Deep Learning-Based Lattice-Based Landslide Micromotion Monitoring and Early Warning Method and Equipment

Abstract

This invention discloses a deep learning-based lattice-type landslide micro-motion monitoring and early warning method and device, comprising the following steps: a large-area lattice-type embedded terminal device is deployed to form a sensor network; the terminal performs sleep monitoring, automatically waking up and collecting raw sensor data upon detecting vibration and tilt changes; the terminal uploads the raw sensor data to a centralized acquisition device via a wireless communication module; the centralized acquisition device aggregates and preprocesses the data, generating landslide micro-motion data, an external field input set, and a topology initialization parameter set; the data processing and analysis system calls an improved CCNN network for comprehensive evaluation, generating an early warning level, a target area index, and trigger commands; if the comprehensive evaluation results indicate landslide instability and increased displacement, an early warning message is issued. This invention improves the accuracy and reliability of landslide early warning.

Description

Technical Field

[0001] This invention relates to the field of landslide disaster monitoring, and in particular to a method and equipment for lattice-based landslide micro-motion monitoring and early warning based on deep learning. Background Technology

[0002] Landslides are one of the most widespread and serious geological hazards. To identify potential landslide hazards in advance, existing methods mostly rely on deploying fixed monitoring stations or inclinometer boreholes to collect data such as displacement, vibration, and pore water pressure. These methods have limited coverage and cannot reflect the overall stability of the landslide body in a timely and comprehensive manner. Furthermore, surface monitoring devices are susceptible to wind and rain erosion or human-caused damage, resulting in insufficient long-term operational reliability.

[0003] At the data processing level, traditional methods often employ statistical models or shallow machine learning algorithms, relying solely on single-point or low-dimensional features for threshold determination. This approach is insufficient for handling multi-source inputs in complex environments, frequently resulting in high false alarm rates, delayed warnings, or missed warnings. Furthermore, the spatial topological relationships between monitoring points and the influence of terrain are often not adequately considered, making it difficult to characterize the correlations between points, edges, and surfaces within the landslide body, leading to inaccurate reflections of landslide evolution mechanisms.

[0004] Furthermore, existing methods generally rely on fixed thresholds or simple redundant verification mechanisms when dealing with uncertainties, failing to dynamically adapt to rapid changes in external conditions such as rainfall intensity and pore water pressure. This results in insufficient stability and real-time performance of early warning results. These shortcomings limit the effectiveness of landslide disaster monitoring technology in complex environments and highlight the urgent need for more accurate and robust early warning methods.

[0005] Therefore, how to provide a deep learning-based lattice-type landslide micro-motion monitoring and early warning method and equipment is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based lattice-based method and device for monitoring and early warning of landslide micro-motions. This invention collects multi-source data such as displacement, velocity, acceleration, and pore water pressure by burying terminal devices over a large area. It combines point coordinates and topographic information to generate landslide micro-motion data, an external field input set, and a topology initialization parameter set. An improved CCNN network is constructed to realize unit complex modeling, spatiotemporal physical gating transmission, physical consistency constraints, and multi-scale uncertainty early warning. It has the advantages of high early warning accuracy, strong real-time performance, and good stability.

[0007] The deep learning-based lattice-type landslide micro-motion monitoring and early warning method according to embodiments of the present invention includes the following steps:

[0008] According to the pre-designed grid and dot matrix scheme, multiple terminal devices are buried in the underground of the landslide body and potential landslide area in a large-area dot matrix pattern, and the antennas are brought out to the surface.

[0009] Under normal circumstances, the terminal device is in a low-power sleep mode and performs sleep monitoring. When vibration or tilt changes are detected to exceed a preset threshold, it will automatically wake up and collect raw sensor data.

[0010] The terminal device transmits the raw sensor data to the centralized data acquisition device via a wireless communication module to complete data aggregation;

[0011] The centralized acquisition equipment completes the aggregation of raw sensor data and performs data preprocessing to generate landslide micro-motion data, external field input set and topology initialization parameter set;

[0012] The data processing and analysis system, based on landslide micro-motion data, external field input set and topology initialization parameter set, calls an improved CCNN network to comprehensively evaluate the stability of the landslide body and generate early warning level, target area index and trigger command;

[0013] If the analysis results show that the landslide is unstable and the displacement is intensifying, an early warning will be issued to the relevant authorities and the public.

[0014] Optionally, the generation of the landslide micro-motion data, the external field input set, and the topology initialization parameter set includes:

[0015] Based on the collected raw sensor data, the time alignment, unit unification and range verification were completed by combining the point coordinates and terrain data, and missing sections and abnormal points were eliminated.

[0016] The point coordinates refer to the spatial position parameters of the terminal devices deployed within the monitoring area;

[0017] The terrain data refers to parameters describing the terrain morphology within the monitored area;

[0018] Based on the time-aligned displacement, velocity, acceleration, and pore water pressure, noise reduction, interpolation, amplitude normalization, and scale standardization are performed. The data are organized into a tensor structure according to the terminal device number and acquisition time to generate landslide micromotion data.

[0019] Displacement, velocity, acceleration, pore water pressure, and rainfall are synchronously packaged into a unified input structure according to the acquisition time, while retaining the units of physical quantities and sampling interval markings to generate an external field input set;

[0020] The rainfall data was collected by the backend processing center;

[0021] Based on the location coordinates and terrain data, the adjacency relationship between monitoring points is calculated. An adjacency matrix is ​​generated according to the preset neighborhood radius and the number of nearest neighbors. Edge weight parameters and surface weight parameters are generated by combining the terrain elevation difference and spatial distance, and a set of topology initialization parameters is generated.

[0022] Optionally, the invocation of the improved CCNN network includes:

[0023] An improved CCNN network is constructed, comprising: a unit complex construction unit, a spatiotemporal physical gated message passing unit, a physical consistency constraint unit, and a multi-scale uncertainty early warning unit;

[0024] In the element complex construction unit, the element complex structure is established based on the landslide micro-motion data and the set of topological initialization parameters, and the element complex input structure is generated.

