Geological disaster monitoring and early warning system based on space-air-ground integration

Abstract

The invention discloses a geological disaster monitoring and early warning system based on space-air-ground integration, and relates to the technical field of integrated monitoring, and the system comprises the steps: carrying out the dynamic hierarchical transmission based on a compressed feature package with a risk tag, analyzing the compressed feature package, and generating a cross-platform task scheduling instruction through a chaos genetic optimization algorithm; according to the cross-platform task scheduling instruction, the space-air-ground computing power nodes cooperatively execute a computing task to generate a geological disaster risk index; based on the geological disaster risk index, multi-channel early warning broadcasting is carried out, network slice bandwidth allocation is dynamically optimized, and an early warning broadcasting instruction and a resource adjustment report are generated; according to the method, the features are extracted through the lightweight Transform, the output dimension is dynamically adjusted, data compression and risk pre-judgment are achieved, and the method has the effects of reducing the transmission load, being used for edge intelligent processing and achieving the effect of improving the response efficiency.

Description

Technical Field

[0001] This invention relates to the field of integrated monitoring technology, and in particular to a geological disaster monitoring and early warning system based on an integrated air-space-ground system. Background Technology

[0002] The construction of next-generation mobile communication core networks and access networks is a key infrastructure for promoting the development of integrated air-space-ground monitoring systems. Through the fusion of multi-source data, including satellite remote sensing, UAV aerial surveys, and ground sensors, comprehensive coverage and dynamic monitoring of geological disaster risks have been achieved. For example, synthetic aperture radar interferometry can identify surface deformation areas through satellite data, while the BeiDou high-precision positioning system provides millimeter-level displacement monitoring capabilities for geological disasters such as landslides and debris flows. Furthermore, with the development of 5G communication technology and edge computing, real-time transmission and low-latency processing of geological disaster monitoring data have become possible, further improving the response efficiency of early warning systems.

[0003] Traditional monitoring data spatiotemporal alignment relies on fixed reference stations or static calibration methods, which are ill-suited to the timestamp discrepancies and spatial coordinate drift of multi-source data during the dynamic evolution of geological disasters. This results in insufficient spatiotemporal consistency of the input data for feature extraction models, thereby affecting the accuracy of risk assessment. Existing risk early warning systems mostly employ a single task scheduling strategy, failing to fully integrate the sudden nature of geological disasters with the dynamic allocation requirements of network resources. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a geological disaster monitoring and early warning system based on an integrated air-space-ground system to solve the problems of insufficient spatiotemporal data consistency and low task scheduling flexibility in traditional monitoring systems.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a geological disaster monitoring and early warning system based on an integrated air-space-ground system, comprising: The data acquisition module collects integrated air-space-ground data and preprocesses it, generating a spatiotemporal alignment monitoring dataset through a dynamic benchmark station difference allocation method. The data compression module uses a lightweight Transformer feature extraction model to perform hierarchical compression on the spatiotemporal alignment monitoring dataset, and dynamically adjusts the output dimension in conjunction with preliminary risk assessment rules to generate compressed edge feature packages with risk labels. The instruction generation module performs dynamic hierarchical transmission based on compressed feature packets with risk labels, parses the compressed feature packets, and generates cross-platform task scheduling instructions through a chaotic genetic optimization algorithm. The collaborative execution module, based on cross-platform task scheduling instructions, enables space-air-ground computing nodes to collaboratively execute computing tasks and generate a geological disaster risk index. The early warning broadcast module, based on the geological disaster risk index, conducts multi-channel early warning broadcasts, dynamically optimizes network slice bandwidth allocation, and generates early warning broadcast instructions and resource adjustment reports.

[0007] As a preferred embodiment of the integrated air-space-ground geological disaster monitoring and early warning system described in this invention, the integrated air-space-ground data includes surface deformation raster data, three-dimensional spatial coordinate set, tilt acceleration and soil moisture content data; The preprocessing includes outlier filtering, multi-source coordinate system normalization, and time stamp alignment.

[0008] As a preferred embodiment of the integrated air-space-ground geological disaster monitoring and early warning system described in this invention, the generation of the spatiotemporally aligned monitoring dataset refers to generating the spatiotemporally aligned monitoring dataset based on the preprocessed integrated air-space-ground data through a dynamic reference station difference allocation method.

[0009] As a preferred embodiment of the integrated air-space-ground geological disaster monitoring and early warning system described in this invention, the step of performing layered compression of the spatiotemporally aligned monitoring dataset using a lightweight Transformer feature extraction model is as follows: Based on the spatiotemporal alignment monitoring dataset, the dataset is standardized and divided into blocks according to data type to generate a fragmented dataset. Low-dimensional feature vectors are extracted from the fragmented dataset using a lightweight Transformer model. The original edge feature vector set is generated by integrating multimodal feature physical alignment with a spatiotemporal benchmark.

