Landslide safety rapid assessment system based on microservices and containerization

Abstract

The application relates to the technical field of landslide monitoring, and discloses a landslide safety rapid evaluation system based on micro services and containerization, which comprises a spatial discretization module, a safety evaluation module, an abnormality identification module, an asynchronous scheduling module and a dynamic rendering module.The spatial discretization module divides a landslide body geometric unit into a multilayer three-dimensional point cloud set to calculate a three-dimensional deformation displacement scalar.The safety evaluation module calculates a displacement gradient statistical index and combines a mechanical mapping model to calculate a landslide global equivalent safety coefficient.The abnormality identification module extracts a deformation active area, outputs a high-displacement abnormal feature point cluster and a three-dimensional space position coordinate.The asynchronous scheduling module instantiates an algorithm container image into a container application entity to generate a background asynchronous data processing task for data storage.The dynamic rendering module performs dynamic updating and superposition rendering in a digital twin three-dimensional scene to realize real-time evaluation.The application realizes high-strength calculation process decoupling, accurately locates and intuitively displays the damage state of the landslide interior.

Description

Technical Field

[0001] This invention relates to the field of landslide monitoring technology, specifically a rapid landslide safety assessment system based on microservices and containerization. Background Technology

[0002] Landslides are a common geological hazard, and real-time monitoring and deformation assessment play an important role in disaster prevention and mitigation. With the development of space exploration and sensing technology, it has become commonplace to acquire large-scale three-dimensional point cloud data of landslide geometric units, and the concept of digital twins has gradually been introduced into the dynamic display and safety assessment of landslide deformation.

[0003] Existing landslide deformation assessment systems still have design limitations in practical applications. Conventional systems often employ highly coupled synchronous data processing architectures, which can easily bear excessive computational loads when faced with high-frequency, complex 3D spatial deformation data, leading to data processing blockages and slow overall response. Furthermore, existing deformation assessment algorithms typically focus on analyzing direct changes in surface displacement, failing to effectively establish a spatial mapping relationship between surface deformation data and the deformation state of deep structural surfaces within the landslide. This makes it difficult to accurately quantify the degree of non-uniform deformation within the landslide and locate deep-seated failure units. In the results presentation stage, conventional 3D visualization methods mainly perform simple deformation or displacement mapping rendering on the landslide surface mesh, failing to provide a perspective view of localized abnormal deformation states within the landslide body. Due to the lack of a collaborative rendering mechanism between deep failure features and the 3D scene, the system struggles to intuitively display the locations of hidden instability and failure points within the landslide, resulting in delayed feedback and limited perspective, failing to meet the business requirements for real-time dynamic assessment of complex landslides. Therefore, this invention proposes a rapid landslide safety assessment system based on microservices and containerization to address the shortcomings of existing technologies. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a rapid landslide safety assessment system based on microservices and containerization. This system solves the problems of existing landslide monitoring systems, such as high computational load leading to processing blockage when dealing with complex three-dimensional spatial deformation data, difficulty in accurately quantifying and locating the deep structural damage state inside the landslide, and the lag in displaying potential instability points inside the landslide due to limitations in surface rendering, making them difficult to observe intuitively.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a rapid landslide safety assessment system based on microservices and containerization, comprising: The spatial discretization module is used to divide the geometric units of the landslide body into a multi-layer three-dimensional point cloud set, and to calculate the three-dimensional deformation displacement scalar of the spatial nodes by integrating displacement information collected by external surface monitoring equipment. The safety assessment module is used to calculate the displacement gradient statistical index of a multi-layer three-dimensional point cloud set based on the three-dimensional deformation displacement scalar, and input the displacement gradient statistical index into the pre-built mechanical mapping model to calculate the global equivalent safety factor of the landslide. The anomaly identification module is used to extract deformation-active regions from multi-layer 3D point cloud sets, perform point-level filtering on deformation-active regions based on density clustering rules, and output high displacement anomaly feature point clusters and 3D spatial coordinates. The asynchronous scheduling module is used to instantiate the algorithm container image into an independently running container application entity to generate background asynchronous data processing tasks, store the three-dimensional spatial coordinates of the landslide global equivalent safety factor and the high displacement anomaly feature point cluster to the system structured database, and retrieve and read the three-dimensional spatial coordinates of the landslide global equivalent safety factor and the high displacement anomaly feature point cluster from the system structured database. The dynamic rendering module is used to retrieve and read the three-dimensional spatial coordinates of the landslide's global equivalent safety factor and high displacement anomaly feature point clusters from the system's structured database, and perform dynamic updates and overlay rendering in the digital twin three-dimensional scene for real-time evaluation.

[0007] Preferably, the spatial discretization module divides the geometric units of the landslide body into multi-layer three-dimensional point cloud sets according to the preset three-dimensional grid resolution; The spatial discretization module acquires surface deformation point cloud data from external surface monitoring equipment as displacement information, extracts the three-dimensional displacement vectors corresponding to the spatial nodes on the landslide surface, and inputs the three-dimensional displacement vectors into a pre-constructed internal structural deformation inference model. The internal structural deformation inference model is a deep layered generation variational autoencoder model. The three-dimensional deformation displacement vectors of each spatial node in the multi-layer three-dimensional point cloud set are calculated through the deep layered generation variational autoencoder model, and the three-dimensional deformation displacement scalar is calculated based on the three-dimensional deformation displacement vectors.

[0008] Preferably, the asynchronous scheduling module packages the system logic code into an algorithm container image and distributes it to the computing cluster nodes, instantiates it into an independently running container application entity, and configures the service call interface; The asynchronous scheduling module has a built-in time scheduling engine. When the time difference condition meets the preset time interval threshold, it triggers the generation of a background asynchronous data processing task. This task sends a computation request signal to the container application entity through a service call interface to extract the 3D deformation displacement scalar and perform internal security assessments and anomaly identification calculations. The microservice architecture separates and encapsulates complex spatial point cloud computing, ensuring that front-end responses are not blocked by back-end data processing.

[0009] Preferably, the safety assessment module performs data scaling preprocessing on the three-dimensional deformation displacement scalar to obtain normalized displacement features; The safety assessment module extracts a set of spatial neighboring points for spatial nodes, calculates the absolute value of the displacement difference between the spatial node and each neighboring spatial node in the set of spatial neighboring points based on the three-dimensional deformation displacement scalar, divides the absolute value of the displacement difference by the three-dimensional Euclidean distance to obtain the local node deformation gradient rate, and performs arithmetic mean processing on the local node deformation gradient rate to obtain the local displacement gradient as a statistical index of displacement gradient.

[0010] Preferably, the safety assessment module divides the multi-layer 3D point cloud set into multiple deformable clustering units using the K-means unsupervised clustering algorithm, extracts each local displacement gradient within the deformable clustering unit as a statistical index of displacement gradient, calculates the arithmetic mean of each local displacement gradient, extracts the level average displacement gradient index of the corresponding level, and splices and combines the level average displacement gradient indices of multiple discrete point cloud levels in depth order to form a multi-layer average displacement gradient sequence. The mechanical mapping model adopts a gradient booster model architecture, which inputs the multi-layer average displacement gradient sequence into the internal independent regression decision tree to perform node splitting calculation. The output prediction residuals of each regression decision tree are summed and accumulated according to the preset weight parameters to obtain the global equivalent safety factor of the landslide.

[0011] Preferably, the anomaly identification module extracts the three-dimensional deformation displacement scalar and local displacement gradient of all spatial nodes in the multi-layer three-dimensional point cloud set to delineate initial clusters, calculates the average displacement gradient value of the initial clusters, and delineates the initial clusters with average displacement gradient values ​​higher than or equal to the overall median as deformation active regions. The anomaly identification module applies principal component analysis to solve the covariance matrix of the deformation-active region, extracts the feature vectors whose cumulative variance contribution rate is greater than the preset contribution threshold as principal components, and projects at least one principal feature dimension to complete the mesoscopic dimensionality reduction processing of the deformation-active region.

[0012] Preferably, the anomaly detection module calculates the mutual reachability distance between each spatial node within the deformation active region based on the main feature dimension; The anomaly identification module constructs a minimum spanning tree using the reachability distance as the weight of graph edge connections, gradually reduces the connections to form a hierarchical clustering tree structure, extracts the subtree branch with the highest stability metric as an independent optimal cluster, and isolates the extracted independent optimal clusters as non-consistent anomaly deformation units.

[0013] Preferably, the anomaly identification module takes the spatial nodes inside the non-uniform anomaly deformation unit as the target spatial nodes, extracts the set of nearest neighbors for the target spatial nodes, calculates the local reachability distance and performs an arithmetic mean summation, and defines the reciprocal of the summation mean as the local reachability density. The anomaly identification module calculates the ratio of the local reachability density of each neighboring spatial node to the local reachability density of the target spatial node, and performs an arithmetic average of all ratios to obtain the local anomaly factor. Target spatial nodes with local anomaly factors greater than or equal to the local anomaly factor judgment threshold are extracted and marked as high-displacement anomaly feature point clusters, and their three-dimensional spatial coordinates are extracted. The anomaly identification mechanism constructs a hierarchical calculation process from macroscopic to mesoscopic levels to microscopic levels, improving the accuracy of identifying internal anomaly feature nodes.

