby obtaining the second training sample.

[0136] To eliminate outliers (i.e., outlier sampling points) caused by instrument errors, lightning interference, or human error, a Hampel filter is used for outlier detection and removal. This method is based on statistical judgment of the median and absolute median difference (MAD) within a sliding window. When a sampling point deviates from the local median by more than a preset multiple (e.g., 3 times MAD), it is identified as an outlier and replaced with the median in the window. This preserves the local structure of the signal while suppressing extreme noise, thus obtaining a third training sample.

[0137] To further reduce high-frequency random noise and inherent measurement errors in the data, wavelet denoising was performed on the third training sample. For example, the db4 wavelet basis function was selected to perform multi-scale decomposition of the displacement and water pressure sequences, the high-frequency wavelet coefficients were suppressed by the threshold shrinkage algorithm, and the signal was reconstructed to achieve a balance between noise suppression and feature preservation, thus generating the fourth training sample.

[0138] Because the sampling frequencies of deformation features and environmental stress data in the original data are inconsistent (e.g., displacement is recorded daily, while water pressure is recorded hourly), in order to meet the synchronization requirements of the model input, the two sets of modal data in the fourth training sample are resampled according to a unified time axis. The average aggregation method is used to downsample the high-frequency data (e.g., hourly water pressure) to the same sampling interval as the low-frequency data (e.g., daily displacement) (e.g., one point every 3 days), ensuring that the displacement and water pressure data correspond one-to-one at each time stamp, forming the final training sample that is structurally regular, noise-suppressed, missing data repaired, and physically consistent.

[0139] Figure 5 This is a flowchart of a landslide displacement prediction method according to an embodiment of this application, such as... Figure 5 As shown, the method includes the following steps:

[0140] Step S501: Obtain landslide monitoring data.

[0141] Step S502: The landslide monitoring data is analyzed using a landslide displacement prediction model based on a denoised diffusion probability network to obtain predicted landslide displacement values. The landslide displacement prediction model based on the denoised diffusion probability network is... Figure 1 The landslide displacement prediction model shown is trained using the training method of the denoised diffusion probability network.

[0142] The specific implementation steps of the landslide displacement prediction method are explained in detail below.

[0143] Step S1, Data Preprocessing. The original monitoring dataset is preprocessed as follows: First, data segments containing a large number of consecutive missing values ​​are removed; then, linear interpolation is used to complete occasional missing data points; next, Hampel filters are used to remove outliers; finally, wavelet denoising methods are applied to suppress noise in the data.

[0144] Step S2, Data Alignment and Resampling. The pore water pressure data and landslide displacement data are aligned to the same time resolution. An average resampling method is used to ensure a one-to-one correspondence between the two types of data at each time point. The sampling interval is set according to specific research needs, typically 1 day or 3 days.

[0145] Step S3, Dataset Construction and Partitioning. The time series data is converted into fixed-length data slices using the sliding window method. Each slice contains two parts: the first part is the input window, representing the historical sequences known to the model; the second part is the prediction window, representing the future sequences to be predicted, whose ground truth values ​​will be used for model training and evaluation. All slices are divided into training and test sets according to a preset ratio, and normalization is performed on each set.

[0146] Step S4: Train the landslide acceleration probability estimation module. Using pore water pressure data from the training set as input, train a binary classification module. Its label (ground value) is defined based on the landslide displacement data: if the displacement exceeds 0.1 cm within a single sampling interval, it is marked as 1 (representing an acceleration event); otherwise, it is marked as 0. Training uses BCELoss as the loss function, and the module's weight parameters are saved after completion.

[0147] Step S5: Train the MG-DS model. Load and freeze the parameters of the landslide acceleration probability estimation module trained in the previous step. Using the pore water pressure and landslide displacement in the training set as input, train the MG-DS model for prediction, using L1 loss as the loss function. During training, Gaussian noise is gradually added to the prediction window portion of the input data. The model captures its dynamic characteristics by learning the difference in data distribution before and after noise addition.

[0148] Step S6, Model Prediction. During the testing phase, the test set data is input into the trained model. At this time, the data within the prediction window of the test samples is replaced with Gaussian noise. The model generates the final landslide displacement prediction sequence from the noise through an iterative denoising process.