[0010] As a preferred embodiment of the integrated air-space-ground geological disaster monitoring and early warning system described in this invention, the steps for dynamically adjusting the output dimension and generating a compressed edge feature package with risk labels, based on preliminary risk assessment rules, are as follows: Based on the original edge feature vector group, risk labels are bound in combination with the preliminary risk assessment rules and the output dimension is adjusted to generate edge feature vectors with risk labels. The risk-labeled edge feature vectors are categorized and compressed by dimension, and metadata is added to generate a compressed edge feature package with risk labels.

[0011] As a preferred embodiment of the integrated air-space-ground geological disaster monitoring and early warning system of the present invention, the dynamic hierarchical transmission based on the compressed feature packet with risk label refers to reading the risk label field in the header of the compressed edge feature packet with risk label, performing dynamic hierarchical transmission, and generating the compressed edge feature packet.

[0012] As a preferred embodiment of the integrated air-space-ground geological disaster monitoring and early warning system described in this invention, the steps for parsing and compressing the feature package and generating cross-platform task scheduling instructions using a chaotic genetic optimization algorithm are as follows: Decompress the compressed edge feature package, extract metadata, and reconstruct the edge feature vector to generate decompressed edge feature vectors and metadata dataset; Based on the decompressed edge feature vector and metadata, genetic evolution operations are performed through competitive selection, single-point crossover, and uniform mutation to generate cross-platform task scheduling instructions.

[0013] As a preferred embodiment of the integrated space-air-ground geological disaster monitoring and early warning system described in this invention, the steps for generating a geological disaster risk index by having space-air-ground computing nodes collaboratively execute computational tasks according to cross-platform task scheduling instructions are as follows. Execute cross-platform task scheduling instructions and dynamically create edge Pod instances via the Kubernetes API to generate initialized compute container instances; The decompressed edge feature vectors are dimension-aligned and input into the initialized computation container instance to generate a computation container that loads the dimension-aligned edge feature vectors. Based on the computational container of the loaded dimension-aligned edge feature vector, the disaster risk quantification calculation is performed in air, space and ground respectively, and disaster risk triplet is generated. Based on the disaster risk triplet, the risk index is calculated by fusion through dynamic confidence weighting to generate the geological disaster risk index.

[0014] As a preferred embodiment of the integrated air-space-ground geological disaster monitoring and early warning system described in this invention, the steps for multi-channel early warning broadcasting based on the geological disaster risk index are as follows: Based on the geological disaster risk index, the early warning level is determined according to the early warning threshold, and the determination timestamp is recorded to generate an early warning level identifier. Based on the warning level identifier, the core fields of the warning message are constructed and encoded into a multi-protocol format to generate multi-channel warning messages.

[0015] As a preferred embodiment of the integrated air-space-ground geological disaster monitoring and early warning system described in this invention, the steps for dynamically optimizing network slice bandwidth allocation and generating early warning broadcast commands and resource adjustment reports are as follows: By combining multi-channel early warning messages with real-time network load, bandwidth adjustment is calculated, uRLLC slice bandwidth is dynamically reconstructed, non-urgent service resources are downgraded, and network slice reconfiguration instructions are generated. Based on multi-channel early warning messages and network slice reconfiguration instructions, early warnings are issued and resource adjustments are recorded, generating early warning broadcast records and resource adjustment reports.

[0016] The beneficial effects of this invention are as follows: by extracting features and dynamically adjusting the output dimension through a lightweight Transformer, data compression and risk prediction are achieved, which reduces the transmission load and is used for edge intelligent processing to improve response efficiency; by generating scheduling instructions through a chaotic genetic algorithm, dynamic coordination of computing resources is achieved, which optimizes task allocation and is used for efficient cross-platform computing to shorten the early warning latency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a geological disaster monitoring and early warning system based on an integrated air-space-ground system.

[0019] Figure 2 This is a flowchart for lightweight Transformer feature extraction.

[0020] Figure 3 A flowchart for generating cross-platform task scheduling instructions for a chaotic genetic optimization algorithm.

[0021] Figure 4 A flowchart illustrating the collaborative execution of computing tasks by space-air-ground computing nodes. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4This is one embodiment of the present invention, which provides a geological disaster monitoring and early warning system based on an integrated air-space-ground system, comprising the following steps: The data acquisition module collects integrated air-space-ground data and preprocesses it, generating a spatiotemporal alignment monitoring dataset through a dynamic benchmark station difference allocation method. The integrated air-space-ground data includes surface deformation raster data, three-dimensional spatial coordinate sets, tilt acceleration, and soil moisture content data; It should be noted that the space layer acquires interferometric images through synthetic aperture radar satellites, and generates surface deformation raster data through differential interferometry; the sky layer uses vertical take-off and landing UAVs equipped with lidar to scan high-risk slopes and generate a three-dimensional spatial coordinate set; the ground layer deploys an Internet of Things sensor network to collect tilt acceleration data and soil moisture content data.

