Unmanned aerial vehicle detection method and system based on visible light polarization imaging

Abstract

The invention relates to the technical field of unmanned aerial vehicle detection and identification, in particular to an unmanned aerial vehicle detection method and system based on visible light polarization imaging. According to the method, a channel state sequence with aligned time sequences is constructed by obtaining multi-source environment sensing data, cross-layer optimization parameters are dynamically calculated and decomposed into calculation and transmission characteristic quantities, and edge nodes are scheduled to cooperatively generate resource scheduling instructions; physical resources are matched to construct a virtual instance topology, a control instruction is generated through online reasoning of a dual-channel graph neural network training model, and finally the control instruction is fed back to the polarization sensor to achieve link calibration. According to the system, through fusion of polarization imaging and cross-layer optimization technologies, the accuracy and real-time performance of unmanned aerial vehicle detection in a complex environment are remarkably improved, meanwhile, computing resource consumption is reduced, and efficient spectrum multiplexing and power adaptive adjustment in a dynamic environment are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of drone detection and identification, and in particular to a drone detection method and system based on visible light polarization imaging. Background Art

[0002] The widespread use of unmanned aerial vehicles (UAVs) poses potential security threats, and the need for efficient and accurate detection is becoming increasingly urgent. Existing UAV detection technologies face the following challenges: 1. Limitations of single / dual-modal perception: Single visible light imaging is susceptible to interference from lighting changes (glare, backlight), weather conditions (fog, haze), camouflage, and complex backgrounds. Detection rate and robustness are significantly reduced in low-contrast or adverse weather conditions. Traditional infrared imaging is sensitive to temperature and susceptible to interference from solar radiation and heat sources. It also lacks resolution for detecting small targets (such as small drones) at long distances. While single millimeter-wave radar offers certain penetration capabilities and all-weather operation, it lacks close-range imaging resolution and the ability to extract target feature information (such as shape and texture). It has difficulty detecting small, low-observable (LowRCS) drones and non-metallic targets and is susceptible to ground clutter interference.

[0003] 2. Deficiencies in spatiotemporal modeling and dynamic response: Existing methods are insufficient in modeling continuous spatiotemporal contextual information when dealing with high-speed maneuvering targets such as drones. Traditional sequence processing methods (such as Kalman filtering and basic RNNs) or static image analysis struggle to effectively capture the complex spatiotemporal evolution of target motion (such as maneuver trajectories and posture changes) and its dynamic interactions with the environment.

[0004] 3. Resource constraints and real-time bottlenecks: Multimodal perception and complex model calculations impose enormous computational overhead. Existing systems often use static resource allocation strategies (such as fixed computational frame rates and preset fusion modes), which are unable to adapt to dynamic changes in environmental complexity, target threat levels, and system load. Summary of the Invention

[0005] The present invention provides a drone detection method and system based on visible light polarization imaging to solve the problem of how to fuse multi-source heterogeneous data of visible light polarization imaging and millimeter-wave radar in complex dynamic environments, and realize high-precision real-time response of the drone detection system through spatiotemporal joint modeling and dynamic resource optimization.

[0006] In order to solve the above technical problems, the present invention provides a drone detection method based on visible light polarization imaging, comprising: The system acquires raw signals from multi-source environmental perception devices, performs non-uniformity correction and noise reduction processing, generates environmental perception data including a polarization matrix, point cloud, and thermal map through spatiotemporal registration, and constructs a time-aligned channel state sequence. The multi-source environmental perception device includes a visible light polarization imaging sensor array, a millimeter-wave radar, and an infrared thermal imager. The visible light polarization imaging sensor synchronously collects raw light intensity signals from the target area in four specific polarization directions to construct a raw data set containing complete polarization state information. The millimeter-wave radar acquires target reflection point cloud data at a preset scanning frequency, and the infrared thermal imager synchronously generates a thermal radiation intensity distribution map. Analyze the channel state sequence to extract channel gain, interference strength, and noise power parameters, dynamically calculate the power adjustment value based on quality of service constraints, synchronously perform spectrum reuse and time slot allocation, and jointly encode to generate cross-layer optimization parameters; Decompose cross-layer optimization parameters into calculated load characteristics and transmission demand characteristics, schedule edge nodes to implement multi-point coordination and link switching, and output prioritized resource scheduling instructions; Match available units in the physical resource pool, build secure and isolated virtual container templates and wavelength-graded virtual channels, deploy instances through atomic transactions, and generate virtual instance topologies with topology constraint verification. The virtual instance topology is converted into a normalized feature vector sample set, a dual-channel graph neural network is used to train a multi-task model, and online inference is used to generate power coefficient and resource weight matrix instructions. The process of generating power coefficient and resource weight matrix instructions by online reasoning includes: Receive the virtual instance topology and calculate the topology deviation:

[0007] in, is the topological deviation; is the total number of nodes in the virtual instance topology; is the current state vector; is the historical benchmark vector; is the Chebyshev norm;

[0008] Output deviation ,when When triggering the online fine-tuning of the model, is the topology deviation threshold; Furthermore, the parameter adjustment instruction generates:

[0009]

[0010] in, is the power adjustment amount; is the truncation function; is the weight matrix for power adjustment; is the global eigenvector; is a new resource weight allocation scheme; softmax(): normalized exponential function; is the weight matrix of resource weights; Output power adjustment and a new resource weight allocation scheme ; The power opcode and sparse matrix weight in the instruction are parsed by bit pattern matching, the transmitter power register and resource pool quota are updated, and the power value is fed back to the polarization sensor calibration link.

[0011] Furthermore, converting the virtual instance topology into a normalized feature vector sample set includes: The system periodically obtains the virtual instance topology historical dataset through the distributed storage interface, and the dataset is stored in the column database in the form of time series; The virtual instance topology includes a node attribute matrix and an edge connection matrix. The node attribute matrix records resource type identifiers, real-time computing unit occupancy, actual memory allocation values, and network bandwidth quotas. The edge connection matrix stores the globally unique identifiers of logical links, transmission delay sliding window measurements, and packet loss rate statistics. A topology snapshot is captured at a fixed sampling frequency, and the power control coefficient and resource allocation weight factor fed back at that moment are synchronously associated to form a spatiotemporally associated data unit.

[0012] Furthermore, according to the above drone detection method, it also includes: Perform one-hot encoding conversion on discrete resource type identifiers to generate binary feature vectors; The computing unit occupancy rate and memory allocation value are linearly mapped to the normalized interval using the minimum and maximum scaling algorithm; Applying a logarithmic transformation function to the transmission delay measurements compresses the data dynamic range.

[0013] Furthermore, according to the above drone detection method, it also includes: For each virtual node, calculate the arithmetic mean of the transmission delay and the maximum statistical value of the packet loss rate of all its associated logical links to generate the edge feature vector; The edge feature vector and the node attribute vector are connected end to end in the feature dimension to generate a topological state vector with unified dimension.

[0014] Furthermore, the dual-channel graph neural network is used to train the multi-task model, including: The model architecture adopts a dual-channel graph neural network design; the first channel integrates a graph attention mechanism to dynamically adjust the influence coefficients of neighboring nodes, and the second channel deploys multi-layer graph convolution operators to extract deep correlation features of logical links; The outputs of the two channels are subjected to feature concatenation in the fusion layer, concatenating the node feature vector and the edge feature vector into a joint representation vector.

[0015] Furthermore, according to the above drone detection method, it also includes: The training process implements a dynamic sample weighting strategy; abnormal samples that exceed the service quality indicator threshold are given a triple sampling weight; The loss function contains three optimization objectives: the delay prediction branch adopts the Huber loss function, the resource allocation branch adopts the cross entropy loss function, and the topological stability branch calculates the cosine similarity of the feature vectors of adjacent time slices.

[0016] Furthermore, according to the above-mentioned drone detection method, the method further includes: applying a sliding time window strategy in the model validation phase; selecting 24 consecutive hours of data as a training set and the subsequent 6 hours of data as an independent validation set; When the topology reconstruction error on the validation set increases for three consecutive evaluation cycles, the early stopping mechanism is triggered to roll back to the historical optimal weight state.

[0017] Furthermore, the online inference instructions for generating power coefficients and resource weight matrices include: Receive virtual instance topology data stream and convert node attribute matrix and edge connection matrix into normalized feature vector; A lightweight inference engine is used to load the model, and the graph convolution layer and the fully connected layer are merged into a single computing unit through operator fusion optimization.

[0018] Furthermore, according to the above drone detection method, it also includes: The online inference process uses a two-stage pipeline architecture; the first stage calculates the dynamic deviation of the current topology from the historical baseline state, and the second stage imposes a boundary constraint of -3dB to +3dB on the power adjustment coefficient; The parameter adjustment instruction adopts a binary coding format; the first sixteen bits store the target transmitting node identifier, the middle thirty-two bits record the quantized value of the power adjustment coefficient, and the last sixty-four bits are sparsely coded to store the resource allocation weight matrix.