[0149] Step S7, Interpretability Analysis. The output of the MG-DS model is used for interpretation analysis. This model can simultaneously provide the landslide acceleration probability and displacement trend components. These outputs help to understand the basis of the model's predictions and further reveal the intrinsic movement patterns of landslides.

[0150] To make the technical solution provided in this application more feasible, the Johnson Creek landslide and the Carmel Knoll landslide are used as examples to illustrate the technical solution provided in this application in detail below.

[0151] The Johnson Creek landslide (44.73°N, 124.05°W) and Carmel Knoll landslide (44.70°N, 124.05°W) are located near Beverly Beach State Park on the central coast of Oregon, with estimated volumes of approximately 1.5 × 10⁻⁶. 6 m³ and 6.5×10 5 The average surface slope angles are 7° and 8°, and the average depths are approximately 23 meters and 18 meters, respectively.

[0152] Figure 6 This is a schematic diagram of raw landslide monitoring data according to an embodiment of this application. Figure 6 The original monitoring data for the Johnson Creek and Carmel Knoll landslides are shown. Data for both landslides includes pore water pressure and landslide displacement data.

[0153] Field investigations indicate that both landslides are located within the Astoria Formation, which is primarily composed of lightly cemented layered sandstone sediments. The 1–20 mm thick base shear zones of both landslides consist of similarly highly weathered clayey silt, forming smooth sliding surfaces. Except for rotational movement at the toe, both landslides moved downhill along the subplanar shear zone. This rotational movement stabilized the landslides during movement through gravity, although wave action intermittently eroded some of the stabilizing toe material. Similar to many regions globally, in this part of the Oregon coast, increased pore water pressure due to precipitation infiltration has been identified as a major factor triggering landslide activity.

[0154] for Figure 6 The raw landslide monitoring data shown first requires preprocessing. Specifically, data filtering removes a large number of missing data segments. Then, linear interpolation, Hampel filtering, and wavelet denoising are used to improve data quality. Average resampling is applied, with both datasets resampled to a 3-day interval to reduce the impact of potential short-term errors.

[0155] Next, the training and testing sets were divided. The landslide monitoring data was constructed into data slices of a specified length using a sliding window. To capture the complete acceleration process, the input window length was set for each landslide: 82 time points for Johnson Creek and 96 time points for Carmel Knoll.

[0156] The prediction window was fixed at 36 time points. Model training was then conducted, first training the landslide acceleration probability module, freezing the module weights, and then training the entire model. After training, the model was used to perform predictions on a test set. Using the model weights obtained during training, pore water pressure data and landslide displacement data were input, the model was run to obtain the prediction results.

[0157] Test results for the Johnson Creek and Carmel Knoll landslides are as follows: Figure 7 As shown. Among them, Figure 7 (a) and (c) in the figure show the performance metrics of each model, including root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). Figure 7 (b) and (d) in the text are examples comparing the model prediction results.

[0158] The predictions of MG-DS were compared with those of baseline deep learning models. The baseline models used included temporal convolutional networks, Seq2Seq-LSTM, and Seq2Seq-Transformer. All experiments used monitoring data from the Johnson Creek and Carmel Knoll landslides. All baseline models used the same input-output window configuration as MG-DS.

[0159] Figure 7 Images (a) and (c) in this paper demonstrate the overall prediction performance of the MG-DS model provided in this application and the baseline model on Johnson Creek and Carmel Knoll, respectively. Overall, the MG-DS model provided in this application achieves better performance than the baseline model. Figure 7 Examples (b) and (d) in the table provide a comparative example of the predicted sequences. For the relatively regular Johnson Creek displacement evolution, all models reasonably predict the final cumulative displacement. In contrast, the Carmel Knoll model exhibits a more irregular and non-periodic displacement sequence, with a significant deviation between baseline predictions and observations. Figure 7 As can be seen from (d) in this application, although the MG-DS provided by this application still shows errors over the entire sequence, this application recovers the cumulative displacement more accurately and performs better, especially in motion timing prediction.

[0160] Figure 8 An example diagram of a trend component and acceleration probability component according to an embodiment of this application. Figure 8 This demonstrates the relationship between the trend component and acceleration probability component, incremental displacement, and pore water pressure, derived from the MG-DS model. (a), (b), and (c) are case studies of the Johnson Creek landslide. (d), (e), and (f) are case studies of the Carmel Knoll landslide.