[0025] In the spatiotemporal physical gated message passing unit, the receiving unit's complex input structure and external field input set construct temporal and physical gates, perform cross-dimensional message passing and feature updates, and generate multi-layer intermediate feature representations;

[0026] In the physical consistency constraint unit, a multi-layer intermediate feature representation and an external field input set are received. Based on the Mohr-Coulomb criterion and the preset energy constraint, the constraint residual set is calculated to generate a physically consistent feature representation.

[0027] In the multi-scale uncertainty early warning unit, the physical uniform feature representation is received, confidence gating is performed and multi-scale unit complex aggregation is carried out, and the early warning level, target area index and trigger command are generated by combining the constraint residual set.

[0028] Optionally, the generation of the unit complex input structure includes:

[0029] In the unit complex construction unit, the landslide micro-motion data and the set of topology initialization parameters are read, and an index mapping is established according to the sensor number and time step to generate an aligned data index.

[0030] Based on the point coordinates and alignment data index, a set of point cells is generated, and the adjacency matrix in the topology initialization parameter set is read to form adjacency relationship data;

[0031] Based on adjacency data and terrain data, and with the addition of edge weight parameters and surface weight parameters, a set of weighted edge cells and a set of surface cells are generated, the directed relationships are determined, and the boundary relationships and co-boundary relationships are generated.

[0032] Record the set of face unit vertices composed of point units for each face unit;

[0033] The edge weight parameters characterize the spatial and terrain relationships between point pairs;

[0034] The area weight parameters characterize the overall geometric and topographic features of the region;

[0035] Node features are extracted from landslide micro-motion data, edge features are calculated by weighting the adjacency relationship data and edge weight parameters, and surface features are generated by aggregating the vertex set of surface units and surface weight parameters.

[0036] The set of point cells, edge cells, face cells, adjacency data, boundary relationships, co-boundary relationships, edge weight parameters, face weight parameters, as well as node features, edge features, and face features are summarized and packaged to generate a complex cell input structure.

[0037] Optionally, the generation of the multi-layer intermediate feature representation includes:

[0038] In the spatiotemporal physical gated message passing unit, the receiving unit has a complex input structure and an external field input set;

[0039] Time-gating and physical-gating are constructed. Time-gating is based on the time step data in the unit complex input structure and the external field input set to establish a dynamic adjustment function to control the update rate of features at different time levels. Physical-gating is based on the rainfall and pore water pressure in the external field input set to establish a threshold control function to adjust the propagation intensity of features under external field driving conditions.

[0040] Cross-dimensional message passing and feature updates are performed. The node features of point units are passed to edge units through adjacency relationships and weighted by edge weight parameters. The edge features of edge units are passed to surface units through boundary relationships and aggregated by surface weight parameters. At the same time, the surface features of surface units are fed back to edge units and point units through co-boundary relationships, forming a cross-dimensional feature flow. In each passing process, time gating and physical gating are introduced as adjustment factors to control the feature update amplitude.

[0041] In the process of cross-dimensional message passing, an update function is defined to perform time-gated and physical-gated weighted aggregation on the features of the previous layer of each dimension unit to generate the intermediate feature representation of the next layer, and the intermediate feature representations of each dimension unit are stacked hierarchically to form a multi-layer intermediate feature representation;

[0042] The intermediate feature representations of each level are stacked to generate multi-level intermediate feature representations.

[0043] Optionally, the update function is defined as:

[0044] ;

[0045] in, Representation unit In the The feature vector of the layer, Representation and Unit Adjacent units In the The feature vector of the layer, Representation unit The set of neighbors, Indicates the first The feature transformation matrix of the layer, Represents the time-gated function. Represents the physical gate function. This represents a non-linear activation function.

[0046] Optionally, the generation of the physically consistent feature representation includes:

[0047] In the physically consistent constraint unit, read the multi-layer intermediate feature representation and the external field input set;

[0048] Based on the edge features in the multi-layer intermediate feature representation, the predicted shear stress is calculated and a constraint function is constructed. The predicted shear stress of the edge element (i,j) is defined as follows:

[0049] ;

[0050] in, This represents the predicted shear stress value of the edge element (i,j) calculated at time step t. Represents the edge feature vector. Represents the dynamic mapping coefficient vector. This represents the transpose of the dynamic mapping coefficient vector;

[0051] Consistency correction is performed on the multi-layer intermediate feature representation with the constraint function as the objective, and residual-driven updates are performed on the unit feature vectors in each layer according to the gradient descent strategy to generate physically consistent feature representations.

[0052] Optionally, the constraint function is defined as:

[0053] ;

[0054] in, This represents the global constraint function value obtained at time step t after weighting and combining the Mohr-Coulomb constraint residuals and the energy constraint residuals. Representing edge cells At time step The predicted amount of shear stress, This represents the cohesive force parameter. This represents the internal friction angle parameter. Indicates effective and positive force. This represents the moving average of rainfall in the set of external inputs. Represents the energy mapping coefficient. Representing edge cells The edge feature vectors, Represents the energy proxy mapping coefficient vector. This represents the transpose of the energy proxy mapping coefficient vector. Indicates non-negative weights. This indicates taking the absolute value.

[0055] Optionally, the generation of the warning level, target area index, and trigger command includes:

[0056] In the multi-scale uncertainty early warning unit, the physical uniformity feature representation is received, and the node features, edge features and surface features are segmented and statistically analyzed in the time dimension to generate the mean estimation sequence and the variance estimation sequence.

[0057] Confidence gating is performed on the mean estimation sequence and variance estimation sequence. Based on the variance estimation sequence, the sample variance within the same time window is taken as the uncertainty level. The corresponding features are controlled according to the uncertainty level to generate a confidence calibration feature set.

[0058] The confidence calibration feature set is input into the multi-scale unit complex aggregation process. Point units, edge units and surface units are grouped and aggregated according to different spatial and temporal scales. During aggregation, the consistency of topological relationships is maintained and the composite feature vector across scales is calculated to generate a multi-scale aggregated feature representation.

[0059] By combining the constrained residual set with the multi-scale aggregated feature representation, a calculation rule for the composite early warning index is established. The physical consistency features of each spatial location and time period are weighted and superimposed to form a composite early warning index set. The early warning time window set is determined according to the index change trend.

[0060] Based on the set of composite early warning indices and the set of early warning time windows, early warning levels are generated according to the classification standards, and the target area index and triggering instructions are determined.