[0026] Preprocessing includes outlier filtering, multi-source coordinate system normalization, and time stamp alignment; It should be noted that outliers in the tilt acceleration data are filtered using a three-standard-deviation criterion; the general geodetic coordinate system of the surface deformation raster data, the local coordinate system of the three-dimensional spatial coordinate set, and the encrypted coordinate system of the soil moisture content data are transformed to an identifiable geodetic coordinate system; the time stamps are aligned using navigation satellite timing, and missing time period data are compensated by forward interpolation to generate preprocessed integrated air-space-ground data.

[0027] Based on the preprocessed integrated air-space-ground data, a spatiotemporal alignment monitoring dataset is generated using the dynamic reference station difference allocation method.

[0028] Furthermore, a mobile reference station is deployed in the monitoring area to receive global navigation satellite signals, generate real-time precise single-point positioning correction parameters, perform geometric fine correction on the surface deformation raster data, and complete absolute georegistration between the three-dimensional spatial coordinate set and the mobile reference station observation data through post-processing dynamic positioning methods; the tilt acceleration data and soil moisture content data are compensated for time delay using inverse Kalman filtering; finally, a spatiotemporally aligned monitoring dataset is generated.

[0029] It should be noted that the post-processing dynamic positioning method is executed as follows: the original observation values ​​of the global navigation satellite are recorded synchronously at the mobile reference station and the UAV; the integer ambiguity parameters of the carrier phase observation values ​​are determined by the joint adjustment of the observation data of the reference station and the UAV; and the three-dimensional spatial coordinate set is transformed from the local coordinate system to the national geodetic coordinate system by the seven-parameter Bursa transform method.

[0030] The data compression module uses a lightweight Transformer feature extraction model to perform hierarchical compression on the spatiotemporal alignment monitoring dataset, and dynamically adjusts the output dimension in conjunction with preliminary risk assessment rules to generate compressed edge feature packages with risk labels. Based on the spatiotemporal alignment monitoring dataset, the dataset is standardized and divided into blocks according to data type to generate a fragmented dataset. Furthermore, based on the spatiotemporally aligned monitoring dataset, the surface deformation raster data is divided into geospatial blocks, the three-dimensional spatial coordinate set is divided into rasterized blocks, and the dip acceleration data and soil moisture content data are divided into time window blocks to generate a segmented dataset.

[0031] It should be noted that data type refers to the type of data contained in the spatiotemporal aligned monitoring dataset, including surface deformation raster data, three-dimensional spatial coordinate sets, dip acceleration, and soil moisture content data. Geospatial segmentation involves cutting the surface deformation raster data into fixed-size pixel blocks and recording the geographic boundary coordinates of each block to generate an image block set; rasterization segmentation converts the three-dimensional spatial coordinate set into regular grid nodes, calculates the maximum elevation value within each grid, and generates an elevation raster set; temporal window segmentation involves sliding and cropping the dip acceleration and soil moisture content data at fixed intervals to generate a time-series window set.

[0032] Low-dimensional feature vectors of the fragmented dataset are extracted using a lightweight Transformer model. The original edge feature vector set is generated by physical alignment of multimodal features and integration with spatiotemporal benchmarks. Furthermore, the lightweight Transformer model processes the pieced dataset. The image patch set is input into the SAR processing layer to extract the deformation displacement vector and phase coherence from the surface deformation raster data patch; the elevation raster set is input into the LiDAR processing layer to generate surface roughness and slope angle; the time window set is input into the sensor processing layer to generate the rate of change of soil moisture content and acceleration variance; finally, the data are integrated to generate the original edge feature vector set.

[0033] Feature extraction process for surface deformation raster data blocks: Differential interferometry is performed on the surface deformation raster data blocks. An interferometric phase map is generated by registering master-slave images. The terrain phase is removed using external digital elevation nodes. The globally optimal unwrapping path is solved using a minimum cost flow algorithm. The unwrapped phase is converted into a deformation displacement vector. Simultaneously, pixel-level coherence is obtained based on a sliding window, and the deformation displacement vector and phase coherence are output. Feature extraction process for LiDAR processing layer: Spatial gradient is solved on the input elevation raster set. A 3×3 neighborhood window is taken with each raster node as the center. The standard deviation of the elevation values ​​within the window is obtained as the surface roughness. The partial derivatives in the X and Y directions are obtained simultaneously. The slope angle is synthesized using the arctangent function, and the surface roughness and slope angle at the raster node level are output. The sensor processing layer feature extraction operation process is as follows: the input time window set is smoothed by moving average filtering of soil moisture content data, and the ratio of the difference in moisture content between adjacent time steps to the time interval is obtained to generate the rate of change of moisture content; the raw data of the triaxial accelerometer is used to synthesize the resultant acceleration, the variance of the resultant acceleration within the window is obtained, and the windowed rate of change of moisture content and acceleration variance are generated.