[0161] Based on the comparative experiments described above, this embodiment further investigates how the trend component and acceleration probability component improve the interpretability of the model. The trend component aims to characterize an approximately linear displacement accumulation process driven by long-term geological factors (such as gravitational stress redistribution). The trend extracted by the model reflects the change in incremental displacement ( Figure 8 (a) and (d) in the figure show that MG-DS effectively learned the evolutionary patterns of landslides during training. The acceleration probability component aims to reflect the moderating effect of pore water pressure on landslide displacement by reflecting the differences in landslide sensitivity under changing hydrological conditions. Figure 8 As shown in (b) and (e), this component captures the intermittent behavior of landslide motion and characterizes the pore water pressure-displacement relationship for each landslide. By adjusting the predicted motion state, the acceleration probability component provides key constraints for reproducing the acceleration-deceleration process characteristic of hydrologically driven slow landslides.

[0162] Figure 9 A structural diagram of a training device for a landslide displacement prediction model based on a denoised diffusion probability network, according to an embodiment of this application, is shown below. Figure 9 As shown, the device includes:

[0163] The acquisition module 91 is used to acquire training samples, wherein the training samples include at least: deformation feature data for characterizing the geometric evolution of the landslide, and environmental stress data for characterizing the internal state of the landslide.

[0164] The first determining module 92 is used to determine the landslide acceleration probability based on the training samples.

[0165] The encoding module 93 is used to encode the deformation feature data and environmental stress data respectively to obtain displacement features and pressure features, and then fuse the displacement features and pressure features to obtain cross-modal features.

[0166] The estimation module 94 is used to superimpose the disturbance signal with the training sample to generate the sample to be restored, and use the landslide displacement prediction model based on the denoising diffusion probability network to estimate the disturbance component in the sample to be restored under the guidance of cross-modal features to obtain the initial landslide displacement prediction value.

[0167] The second determining module 95 is used to determine the predicted landslide displacement value based on the initial predicted landslide displacement value and the landslide acceleration probability.

[0168] The optimization module 96 is used to construct a loss function based on the deviation between the predicted landslide displacement value and the disturbance signal, and optimize the parameters of the landslide displacement prediction model based on the denoising diffusion probability network based on the loss function until the preset stopping condition is met.

[0169] Optionally, the landslide acceleration probability is determined based on the training samples, specifically including the following steps: Using a value-excited filtering layer, feature enhancement processing is performed on the environmental stress data to increase the weight of environmental stress data above a preset threshold in the feature space and decrease the weight of environmental stress data below a preset threshold in the feature space, resulting in first environmental stress data. The feature space is used to characterize the evolution of preset feature components in the environmental stress data. A multi-scale feature extraction module is used to process the first environmental stress data to obtain temporal dynamic features at different scales. The multi-scale feature extraction module includes at least two parallel convolutional branches, each using a different kernel size. The temporal dynamic features at different scales are concatenated to obtain multi-scale fused features. A convolutional layer with channel attention is used to process the multi-scale fused features to adaptively allocate weights for features at different scales, resulting in attention-weighted features. An activation function is used to process the attention-weighted features to obtain the landslide acceleration probability at each sampling time.

[0170] Optionally, displacement features and pressure features are fused to obtain cross-modal features, specifically including the following steps: Displacement features and pressure features are interactively fused through two symmetrical and parallel encoding paths; wherein the two symmetrical and parallel encoding paths include a first encoding path and a second encoding path; in the first encoding path, displacement features are used as the main query vector and pressure features as the key value vector, and the dependency relationship between displacement mode and pressure mode is extracted through a cross-modal attention mechanism to obtain a first fused feature; the first fused feature is subjected to feature enhancement processing to obtain a first target fused feature; in the second encoding path, pressure features are used as the main query vector and displacement features as the key value vector, and the dependency relationship between pressure mode and displacement mode is extracted through a cross-modal attention mechanism to obtain a second fused feature; the second fused feature is subjected to feature enhancement processing to obtain a second target fused feature; the first target fused feature and the second target fused feature are concatenated to obtain an initial cross-modal feature; the initial cross-modal feature is mapped to a pre-set representation space to obtain the cross-modal feature.