[0061] Optional, deep learning-based lattice-type landslide micro-motion monitoring and early warning equipment includes:

[0062] Terminal devices, centralized data acquisition equipment, and data processing and analysis systems;

[0063] The terminal device includes: a sensor unit for detecting micro-vibration signals and tilt change signals; a low-power processor and a data acquisition module for maintaining ultra-low power monitoring in sleep mode and automatically waking up and performing data acquisition when the micro-vibration amplitude and tilt angle increment exceed a preset threshold; a wireless communication module for uploading processed data to a centralized acquisition device when acquisition is triggered; and a power supply unit for supplying power to the system.

[0064] The centralized data acquisition device is used to automatically scan and poll the status of each terminal device, receive the data uploaded by each terminal, and perform summary processing.

[0065] The data processing and analysis system is used to analyze and issue early warnings based on the aggregated data.

[0066] The beneficial effects of this invention are:

[0067] First, by combining a dot-matrix array of sensors with deep learning modeling, this invention breaks through the limitation of the limited coverage of traditional single-point monitoring, and achieves a comprehensive characterization of the multi-level correlation features of landslide points, edges, and surfaces, thereby enabling earlier identification of potential risks.

[0068] Secondly, this invention introduces a spatiotemporal physical gating mechanism and the Mohr-Coulomb criterion constraint into deep learning networks, which solves the problems of insufficient response to external inputs and disconnection from geological mechanisms in existing methods, effectively improving the reliability and physical consistency of model predictions.

[0069] Furthermore, this invention utilizes uncertainty modeling and multi-scale aggregation mechanisms to significantly reduce the probability of false alarms and missed alarms, enhancing the stability and robustness of early warning results while maintaining real-time performance, thus providing a more accurate and efficient technical means for monitoring and preventing landslide disasters in complex environments. Attached Figure Description

[0070] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0071] Figure 1 This is an overall flowchart of the deep learning-based lattice-type landslide micro-motion monitoring and early warning method and equipment proposed in this invention;

[0072] Figure 2 This is a schematic diagram of the improved CCNN network structure in this invention;

[0073] Figure 3 This is a schematic diagram of the joint mechanism of physical consistency constraint and multi-scale uncertainty early warning in this invention. Detailed Implementation

[0074] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0075] refer to Figure 1-3 A deep learning-based lattice-type landslide micro-motion monitoring and early warning device includes:

[0076] Terminal devices, centralized data acquisition equipment, and data processing and analysis systems;

[0077] The terminal device is buried 0.5 to 1 meter underground in the monitoring area, and includes:

[0078] The sensor unit, including micro-vibration sensors and tilt sensors, is used to detect subtle vibrations, displacements, and tilt changes on the ground and underground in real time. When external disturbances or abnormal activities are detected, the sensors generate corresponding electrical signals.

[0079] The low-power processor and data acquisition module are used to maintain ultra-low power monitoring in sleep mode and automatically wake up and perform data acquisition when the micro-vibration amplitude and tilt angle increment exceed the preset threshold. They receive signals from the sensor unit and perform preliminary information processing. Once the trigger condition is met (e.g., the vibration amplitude exceeds the threshold), the processor sends a reporting signal to the wireless communication module to start the data transmission process.

[0080] The wireless communication module is used to upload the processed data to the centralized acquisition equipment when the acquisition is triggered. The terminal device is buried 0.5 to 1 meter underground. The antenna is brought out to the surface and has a waterproof and moisture-proof structure to ensure that the signal transmission is not severely interfered with by the underground environment.

[0081] The power supply unit adopts an ultra-low power consumption design, combined with a long-life battery to power the system, and the main body of the terminal is made of biodegradable materials;

[0082] The power supply unit further includes a micro-motion energy harvesting module and a solar auxiliary module, which are used to achieve energy self-supply in an environment where the terminal is buried underground for a long time.

[0083] The terminal of this device uses biodegradable materials to make the main shell. After the end of its service life or damage, it can be recycled in a unified manner to reduce the impact on the environment. Some electronic components of the terminal can be regenerated or centrally processed to reduce pollution to the ecological environment. In addition, when the terminal device is detected to have insufficient power, abnormal function or exceed its service life, it can be remotely marked and located through centralized data collection equipment to facilitate maintenance or replacement.

[0084] The centralized data acquisition device:

[0085] The centralized data acquisition equipment is placed in a safe location within the landslide monitoring area and is equipped with a highly sensitive wireless receiving module;

[0086] The device can automatically scan or poll the status of each terminal device. Once a data trigger signal is detected, it will start the centralized acquisition mode and receive the data uploaded by each terminal.

[0087] Data storage and processing modules: These modules summarize, denoise, and perform some data analysis on the collected multi-point data, and then upload the processed monitoring information to the remote monitoring center or cloud server.

[0088] The power supply can be either mains power or solar panels to ensure stable equipment operation.

[0089] The centralized data acquisition equipment also includes a remote marking function, which is used to automatically record and locate when the terminal device is found to have insufficient power or abnormal function, so as to facilitate remote maintenance and replacement.

[0090] The data processing and analysis system is used to analyze and provide early warnings for the aggregated data;

[0091] In this embodiment, the devices are connected via the following method:

[0092] Large-area dot matrix deployment: According to the pre-designed grid or dot matrix scheme, multiple terminal devices are buried 0.5 to 1 meter underground in the landslide body and potential landslide area, and the antennas are brought out to the surface;

[0093] Sleep monitoring and threshold-triggered data acquisition: Under normal circumstances, the terminal device is in a low-power sleep mode. When vibration or tilt changes are detected to exceed a preset threshold, it will automatically wake up and acquire raw sensor data.

[0094] Data aggregation: The terminal device sends the raw sensor data to the centralized data acquisition equipment via the wireless communication module;

[0095] Data preprocessing: The centralized acquisition equipment completes the aggregation of raw sensor data and performs data preprocessing to generate landslide micro-motion data, external field input set and topology initialization parameter set;

[0096] Landslide risk identification: The data processing and analysis system, based on landslide micro-motion data, external field input set and topology initialization parameter set, calls an improved CCNN network to comprehensively evaluate the stability of the landslide body and generate warning level, target area index and trigger command;

[0097] Warning: If the analysis results show that the landslide body is unstable and the displacement is intensifying, a warning message will be issued to the relevant authorities and the public.