[0034] It should be noted that the lightweight Transformer model construction process is as follows: the architecture is based on the standard Transformer encoder and is a lightweight modification. The core innovations are the compressed multi-head attention mechanism and the adjustable dimension output structure. The input layer is designed with independent branches to process heterogeneous data sources from air, space, and ground: the SAR branch processes surface deformation raster data blocks, the LiDAR branch processes elevation raster sets, and the sensor branch processes temporal window sets. The encoder feedforward network uses depthwise separable convolutions to replace fully connected layers, reducing the number of parameters.

[0035] Hardware adaptation is optimized for the edge by obtaining the attention operation parameters of layer fusion and combining them with dynamic quantization to compress node weights. It integrates hardware-level dimension switching nodes to ensure runtime flexibility, generates cross-platform files compatible with mainstream edge chips, and achieves low-latency inference.

[0036] The lightweight Transformer model training employs a progressive strategy: the first stage involves pre-training the feature extraction capabilities of each data branch modally; the second stage involves joint fine-tuning to freeze the underlying parameters and optimize the multi-task loss function to achieve cross-modal feature fusion; and the third stage involves knowledge distillation to transfer complex teacher node knowledge to the lightweight Transformer model. Ultimately, high-precision feature extraction and real-time response capabilities are achieved on the validation set, meeting the needs of geological disaster monitoring scenarios.

[0037] Based on the original edge feature vector group, risk labels are bound in combination with the preliminary risk assessment rules and the output dimension is adjusted to generate edge feature vectors with risk labels. Furthermore, based on the deformation displacement vector components and slope angle components in the original edge feature vector group, a preliminary risk assessment rule is executed to generate risk labels; the feature vector dimension is dynamically adjusted according to the risk labels, the dimension of high-risk labels is expanded to supplement time-frequency domain decomposition features, the dimension of medium-risk labels is maintained by adding first-order derivative features, and the dimension of low-risk labels is compressed to retain only the basic feature components; finally, edge feature vectors with risk labels are generated.

[0038] It should be noted that the preliminary risk assessment rules are set by fitting the deformation displacement vector, slope angle and rainfall combination to the measured geological disaster case data (for example, when any of the following conditions are met, the deformation displacement vector is ≥5mm and the slope angle is ≥30°, and the cumulative rainfall is ≥250mm, it is judged as a high-risk label; when the following single conditions are met, 2mm < deformation displacement vector <5mm, 15° < slope angle <30°, 150mm < cumulative rainfall <250mm, it is judged as a medium-risk label; when the following conditions are met simultaneously, the deformation displacement vector is <2mm, the slope angle is <15°, and the cumulative rainfall is <150mm, it is judged as a low-risk label).

[0039] The extended dimension supplements the time-frequency domain decomposition features. Multi-scale wavelet packet decomposition is performed on the time-series components of the original edge feature vector to extract the energy distribution features of each sub-band, obtaining the relative energy value of the dominant frequency band to generate an energy entropy feature vector. Principal component analysis is used to compress and retain the main information features, which are then concatenated with the original spatial feature vectors in the original edge feature vector group to form a high-dimensional extended feature vector, achieving cross-domain fusion of time-frequency and spatial domain features. The dimension is maintained by adding first-order derivative features. Time-series differencing is performed on the original edge feature vector group to generate displacement change rate and slope change rate. Feature components with the lowest contribution in the original spatial feature vector are selected by ranking feature importance, and the newly added first-order derivative features replace the selected low-contribution features, keeping the total dimension unchanged, achieving iterative updates within the dimension of dynamically changing features. The compression dimension retains only the basic feature components. When the risk level is determined to be low based on the risk label, the core basic feature components are extracted from the original edge feature vector group, all derived extended features are discarded, and the core basic feature components are encapsulated into a low-dimensional feature vector in a fixed order. The part that is insufficient for the target dimension is filled with zero values ​​to the target length, and a low-dimensional feature vector containing only basic physical quantities is generated.

[0040] The risk-labeled edge feature vectors are categorized and compressed by dimension, and metadata is added to generate a compressed edge feature package with risk labels.

[0041] Specifically, a compression algorithm is selected based on the risk label: the Brotli compression algorithm is used for high-risk labels, the Zstandard compression algorithm is used for medium-risk labels, and Delta+entropy encoding is used for low-risk labels; a metadata header (including risk label, feature dimension, BeiDou time stamp, and geographic bounding box field) is added to generate a compressed edge feature package with risk labels.

[0042] It should be noted that the Brotli compression algorithm operates by using LZ77 sliding window matching and Huffman coding based on a geological dictionary to achieve a high compression ratio. The Zstandard compression algorithm employs finite-state entropy coding and sequence compression, dynamically training the dictionary to optimize the compression efficiency of mid-dimensional data. The Delta+ entropy coding operation performs arithmetic entropy coding after differencing adjacent values ​​of low-dimensional feature vectors to achieve lossless compression. This compressed transmission is designed to address the bandwidth bottleneck of air-space-ground networks and facilitate transmission.