[0171] Optionally, a landslide displacement prediction model based on a denoising diffusion probability network is used to estimate the disturbance component in the sample to be restored under the guidance of cross-modal features, thereby obtaining the initial landslide displacement prediction value. Specifically, this includes the following steps: In the iterative denoising step of the landslide displacement prediction model based on a denoising diffusion probability network, the following sub-steps are executed: The temporal dependency within the sample to be restored in the current iteration is modeled using a self-attention mechanism; trend component features and residual component features corresponding to the disturbance component in the sample to be restored are extracted and predicted respectively through parallel sub-branches; cross-modal features are injected into the sub-branches as physical constraint signals using a cross-attention mechanism to guide the denoising direction for the trend component features and residual component features; the fusion ratio of trend component features and residual component features is dynamically adjusted using a gating unit to generate the intermediate prediction sample corresponding to the current iteration; the iterative denoising step is repeated to gradually restore the real data in the sample to be restored through multiple iterations until the preset stopping condition is met, and the intermediate prediction sample obtained in the last iteration is determined as the initial landslide displacement prediction value.

[0172] Optionally, by using parallel sub-branches, the trend component features and residual component features corresponding to the perturbation components in the sample to be restored are extracted and predicted respectively. Specifically, this includes the following steps: by using parallel first and second sub-branches, the trend component features and residual component features corresponding to the perturbation components in the sample to be restored are extracted and predicted respectively; in the first sub-branch, polynomial fitting is used to characterize the trend component features in the landslide movement; in the second sub-branch, a neural network is used to capture the residual component features between historical data and predicted data.

[0173] Optionally, a cross-attention mechanism is used to inject cross-modal features as physical constraint signals into sub-branches to guide the denoising direction for trend component features and residual component features. Specifically, this includes the following steps: using the cross-attention mechanism, with the sample to be restored in the current round as the query vector and the cross-modal features as the key vector and value vector, the correlation weight distribution of the sample to be restored relative to the environmental stress data at each sampling time is calculated; based on the correlation weight distribution, the representation of the sample to be restored is weighted and corrected to adjust the denoising update direction of the sample to be restored; wherein, through weighted correction, the sample to be restored is constrained to maintain physical evolution consistency with the environmental stress data on the denoising trajectory, and the landslide displacement prediction model based on the denoising diffusion probability network is guided to generate a target evolution path that matches the changing pattern of the environmental stress data.

[0174] Optionally, obtaining training samples specifically includes the following steps: obtaining original training samples; removing data segments with consecutive missing values ​​exceeding a preset threshold from the original training samples to obtain a first training sample; filling in missing sampling points in the first training sample to obtain a second training sample; identifying and removing outlier sampling points in the second training sample to obtain a third training sample; denoising the third training sample to obtain a fourth training sample; and resampling the deformation feature data and environmental stress data in the fourth training sample at the same multiple sampling times to obtain training samples.

[0175] It should be noted that the above Figure 9 The modules in can be program modules (e.g., a set of program instructions that implements a specific function) or hardware modules. For the latter, they can be represented in the following forms, but are not limited to these: each of the above modules is represented by a processor, or the functions of each of the above modules are implemented by a processor.

[0176] It should be noted that, Figure 9 Preferred embodiments of the shown examples can be found in [reference needed]. Figure 1 The relevant descriptions of the embodiments shown will not be repeated here.

[0177] Figure 10 A hardware block diagram of a computer terminal for implementing a training method for a landslide displacement prediction model based on a denoising diffusion probability network is shown. Figure 10 As shown, the computer terminal 100 may include one or more processors (shown as 1002a, 1002b, ..., 1002n in the figure) (the processor may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 1004 for storing data, and a transmission module 10010 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 10 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, the computer terminal 100 may also include... Figure 10 The more or fewer components shown, or having the same Figure 10 The different configurations shown.

[0178] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be implemented wholly or partially as software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be wholly or partially integrated into any other element in the computer terminal 100. As involved in the embodiments of this application, the data processing circuit serves as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0179] The memory 1004 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the training method of the landslide displacement prediction model based on a denoised diffusion probability network in this embodiment of the application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 1004, thereby realizing the above-mentioned training method of the landslide displacement prediction model based on a denoised diffusion probability network. The memory 1004 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1004 may further include memory remotely located relative to the processor, and these remote memories can be connected to the computer terminal 100 via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0180] The transmission module 10010 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 100. In one example, the transmission module 10010 includes a network interface controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 10010 may be a radio frequency (RF) module, used for data interaction with external networks or devices via wired and / or wireless network connections.