[0098] In this embodiment, the generation of the landslide micro-motion data, the external field input set, and the topology initialization parameter set includes:

[0099] Collect raw data on sensor displacement, velocity, acceleration and pore water pressure, and combine them with point coordinates and topographic data to complete time alignment, unit unification and range verification, and eliminate missing sections and abnormal points;

[0100] The point coordinates refer to the spatial position parameters of the terminal devices deployed within the monitoring area;

[0101] The terrain data refers to parameters describing the terrain morphology within the monitored area;

[0102] Rainfall data is collected by a back-end processing center, which operates a data processing and analysis system.

[0103] Point coordinates refer to the spatial position parameters of sensors deployed within the monitoring area. They are used to describe the precise position of the sensors in three-dimensional space. Point coordinates are usually obtained using surveying methods such as the Global Positioning System or a total station and are represented in a unified coordinate system. Each monitoring point corresponds to a coordinate triplet, which is used to uniquely identify the point in space. In this invention, point coordinates are not only used to determine the physical distribution of sensors within the area, but also to calculate the adjacency relationship and spatial distance between sensors, thereby providing a geometric basis for the generation of point units, edge units, and surface units in the unit complex structure.

[0104] Topographic data refers to parameters describing the topographic morphology of the monitoring area, such as surface elevation, slope, and aspect. It is usually obtained by digital elevation model, remote sensing mapping, or lidar scanning. Topographic data can characterize the undulation features of the landslide body and its surrounding environment and reflect the terrain differences in the area. In this invention, topographic data is combined with point coordinates to correct the spatial adjacency relationship between sensors and to calculate the elevation difference and slope information between edge units and surface units, so that the topology initialization parameter set can accurately reflect the influence of topography on the landslide incubation and propagation path.

[0105] Based on the time-aligned displacement, velocity, acceleration, and pore water pressure, noise reduction, interpolation, amplitude normalization, and scale standardization are performed. The data are organized into a tensor structure according to the sensor number and acquisition time to generate landslide micromotion data.

[0106] Displacement, velocity, acceleration, pore water pressure, and rainfall are synchronously packaged into a unified input structure according to the acquisition time, while retaining the units of physical quantities and sampling interval markings to generate an external field input set;

[0107] Based on the coordinates of the monitoring points and the terrain data, the spatial relationships between the monitoring points are analyzed. First, the neighborhood radius and the number of nearest neighbors are set. By calculating the spatial distance between any two monitoring points, it is determined whether the two points are within the neighborhood radius or belong to the nearest neighbor set. When the conditions are met, the connection relationship between the monitoring points is established, and all connection relationships are organized into an adjacency matrix. On the basis of the adjacency matrix, a comprehensive metric of terrain elevation difference and spatial distance is introduced, and edge weight parameters are assigned to each connection relationship to characterize the coupling strength between the monitoring points. On this basis, geometric calculations are performed on the area enclosed by multiple monitoring points to obtain the area information and average slope information of the area. The two are combined to form the surface weight parameter to characterize the overall terrain features of the area. Finally, the adjacency matrix, edge weight parameters, surface weight parameters, and boundary relationships and co-boundary relationships constructed by the relationships between points and edges, and edges and surfaces are integrated to form the topological parameters for initializing the unit complex structure, generating a set of topological initialization parameters.

[0108] In this embodiment, the generation of the unit complex input structure includes:

[0109] In the unit complex construction unit, the landslide micro-motion data and the set of topology initialization parameters are read, and an index mapping is established according to the sensor number and time step to generate an aligned data index for unit construction.

[0110] Based on the point coordinates and alignment data index, a corresponding point unit is established for each monitoring point according to the spatial coordinates of the monitoring point. The standardized monitoring quantity in the alignment data index is added as the initial feature of the point unit to generate a set of point units containing spatial location and time series features. The adjacency matrix in the topology initialization parameter set is read to form adjacency relationship data. At the same time, the edge weight parameter and surface weight parameter are loaded for subsequent construction.

[0111] Based on the adjacency relationship data and terrain data, edge units are generated according to the adjacent monitoring point pairs in the adjacency relationship data and edge weight parameters are added. Closed point sets are identified based on the edge unit set and point coordinates to form surface units and surface weight parameters are added. Weighted edge unit set and surface unit set are generated. Directed relationships are determined according to the slope aspect information in the terrain data and boundary relationships and co-boundary relationships are generated.

[0112] For each face element, record the set of face element vertices consisting of several point elements, and store this set in association with the face weight parameters;

[0113] The edge weight parameter characterizes the spatial and terrain relationship between point pairs, measures the connection strength between two monitoring points, and is used in edge units to adjust the propagation weight of features between points during the message transmission process of the network.

[0114] The surface weight parameter characterizes the overall geometric and topographic features of the region, representing the overall feature strength of a region enclosed by multiple monitoring points (e.g., a triangular face composed of three points). It is used for surface units and, in the propagation of higher-order features of the network, is used to control the strength of regional synergistic effects.

[0115] Node features are extracted from landslide micro-motion data, edge features are calculated by weighting the adjacency relationship data and edge weight parameters, and surface features are generated by aggregating the vertex set of surface units and surface weight parameters, thus forming node features, edge features and surface features;

[0116] When calculating edge features, for each monitoring point pair marked as connected based on adjacency relationship data, the corresponding edge weight parameters, the node feature vectors of the two points, and the point coordinates are read. First, the difference and magnitude of the node feature vectors of the two points are calculated to obtain the feature difference and the magnitude of the feature difference. Then, the average value of the node feature vectors of the two points is calculated to obtain the feature mean. At the same time, the spatial distance and unit direction vector are obtained from the coordinates of the two points. The direction is uniformly defined as pointing from the monitoring point with the smaller number to the monitoring point with the larger number. Then, the edge weight parameters are multiplied by the feature difference to obtain the first set of results and defined as the difference component. The edge weight parameters are multiplied by the feature mean to obtain the second set of results and defined as the mean component. The magnitude of the edge weight parameters and the feature difference is multiplied by the unit direction vector to obtain the third set of results and defined as the directional gradient component. The three sets of results are spliced ​​together in a fixed order to form the edge features of the point pair. The above steps are repeated for all connected point pairs and collected into an edge feature set, which serves as the edge feature channel of the unit complex input structure for the spatiotemporal physical gating message transmission unit to call.