[0043] The instruction generation module performs dynamic hierarchical transmission based on compressed feature packets with risk labels, parses the compressed feature packets, and generates cross-platform task scheduling instructions through a chaotic genetic optimization algorithm. Read the risk label field from the header of the compressed edge feature packet with risk label, perform dynamic hierarchical transmission, and generate a compressed edge feature packet; Furthermore, the risk label field in the header of the compressed edge feature packet with risk label is parsed to trigger a hierarchical transmission strategy. High-risk labels call the 5G core network NSSF interface to create a uRLLC dedicated slice, medium-risk labels are allocated eMBB slices to guarantee bandwidth, and low-risk labels are included in the mMTC batch transmission queue. The wireless signal strength is monitored in real time, and if it is lower than -90dBm, BeiDou RDSS short message transmission is switched. Finally, a compressed edge feature packet with the transmission strategy scheduling completed is generated.

[0044] Decompress the compressed edge feature package, extract metadata, and reconstruct the edge feature vector to generate decompressed edge feature vectors and metadata dataset; Furthermore, the risk label field and feature dimension field of the compressed edge feature packet header with risk labels are read, and the decompression algorithm is selected according to the label value. High-risk labels are decompressed using Brotli, medium-risk labels are decompressed using Zstandard, and low-risk labels are decoded using Delta+entropy. After decompression, the binary stream is restored to a floating-point edge feature vector according to the dimension field. The packet header metadata is extracted simultaneously to generate decompressed edge feature vectors and metadata datasets.

[0045] Based on the decompressed edge feature vector and metadata, genetic evolution operations are performed through competitive selection, single-point crossover, and uniform mutation to generate cross-platform task scheduling instructions.

[0046] Specifically, based on the decompressed edge feature vector and the metadata, candidate scheduling schemes are initialized, and an initial population is generated using Tent chaotic mapping; competitive selection is performed to screen individuals with high fitness, and a fitness function is defined; single-point crossover is performed on the selected individuals to distribute gene fragments, and then uniform mutation perturbation resource allocation parameters are applied to generate cross-platform task scheduling instructions.

[0047] It should be noted that the Tent chaotic mapping method is a chaotic node that generates a pseudo-random sequence through a deterministic piecewise linear function. The iterative formula is as follows: it utilizes the tiny differences in the microsecond portion of the BeiDou time stamp in the metadata set to generate a highly non-repeating ergodic sequence, which is used to initialize the population in the genetic algorithm to overcome the local clustering limitation of traditional random number generators and improve the global search efficiency.

[0048] The fitness function definition process is as follows: Based on the feature dimension parameters of the decompression edge feature vector and its metadata set, a multidimensional cost function is constructed; feature dimension values ​​are extracted from the metadata set as input, and the task execution time is determined by the relational formula fitted by the measured data; the current energy consumption parameters are obtained from the platform resource database; the data transmission communication overhead is read from the network status interface; and finally, the fitness evaluation value of each individual is output as a quantitative basis for genetic evolution selection.

[0049] The competitive selection process proceeds as follows: Individuals are randomly selected from the current population to form a competitive group. The fitness value of each individual within the group is obtained, and the individual with the highest fitness value is selected as the winner and retained for the next generation. This process is repeated until the new population size reaches a set value (example value: 50, determined according to the population size parameter in the cross-platform task scheduling instruction), thus implementing an evolutionary selection mechanism of survival of the fittest. The single-point crossover and gene fragment allocation process proceeds as follows: A crossover point is randomly selected between the gene sequences of the two chosen parent individuals, and all gene fragments after that point are exchanged. The gene sequence of parent individual one contains computation type, algorithm type, and resource requirement parameters, while the gene sequence of parent individual two contains different computation types, algorithm types, and resource requirement parameters. After crossover, two offspring individuals are generated. The new individuals inherit the allocation scheme before the crossover point and the resource allocation scheme after the crossover point from their parents. A conflict detection mechanism is executed; when the computation type and resource requirement parameters in the offspring are incompatible, they are replaced with a compatible resource type, and a feasible offspring individual is output.

[0050] The collaborative execution module, based on cross-platform task scheduling instructions, enables space-air-ground computing nodes to collaboratively execute computing tasks and generate a geological disaster risk index. Execute cross-platform task scheduling instructions and dynamically create edge Pod instances via the Kubernetes API to generate initialized compute container instances; Specifically, in executing cross-platform task scheduling instructions, a container instance creation request is initiated to the target edge node through the Kubernetes API, requesting graphics processor or programmable gate array hardware acceleration resources, mounting pre-trained node file storage volumes, configuring environment variable parameters including geographical bounding boxes and timestamps, and the container orchestration scheduler dynamically allocates computing resources and generates edge container instances based on the real-time resource indicators periodically reported by each edge node, outputting a computing container instance that has completed environment initialization.