[0014] Preferably, the asynchronous scheduling module uses a hybrid storage strategy to write the global equivalent safety factor of the landslide and its three-dimensional spatial coordinates into the system's structured database to build a preheating cache; The dynamic rendering module establishes a two-way communication channel based on the network socket protocol, encapsulates the global equivalent safety factor of the landslide into a streaming data packet and actively pushes it to the client front-end engine. The client front-end engine then extracts the evaluation timestamp from the streaming data packet and dynamically adds two-dimensional coordinate nodes to the global equivalent safety factor of the landslide and the evaluation timestamp in the chart component, and connects them with line segments to generate a safety factor change curve.

[0015] Preferably, the dynamic rendering module uses an affine transformation matrix containing three-dimensional rotational parameter components and three-dimensional translational parameter vector components to convert the three-dimensional spatial coordinates of the high displacement anomaly feature point cluster as local spatial coordinates into three-dimensional world coordinates that can be recognized by the front-end three-dimensional rendering engine. The dynamic rendering module performs hybrid rendering of the alpha channel of the surface mesh material of the landslide body in the digital twin 3D scene according to the pre-set transparency parameters. It establishes heat map gradient mapping rules based on the upper and lower thresholds of the local anomaly factor values ​​to generate color materials. Based on the 3D world coordinates, the color materials are assigned to the instantiated 3D spatial node model, and the generated highlighted marker tiles are superimposed and rendered in the non-consistent abnormal deformation areas corresponding to each level inside the landslide model to complete real-time evaluation.

[0016] This invention provides a rapid landslide safety assessment system based on microservices and containerization. It offers the following advantages: 1. This invention packages the system logic code into an algorithm container image and instantiates it into an independently running container application entity. It uses the time scheduling engine built into the asynchronous scheduling module to trigger background asynchronous data processing tasks, thereby isolating the internal security assessment and abnormal feature identification operations from the operating system's runtime environment. This achieves asynchronous decoupling between the high-intensity calculation process and the front-end communication response process of the landslide deformation real-time assessment system, avoiding computational blocking phenomena during the processing of complex three-dimensional spatial deformation data, and improving the data processing efficiency and operational stability of the landslide deformation real-time assessment system.

[0017] 2. This invention calculates the three-dimensional deformation displacement scalar of each spatial node in a multi-layer three-dimensional point cloud set through an internal structural deformation inference model. Based on the three-dimensional deformation displacement scalar, the absolute value of the displacement difference and the three-dimensional Euclidean distance are calculated to obtain the local node deformation gradient rate. Combined with the local reachability density, local anomaly factors are obtained. Target spatial nodes with local anomaly factors greater than or equal to the local anomaly factor judgment threshold are extracted and marked as high displacement anomaly feature point clusters. The mapping relationship between surface deformation point cloud data and the deformation state of the internal structure of the landslide is established, the degree of non-uniform deformation inside the landslide is quantified, and the specific location of non-uniform anomaly deformation units inside the landslide is realized, thereby improving the accuracy of identifying the failure state of the internal structure of the landslide.

[0018] 3. This invention utilizes a dynamic rendering module to convert the three-dimensional spatial coordinates of high displacement anomaly feature point clusters into three-dimensional world coordinates recognized by the front-end three-dimensional rendering engine using an affine transformation matrix. In the digital twin three-dimensional scene, the alpha channel of the landslide surface mesh material is mixed and rendered according to the transparency parameter. Based on the local anomaly factor values, a heat map gradient mapping rule is established to generate color and material to assign to the instantiated three-dimensional spatial node model. Through the collaborative rendering of the transparent landslide surface and the superimposed internal highlighted marker blocks, the location of potential instability and failure points inside the landslide is displayed in a three-dimensional perspective, solving the problems of delayed display of assessment results and difficulty in intuitively observing the internal failure state. Attached Figure Description

[0019] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a schematic diagram of the overall method flow of the present invention; Figure 3 This is a schematic diagram of the abnormal feature point cluster identification process of the present invention; Figure 4 This is a schematic diagram illustrating the process of dynamic rendering of a 3D scene and asynchronous communication with the system according to the present invention. Figure 5 This is a schematic diagram of a comparative test of the concurrent processing latency of the system according to the present invention; Figure 6 This is a schematic diagram comparing the accuracy of abnormal feature recognition in this invention; Figure 7 This is a schematic diagram comparing the accuracy of the global equivalent safety factor of the present invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 This invention provides a rapid landslide safety assessment system based on microservices and containerization, including: a spatial discretization module, a safety assessment module, an anomaly identification module, an asynchronous scheduling module, and a dynamic rendering module.

[0022] The spatial discretization module is used to extract the three-dimensional terrain features of the target area and separate the geometric units of the landslide body. The spatial discretization module divides the geometric units of the landslide body into a multi-layer three-dimensional point cloud set. The spatial discretization module receives surface deformation monitoring data, generates a displacement scalar sequence of discrete points inside and outside the landslide body, and establishes a basic deformation data field.

[0023] The safety assessment module is used to perform standardization processing on the basic deformation data field. The safety assessment module calculates the displacement gradient statistical index of the corresponding level from the multi-layer three-dimensional point cloud set. The safety assessment module processes the displacement gradient statistical index by inputting a pre-built mechanical mapping model, calculates the global equivalent safety factor of the landslide, and outputs the risk level sequence.

[0024] The anomaly identification module is used to perform feature dimensionality reduction and spatial partitioning operations on the basic deformation data field. The anomaly identification module extracts deformation active regions from the multi-layer 3D point cloud set. Based on density clustering rules, the anomaly identification module isolates non-consistent deformation units within the deformation active regions. The anomaly identification module performs point-level filtering on non-consistent deformation units through density difference calculation and outputs the 3D coordinates and attribute information of high displacement anomaly feature point clusters.

[0025] The asynchronous scheduling module is used to package the algorithm runtime environment and parameters corresponding to the security assessment module and the anomaly identification module into a system image. The asynchronous scheduling module deploys application containers and configures service call interfaces on the system computing nodes. The asynchronous scheduling module generates background asynchronous computing tasks according to preset time nodes. The asynchronous scheduling module receives the computing results of the background asynchronous computing tasks and writes them into the system structured database.

[0026] The dynamic rendering module is used to establish a two-way data connection channel between the system's structured database and the front-end rendering engine. The dynamic rendering module obtains the global equivalent safety factor of the landslide and the three-dimensional coordinates of the high displacement anomaly feature point clusters. The dynamic rendering module performs spatial position color mapping on the displacement distribution state of the multi-layer three-dimensional point cloud set and the high displacement anomaly feature point clusters in the three-dimensional terrain scene.

[0027] See attached document Figure 2 This invention provides a real-time landslide deformation assessment method based on digital twins and microservices, comprising the following steps: S10, the spatial discretization module receives the three-dimensional geological model data of the target area, delineates the geometric boundary of the landslide body, divides the internal structure of the landslide body from the surface to the deep layer into multiple discrete point cloud levels, integrates the displacement information collected by external surface monitoring equipment, and calculates the three-dimensional deformation displacement scalar of each spatial node in multiple discrete point cloud levels. S20, the asynchronous scheduling module reads the system logic code and dependent library files, builds the algorithm container image, distributes the algorithm container image to the computing cluster nodes to start the instance application, generates data processing tasks according to the timed trigger instructions, and calls the instance application to extract the three-dimensional deformation displacement scalar and perform calculations. S30, the safety assessment module performs normalization processing on the three-dimensional deformation displacement scalar, calculates the local displacement gradient at different levels according to the spatial neighborhood distance, inputs the local displacement gradient into the pre-built machine learning model, and calculates the global equivalent safety factor of the landslide. S40, the anomaly identification module performs spatial clustering operations on spatial nodes and local displacement gradients at multiple discrete point cloud levels, delineates the boundaries of active deformation regions, applies a hierarchical clustering network based on core distance to the active deformation regions, isolates abnormal deformation units, calculates the local reachability density for spatial nodes in the abnormal deformation units, and selects the set of coordinate points that deviate from the predetermined density as high displacement anomaly feature point clusters. S50, the asynchronous scheduling module stores the calculation results to the system's structured database, the dynamic rendering module starts a two-way communication channel to transmit streaming data to the client, draws the landslide global equivalent safety factor change curve in the constructed three-dimensional terrain scene, and superimposes the marker blocks at the corresponding positions of multiple discrete point cloud layers based on the three-dimensional coordinates of the high displacement anomaly feature point clusters.