[0117] When generating surface features based on the vertex set and surface weight parameters of surface units, the vertex set corresponding to each surface unit in the surface unit set is first read. The node features of each point unit in the vertex set are extracted and aligned in time step to form a vertex feature sequence. Then, the vertex feature sequence is weighted and aggregated according to the surface weight parameters. The surface weight parameters include weight values ​​formed by the combination of the geometric area and average slope of the surface unit, which are used to adjust the contribution of each vertex feature to the overall surface. In the weighted aggregation process, the weighted average of the vertex feature sequence is first calculated as the basic component, and then the product of the vertex feature difference and the average slope is calculated as the direction component. The two types of components are normalized according to the surface weight parameters and then concatenated to generate surface features. Finally, the surface features of all surface units are organized into a surface feature set, and the structure is kept consistent with the node feature set and edge feature set for subsequent steps.

[0118] The set of point cells, edge cells, face cells, adjacency data, boundary relationships, co-boundary relationships, edge weight parameters, face weight parameters, as well as node features, edge features, and face features are summarized and packaged to generate a complex cell input structure.

[0119] In this embodiment, the generation of the multi-layer intermediate feature representation includes:

[0120] In the spatiotemporal physical gated message transmission unit, the receiving unit receives the complex input structure and the external field input set, reads the node features, edge features and surface features of the point unit, edge unit and surface unit, and synchronously loads the displacement, velocity, acceleration, pore water pressure and rainfall in the external field input set;

[0121] Temporal gating and physical gating are constructed. Temporal gating is based on the time step data in the unit complex input structure and the external field input set to establish a dynamic adjustment function, which is used to control the update rate of features at different time levels. Physical gating is based on the rainfall and pore water pressure in the external field input set to establish a threshold control function, which is used to adjust the propagation intensity of features under external field driving conditions.

[0122] Define the time gating function as follows:

[0123] ;

[0124] ;

[0125] in, The moving average of rainfall is calculated using historical rainfall data. The result is obtained by recursively calculating the smoothing result from the previous time step. This represents the time interval between adjacent sampling points, used to control the time scale of feature updates. This represents the smoothing coefficient, used to adjust the ratio between historical information and new input. Indicates historical rainfall. This represents a time-gating function used to control the update rate of features at different time steps. This represents the time decay scale, used to characterize how quickly a feature decays over time. This represents the normalization constant for rainfall, used to ensure the stability of the numerical range. Represents dynamic coefficients, used to control the weights of different factors on the gating function. This represents a non-linear activation function used to map the output to... Interval, time-gated functions are used to dynamically adjust the feature update amplitude based on time intervals and rainfall intensity during cross-dimensional message passing;

[0126] Define the physical gate function as follows:

[0127] ;

[0128] ;

[0129] in, This represents the physical gating function, used to control the propagation intensity of features under different external field driving conditions; Representing edge cells The mean pore water pressure is the average pore water pressure of the two point elements at the current time step. This represents the pore water pressure threshold, used to determine whether the pore water pressure has reached the triggering condition. This represents the moving average of rainfall. This represents the rainfall threshold, used to determine whether rainfall has reached a critical condition. This represents the edge weight parameters in the topology initialization parameter set, used to describe the coupling strength between point elements; This represents the edge weight normalization constant, used to normalize the weights of different edges; This represents a dynamic coefficient used to adjust the weights of different factors in the gating process. This represents a non-linear activation function used to ensure that the output is within a certain range. Interval; Physical gating functions are used to amplify message intensity when rainfall and pore water pressure exceed a set threshold during cross-dimensional message transmission, and adjust the propagation intensity in combination with edge weight parameters, thereby achieving dynamic control consistent with geological and physical conditions;

[0130] Cross-dimensional message passing and feature updates are performed. The node features of point units are passed to edge units through adjacency relationships and weighted by edge weight parameters. The edge features of edge units are passed to surface units through boundary relationships and aggregated by surface weight parameters. At the same time, the surface features of surface units are fed back to edge units and point units through co-boundary relationships, forming a cross-dimensional feature flow. In each passing process, time gating and physical gating are introduced as adjustment factors to control the feature update amplitude.

[0131] In the process of cross-dimensional message passing, an update function is defined to perform time-gated and physical-gated weighted aggregation on the features of the previous layer of each dimension unit to generate the intermediate feature representation of the next layer, and the intermediate feature representations of each dimension unit are stacked hierarchically to form a multi-layer intermediate feature representation;

[0132] The intermediate feature representations of each level are stacked to generate multi-layer intermediate feature representations, which are then output to the physical consistency constraint unit.

[0133] In this embodiment, the update function is defined as:

[0134] ;

[0135] in, Representation unit In the The feature vector of a layer is the output obtained through cross-dimensional message passing and gating updates. Representation and Unit Adjacent units In the The feature vectors of the layer are used as input for message passing. Representation unit The set of neighbors, containing units that are adjacent to it. Indicates the first The feature transformation matrix of a layer is used to linearly map the input feature vector to a new feature space. This represents a time-gating function used to adjust the update magnitude of features across different time dimensions based on time-step information in the external input set. This represents a physical gate function used to regulate the unit based on rainfall and pore water pressure in the external field input set. With unit The strength of message passing between them This represents a nonlinear activation function used to enhance the model's expressive power and generate nonlinear feature representations;

[0136] In this embodiment, the generation of the physically consistent feature representation includes:

[0137] In the physical consistency constraint unit, the multi-layer intermediate feature representation and the external field input set are read. The multi-layer intermediate feature representation comes from the output of the spatiotemporal physical gating message passing unit stacked in the dimensions of point unit, edge unit and surface unit. The external field input set contains time step data of displacement, velocity, acceleration, pore water pressure and rainfall.