[0019] Furthermore, a drone detection system based on visible light polarization imaging is applied to any of the above methods, comprising: An environmental perception module is configured to acquire raw signals from multiple source devices, perform non-uniformity correction and noise reduction, generate environmental perception data including polarization matrix, point cloud, and heat map through spatiotemporal registration, and construct a channel state sequence; a cross-layer optimization module configured to parse the channel state sequence to extract channel gain, interference strength, and noise power parameters, calculate power adjustment values ​​based on quality of service constraints, perform spectrum reuse and time slot allocation, and generate cross-layer optimization parameters; A resource scheduling module is configured to decompose cross-layer optimization parameters into calculation load characteristic quantities and transmission demand characteristic quantities, schedule edge nodes to implement multi-point coordination and link switching, and output resource scheduling instructions; A virtualization module is configured to match available units in the physical resource pool, build virtual container templates and virtual channels, deploy instances, and generate virtual instance topology; A model inference module, configured to convert the virtual instance topology into a set of feature vector samples, train the model using a dual-channel graph neural network, and generate power coefficient and resource weight matrix instructions; The execution feedback module is configured to parse the power operation code and sparse matrix weight in the instruction, update the power register and resource pool quota, and feed back the power value to the polarization sensor calibration link.

[0020] The key innovations of the present invention include: (1) A unified feature representation framework was constructed that can effectively fuse visible light polarization imaging (providing material, shape, and contour information after interference suppression) and millimeter wave radar (providing speed, distance, and penetration information) with two types of data with very different physical properties. This framework overcomes the problem of data heterogeneity and achieves deep complementarity and synergy at the information level.

[0021] (2) A joint learning architecture for spatiotemporal features was designed. This architecture organically combines a model that is good at capturing long-sequence temporal dependencies with a model that is good at extracting spatial global correlations, forming a synergistic effect. It achieves end-to-end, integrated modeling and analysis of target spatiotemporal features, and accurately captures the motion trajectory and spatial distribution evolution of dynamic targets.

[0022] (3) An intelligent dynamic allocation strategy for computing resources is proposed. This strategy dynamically adjusts the resource input of different data processing links (such as area selection, feature extraction depth, and model complexity) based on the real-time perception of environmental complexity (such as scene clutter and number of targets), target priority assessment (such as threat level and confidence), and the current resource load of the system, thereby maximizing the overall processing efficiency while ensuring the accuracy of core area and target detection.

[0023] The following are its main beneficial effects: (1) The core of this invention solves the problem that the detection performance of traditional single sensors (such as ordinary visible light cameras) is significantly reduced in complex backgrounds (such as clouds, strong light, weak light, and interference from swaying leaves) and dynamic scenes (high-speed drones, small targets, and multiple postures). By fusing visible light polarization imaging with millimeter-wave radar information, it can effectively utilize polarization characteristics to suppress background clutter, enhance target contours, and identify material characteristics. Combined with the precise speed and distance information of millimeter-wave radar, it forms a complementary advantage. This multimodal fusion strategy greatly improves the system's target recognition capability and overall detection confidence under harsh lighting and complex background interference, and significantly reduces missed detection and false detection rates.

[0024] (2) Through a spatiotemporal joint modeling approach, the present invention can simultaneously and collaboratively learn the target's spatial distribution characteristics (e.g., shape, size, and position) and its dynamic trajectory characteristics (e.g., speed, direction, and attitude changes) over time. This deep spatiotemporal fusion mechanism enables the system to not only accurately identify the target's current state but also effectively predict its motion trends, enabling stable and continuous tracking of fast-moving, maneuvering small UAV targets.

[0025] (3) In scenarios where large amounts of multi-source heterogeneous data (high-resolution polarization images, radar point clouds) need to be processed and complex spatiotemporal modeling needs to be performed, traditional systems often experience processing delays due to computing resource limitations, making them unable to meet the stringent real-time requirements of drone detection. The dynamic resource optimization mechanism introduced in this invention can intelligently adjust the depth and breadth of the data processing flow based on the complexity of the current environment, the target threat level, and the computing resource status. For example, processing can be simplified in resource-constrained or low-threat areas, while resources can be concentrated for detailed analysis in high-threat areas or when targets appear. This on-demand allocation strategy ensures that the system can maintain low-latency, high-frame-rate real-time detection and response capabilities even in complex and changing high-load environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flowchart of a method for detecting drones based on visible light polarization imaging provided in an embodiment of the present application; Figure 2 This is a structural block diagram of a drone detection system based on visible light polarization imaging provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] Example 1: Reference Figure 1 , is a flow chart of a method for detecting drones based on visible light polarization imaging provided by an embodiment of the present invention. The flow chart may include at least steps S100-S600: S100. Obtain the original signals of multi-source environmental perception devices, perform non-uniformity correction and noise reduction processing, generate environmental perception data including polarization matrix, point cloud, and heat map through spatiotemporal registration, and construct a time-aligned channel state sequence.

[0028] S200, parse the channel state sequence to extract channel gain, interference intensity, and noise power parameters, dynamically calculate the power adjustment value in combination with QoS constraints, synchronously perform spectrum multiplexing and time slot allocation, and jointly encode to generate cross-layer optimization parameters.