[0181] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows the user to interact with the user interface of the computer terminal 100. Furthermore, the computer terminal 100 can be connected to a cursor control device and a keyboard via an input / output interface. The cursor control device (e.g., a mouse, trackball, or touchpad) is primarily used to transmit directional information and command selections to the terminal and control the movement of the cursor on the display; the keyboard is primarily used to receive letters, numbers, and other control commands input by the user, thereby enabling interaction between the user and the computer terminal 100.

[0182] It should be noted here that, in some optional embodiments, the above... Figure 10 The computer terminal shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 10 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computer terminal.

[0183] This application also provides a non-volatile storage medium, which includes a stored program. When the program runs, it controls the device where the storage medium is located to execute the above-described training method for the landslide displacement prediction model based on a denoising diffusion probability network.

[0184] A non-volatile storage medium performs the following functions: identifying the discrepancies between signals, constructing a loss function, and optimizing the parameters of a landslide displacement prediction model based on a denoising diffusion probability network based on the loss function until a preset stopping condition is met to obtain training samples. The training samples include at least: deformation feature data characterizing the geometric evolution of the landslide, and environmental stress data characterizing the internal state of the landslide; determining the landslide acceleration probability based on the training samples; encoding the deformation feature data and environmental stress data separately to obtain displacement features and pressure features, and fusing the displacement features and pressure features to obtain cross-modal features; superimposing the disturbance signal with the training samples to generate a sample to be restored, and using the landslide displacement prediction model based on a denoising diffusion probability network, guided by the cross-modal features, estimating the disturbance component in the sample to be restored to obtain an initial landslide displacement prediction value; determining the landslide displacement prediction value based on the initial landslide displacement prediction value and the landslide acceleration probability; and determining the landslide displacement prediction value based on the landslide displacement prediction value and the disturbance.

[0185] This application also provides an electronic device, including: a memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program executes the above-described training method for a landslide displacement prediction model based on a denoising diffusion probability network.

[0186] The processor runs a program that performs the following functions: identifying the discrepancy between signals, constructing a loss function, and optimizing the parameters of a landslide displacement prediction model based on a denoising diffusion probability network based on the loss function until a preset stopping condition is met to obtain training samples. The training samples include at least: deformation feature data characterizing the geometric evolution of the landslide, and environmental stress data characterizing the internal state of the landslide; determining the landslide acceleration probability based on the training samples; encoding the deformation feature data and environmental stress data respectively to obtain displacement features and pressure features, and fusing the displacement features and pressure features to obtain cross-modal features; superimposing the disturbance signal with the training samples to generate a sample to be restored, and using the landslide displacement prediction model based on a denoising diffusion probability network, guided by the cross-modal features, estimating the disturbance component in the sample to be restored to obtain an initial landslide displacement prediction value; determining the landslide displacement prediction value based on the initial landslide displacement prediction value and the landslide acceleration probability; and determining the landslide displacement prediction value based on the landslide displacement prediction value and the disturbance.