[0138] Based on the edge features in the multi-layer intermediate feature representation, the predicted shear stress is calculated and a constraint function is constructed. The predicted shear stress of the edge element (i,j) is defined as follows:

[0139] ;

[0140] in, This represents the predicted shear stress value of the edge element (i,j) calculated by the improved CCNN network based on edge features at time step t. Represents the edge feature vector. Represents the dynamic mapping coefficient vector;

[0141] The consistency correction of the multi-layer intermediate feature representation is performed with the constraint function as the objective, and the residual-driven update of the unit feature vector in each layer is performed according to the gradient descent strategy to generate physically consistent feature representation.

[0142] The residual-driven update process refers to the process of comparing the constraint function with the constraint residuals calculated from the multi-layer intermediate feature representation and the external field input set, reading the unit feature vectors one by one in each layer, evaluating their deviation from the Mohr-Coulomb criterion and the preset energy constraints, and gradually correcting the feature vectors according to the residual size. The correction method includes adding or subtracting the components of the feature vectors numerically, and limiting the range of change during the adjustment process so that it gradually approaches the interval that satisfies the constraint conditions. Through multiple iterations, the gap between the features and the physical constraints is reduced layer by layer. Finally, the process stops when the residual converges to the threshold range, and the feature set after residual-driven correction is obtained.

[0143] The process of generating physically consistent feature representations refers to organizing the node features, edge features, and surface features updated by residual driving into new sets, and performing normalization and scale correction within the sets to ensure that the numerical range and stability of features in different dimensions remain consistent. Then, the node feature set, edge feature set, and surface feature set are recombined to form an updated cross-dimensional feature structure, which is used to replace the original multi-layer intermediate feature representation. This new cross-dimensional feature structure is the physically consistent feature representation. It numerically satisfies the requirements of the Mohr-Coulomb criterion and energy constraints, ensuring that the calculations performed in the subsequent multi-scale uncertainty early warning unit inherit the expressive power of the deep learning model and conform to the basic laws of geological and physical conditions.

[0144] Output the physically consistent feature representation and the set of constraint residuals obtained by constraint function decomposition, which can be called by the multi-scale uncertainty early warning unit.

[0145] In this embodiment, the constraint function is defined as:

[0146] ;

[0147] in, This represents the global constraint function value obtained at time step t by weighting the Mohr-Coulomb constraint residual and the energy constraint residual. It is used to measure whether the multi-layer intermediate feature representation satisfies the physical constraint conditions and uniformly characterizes the degree of deviation between the two types of constraints. Representing edge cells At time step The predicted shear stress is calculated from the edge feature vectors in the multi-layer intermediate feature representation. This represents the cohesion parameter, used to characterize the shear strength constant of soil. This represents the internal friction angle parameter, used to characterize the frictional effect during shear failure. Indicates effective and positive force, among which For the normal stress parameter in the preset energy constraint, This represents the average pore water pressure at both ends of the edge element. This represents the moving average of rainfall in the external input set, used to reflect the short-timescale rainfall-driven effect. This represents the energy mapping coefficient, used to adjust the contribution ratio of rainfall and pore water pressure to the energy-constrained residuals. Representing edge cells The edge feature vectors, This represents the energy proxy mapping coefficient vector, used to map edge feature vectors to energy proxy quantities. This indicates a non-negative weight, used to balance the relative influence of the Mohr-Coulomb constraint residual and the energy constraint residual. This indicates taking the absolute value, used to measure the magnitude of the deviation between the prediction result and the constraint conditions;

[0148] In this embodiment, the generation of the warning level, target area index, and trigger command includes:

[0149] In the multi-scale uncertainty early warning unit, the physical uniform feature representation is received, and the node features, edge features and surface features are segmented and statistically analyzed in the time dimension to generate the mean estimation sequence and the variance estimation sequence. The mean estimation sequence is used to characterize the central trend of the features under different time windows, and the variance estimation sequence is used to characterize the degree of fluctuation of the features within the time window.

[0150] Confidence gating is performed based on the mean estimation sequence and the variance estimation sequence. According to the variance estimation sequence, the sample variance within the same time window is taken as the uncertainty level. The variance value is normalized and mapped to determine the confidence gating. When the uncertainty level is higher than the preset threshold, the corresponding feature is suppressed. When the uncertainty level is lower than the threshold, the corresponding feature is enhanced, and a confidence calibration feature set is generated.

[0151] The confidence calibration feature set is input into the multi-scale unit complex aggregation process. Point units, edge units and surface units are grouped and aggregated according to different spatial and temporal scales. During aggregation, the consistency of topological relationships is maintained and the composite feature vector across scales is calculated to generate a multi-scale aggregated feature representation.

[0152] By combining the constrained residual set with the multi-scale aggregated feature representation, a calculation rule for the composite early warning index is established. The physical consistency features of each spatial location and time period are weighted and superimposed to form a composite early warning index set. The early warning time window set is determined according to the index change trend.

[0153] Based on the set of composite early warning indices and the set of early warning time windows, an early warning level is generated according to the grading standard, and the target area index and trigger command are determined and output as the early warning result for subsequent monitoring system to call.

[0154] Example 1:

[0155] To verify the feasibility of this invention in practice, it was applied to a landslide stability monitoring and early warning scenario in a geologically hazardous area. In this scenario, multiple potential sliding zones exist on the surface and inside the landslide. Changes in rainfall and pore water pressure often accelerate the evolution of landslides. Traditional monitoring methods rely on a small number of inclinometer boreholes or fixed monitoring stations, making it difficult to achieve comprehensive coverage and resulting in insufficient accuracy in early landslide warnings.

[0156] In practical applications, this invention employs a dot-matrix array of buried sensor terminals to construct a monitoring network covering the entire landslide area. The burial depth ranges from 0.5 to 1 meter, avoiding interference from the surface environment. Simultaneously, a low-power processor and energy harvesting module ensure long-term stable operation of the equipment. Upon triggering the acquisition conditions, the terminals collect data in real time, including displacement, velocity, acceleration, pore water pressure, and rainfall, and upload this data to a centralized acquisition device for preprocessing. This data is then input into an improved CCNN network to complete unit complex modeling, spatiotemporal physical gating message passing, physical consistency constraints, and multi-scale uncertainty early warning. During this process, the multi-layered topology of points, edges, and surfaces is effectively utilized. External field inputs such as rainfall and pore water pressure are dynamically adjusted through physical gating, enabling the model to capture the micro-motion evolution trend of the landslide body in real time.