[0028] See attached document Figure 2 For step S10 of the landslide deformation real-time assessment method based on digital twins and microservices, the following steps are specifically performed: The spatial discretization module receives the three-dimensional geological model data of the target area, delineates the geometric boundary of the landslide body, divides the internal structure of the landslide body from the surface to the deep layer into multiple discrete point cloud levels, integrates the displacement information collected by external surface monitoring equipment, and calculates the three-dimensional deformation displacement scalar of each spatial node in the multiple discrete point cloud levels. The specific implementation process includes the following sub-steps: S110, the spatial discretization module acquires the three-dimensional geological model data of the target area, which includes digital elevation model and three-dimensional stratigraphic information. Based on the potential sliding surface and the topographic slope change boundary determined by geological exploration, the spatial discretization module establishes the spatial range of the landslide body, thereby delineating the geometric boundary of the landslide body. The spatial discretization module separates the three-dimensional spatial volume inside the geometric boundary of the landslide body from the three-dimensional geological model data of the target area, generating independent geometric units of the landslide body.

[0029] S120, the spatial discretization module performs spatial discretization operation on the geometric unit of the landslide body. According to the preset three-dimensional mesh resolution, the spatial discretization module divides the internal volume structure of the geometric unit of the landslide body into multiple discrete point cloud levels extending from the ground surface to the deep sliding surface. The multiple discrete point cloud levels include shallow point cloud sets, middle point cloud sets and deep sliding surface point cloud sets. Each discrete point cloud level is composed of multiple spatial nodes. The spatial nodes record three-dimensional spatial coordinates. For the setting of the three-dimensional mesh resolution and the generation rules of the discrete spatial nodes, those skilled in the art can use conventional octree voxelization or finite element mesh generation algorithms.

[0030] S130, the spatial discretization module receives displacement information from external surface monitoring equipment. The displacement information includes surface deformation point cloud data and corresponding time series information obtained by interferometric synthetic aperture radar. The spatial discretization module performs spatial registration transformation between the coordinate system of the interferometric synthetic aperture radar surface deformation point cloud data and the three-dimensional spatial position coordinate system of the landslide body geometric unit to unify the spatial reference. The spatial discretization module extracts the three-dimensional displacement vector of the surface corresponding to the spatial node of the landslide body based on the spatial registration transformation result.

[0031] S140, the spatial discretization module constructs the basic deformation data field. This module uses the three-dimensional displacement vector of the landslide surface as the spatial constraint boundary condition, inputting it into a pre-constructed internal structure deformation inference model. This model is a deep-layered variational autoencoder model. It is pre-trained using historical landslide monitoring data and finite element physical simulation data to learn the spatial co-mapping relationship between the geometric deformation of the landslide surface and the deformation of internal structural surfaces at various depths. Based on the three-dimensional displacement vector of the landslide surface, the model calculates the three-dimensional deformation displacement vector of spatial nodes within the landslide body at corresponding timestamps in multiple discrete point cloud layers. For each spatial node in the multiple discrete point cloud layers, the spatial discretization module calculates the three-dimensional deformation displacement scalar based on the calculated three-dimensional deformation displacement vector. The mathematical expression for calculating the three-dimensional deformation displacement scalar is as follows: ; in, Indicates the first Three-dimensional deformation displacement scalar of each spatial node; Indicates the first Displacement components of a spatial node in the X-axis direction of a three-dimensional coordinate system; Indicates the first The displacement components of a spatial node in the Y-axis direction of the three-dimensional coordinate system; Indicates the first The displacement components of a spatial node in the Z-axis direction of the three-dimensional coordinate system.

[0032] The range and unit of the above displacement components are usually determined based on the monitoring accuracy of the actual geological engineering, specifically set to the millimeter or centimeter level. The spatial discretization module integrates the spatial position coordinates of multiple discrete point cloud layers with the corresponding three-dimensional deformation displacement scalars to generate a basic deformation data field that characterizes the overall spatial deformation distribution of the landslide.

[0033] See attached document Figure 2 For step S20 of the landslide deformation real-time assessment method based on digital twins and microservices, the specific steps are as follows: The asynchronous scheduling module reads the system logic code and dependency library files, builds an algorithm container image, distributes the algorithm container image to the computing cluster nodes to start the instance application, generates data processing tasks according to the timed trigger instructions, and calls the instance application to extract the three-dimensional deformation displacement scalar and perform calculations. The specific implementation process includes the following sub-steps: S210, the asynchronous scheduling module reads the system logic code and dependent library files to build an algorithm container image. The system logic code includes deformable data preprocessing scripts, multi-scale spatial clustering algorithm scripts, and machine learning model inference scripts. The dependent library files contain the underlying mathematical calculation framework and network communication components required for the system logic code to run. The asynchronous scheduling module packages and isolates the system logic code, dependent library files, and system runtime environment variables according to the preset environment configuration file to generate a standardized algorithm container image.

[0034] The environment configuration file defines the build layers of the algorithm container image. Logically, the algorithm container image is divided into a base layer, a dependency layer, and an application layer. The base layer contains the operating system runtime and programming language interpreter, the dependency layer contains the underlying mathematical computation framework, and the application layer contains the system logic code. This layered build principle physically isolates the complex security assessment algorithm from its dependent environment, preventing algorithm updates from interfering with the underlying system. The syntax structure of the environment configuration file and the image packaging and compilation process can be implemented using conventional Docker build tools by those skilled in the art. Image compilation and packaging are well-known technologies in this field and will not be elaborated here.

[0035] In S220, the asynchronous scheduling module distributes the algorithm container image to the computing cluster nodes to start instance applications. The asynchronous scheduling module pushes the completed algorithm container image and centrally stores it in the system's private image repository. The container daemon deployed on the target computing cluster node pulls the algorithm container image from the private image repository. The container daemon allocates independent computing memory resources and network communication ports on the target computing cluster node, and instantiates the algorithm container image into an independently running container application entity. The asynchronous scheduling module configures a service call interface for the container application entity. The service call interface is represented by an application programming interface protocol, which is used to establish a standardized data transmission and instruction issuance channel between the system main process and the container application entity.

[0036] S230, the asynchronous scheduling module generates data processing tasks according to timed trigger instructions. The asynchronous scheduling module is internally configured with a time scheduling engine, which continuously monitors the system runtime and the frequency of external monitoring data access. The time scheduling engine determines whether to generate a processing task based on the time difference condition. The time trigger condition expression is as follows: ; in, Indicates the current monitoring time point of the system; Indicates the time point of the last execution of the background asynchronous data processing task; This represents the preset time interval threshold. Based on the historical deformation evolution rate of the landslide body and the acquisition frequency of external sensors, the interval is usually set to 12 to 24 hours in engineering applications.

[0037] When the above time-triggered condition expression is met, or when new 3D deformation displacement scalar data is detected being written to the underlying database, the time scheduling engine triggers a timed trigger command. The asynchronous scheduling module generates a background asynchronous data processing task containing the underlying data index identifier based on the timed trigger command. The asynchronous scheduling module pushes the background asynchronous data processing task to the system's task message queue, where it is in a waiting scheduling state.

[0038] S240, the asynchronous scheduling module calls the instance application to extract the 3D deformation displacement scalar and perform calculations. The asynchronous scheduling module extracts the background asynchronous data processing task at the top of the sequence from the task message queue. The asynchronous scheduling module sends a calculation request signal and the data access route of the basic deformation data field to the container application entity in the running state through the service call interface. After receiving the calculation request signal, the container application entity extracts the 3D deformation displacement scalar of the multi-layer 3D point cloud set in the independent container running environment according to the data access route, and executes the internal security assessment and abnormal feature identification calculation logic. The application container isolation execution mechanism avoids the massive point cloud clustering operation from occupying the computing resources of the main system program. The main system control process continues to respond to the front-end user interaction operation commands during the calculation cycle of the container application entity's execution calculation logic, realizing asynchronous decoupling of the high-intensity algorithm calculation process and the main platform operation process.

[0039] See attached document Figure 2 For step S30 in the real-time landslide deformation assessment method based on digital twins and microservices, the following steps are specifically performed: The safety assessment module performs normalization processing on the three-dimensional deformation displacement scalar and calculates the local displacement gradient at different levels according to the spatial neighborhood distance. The specific implementation process includes the following sub-steps: S310, the safety assessment module performs normalization processing on the three-dimensional deformation displacement scalars. The module extracts the three-dimensional deformation displacement scalars of each spatial node within a multi-layered three-dimensional point cloud set from the basic deformation data field. Due to the different boundary constraints experienced by the landslide medium at different geological depth levels, there is a significant difference in the magnitude of the original three-dimensional deformation displacement scalars between deep structures and shallow surfaces. Directly inputting data with such magnitude differences into subsequent clustering algorithms can easily lead to model fitting bias in high-value regions. The safety assessment module employs a data standardization transformation principle to perform independent data scaling preprocessing on the point cloud data at each level. This data processing mechanism converts absolute displacement values ​​into relative changes deviating from the average deformation level of the corresponding level, eliminating the dimensional influence of spatial location and depth. The safety assessment module calculates the arithmetic mean and standard deviation of the three-dimensional deformation displacement scalars for specific discrete point cloud levels. The module also calculates the normalized displacement characteristics using the following mathematical formula: ; in, Indicates the first Normalized displacement characteristics of spatial nodes after data standardization processing; Indicates the first Three-dimensional deformation displacement scalar of each spatial node; This represents the arithmetic mean of the three-dimensional deformation displacement scalars of all spatial nodes within the target discrete point cloud layer; It represents the standard deviation of the three-dimensional deformation displacement scalar of all spatial nodes within the target discrete point cloud layer.