[0187] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0188] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0189] In the above embodiments of this application, the information collected is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with relevant laws, regulations and standards, take necessary protective measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0190] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0191] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0192] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0193] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0194] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A training method for a landslide displacement prediction model based on a denoised diffusion probability network, characterized in that, include: Acquire training samples, wherein the training samples include at least: deformation feature data for characterizing the geometric evolution of landslides and environmental stress data for characterizing the internal state of landslides, wherein the deformation feature data includes at least landslide displacement data and the environmental stress data includes at least pore pressure data. Based on the training samples, determine the probability of landslide acceleration; The deformation feature data and environmental stress data are encoded separately to obtain displacement features and pressure features, and the displacement features and pressure features are fused to obtain cross-modal features; The disturbance signal is superimposed on the training sample to generate the sample to be restored. Then, the disturbance component in the sample to be restored is estimated by using the landslide displacement prediction model based on the denoising diffusion probability network under the guidance of the cross-modal features, so as to obtain the initial landslide displacement prediction value. Determining the landslide displacement prediction value based on the initial landslide displacement prediction value and the landslide acceleration probability includes: multiplying the initial landslide displacement prediction value and the landslide acceleration probability element by element to obtain the landslide displacement prediction value. Based on the deviation between the predicted landslide displacement and the disturbance signal, a loss function is constructed, and based on the loss function, the parameters of the landslide displacement prediction model based on the denoising diffusion probability network are optimized until a preset stopping condition is met. Using the landslide displacement prediction model based on the denoising diffusion probability network, guided by the cross-modal features, the disturbance component in the sample to be restored is estimated to obtain the initial landslide displacement prediction value. This includes: in the iterative denoising step of the landslide displacement prediction model based on the denoising diffusion probability network, the following sub-steps are executed: Modeling the temporal dependency within the sample to be restored in the current iteration using a self-attention mechanism; extracting and predicting the trend component features and residual component features corresponding to the disturbance component in the sample to be restored through parallel sub-branches; injecting the cross-modal features as physical constraint signals into the sub-branches using a cross-attention mechanism to guide the denoising direction for the trend component features and the residual component features; dynamically adjusting the fusion ratio of the trend component features and the residual component features using a gating unit to generate an intermediate prediction sample corresponding to the current iteration; repeatedly executing the iterative denoising step to gradually restore the real data in the sample to be restored through multiple iterations until a preset stopping condition is met, and determining the intermediate prediction sample obtained in the last iteration as the initial landslide displacement prediction value.

2. The method according to claim 1, characterized in that, Based on the training samples, the probability of landslide acceleration is determined, including: The environmental stress data is enhanced by using a value-excited filter layer to increase the weight of environmental stress data above a preset threshold in the feature space and decrease the weight of environmental stress data below the preset threshold in the feature space to obtain first environmental stress data. The feature space is used to characterize the evolution law of preset feature components in the environmental stress data. The first environmental stress data is processed using a multi-scale feature extraction module to obtain time dynamic features at different scales. The multi-scale feature extraction module includes at least two parallel convolutional branches, and different convolutional branches use different convolutional kernel sizes. The time dynamic features at different scales are spliced ​​together to obtain multi-scale fused features; The multi-scale fused features are processed using a convolutional layer with channel attention mechanism to adaptively allocate weights to features of different scales, resulting in attention-weighted features. The attention-weighted features are processed by an activation function to obtain the landslide acceleration probability at each sampling time.

3. The method according to claim 1, characterized in that, The displacement feature and the pressure feature are fused to obtain cross-modal features, including: The displacement feature and the pressure feature are interactively fused through two symmetrical and parallel encoding paths; wherein the two symmetrical and parallel encoding paths include a first encoding path and a second encoding path. In the first encoding path, the displacement feature is used as the main query vector and the pressure feature is used as the key vector. The dependency relationship between the displacement mode and the pressure mode is extracted through a cross-modal attention mechanism to obtain the first fused feature. The first fused feature is then subjected to feature enhancement processing to obtain the first target fused feature. In the second encoding path, the pressure feature is used as the main query vector and the displacement feature is used as the key vector. The dependency relationship between the pressure mode and the displacement mode is extracted through a cross-modal attention mechanism to obtain the second fusion feature. The second fusion feature is then subjected to feature enhancement processing to obtain the second target fusion feature. The first target fusion feature and the second target fusion feature are concatenated to obtain the initial cross-modal feature; The initial cross-modal features are mapped to a pre-defined representation space to obtain the cross-modal features.

4. The method according to claim 1, characterized in that, Through parallel sub-branches, the trend component features and residual component features corresponding to the perturbation components in the sample to be restored are extracted and predicted respectively, including: Through parallel first and second sub-branches, the trend component features and residual component features corresponding to the perturbation components in the sample to be restored are extracted and predicted, respectively. In the first sub-branch, the trend component characteristics of the landslide motion are characterized by polynomial fitting. In the second sub-branch, residual component features between historical data and predicted data are captured using a neural network.