[0157] To verify the effectiveness, a long-term observation was conducted comparing the traditional threshold-based risk assessment method with the method proposed in this invention. In the experiment, 120 sensor terminals were deployed in the area, with a total sampling period of 180 days. Traditional methods determine risk based on fixed thresholds, while the method of this invention utilizes a deep learning model to comprehensively model multi-source inputs and outputs continuous risk levels and target area indices. During the experiment, the advance warning time, accuracy, false alarm rate, and stability of this invention were significantly improved. The following table shows the comparison of experimental results.

[0158] Table 1. Performance comparison of different methods in landslide micromotion monitoring and early warning

[0159] ;

[0160] As shown in Table 1, the early warning lead time of this invention reaches 36 hours, which is about 24 hours longer than the traditional method. This allows more time for personnel evacuation and emergency response before a disaster occurs. The early warning accuracy rate increased from 72.5% to 86.3%, the false alarm rate decreased from 18.2% to 7.6%, and the false alarm rate decreased from 14.7% to 5.8%, indicating that this invention is more reliable in identifying potential landslide risks. The fluctuation range of real-time data stability decreased from ±8.5% to ±3.3%, indicating that the model maintained stronger consistency during long-term operation and effectively suppressed interference caused by noise and local disturbances. Equipment energy consumption decreased from 95mW to 44mW, a reduction of more than half, thanks to the low-power design and energy harvesting mechanism of the terminal. Network coverage integrity increased from 63% to 90%, proving that the point-matrix deployment method can achieve large-scale monitoring and comprehensively capture the overall change characteristics of the landslide body, rather than relying solely on a limited number of observation points.

[0161] The performance improvement is mainly due to the following factors: First, the use of unit complex modeling enables the simultaneous capture of multi-layer topological features of sensor points, edges, and surfaces, allowing for the reconstruction of the overall evolution process of the landslide body. Second, the introduction of a spatiotemporal physical gating mechanism dynamically couples the external field input with the monitoring data, enhancing the model's sensitivity to changes in rainfall and pore water pressure. Third, physical consistency constraints effectively ensure that the model output conforms to geomechanical laws, avoiding the "black box" problem that often occurs in deep learning models. Fourth, uncertainty modeling and multi-scale aggregation processes improve the stability of the early warning results, reducing misjudgments caused by noise or local anomalies. These factors combined make the method of this invention superior to traditional methods in terms of early warning accuracy, real-time performance, and stability, demonstrating significant creativity and innovation.

[0162] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based lattice-type landslide micro-motion monitoring and early warning method, characterized in that, Includes the following steps: According to the pre-designed grid and dot matrix scheme, multiple terminal devices are buried in the underground of the landslide body and potential landslide area in a large-area dot matrix pattern, and the antennas are brought out to the surface. Under normal circumstances, the terminal device is in a low-power sleep mode and performs sleep monitoring. When vibration or tilt changes are detected to exceed a preset threshold, it will automatically wake up and collect raw sensor data. The terminal device transmits the raw sensor data to the centralized data acquisition device via a wireless communication module to complete data aggregation; The centralized acquisition equipment completes the aggregation of raw sensor data and performs data preprocessing to generate landslide micro-motion data, external field input set and topology initialization parameter set; The data processing and analysis system, based on landslide micro-motion data, external field input set and topology initialization parameter set, calls an improved CCNN network to comprehensively evaluate the stability of the landslide body and generate early warning level, target area index and trigger command; If the analysis results show that the landslide is unstable and the displacement is intensifying, an early warning will be issued to the relevant authorities and the public. The invocation of the improved CCNN network includes: An improved CCNN network is constructed, comprising: a unit complex construction unit, a spatiotemporal physical gated message passing unit, a physical consistency constraint unit, and a multi-scale uncertainty early warning unit; In the element complex construction unit, the element complex structure is established based on the landslide micro-motion data and the set of topological initialization parameters, and the element complex input structure is generated. In the spatiotemporal physical gated message passing unit, the receiving unit's complex input structure and external field input set construct temporal and physical gates, perform cross-dimensional message passing and feature updates, and generate multi-layer intermediate feature representations; In the physical consistency constraint unit, it receives multi-layer intermediate feature representations and external field input sets, calculates constraint residual sets based on the Mohr-Coulomb criterion and preset energy constraints, and generates physically consistent feature representations. In the multi-scale uncertainty early warning unit, the physical uniform feature representation is received, confidence gating is performed and multi-scale unit complex aggregation is carried out, and the early warning level, target area index and trigger command are generated by combining the constraint residual set.

2. The deep learning-based lattice-type landslide micro-motion monitoring and early warning method according to claim 1, characterized in that, The generation of the landslide micro-motion data, the external field input set, and the topology initialization parameter set includes: Based on the collected raw sensor data, the time alignment, unit unification and range verification were completed by combining the point coordinates and terrain data, and missing sections and abnormal points were eliminated. The point coordinates refer to the spatial position parameters of the terminal devices deployed within the monitoring area; The terrain data refers to parameters describing the terrain morphology within the monitored area; Based on the time-aligned displacement, velocity, acceleration, and pore water pressure, noise reduction, interpolation, amplitude normalization, and scale standardization are performed. The data are organized into a tensor structure according to the terminal device number and acquisition time to generate landslide micromotion data. Displacement, velocity, acceleration, pore water pressure, and rainfall are synchronously packaged into a unified input structure according to the acquisition time, while retaining the units of physical quantities and sampling interval markings to generate an external field input set; The rainfall data was collected by the backend processing center; Based on the location coordinates and terrain data, the adjacency relationship between monitoring points is calculated, and an adjacency matrix is ​​generated according to the preset neighborhood radius and the number of nearest neighbors. The edge weight parameters and surface weight parameters are generated by combining the terrain elevation difference and spatial distance, and a set of topology initialization parameters is generated.