[0040] S320, the safety assessment module calculates the local displacement gradient at different levels according to the spatial neighborhood distance. The absolute displacement value of a single node cannot fully reflect the shear displacement phenomenon that occurs on the internal structural surface of the landslide. The safety assessment module uses the spatial relative deformation rate to characterize the degree of deformation change inside the landslide medium. In order to cope with the retrieval and computing pressure of large-scale point cloud data, the safety assessment module reads the spatial location coordinates of multi-layer three-dimensional point cloud sets to construct a KD Tree (tree data structure) spatial index structure.

[0041] The KD Tree spatial index structure calculates the data variance along different coordinate axes and vertically divides the three-dimensional physical space downwards along the median node of the dimension with the largest variance, generating a binary tree network that supports fast nearest neighbor retrieval. The security assessment module establishes three-dimensional spatial topological relationships between nodes based on the KD Tree spatial index structure. Using each spatial node as the center of a sphere, a three-dimensional spherical search space is constructed with a preset spatial search radius. For the value parameter of the spatial search radius, those skilled in the art can dynamically set it according to the average spacing of the three-dimensional grid of the discrete point cloud. The value range of the spatial search radius is usually set to 2 to 3 times the average spacing of the point cloud grid to ensure that the three-dimensional spherical search space contains sufficient structural calculation reference nodes.

[0042] S330, the safety assessment module calls the search interface of the KD Tree spatial index structure to extract all neighboring spatial nodes covered within the three-dimensional spherical search space to form a spatial neighborhood point set. The safety assessment module retrieves the three-dimensional deformation displacement scalar and three-dimensional spatial coordinates of each neighboring spatial node within the spatial neighborhood point set. The safety assessment module calculates the absolute value of the displacement difference between the target spatial node and each neighboring spatial node, and calculates the three-dimensional Euclidean distance between the target spatial node and each neighboring spatial node. The safety assessment module divides the absolute value of the displacement difference by the three-dimensional Euclidean distance to calculate the local node deformation gradient rate. The safety assessment module performs an arithmetic mean on the deformation gradient rates of all local nodes within the three-dimensional spherical search space to obtain the local displacement gradient. The mathematical calculation formula is as follows: ; in, Indicates the first Local displacement gradient of each spatial node; Indicates the first The set of spatial neighborhood points corresponding to each spatial node within the spatial search radius; This represents the total number of nodes contained in the spatial neighborhood point set; Represents the first point in the spatial neighborhood set. One neighboring spatial node; Indicates the first Three-dimensional deformation displacement scalar of each spatial node; Indicates the first Three-dimensional deformation displacement scalar of a neighboring spatial node; Indicates the first The spatial node and the first The three-dimensional Euclidean distance scalar between neighboring spatial nodes.

[0043] Local displacement gradient As a core spatial characteristic index characterizing the intensity of non-uniform deformation within a landslide, this index is equivalent in geophysical terms to an approximate value of shear strain in a local area. It represents the local displacement gradient when the local deformation rate changes significantly over a short distance. The sudden increase in values ​​reveals the spatial location of potential sliding surfaces and structural weakening zones.

[0044] S340, the safety assessment module constructs a multi-layer average displacement gradient sequence based on local displacement gradients. Since the multi-layer 3D point cloud set exhibits high-dimensional discrete spatial distribution characteristics, directly inputting it into a machine learning model would result in excessively high feature dimensions. Therefore, the safety assessment module applies the K-means unsupervised clustering algorithm within each discrete point cloud layer of the multi-layer 3D point cloud set. To improve clustering efficiency, the module uses a pre-built KD Tree spatial index structure to obtain sample nearest neighbor relationships during the cluster center distance search phase of the K-means unsupervised clustering algorithm. Based on the numerical differences in the 3D deformation displacement scalars, the module divides each discrete point cloud layer into multiple deformation clustering units with consistent internal displacement changes. The module extracts the local displacement gradients of all spatial nodes within each deformation clustering unit. It calculates the arithmetic mean of all local displacement gradients within the deformation clustering unit, extracting the corresponding level's level average displacement gradient index. Finally, the module concatenates the level average displacement gradient indices of multiple discrete point cloud layers according to depth order, generating a structured multi-layer average displacement gradient sequence.

[0045] The S350 safety assessment module reads a pre-built mechanical mapping model. This model employs a gradient boosting machine architecture. Traditional numerical mechanics simulation methods are computationally expensive and rely on idealized physical parameter assumptions. The pre-built mechanical mapping model uses machine learning mechanisms to establish a mapping channel between the actual deformation response characteristics of multiple layers of landslides and the finite element mechanical state at the data structure level, thereby achieving real-time safety assessment under complex conditions. The training sample data for the pre-built mechanical mapping model comes from the finite element strength reduction calculation results. The system pre-constructs a finite element model sample set covering different combinations of geotechnical mechanical parameters.

[0046] The finite element model proportionally reduces the cohesion and internal friction angle parameters of the soil and rock mass using a uniform strength reduction factor. The system records the multi-layer finite element deformation field data and the critical strength reduction factor when the finite element model reaches the critical instability state. The system extracts the finite element average displacement gradient sequence from the multi-layer finite element deformation field data using clustering statistics as training input feature samples. The system uses the critical strength reduction factor as the training output label sample. The system iteratively trains the gradient booster model using the training input feature samples and the training output label samples. The gradient booster model employs a progressive additive modeling mechanism, fitting the residual data of the previous round's prediction calculation results in each algorithm iteration, continuously optimizing the model towards reducing the overall loss function, and establishing a quantitative nonlinear mapping relationship between the overall deformation state of the landslide and the strength reduction safety level.

[0047] The S360 safety assessment module inputs the multi-layer average displacement gradient sequence into the trained gradient booster model for inference calculation. The gradient booster model contains multiple independent regression decision trees. It assigns the multi-layer average displacement gradient sequence to each regression decision tree for node splitting and numerical mapping calculations, deriving the output prediction residuals of each regression decision tree. The gradient booster model then sums and accumulates these output prediction residuals according to preset weight parameters to calculate the global equivalent safety factor of the landslide. The mathematical formula for calculating the global equivalent safety factor of the landslide is as follows: ; in, This represents the global equivalent safety factor for a landslide. This represents the initial prediction baseline constant of the gradient booster model; This represents the total number of regression decision trees contained within the gradient booster model; Indicates the index of the regression decision tree; This represents the multi-level average displacement gradient sequence input to the gradient booster model; Indicates the first The regression decision tree outputs the predicted residuals based on the multi-level average displacement gradient sequence; Indicates the first The learning rate and weight parameters corresponding to each regression decision tree. The value range is set to between 0.01 and 0.1. Setting a smaller learning rate weight parameter can prevent overfitting in the gradient booster model.

[0048] Global equivalent safety factor of landslide This reflects the equivalent stability level of the landslide body under its current actual deformation response characteristics. The safety assessment module is based on the global equivalent safety factor of the landslide. The system outputs a risk level sequence based on the numerical values. The threshold for dividing the risk level sequence is obtained through confidence interval analysis of historical equivalent safety factor calculation samples. The safety assessment module sets the minimum physical baseline lower limit of the global equivalent safety factor for landslides to 1.0. The module extracts the 97.5th percentile value after sorting the historical equivalent safety factor calculation samples as the robust upper limit to eliminate the interference of extreme outliers on the interval division. The safety assessment module divides the area between the minimum physical baseline lower limit and the robust upper limit into four numerical intervals based on quantiles, corresponding to high-risk, medium-risk, low-risk, and no-risk levels, respectively.

[0049] See attached document Figure 3For step S40 of the landslide deformation real-time assessment method based on digital twins and microservices, the following steps are specifically performed: The anomaly identification module performs spatial clustering operations on spatial nodes of multiple discrete point cloud layers and local displacement gradients to delineate the boundaries of deformation active regions. A hierarchical clustering network based on core distance is applied to the deformation active regions to isolate non-consistent abnormal deformation units. The specific implementation process includes the following sub-steps: S410, the anomaly detection module performs spatial clustering operations on spatial nodes and local displacement gradients at multiple discrete point cloud levels. The module uses K-means unsupervised clustering as the basic architecture for this operation. It extracts the 3D deformation displacement scalars and local displacement gradients of all spatial nodes at multiple discrete point cloud levels as clustering input features. Based on the numerical similarity of these input features, the module divides the 3D space of the landslide body into multiple initial clusters. It calculates the average displacement gradient value of each initial cluster, designating clusters with average displacement gradient values ​​lower than the overall median as stable regions, and clusters with average displacement gradient values ​​higher than or equal to the median as stable regions. The initial clusters equal to the overall median are defined as the deformation active regions. The spatial nodes in the stable regions have continuous planar distribution characteristics, and the local displacement gradient is in the preset low level range. The spatial nodes in the deformation active regions are mostly clustered at the bottom of the slope and the two side edges, and the local displacement gradient shows obvious abrupt increase. The anomaly identification module extracts the spatial boundary coordinates of the deformation active regions and filters out the spatial nodes in the stable regions to achieve a coarse division of the boundary of the overall deformation field of the landslide. For the center iteration optimization process of the K-means unsupervised clustering algorithm, those skilled in the art can use the conventional minimum square error criterion to achieve it. The iterative update of cluster centers is a well-known technology in this field and will not be described in detail here.