[0051] The process for requesting graphics processing unit (GPU) or programmable gate array (FPGA) hardware acceleration resources should be explained as follows: The hardware resource requirement field is declared in the container creation request, specifying the accelerator type and quantity parameters. The container orchestration scheduler calls the device plugin interface to query the edge node hardware resource pool status, dynamically allocates physical devices that meet the requirements, and maps device access permissions to the container instance, achieving exclusive binding and secure isolation of hardware acceleration resources. The environment variable configuration process involves extracting the geographic bounding box field and the BeiDou time stamp field from the metadata dataset, converting them to key-value pair format, and injecting the environment variable list when creating the Pod via the Kubernetes API. This synchronously sets the container timezone parameters, ensuring that the spatial range and time base parameters are loaded when the container instance starts.

[0052] The decompressed edge feature vectors are dimension-aligned and input into the initialized computation container instance to generate a computation container that loads the dimension-aligned edge feature vectors. Specifically, based on the decompressed edge feature vectors and metadata dataset, zero-padding alignment is performed on the vectors; the dimension-aligned edge feature vectors and metadata are mounted to the path of the initialized compute container instance via a Kubernetes shared storage volume; the container startup script loads the vectors into the memory buffer and generates a compute container loaded with the dimension-aligned edge feature vectors.

[0053] The zero-padding alignment process is as follows: The original dimension value of the decompressed edge feature vector is read. When the original dimension value is less than the target dimension value (which refers to the input dimension requirement of each processing layer in the lightweight Transformer model, dynamically determined according to the risk level), consecutive floating-point zero values ​​are padded to the end of the decompressed edge feature vector to generate a feature vector of the target dimension length. After padding, the physical meaning and position index of the original feature components at the beginning of the vector remain unchanged, ensuring the correct mapping relationship of the input layer neurons. The dimension-aligned edge feature vector and metadata mounting process is as follows: The dimension-aligned edge feature vector is serialized into a binary file, and the metadata is converted into a JSON file. The storage volume mount path is declared in the compute container instance configuration file. When the container starts, the feature vector file and metadata file under the storage volume mount path are loaded into the memory buffer to achieve data injection. The loading of the edge feature vector into the memory buffer process is as follows: The container startup script calls the memory mapping function to map the dimension-aligned edge feature vector binary file under the mount path to a user-space virtual address, setting read-only access permissions to prevent tampering. Simultaneously, the metadata file is loaded into the heap memory structure, and the buffer pointer is initialized to point to the memory starting address of the edge feature vector, achieving a zero-copy data access mechanism.

[0054] Based on the computational container of the loaded dimension-aligned edge feature vector, the disaster risk quantification calculation is performed in air, space and ground respectively, and disaster risk triplet is generated. Specifically, the satellite edge computing node executes the InSAR phase solution container, inputs the terrain parameters in the dimension-aligned edge feature vector, and calculates the slope stability coefficient using geotechnical mechanics formulas; the UAV MEC node executes the LSTM landslide prediction container, inputs the full 256-dimensional feature vector, and outputs the landslide probability through the Sigmoid activation function; the ground edge gateway executes the sensor fusion container, inputs physical quantity features, and calculates the ground motion risk coefficient; finally, the disaster risk triplet is output synchronously.

[0055] The formula for calculating the slope stability coefficient is as follows: ; in, This represents the slope stability coefficient. This represents the cohesive component of soil and rock mass that resists shear failure. This represents the weight per unit volume of rock and soil. This represents the vertical distance from the potential sliding surface of the slope to the ground surface. Indicates the slope angle, the angle between the slope surface and the horizontal plane. This indicates the pressure generated by groundwater in the pores of rock and soil. Parameters representing the frictional characteristics of soil and rock masses resisting shear failure; The formula for calculating the ground motion risk coefficient is as follows: ; in, Indicates the risk factor for ground sports. This indicates the rate of change of soil volumetric moisture content per unit time. This represents the variance of the triaxial accelerometer measurements.

[0056] It should be noted that 0.7 in the expression is the dominant weight of moisture content change, and 0.3 is the auxiliary weight of vibration energy. These weighting coefficients are set through statistical analysis of historical geological disaster data and verification by engineering experience.

[0057] The cohesive component of the soil and rock mass resisting shear failure and the frictional characteristic parameters of the soil and rock mass resisting shear failure were obtained through soil and rock mechanics tests. The weight of the soil and rock mass per unit volume was measured by field sampling. The vertical distance between the potential sliding surface of the slope and the ground surface was obtained through borehole and geophysical inversion. The slope angle was calculated by LiDAR point cloud slope solution. The pressure generated by groundwater in the pores of the soil and rock mass was monitored in real time by a piezometer. The rate of change of soil volumetric water content per unit time was calculated by time-series difference calculation of soil water content sensor. The variance of the triaxial accelerometer measurement value was calculated by the variance of the triaxial accelerometer measurement value.