5. The method according to claim 1, characterized in that, By utilizing a cross-attention mechanism, the cross-modal features are injected as physical constraint signals into the sub-branch to guide the denoising direction for the trend component features and the residual component features, including: Using the cross-attention mechanism, the sample to be restored in the current round is used as the query vector, and the cross-modal features are used as the key vector and value vector to calculate the association weight distribution of the sample to be restored with respect to the environmental stress data at each sampling time. Based on the associated weight distribution, the representation of the sample to be restored is weighted and corrected to adjust the denoising update direction of the sample to be restored; wherein, through the weighted correction, the sample to be restored is constrained to maintain physical evolution consistency with the environmental stress data on the denoising trajectory, and the landslide displacement prediction model based on the denoising diffusion probability network is guided to generate a target evolution path that matches the changing pattern of the environmental stress data.

6. The method according to claim 1, characterized in that, Obtain training samples, including: Obtain the original training samples; The first training sample is obtained by removing data segments with consecutive missing values ​​exceeding a preset threshold from the original training sample. The missing sampling points in the first training sample are filled in to obtain the second training sample; Outlier sampling points in the second training sample are identified and removed to obtain the third training sample; The third training sample is denoised to obtain the fourth training sample; The deformation feature data and environmental stress data in the fourth training sample are resampled at the same multiple sampling times to obtain the training sample.

7. A landslide displacement prediction method, characterized in that, include: Obtain landslide monitoring data; The landslide monitoring data is analyzed using a landslide displacement prediction model based on a denoised diffusion probability network to obtain predicted landslide displacement values. The landslide displacement prediction model based on a denoised diffusion probability network is obtained by training the landslide displacement prediction model based on a denoised diffusion probability network as described in any one of claims 1 to 6.

8. A training device for a landslide displacement prediction model based on a denoised diffusion probability network, characterized in that, include: An acquisition module is used to acquire training samples, wherein the training samples include at least: deformation feature data for characterizing the geometric evolution of landslides and environmental stress data for characterizing the internal state of landslides, wherein the deformation feature data includes at least landslide displacement data and the environmental stress data includes at least pore pressure data. The first determining module is used to determine the landslide acceleration probability based on the training samples; The encoding module is used to encode the deformation feature data and environmental stress data respectively to obtain displacement features and pressure features, and to fuse the displacement features and pressure features to obtain cross-modal features; The estimation module is used to superimpose the disturbance signal with the training sample to generate the sample to be restored, and use the landslide displacement prediction model based on the denoising diffusion probability network to estimate the disturbance component in the sample to be restored under the guidance of the cross-modal features to obtain the initial landslide displacement prediction value. The second determining module is used to determine the landslide displacement prediction value based on the initial landslide displacement prediction value and the landslide acceleration probability, including: multiplying the initial landslide displacement prediction value and the landslide acceleration probability element by element to obtain the landslide displacement prediction value. The optimization module is used to construct a loss function based on the deviation between the predicted landslide displacement value and the disturbance signal, and optimize the parameters of the landslide displacement prediction model based on the denoising diffusion probability network based on the loss function until a preset stopping condition is met. The estimation module is further configured to perform the following steps: in the iterative denoising step of the landslide displacement prediction model based on the denoising diffusion probability network, the following sub-steps are performed: modeling the temporal dependency within the sample to be restored in the current round using a self-attention mechanism; extracting and predicting the trend component features and residual component features corresponding to the disturbance components in the sample to be restored through parallel sub-branches; injecting the cross-modal features as physical constraint signals into the sub-branches using a cross-attention mechanism to guide the denoising direction for the trend component features and the residual component features; dynamically adjusting the fusion ratio of the trend component features and the residual component features using a gating unit to generate intermediate prediction samples corresponding to the current iteration round; repeating the iterative denoising step to gradually restore the real data in the sample to be restored through multiple iterations until a preset stopping condition is met, and determining the intermediate prediction sample obtained in the last iteration as the initial landslide displacement prediction value.

9. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein, when the program is executed, it controls the device where the non-volatile storage medium is located to execute the training method of the landslide displacement prediction model based on a denoised diffusion probability network as described in any one of claims 1 to 6.

10. An electronic device, characterized in that, include: A memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the training method for a landslide displacement prediction model based on a denoised diffusion probability network as described in any one of claims 1 to 6.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method of the landslide displacement prediction model based on the denoising diffusion probability network as described in any one of claims 1 to 6.

Images