[0050] S420, the anomaly identification module extracts mesoscopic dimensionality reduction features within the deformation-active region. The module summarizes the original three-dimensional deformation displacement scalars, local displacement gradients, and local displacement variances of spatial nodes within the deformation-active region to form an initial high-dimensional feature space. Since the initial high-dimensional feature space contains redundant data dimensions, directly inputting it into subsequent density clustering algorithms would cause distance metric distortion. Therefore, the anomaly identification module applies principal component analysis (PCA) to the initial high-dimensional feature space. It calculates the covariance matrix of the initial high-dimensional feature space and performs eigenvalue decomposition on the covariance matrix, obtaining all eigenvalues ​​and eigenvectors. The module calculates the cumulative variance contribution rate in descending order of eigenvalues ​​and extracts eigenvectors with a cumulative variance contribution rate greater than a preset contribution threshold as principal components. The preset contribution threshold is set between 90% and 95% to ensure that most of the original deformation information is retained. PCA projects the initial high-dimensional feature space onto a minimum principal feature dimension. This minimum principal feature dimension primarily retains the core variables characterizing the deformation amplitude level and the rate of change in deformation space, thus completing the mesoscopic dimensionality reduction processing of the deformation-active region.

[0051] The S430 anomaly detection module applies a core distance-based hierarchical clustering network to the active deformation region. This network employs a density-based noise-applied spatial clustering algorithm architecture (HDBSCAN). Based on the principal feature dimension after mesoscopic dimensionality reduction, the anomaly detection module calculates the reachability distance between any two spatial nodes within the active deformation region. This reachability distance overcomes the parameter sensitivity issue of traditional density-based clustering algorithms. Physically, it is equivalent to establishing a density buffer zone between spatial nodes. When a node is in a sparse region, the calculated distance is forcibly amplified to the core distance to prevent sparse noise nodes from being incorrectly connected to high-density anomalous deformation clusters. The mathematical expression for calculating the reachability distance is as follows: ; in, Representing spatial nodes With spatial nodes The mutual reachability distance between them; This parameter represents the preset number of core nodes in the neighborhood. The value range is usually set between 10 and 50. The specific value is determined based on the overall grid density of the landslide point cloud data. The higher the density, the larger the value. Representing spatial nodes Core distance, Defined as a spatial node Sort by distance Euclidean distance between each neighborhood core node; Representing spatial nodes Core distance; Representing spatial nodes With spatial nodes The original three-dimensional Euclidean distance between them. The mutual reachability distance mechanism increases the spatial distance between nodes in the low-density background region, enhancing the cohesive distribution characteristics of the high-abnormal deformation node cluster.

[0052] S440, the anomaly identification module isolates non-consistent abnormal deformation units based on their mutual reachability. The anomaly identification module uses spatial nodes within the deformation active region as graph vertices and mutual reachability as the connection weights of graph edges to construct a minimum spanning tree containing all graph vertices. The anomaly identification module gradually reduces the connections in the minimum spanning tree according to the connection weights of graph edges from large to small, forming a hierarchical clustered tree structure from dense to sparse. The anomaly identification module performs cluster stability measurement on the hierarchical clustered tree structure according to a preset lower limit threshold for the number of nodes within a cluster. The lower limit threshold for the number of nodes within a cluster is set according to the grid resolution of the landslide digital twin model, and is set to no less than 500 spatial nodes to avoid misjudgment of random disturbances caused by excessively small point clusters.

[0053] The anomaly identification module defines the lifecycle of a cluster as the density distribution span from when the cluster splits from the parent cluster to when it splits again or disappears. The anomaly identification module integrates the density change difference of all spatial nodes within the cluster during the lifecycle to obtain a numerical cluster stability metric. The anomaly identification module extracts the subtree branch with the highest stability metric as the independent optimal cluster and identifies low-density scattered spatial nodes outside the optimal cluster as environmental noise. The extracted independent optimal cluster is the non-uniform anomaly deformation unit. The non-uniform anomaly deformation unit reveals dangerous areas of weakened sliding body strength and stress concentration redistribution under a unified active macro background.

[0054] S450, the anomaly identification module extracts local anomaly detection features within the non-uniform anomaly deformation unit. The non-uniform anomaly deformation unit reflects deformation anomalies at the regional scale. However, within the non-uniform anomaly deformation unit, the deformation evolution characteristics of different spatial nodes are significantly different. A small number of abrupt spatial nodes often correspond to the physical locations of shear zone initiation and structural surface penetration, which are microscopic precursors of landslide evolution from continuous deformation to discontinuous failure. The anomaly identification module extracts the three-dimensional deformation displacement scalar and local displacement gradient of all spatial nodes within the non-uniform anomaly deformation unit. The anomaly identification module combines the three-dimensional deformation displacement scalar and local displacement gradient to form the input feature space for local anomaly detection.

[0055] S460, the anomaly identification module calculates the local reachability density of each spatial node within the non-uniform anomaly deformation unit. The module employs a local anomaly factor algorithm to extract microscopic features. For a target spatial node, the module searches for a preset number of neighboring spatial nodes to form a nearest neighbor set, based on a preset parameter for the number of core neighboring nodes. Sure, The value range is set to 10 to 50 to ensure that the nearest neighbor set contains sufficient background deformation reference nodes, while avoiding the introduction of too many irrelevant distant nodes. The anomaly detection module calculates the local reachability distance between the target spatial node and each neighboring spatial node in the nearest neighbor set. The calculation logic for the local reachability distance is to select the first neighboring spatial node. - The maximum value between the distance and the actual 3D Euclidean distance is used to avoid infinite computational overflow in density calculations caused by spatial nodes being extremely close or even overlapping in the feature space. The anomaly detection module calculates the arithmetic mean of all locally reachable distances within the nearest neighbor set and defines the reciprocal of the summation as the local reachability density. The mathematical expression for calculating the local reachability density is as follows: ; in, Indicates the first spatial nodes Locally achievable density; Indicates the first spatial nodes The corresponding set of nearest neighbors; This represents the total number of nodes in the nearest neighbor set; Represents the first point in the spatial neighborhood set. One neighboring spatial node; Indicates the first Neighboring spatial nodes The k-th distance, i.e., the spatial node Its first The three-dimensional Euclidean distance between each neighboring core node; Indicates the first spatial nodes With the Neighboring spatial nodes The three-dimensional Euclidean distance between them.

[0056] S470, the anomaly identification module calculates the local anomaly factor of each spatial node within the non-uniform anomaly deformation unit. Based on the calculated local reachability density, the module measures the outlier degree of the target spatial node in the local spatial structure. The module extracts the local reachability density of all neighboring spatial nodes within the nearest neighbor set. It then calculates the ratio of the local reachability density of each neighboring spatial node to the local reachability density of the target spatial node. Finally, the module performs an arithmetic average of all ratios to obtain the local anomaly factor of the target spatial node. The mathematical expression for calculating the local anomaly factor is as follows: ; in, Indicates the first spatial nodes Local anomalous factors; Indicates the first Neighboring spatial nodes Locally achievable density; Indicates the first spatial nodes The locally achievable density.

[0057] Local abnormal factors In a physical sense, it characterizes the degree to which a spatial node deviates from the local background deformation trend. If the spatial node's deformation trend is consistent with that of its surrounding nodes in the displacement and gradient feature space, its local reachability density is similar to that of its neighboring nodes, and its local anomaly factor is close to 1.0. If a spatial node undergoes sudden deformation, it will move away from the neighboring node group in the feature space, causing its local reachability density to drop sharply, and its local anomaly factor to be significantly greater than 1.0.

[0058] S480, the anomaly identification module filters the set of coordinate points that deviate from the predetermined density as the high displacement anomaly feature point cluster. The anomaly identification module configures a predetermined local anomaly factor judgment threshold. The local anomaly factor judgment threshold is set according to the normal deformation statistical dispersion in the landslide feature space. The set value is 1.1. This value defines the boundary between normal measurement fluctuation error and significant abnormal physical mutation. The anomaly identification module performs numerical comparison calculations between the local anomaly factors of all spatial nodes and the local anomaly factor judgment threshold.