[0058] Based on the disaster risk triplet, the risk index is calculated by fusion through dynamic confidence weighting to generate the geological disaster risk index.

[0059] Furthermore, based on the disaster risk triplet and combined with the confidence weights of each data source, the risk index fusion calculation formula is as follows: ; in, Indicates the geological disaster risk index. This represents the confidence level of the landslide probability. This represents the confidence level of the stability coefficient. This represents the confidence level of ground motion. landslide probability This represents the reciprocal of the slope stability coefficient. Indicates the risk coefficient of ground motion; It should be noted that 0.4 in the expression is the mandatory minimum weight of the main early warning factor, and 0.3 is the balanced weight of the auxiliary early warning factor, which is determined by inferring the contribution rate of different risk factors in historical landslide cases.

[0060] The early warning broadcast module, based on the geological disaster risk index, conducts multi-channel early warning broadcasts, dynamically optimizes network slice bandwidth allocation, and generates early warning broadcast instructions and resource adjustment reports.

[0061] Based on the geological disaster risk index, the early warning level is determined according to the early warning threshold, and the determination timestamp is recorded to generate an early warning level identifier. Specifically, the warning level is determined based on the warning threshold. When the geological disaster risk index is within the range of the first-level warning threshold, a red warning is issued; when it is within the range of the second-level warning threshold, an orange warning is issued; and when it is within the range of the third-level warning threshold, a yellow warning is issued. The timestamp of the determination is recorded simultaneously, and a warning level identifier is generated. It should be noted that the warning thresholds are set by analyzing the correspondence between the risk index and the occurrence of historical geological disaster events. The range of the first-level warning threshold is usually [0.85, 1), the range of the second-level warning threshold is usually [0.7, 0.85), and the range of the third-level warning threshold is usually [0.55, 0.7].

[0062] Based on the warning level identifier, construct the core fields of the warning message and encode them into a multi-protocol format to generate multi-channel warning messages; Furthermore, based on the warning level identifier, the Tianditu API is called to convert the geographical boundary box into the administrative division name, and the disaster type is determined by combining the geological map; core fields are filled in, including location information, disaster type, and suggested actions; and the messages are encoded according to the protocol specifications: JSON format messages are generated for 4G and 5G channels, compressed binary messages are generated for Beidou RDSS channels, and hexadecimal control codes are generated for public screen protocols, and finally, multi-channel warning messages are output.

[0063] By combining multi-channel early warning messages with real-time network load, bandwidth adjustment amounts are obtained, and uRLLC slice bandwidth is dynamically reconstructed, non-urgent service resources are downgraded, and network slice reconfiguration instructions are generated. Parse the warning level identifier in the multi-channel warning message, combine it with the real-time network load, and execute the bandwidth preemption strategy (e.g., preempt 20% of the maximum available bandwidth of the base station during a red warning, and preempt 10% during an orange warning); dynamically reconstruct the uRLLC slice bandwidth through the 5G core network NSSF interface, and simultaneously downgrade non-emergency service resources (eMBB slice guaranteed bandwidth is reduced to 70%, and mMTC connection number is limited to 80% concurrency); generate network slice reconfiguration instructions.

[0064] Based on multi-channel early warning messages and network slice reconfiguration instructions, early warnings are issued and resource adjustments are recorded, generating early warning broadcast records and resource adjustment reports.

[0065] After the multi-channel early warning message is input, a hierarchical broadcast strategy is executed. Red early warning messages are broadcast at high frequency through 4G and 5G base station information blocks, transmitted at high power through BeiDou short messages, and linked with public screen audio and visual alarms. Simultaneously, network slice reconfiguration instructions are executed to forcibly increase the bandwidth of high-reliability low-latency slices and compress non-emergency service resources. The amount of resource adjustment (such as the number of users affected by downgrading services) is recorded in real time, and an early warning broadcast execution record (including timestamp, channel type, and message summary) and a network resource adjustment report are generated.

[0066] In summary, this invention achieves data compression and risk prediction by using a lightweight Transformer to extract features and dynamically adjust the output dimension, thereby reducing transmission load and improving response efficiency for edge intelligent processing. Furthermore, it generates scheduling instructions using a chaotic genetic algorithm, enabling dynamic coordination of computing resources to optimize task allocation for efficient cross-platform computing and shortening early warning latency.

[0067] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A geological disaster monitoring and early warning system based on integrated air-space-ground systems, characterized in that: include, The data acquisition module collects integrated air-space-ground data and preprocesses it, generating a spatiotemporal alignment monitoring dataset through a dynamic reference station difference allocation method. The data compression module uses a lightweight Transformer feature extraction model to perform hierarchical compression on the spatiotemporal alignment monitoring dataset, and dynamically adjusts the output dimension in conjunction with preliminary risk assessment rules to generate compressed edge feature packages with risk labels. The instruction generation module performs dynamic hierarchical transmission based on compressed feature packets with risk labels, parses the compressed feature packets, and generates cross-platform task scheduling instructions through a chaotic genetic optimization algorithm. The collaborative execution module, based on cross-platform task scheduling instructions, enables space-air-ground computing nodes to collaboratively execute computing tasks and generate a geological disaster risk index. The early warning broadcast module, based on the geological disaster risk index, conducts multi-channel early warning broadcasts, dynamically optimizes network slice bandwidth allocation, and generates early warning broadcast instructions and resource adjustment reports.