3. The deep learning-based lattice-type landslide micro-motion monitoring and early warning method according to claim 1, characterized in that, The generation of the unit complex input structure includes: In the unit complex construction unit, the landslide micro-motion data and the set of topology initialization parameters are read, and an index mapping is established according to the sensor number and time step to generate an aligned data index. Based on the point coordinates and alignment data index, a set of point cells is generated, and the adjacency matrix in the topology initialization parameter set is read to form adjacency relationship data; Based on adjacency data and terrain data, and with the addition of edge weight parameters and surface weight parameters, a set of weighted edge cells and a set of surface cells are generated, the directed relationships are determined, and the boundary relationships and co-boundary relationships are generated. Record the set of face unit vertices composed of point units for each face unit; The edge weight parameters characterize the spatial and terrain relationships between point pairs; The area weight parameters characterize the overall geometric and topographic features of the region; Node features are extracted from landslide micro-motion data, edge features are calculated by weighting the adjacency relationship data and edge weight parameters, and surface features are generated by aggregating the vertex set of surface units and surface weight parameters. The set of point cells, edge cells, face cells, adjacency data, boundary relationships, co-boundary relationships, edge weight parameters, face weight parameters, as well as node features, edge features, and face features are summarized and packaged to generate a complex cell input structure.

4. The deep learning-based lattice-type landslide micro-motion monitoring and early warning method according to claim 1, characterized in that, The generation of the multi-layer intermediate feature representation includes: In the spatiotemporal physical gated message passing unit, the receiving unit has a complex input structure and an external field input set; Time-gating and physical-gating are constructed. Time-gating is based on the time step data in the unit complex input structure and the external field input set to establish a dynamic adjustment function to control the update rate of features at different time levels. Physical-gating is based on the rainfall and pore water pressure in the external field input set to establish a threshold control function to adjust the propagation intensity of features under external field driving conditions. Cross-dimensional message passing and feature updates are performed. The node features of point units are passed to edge units through adjacency relationships and weighted by edge weight parameters. The edge features of edge units are passed to surface units through boundary relationships and aggregated by surface weight parameters. At the same time, the surface features of surface units are fed back to edge units and point units through co-boundary relationships, forming a cross-dimensional feature flow. In each passing process, time gating and physical gating are introduced as adjustment factors to control the feature update amplitude. In the process of cross-dimensional message passing, an update function is defined to perform time-gated and physical-gated weighted aggregation on the features of the previous layer of each dimension unit to generate the intermediate feature representation of the next layer, and the intermediate feature representations of each dimension unit are stacked hierarchically to form a multi-layer intermediate feature representation; The intermediate feature representations of each level are stacked to generate multi-level intermediate feature representations.

5. The deep learning-based lattice-type landslide micro-motion monitoring and early warning method according to claim 1, characterized in that, The generation of the physically consistent feature representation includes: In the physically consistent constraint unit, read the multi-layer intermediate feature representation and the external field input set; Shear stress prediction is calculated and constraint functions are constructed based on edge features in multi-layer intermediate feature representation; Consistency correction is performed on the multi-layer intermediate feature representation with the constraint function as the objective, and residual-driven updates are performed on the unit feature vectors in each layer according to the gradient descent strategy to generate physically consistent feature representations.

6. The deep learning-based lattice-type landslide micro-motion monitoring and early warning method according to claim 1, characterized in that, The generation of the warning level, target area index, and trigger command includes: In the multi-scale uncertainty early warning unit, the physical uniformity feature representation is received, and the node features, edge features and surface features are segmented and statistically analyzed in the time dimension to generate the mean estimation sequence and the variance estimation sequence. Confidence gating is performed on the mean estimation sequence and variance estimation sequence. Based on the variance estimation sequence, the sample variance within the same time window is taken as the uncertainty level. The corresponding features are controlled according to the uncertainty level to generate a confidence calibration feature set. The confidence calibration feature set is input into the multi-scale unit complex aggregation process. Point units, edge units and surface units are grouped and aggregated according to different spatial and temporal scales. During aggregation, the consistency of topological relationships is maintained and the composite feature vector across scales is calculated to generate a multi-scale aggregated feature representation. By combining the constrained residual set with the multi-scale aggregated feature representation, a calculation rule for the composite early warning index is established. The physical consistency features of each spatial location and time period are weighted and superimposed to form a composite early warning index set. The early warning time window set is determined according to the index change trend. Based on the set of composite early warning indices and the set of early warning time windows, early warning levels are generated according to the classification standards, and the target area index and triggering instructions are determined.

7. A deep learning-based lattice-type landslide micro-motion monitoring and early warning device, characterized in that, include: Terminal devices, centralized data acquisition equipment, and data processing and analysis systems; The terminal device includes: a sensor unit for detecting micro-vibration signals and tilt change signals; a low-power processor and a data acquisition module for maintaining ultra-low power monitoring in sleep mode and automatically waking up and performing data acquisition when the micro-vibration amplitude and tilt angle increment exceed a preset threshold; a wireless communication module for uploading processed data to a centralized acquisition device when acquisition is triggered; and a power supply unit for supplying power to the system. The centralized data acquisition device is used to automatically scan and poll the status of each terminal device, receive the data uploaded by each terminal, and perform summary processing. The data processing and analysis system is used to analyze and provide early warnings for the aggregated data; The analysis and early warning of the aggregated data includes using an improved CCNN network based on landslide micro-motion data, field input set, and topology initialization parameter set to comprehensively evaluate the stability of the landslide body, generate early warning level, target area index, and trigger command. If the analysis results show that the landslide body is unstable and the displacement is aggravated, early warning information is issued to the competent authorities and the public. The invocation of the improved CCNN network includes: An improved CCNN network is constructed, comprising: a unit complex construction unit, a spatiotemporal physical gated message passing unit, a physical consistency constraint unit, and a multi-scale uncertainty early warning unit; In the element complex construction unit, the element complex structure is established based on the landslide micro-motion data and the set of topological initialization parameters, and the element complex input structure is generated. In the spatiotemporal physical gated message passing unit, the receiving unit's complex input structure and external field input set construct temporal and physical gates, perform cross-dimensional message passing and feature updates, and generate multi-layer intermediate feature representations; In the physical consistency constraint unit, it receives multi-layer intermediate feature representations and external field input sets, calculates constraint residual sets based on the Mohr-Coulomb criterion and preset energy constraints, and generates physically consistent feature representations. In the multi-scale uncertainty early warning unit, the physical uniform feature representation is received, confidence gating is performed and multi-scale unit complex aggregation is carried out, and the early warning level, target area index and trigger command are generated by combining the constraint residual set.

Images