[0059] When the local anomaly factor of a spatial node is greater than or equal to 1.1, it indicates that the local reachability density of this spatial node is significantly lower than the average density level of the surrounding neighborhood, showing obvious isolation and abrupt change characteristics. The anomaly identification module extracts spatial nodes whose local anomaly factor is greater than or equal to the local anomaly factor judgment threshold, filters and discards ordinary background nodes whose local anomaly factor is less than the local anomaly factor judgment threshold, and aggregates and marks the extracted spatial node set as a high displacement anomaly feature cluster. The high displacement anomaly feature cluster realizes the micro-focusing of the non-uniform anomaly deformation unit at the regional scale to the point cluster scale, reveals the distribution location of the key nodes most likely to fail and slide first inside the landslide, and completes the refined extraction mechanism of macro-micro nested features.

[0060] See attached document Figure 4 For step S50 of the landslide deformation real-time assessment method based on digital twins and microservices, the following steps are specifically performed: The asynchronous scheduling module stores the calculation results in the system's structured database as a preheating cache data disk persistence mechanism, and starts a bidirectional communication channel to transmit the streaming algorithm calculation results to the client without request delay, completing the communication construction process for dynamic rendering of the digital twin 3D scene. The specific implementation process includes the following sub-steps: S510, the asynchronous scheduling module executes algorithm calculation tasks and generates calculation results. The asynchronous scheduling module encapsulates the landslide global equivalent safety factor calculation process and the high displacement anomaly feature point cluster extraction process into independent background calculation tasks. Density clustering operations and multi-dimensional feature dimensionality reduction operations of multi-layer 3D point cloud sets have high computational complexity. When the scale of discrete point cloud nodes is large, using a synchronous calculation mechanism to respond to frontend calls will lead to significant network timeouts. Therefore, the asynchronous scheduling module deploys a timed scheduling engine within the background system. The timed scheduling engine triggers the asynchronous execution of the independent background calculation tasks at preset time nodes (e.g., midnight each day or after the latest displacement monitoring data is aggregated).

[0061] In S520, the asynchronous scheduling module stores the calculation results in the system's structured database to build a preheating cache. After the background independent calculation task is completed, the asynchronous scheduling module summarizes and extracts the calculation results such as the landslide global equivalent safety factor, risk level sequence, boundary parameters of non-consistent abnormal deformation units, and three-dimensional spatial coordinates of high displacement abnormal feature point clusters. For data structures of different dimensions, the asynchronous scheduling module adopts a hybrid storage strategy. The asynchronous scheduling module uses the timestamp as the primary key index of the database and writes the landslide global equivalent safety factor and risk level sequence as standard numerical fields directly into the structured data table.

[0062] For the three-dimensional spatial coordinates and local anomaly factors of high-dimensional variable-length high-displacement anomaly feature point clusters, the asynchronous scheduling module serializes them into JSON format text blocks or binary data blocks and stores them in the database in the form of large object fields. The data persistence mechanism transforms the time-consuming instantaneous calculation process into a static data storage process, forming a preheating cache that supports the rapid loading of digital twin three-dimensional scenes. Under normal non-emergency conditions, when a client initiates an access request, the asynchronous scheduling module directly retrieves and reads the corresponding preheating cache data from the system's structured database and returns it to the front-end engine, avoiding the system from repeatedly starting the inference calculation of the algorithm model and realizing the immediate use of security assessment data.

[0063] The S530 dynamic rendering module establishes a full-duplex communication environment by initiating a bidirectional communication channel. In emergency response scenarios where landslides experience sudden and intensified deformation, fixed-period preheating cache data cannot cover the instantaneously changing latest deformation state of the landslide. Traditional front-end and back-end interaction models rely on the client periodically sending polling requests using the Hypertext Transfer Protocol (HTTP) to obtain updated data. This polling communication mechanism not only consumes a large amount of system network bandwidth but also leads to unavoidable response delays in the data acquisition process. In the traditional HTTP protocol, each data interaction requires carrying a large data packet containing request header information, along with the resource overhead of connection establishment and release. The dynamic rendering module establishes a full-duplex communication environment between the computing server and the client. The bidirectional communication channel of the communication technology adopts the WebSocket protocol (Network Socket Protocol) architecture. The dynamic rendering module receives the protocol upgrade network request sent by the client and completes the protocol handshake verification on the computing server side, thereby establishing and maintaining an uninterrupted persistent network communication link. During the link maintenance, the computing server can push information containing a very small data frame header directly to the client at any time, reducing the network transmission load. For the handshake verification of the WebSocket protocol and the underlying data packet forwarding mechanism of the Hypertext Transfer Protocol, those skilled in the art can use standard network transmission architecture to implement it. The method of establishing the network communication link is a well-known technology in this field and will not be described in detail here.

[0064] In the S540, the dynamic rendering module transmits streaming algorithm calculation results to the client without request delay. When the emergency calculation task completes the identification of multi-layer point cloud anomaly areas and generates the latest algorithm calculation results in the background, the dynamic rendering module bypasses the limitations of traditional client query commands and directly utilizes the pre-established persistent bidirectional communication channel to encapsulate and reorganize the algorithm calculation results, which include point cloud spatial coordinates, deformation displacement scalar values, and local anomaly factor identifiers, into streaming data packets. To adapt to the parsing speed of the front-end 3D rendering engine, the dynamic rendering module performs lightweight encoding processing on the point cloud spatial coordinates, truncates redundant decimal places of floating-point numbers, and integrates the packet content according to the key-value pair format of coordinate point sequence number, coordinate value, and anomaly feature. The dynamic rendering module actively pushes the streaming data packets to the corresponding digital twin 3D scene interface on the client. The communication construction process enables the client front-end engine to continuously receive the evaluation data transmitted by the computing server in an uninterrupted pipeline mode, supporting the dynamic real-time rendering of the model's appearance structure in the digital twin 3D scene without page refresh, ensuring that the actual deformation evolution of the landslide physical entity and the visual feedback of the digital twin 3D scene are strictly synchronized.

[0065] In the S550, the dynamic rendering module parses streaming data packets and plots safety factor change curves. The client-side front-end engine of the dynamic rendering module continuously monitors the bidirectional communication channel. When it receives streaming data packets pushed by the computing server, the dynamic rendering module decodes the streaming data packets, extracts the current assessment timestamp, the global equivalent safety factor value of the landslide, and the corresponding risk level. The dynamic rendering module correlates the global equivalent safety factor value of the landslide with the assessment timestamp and inputs it as a new two-dimensional coordinate node into the chart component of the front-end interface. The chart component dynamically adds two-dimensional coordinate nodes and connects them with line segments on the basis of the original historical data curve, generating a safety factor change curve that progresses with the time axis. Based on the risk level settings, when the dynamic rendering module detects that the risk level has crossed from a low level to a high level, it synchronously changes the color of the rendered lines of the safety factor change curve for that time period. The dynamic rendering module sets a mapping matrix from risk level to color: no risk level corresponds to green, low risk level corresponds to yellow, medium risk level corresponds to orange, and high risk level corresponds to red. This mapping rule provides engineering managers with an intuitive view of the evolution of safety status.

[0066] In the S560, the dynamic rendering module performs a mapping transformation of the 3D coordinate system. The 3D spatial coordinates of the high-displacement anomaly feature point clusters generated by the algorithm are established based on the local relative mathematical space of the landslide body. To ensure that the anomaly feature point clusters correspond to actual geographical locations in the digital twin 3D scene, the dynamic rendering module uses an affine transformation matrix to convert the local spatial coordinates of the high-displacement anomaly feature point clusters into 3D world coordinates that can be recognized by the front-end 3D rendering engine. The dynamic rendering module expands the 3D coordinates into a four-dimensional homogeneous coordinate structure. The introduction of homogeneous coordinates allows the rotation and translation operations, which originally required separate matrix multiplication and addition, to be combined into a single matrix multiplication operation, improving the efficiency of the 3D rendering engine's underlying calculation of vertex transformations. The mathematical expression formula for performing spatial coordinate mapping transformation is as follows: ; in, , , These represent the X-axis, Y-axis, and Z-axis coordinate components in the transformed 3D world coordinate system, respectively. , , Let X, Y, and Z represent the local X-axis, Y-axis, and Z-axis spatial coordinate components of the high displacement anomaly feature point cluster input to the algorithm, respectively; the fourth-order matrix on the right side of the equation is the affine transformation matrix. , , , , , , , and The three-dimensional rotation parameter components in this fourth-order matrix are determined by the initial spatial tilt angle and attitude when the landslide 3D model is imported into the scene engine. , , The translation parameter vector components are 3D translation parameter vector components. The values ​​of the translation parameter vector components are determined by the actual measured offset distance between the virtual geospatial origin of the digital twin 3D scene and the local coordinate origin of the landslide.

[0067] In the S570, the dynamic rendering module overlays marker tiles at corresponding positions in multiple discrete point cloud layers and performs scene rendering updates. High displacement anomaly feature clusters are distributed in various depth layers within the landslide 3D model. Conventional opaque surface rendering modes would obscure the internal structural layer features. The dynamic rendering module enables landslide body transparency processing and internal sectioning processing operations in the digital twin 3D scene. The dynamic rendering module performs mixed rendering of the alpha channel of the landslide body surface mesh material according to the pre-set transparency parameters. The transparency parameter is set between 0.3 and 0.6 to ensure that the overall outer boundary contour of the landslide body can be observed, while allowing the line of sight to penetrate to the discrete point cloud layer inside the landslide.