2. The integrated air-space-ground geological disaster monitoring and early warning system as described in claim 1, characterized in that: The integrated air-space-ground data includes surface deformation raster data, three-dimensional spatial coordinate set, tilt acceleration, and soil moisture content data; The preprocessing includes outlier filtering, multi-source coordinate system normalization, and time stamp alignment.

3. The integrated air-space-ground geological disaster monitoring and early warning system as described in claim 2, characterized in that: The generation of the spatiotemporal alignment monitoring dataset refers to the generation of the spatiotemporal alignment monitoring dataset based on the preprocessed integrated air-space-ground data and through the dynamic reference station difference allocation method.

4. The integrated air-space-ground geological disaster monitoring and early warning system as described in claim 3, characterized in that: The method of using a lightweight Transformer feature extraction model to perform hierarchical compression on the spatiotemporal alignment monitoring dataset involves the following steps. Based on the spatiotemporal alignment monitoring dataset, the dataset is standardized and divided into blocks according to data type to generate a fragmented dataset. Low-dimensional feature vectors are extracted from the fragmented dataset using a lightweight Transformer model. The original edge feature vector set is generated by integrating multimodal feature physical alignment with a spatiotemporal benchmark.

5. The integrated air-space-ground geological disaster monitoring and early warning system as described in claim 4, characterized in that: The steps for dynamically adjusting the output dimension based on the preliminary risk assessment rules to generate a compressed edge feature package with risk labels are as follows: Based on the original edge feature vector group, risk labels are bound in combination with the preliminary risk assessment rules and the output dimension is adjusted to generate edge feature vectors with risk labels. The risk-labeled edge feature vectors are categorized and compressed by dimension, and metadata is added to generate a compressed edge feature package with risk labels.

6. The integrated air-space-ground geological disaster monitoring and early warning system as described in claim 5, characterized in that: The dynamic hierarchical transmission based on compressed feature packets with risk labels refers to reading the risk label field in the header of the compressed edge feature packet with risk labels, performing dynamic hierarchical transmission, and generating compressed edge feature packets.

7. The integrated air-space-ground geological disaster monitoring and early warning system as described in claim 6, characterized in that: The parsed compressed feature package is then used to generate cross-platform task scheduling instructions via a chaotic genetic optimization algorithm. The steps are as follows. Decompress the compressed edge feature package, extract metadata, and reconstruct the edge feature vector to generate decompressed edge feature vectors and metadata dataset; Based on the decompressed edge feature vector and metadata, genetic evolution operations are performed through competitive selection, single-point crossover, and uniform mutation to generate cross-platform task scheduling instructions.

8. The integrated air-space-ground geological disaster monitoring and early warning system as described in claim 7, characterized in that: The process of generating a geological disaster risk index by having space-air-ground computing nodes collaboratively execute computing tasks according to cross-platform task scheduling instructions is as follows: Execute cross-platform task scheduling instructions and dynamically create edge Pod instances via the Kubernetes API to generate initialized compute container instances; The decompressed edge feature vectors are dimension-aligned and input into the initialized computation container instance to generate a computation container that loads the dimension-aligned edge feature vectors. Based on the computational container of the loaded dimension-aligned edge feature vector, the disaster risk quantification calculation is performed in air, space and ground respectively, and disaster risk triplet is generated. Based on the disaster risk triplet, the risk index is calculated by fusion through dynamic confidence weighting to generate the geological disaster risk index.

9. The integrated air-space-ground geological disaster monitoring and early warning system as described in claim 8, characterized in that: The steps for multi-channel early warning broadcasting based on the geological disaster risk index are as follows: Based on the geological disaster risk index, the early warning level is determined according to the early warning threshold, and the determination timestamp is recorded to generate an early warning level identifier. Based on the warning level identifier, the core fields of the warning message are constructed and encoded into a multi-protocol format to generate multi-channel warning messages.

10. The integrated air-space-ground geological disaster monitoring and early warning system as described in claim 9, characterized in that: The steps for dynamically optimizing network slice bandwidth allocation and generating early warning broadcast commands and resource adjustment reports are as follows. By combining multi-channel early warning messages with real-time network load, bandwidth adjustment is calculated, uRLLC slice bandwidth is dynamically reconstructed, non-urgent service resources are downgraded, and network slice reconfiguration instructions are generated. Based on multi-channel early warning messages and network slice reconfiguration instructions, early warnings are issued and resource adjustments are recorded, generating early warning broadcast records and resource adjustment reports.

Images