[0068] S580, the dynamic rendering module performs the overlay identifier tile generation operation. Based on the transformed 3D world coordinates, the dynamic rendering module instantiates a 3D spatial node model within the depth space of the internal discrete point cloud layer corresponding to the digital twin 3D scene. The dynamic rendering module extracts the local anomaly factor identifier values ​​attached to the streaming data packets, establishes a heatmap gradient mapping rule between the anomaly values ​​and the rendering color space, and sets upper and lower thresholds for truncating the local anomaly factor values. For the lower threshold, it sets a preset local anomaly factor judgment threshold; for the upper threshold, it dynamically proposes... The 95th percentile of the local anomaly factor sequence in the current batch of high displacement anomaly feature point clusters is taken as the upper threshold for truncation. The dynamic rendering module converts the anomaly values ​​in this range into a normalization ratio of 0 to 1. The dynamic rendering module maps this normalization ratio to the hue channel of the HSV color model. The lower the normalization ratio value, the more the mapped hue tends to be cool colors such as blue. The higher the normalization ratio value, the more the mapped hue tends to be warm colors such as red. The dynamic rendering module assigns the corresponding color material to the instantiated three-dimensional spatial node model according to the heatmap gradient mapping rules, generating a highlighted marker block.

[0069] The dynamic rendering module overlays the generated highlighted marker tiles onto the non-consistent abnormal deformation areas at each level inside the landslide model. Through a collaborative rendering method that combines a transparent surface layer with internal highlighted color blocks, it achieves the goal of exposing potential instability and failure points inside the landslide to management personnel from a three-dimensional perspective.

[0070] Specific application examples: To verify the effectiveness and feasibility of this invention, this invention targets the K3 landslide, which shows signs of reactivation in a reservoir area. The system has completed the construction of a three-dimensional geological model and a digital twin base. This embodiment illustrates a complete asynchronous scheduling and evaluation cycle during system operation: Spatial discretization and construction of basic deformation data fields: The system extracts target spatial nodes from the shallow point cloud set of the landslide body. By integrating external surface monitoring equipment (such as interferometric synthetic aperture radar) with internal structural deformation simulation models, the three-dimensional coordinate system displacement components of this node at the current timestamp are calculated as follows: , , .

[0071] Based on the mathematical expression formula for calculating the scalar displacement of three-dimensional deformation, the target space nodes are calculated. Three-dimensional deformation displacement scalar : ; Safety assessment calculations: System Extraction The arithmetic mean of the scalar values ​​of the three-dimensional deformation displacement of all spatial nodes within the target discrete point cloud layer is calculated from the shallow point cloud dataset. Standard deviation Calculate the normalized displacement characteristic of this node. : ; Based on the KD Tree index structure, nodes containing neighboring spatial nodes are extracted using a preset spatial search radius. The spatial neighborhood point set is known. Three-dimensional deformation displacement scalar The three-dimensional Euclidean distance between the two Assuming there is only one reference point within the search radius ( Substitute the values ​​into the formula for calculating the local displacement gradient to calculate... : ; This value is significantly higher than the normal deformation background value, indicating that shear displacement is occurring in this local area.

[0072] The system will reduce the dimensionality of the multi-level average displacement gradient sequence. Input a gradient booster model and set the total number of regression decision trees within the model. Initial prediction baseline constant Learning rate and weight parameters , The predicted residual values ​​of the two trees in the model based on the input features are: , Substitute the values ​​into the mathematical formula for calculating the global equivalent safety factor of a landslide: ; The assessment determined that the current global equivalent safety factor of the landslide is 1.15, triggering a medium-risk level warning.

[0073] Anomaly feature point extraction: Entering the microscopic analysis phase of the non-uniform abnormal deformation element, targeting the above-mentioned target space nodes. The system calculates its local reachability density. Simultaneously, the system extracts the arithmetic mean of the local reachability density of all neighborhood nodes within the nearest neighbor set, which is 0.62. This value is then substituted into the mathematical expression for calculating the local anomaly factor: ; Judgment result: The local anomaly factor of 1.55 is greater than the predetermined local anomaly factor judgment threshold of 1.1, therefore... The points were clearly selected and marked as clusters of high displacement anomalies.

[0074] Dynamic rendering and coordinate mapping: In a digital twin engine, it is necessary to... The spatial coordinate components of the X-axis, Y-axis, and Z-axis (defined as) , , Converting to 3D world coordinates, given that the rotation parameter of the affine transformation matrix was set to rotate 90 degrees around the Z-axis when importing the landslide model, and the 3D translation parameter vector components are... , , Substitute the mathematical expression for the spatial coordinate mapping transformation: ; Based on this result, the dynamic rendering module overlays a highlighted marker tile at world coordinates (80, 210, 55) in the digital twin 3D scene.

[0075] Experimental verification and effect comparison: To verify the effectiveness of the embodiments of the present invention and overcome the defects of the prior art, a comparative experiment was conducted between the landslide safety rapid assessment system based on microservices and containerization provided by the present invention and the traditional monolithic monitoring platform under the same data flow and hardware environment. The data results are shown in the accompanying drawings.

[0076] System concurrent processing latency comparison test: See attached document Figure 5 The figure shows the average system response latency when the system processes massive 3D point clouds and performs mechanical calculations under different concurrent request volumes (0 to 5000 requests / second).

[0077] Traditional monolithic monitoring platforms (dashed lines in the diagram): Because the core algorithm is not isolated from the operating environment, massive point cloud clustering competes for main program resources. When the concurrent request volume reaches 1000 times / second, the latency begins to increase exponentially on a quadratic scale; when the concurrency reaches 3000 times / second, the response latency has soared to nearly 4500ms; when the concurrency exceeds 3500 times / second, the latency directly exceeds 6000ms (screen stuttering and timeout packet loss occur).

[0078] The evaluation system of this invention (solid line in the figure): thanks to the asynchronous scheduling module decoupling the computing task container and the WebSocket streaming bidirectional communication built by the dynamic rendering module, the curve of this system always runs close to the horizontal axis throughout the entire concurrent test range of 0 to 5000 times / second. The system response latency is extremely stable, always remaining on the warm-up cache read baseline of 120ms to 150ms, eliminating the concurrent blocking problem.

[0079] Comparison of anomaly feature recognition accuracy: See attached document Figure 6 The figure shows the system’s ability to identify small abnormal mutation points (non-uniform abnormal deformation units) inside the landslide through ROC curves (receiver operating characteristic curves). The horizontal axis represents the false positive rate (false alarm rate), and the vertical axis represents the true positive rate (recall rate / hit rate).

[0080] Traditional threshold alarm (dotted line in the figure): relies on a single absolute displacement for judgment, and cannot identify microscopic precursors before the structural surface is penetrated. Its curve rises slowly, and the area under the curve (AUC) is only about 0.72.

[0081] The method of this invention (solid line in the figure): This invention uses multi-layer K-means unsupervised clustering superimposed with a hierarchical clustering network based on core distance (HDBSCAN), and finally uses the local anomaly factor (LOF) to sink down to the micro-point cluster for judgment. As shown in the figure, its curve is relatively steep in the initial stage. When the false positive rate (FPR) is only 0.05 (i.e., an extremely low false positive rate of 5%), the true positive rate (TPR) has rapidly climbed to about 0.89, and the overall AUC reaches 0.94, which is significantly better than the traditional method, confirming the high sensitivity of this algorithm to capturing anomalies in complex backgrounds.

[0082] Global equivalent safety factor accuracy verification: See attached document Figure 7 The figure verifies the computational accuracy of the machine learning model (gradient booster) of this invention when replacing traditional physical mechanics simulation, and captures the data trend of landslides at 10 consecutive evaluation time points.

[0083] Finite element strength reduction method: As a reference true value, its coefficient steadily decreases from about 1.22 on day 1 to about 1.12 on day 10 over time, but a single calculation takes up to several hours and cannot be used for real-time rendering.

[0084] The gradient booster model / invention (solid line with a circular marker in the figure): Performs second-level inference calculations using a pre-built mechanical mapping model. As shown in the figure, the predicted data points of this invention (e.g., 1.23 on day 1, 1.15 on day 6, and 1.12 on day 9) fluctuate closely around the standard theoretical solution. Within the evaluation period, the maximum absolute error occurs on day 6 (the difference between 1.17 and 1.15 is 0.02), and the maximum relative error is less than 1.8%.

[0085] Conclusion: Under the premise of ensuring that the visual feedback accuracy of engineering management personnel fully meets the standards, this invention successfully leaps from the offline hour level to the online millisecond / second level in obtaining the safety factor, providing core data support capabilities for the request-free and latency-free transmission of streaming data.

Claims

1. A rapid landslide safety assessment system based on microservices and containerization, characterized in that: include: The spatial discretization module is used to divide the geometric units of the landslide body into a multi-layer three-dimensional point cloud set, and to calculate the three-dimensional deformation displacement scalar of the spatial nodes by integrating displacement information collected by external surface monitoring equipment. The safety assessment module is used to calculate the displacement gradient statistical index of the multi-layer three-dimensional point cloud set based on the three-dimensional deformation displacement scalar, and input the displacement gradient statistical index into the pre-constructed mechanical mapping model to calculate the global equivalent safety factor of the landslide. Anomaly identification module is used to extract deformation active regions from the multi-layer 3D point cloud set, perform point-level filtering on the deformation active regions based on density clustering rules, and output high displacement anomaly feature point clusters and 3D spatial coordinates. The asynchronous scheduling module is used to instantiate the algorithm container image into an independently running container application entity to generate a background asynchronous data processing task, and to store the global equivalent safety factor of the landslide and the three-dimensional spatial coordinates of the high displacement anomaly feature point cluster into the system structured database, and to retrieve and read the global equivalent safety factor of the landslide and the three-dimensional spatial coordinates of the high displacement anomaly feature point cluster from the system structured database. The dynamic rendering module is used to retrieve and read the three-dimensional spatial coordinates of the landslide global equivalent safety factor and the high displacement anomaly feature point cluster from the system's structured database, and perform dynamic updates and overlay rendering in the digital twin three-dimensional scene for real-time evaluation.

2. The landslide safety rapid assessment system based on microservices and containerization according to claim 1, characterized in that, The process of dividing the landslide body into multi-layered three-dimensional point cloud sets and integrating displacement information collected by external surface monitoring equipment to calculate the three-dimensional deformation displacement scalar of spatial nodes specifically includes: The spatial discretization module divides the landslide geometric unit into the multi-layer three-dimensional point cloud set according to the preset three-dimensional grid resolution; The spatial discretization module acquires surface deformation point cloud data obtained from external surface monitoring equipment as the displacement information, extracts the surface three-dimensional displacement vectors corresponding to the spatial nodes on the landslide body surface, and inputs the surface three-dimensional displacement vectors into a pre-constructed internal structural deformation inference model. The internal structural deformation inference model is a deep layered generation variational autoencoder model, which calculates the three-dimensional deformation displacement vectors of each spatial node in the multi-layer three-dimensional point cloud set, and calculates the three-dimensional deformation displacement scalar based on the three-dimensional deformation displacement vectors.

3. The landslide safety rapid assessment system based on microservices and containerization according to claim 1, characterized in that, The process of instantiating the algorithm container image into an independently running container application entity to generate a background asynchronous data processing task specifically includes: The asynchronous scheduling module packages the system logic code into an algorithm container image and distributes it to the computing cluster nodes, instantiates it into an independently running container application entity, and configures the service call interface. The asynchronous scheduling module has a built-in time scheduling engine. When the time difference condition meets the preset time interval threshold, it triggers the generation of a background asynchronous data processing task. It sends a calculation request signal to the container application entity through the service call interface to extract the three-dimensional deformation displacement scalar and perform internal security assessment and anomaly feature identification calculation.

4. The landslide safety rapid assessment system based on microservices and containerization according to claim 1, characterized in that, The calculation of the displacement gradient statistical index of the multi-layer three-dimensional point cloud set based on the three-dimensional deformation displacement scalar specifically includes: The safety assessment module performs data scaling preprocessing on the three-dimensional deformation displacement scalar to obtain normalized displacement features; The safety assessment module extracts a set of spatial neighboring points for the spatial node, calculates the absolute value of the displacement difference between the spatial node and each neighboring spatial node in the set of spatial neighboring points based on the three-dimensional deformation displacement scalar, divides the absolute value of the displacement difference by the three-dimensional Euclidean distance to calculate the local node deformation gradient rate, and performs an arithmetic mean on the local node deformation gradient rate to obtain the local displacement gradient as the displacement gradient statistical index.

5. The landslide safety rapid assessment system based on microservices and containerization according to claim 4, characterized in that, The process of inputting the displacement gradient statistical index into a pre-constructed mechanical mapping model to calculate the global equivalent safety factor of the landslide specifically includes: The safety assessment module divides the multi-layer 3D point cloud set into multiple deformable clustering units using the K-means unsupervised clustering algorithm, extracts each local displacement gradient within the deformable clustering unit as the displacement gradient statistical index, calculates the arithmetic mean of each local displacement gradient, extracts the corresponding level average displacement gradient index, and concatenates and combines the level average displacement gradient indices of multiple discrete point cloud levels in depth order to form a multi-layer average displacement gradient sequence. The mechanical mapping model adopts a gradient booster model architecture, which inputs the multi-layer average displacement gradient sequence into an internal independent regression decision tree to perform node splitting calculation. The output prediction residuals of each regression decision tree are summed and accumulated according to preset weight parameters to obtain the global equivalent safety factor of the landslide.

6. The landslide safety rapid assessment system based on microservices and containerization according to claim 4, characterized in that, Extracting the deformation-active region from the multi-layer 3D point cloud set specifically includes: The anomaly identification module extracts the three-dimensional deformation displacement scalar and the local displacement gradient of all spatial nodes in the multi-layer three-dimensional point cloud set to delineate initial clusters, calculates the average displacement gradient value of the initial clusters, and delineates the initial clusters with the average displacement gradient value being higher than or equal to the overall median as the deformation active region. The anomaly identification module applies principal component analysis to solve the covariance matrix of the deformation active region, extracts feature vectors with a cumulative variance contribution rate greater than a preset contribution threshold as principal components, and projects at least one principal feature dimension to complete the mesoscopic dimensionality reduction processing of the deformation active region.

7. The landslide safety rapid assessment system based on microservices and containerization according to claim 6, characterized in that, The point-level filtering of the deformation-active regions based on density clustering rules specifically includes: The anomaly identification module calculates the mutual reachability distance between each spatial node within the deformation active region based on the main feature dimension. The anomaly identification module constructs a minimum spanning tree using the mutually reachable distance as the graph edge connection weight, gradually reduces the connection to form a hierarchical clustering tree structure, extracts the subtree branch with the highest stability metric as an independent optimal cluster, and isolates the extracted independent optimal clusters as non-consistent abnormal deformation units.

8. The landslide safety rapid assessment system based on microservices and containerization according to claim 7, characterized in that, The output high displacement anomaly feature point cluster and three-dimensional spatial coordinates specifically include: The anomaly identification module takes the spatial nodes inside the non-consistent anomaly deformation unit as the target spatial nodes, extracts the set of nearest neighbors for the target spatial nodes, calculates the local reachability distance and performs an arithmetic mean summation, and defines the reciprocal of the summation mean as the local reachability density. The anomaly identification module calculates the ratio of the local reachability density of each of the neighboring spatial nodes to the local reachability density of the target spatial node, and performs an arithmetic average of all ratios to obtain a local anomaly factor. The target spatial nodes whose local anomaly factor is greater than or equal to the local anomaly factor determination threshold are extracted and marked as the high displacement anomaly feature point cluster, and the three-dimensional spatial coordinates of the high displacement anomaly feature point cluster are extracted.

9. The landslide safety rapid assessment system based on microservices and containerization according to claim 1, characterized in that, Storing the global equivalent safety factor of the landslide and the three-dimensional spatial coordinates of the high displacement anomaly feature point cluster into the system's structured database, and performing dynamic updates in the digital twin three-dimensional scene specifically includes: The asynchronous scheduling module uses a hybrid storage strategy to write the global equivalent safety factor of the landslide and the three-dimensional spatial coordinates into the system's structured database to build a preheating cache; The dynamic rendering module establishes a bidirectional communication channel based on the network socket protocol, encapsulates the global equivalent safety factor of the landslide into a streaming data packet and actively pushes it to the client front-end engine, drives the client front-end engine to extract the evaluation timestamp in the streaming data packet, and dynamically adds two-dimensional coordinate nodes to the global equivalent safety factor of the landslide and the evaluation timestamp in the chart component, and connects them with line segments to generate a safety factor change curve.

10. The landslide safety rapid assessment system based on microservices and containerization according to claim 8, characterized in that, Performing dynamic updates and overlay rendering for real-time evaluation in a digital twin 3D scene specifically includes: The dynamic rendering module uses an affine transformation matrix containing three-dimensional rotational parameter components and three-dimensional translational parameter vector components to convert the three-dimensional spatial coordinates of the high displacement anomaly feature point cluster as local spatial coordinates into three-dimensional world coordinates that can be recognized by the front-end three-dimensional rendering engine. The dynamic rendering module performs hybrid rendering of the alpha channel of the landslide surface mesh material in the digital twin 3D scene according to the pre-set transparency parameters. It establishes a heat map gradient mapping rule based on the upper and lower threshold ranges of the local anomaly factor values ​​to generate color materials. Based on the 3D world coordinates, the color materials are assigned to the instantiated 3D spatial node model, and a highlighted marker tile is generated and superimposed on the non-consistent abnormal deformation area corresponding to each level inside the landslide model to complete the real-time evaluation.

Images