[0029] S300. Decompose cross-layer optimization parameters to calculate load characteristic quantities and transmission demand characteristic quantities, schedule edge nodes to implement multi-point coordination and link switching, and output priority-ordered resource scheduling instructions.

[0030] S400: Match available units in the physical resource pool, build secure and isolated virtual container templates and wavelength-graded virtual channels, deploy instances through atomic transactions, and generate virtual instance topologies with topology constraint verification.

[0031] S500: Convert the virtual instance topology into a normalized feature vector sample set, use a dual-channel graph neural network to train a multi-task model, and generate power coefficient and resource weight matrix instructions through online reasoning.

[0032] S600 , parsing the power operation code and sparse matrix weight in the instruction through bit pattern matching, updating the transmitter power register and resource pool quota, and feeding back the power value to the polarization sensor calibration link.

[0033] Step S100 at least includes steps S110-S130: S110: Obtain original signals from multi-source environmental perception devices, perform noise suppression and data calibration, and obtain environmental perception data.

[0034] The multi-source environmental perception equipment deployed in the monitoring area includes three heterogeneous sensor devices: a visible light polarization imaging sensor array, a millimeter-wave radar, and an infrared thermal imager. First, the visible light polarization imaging sensor synchronously collects raw light intensity signals from the target area in four specific polarization directions, constructing a raw dataset containing complete polarization state information. The millimeter-wave radar acquires target reflection point cloud data at a preset scanning frequency, and the infrared thermal imager simultaneously generates a thermal radiation intensity distribution map. Non-uniformity correction is performed on the polarization imaging data: Based on a reference image acquired under uniform illumination using a calibration plate, the response deviation coefficient of each imaging sensor pixel is calculated. A linear transformation algorithm is used to map the raw light intensity signal to a standardized response curve, eliminating inherent sensor response variations. For the millimeter-wave point cloud data, an adaptive range threshold filtering mechanism is used to eliminate multipath interference noise. This mechanism dynamically adjusts the effective detection range threshold in real time based on the radar cross-sectional area, automatically filtering out discrete noise points that exceed the range threshold. The infrared thermal imaging data is subjected to temporal median filtering for noise reduction. The median value of each pixel in the thermal radiation image is taken across multiple frames to effectively suppress transient thermal noise interference.

[0035] To unify the spatiotemporal benchmarks of multi-source data, a precise clock source is provided to all sensing devices via the GPS timing module, establishing a nanosecond-level synchronized time base. Spatial registration is performed using the coordinate points of fixed calibration objects in three-dimensional space. The millimeter-wave point cloud coordinate system is mapped to the polarization imaging coordinate system via affine transformation, completing the spatial alignment of cross-modal data. The final output of environmental perception data contains four standardized dimensions: a spatially registered four-channel polarization intensity matrix, a millimeter-wave scattered point cloud dataset aligned with the polarization imaging coordinate system, a thermal radiation intensity distribution map after time-domain filtering, and a synchronization data identifier containing a unified timestamp index. This environmental perception data is transmitted to the S120 processing module as the sole input source.

[0036] S120: Extract channel characteristic parameters from the environmental perception data, perform time-frequency domain conversion processing, and obtain channel state data.

[0037] Based on the environmental perception data output by S110, the physical layer channel characteristic parameters are extracted from three dimensions. In the visible light polarization imaging dimension, the light intensity vector and polarization characteristic parameters of each pixel are first calculated: the total light intensity parameter and the difference of the orthogonal polarization components are derived from the light intensity values ​​in the four polarization directions, and then the polarization degree index and polarization angle parameter that characterize the polarization characteristics are calculated. The polarization degree index of the full-frame pixels of the imaging screen is gridded and statistically analyzed to generate a polarization degree spatial distribution matrix with a resolution consistent with the original image. In the millimeter wave point cloud dimension, three-dimensional scattering features are extracted for the clustered point cloud clusters: the spatial centroid coordinates of the point cloud cluster, the mean Doppler frequency shift, the scattering intensity dispersion and other characteristic vectors that characterize the physical characteristics of the target are calculated. In the infrared thermal imaging dimension, the temperature gradient changes of adjacent pixels are analyzed along the target movement direction to form a characteristic vector that describes the dynamic characteristics of heat conduction.

[0038] Furthermore, these features are deeply processed in the time-frequency domain. The spatial distribution matrix of the degree of polarization (DOP) is processed using a sliding window Fourier transform. The frequency-domain energy distribution of the DOP parameter at each pixel is calculated within a continuous multi-frame time window, and the low-frequency energy fraction is extracted as a eigenvector representing the time-varying characteristics. The millimeter-wave Doppler shift data is decomposed using a wavelet packet algorithm to extract the energy entropy parameters of the target subband at a specific decomposition level, generating a frequency-domain eigenvector. The infrared heat conduction eigenvector is input into a recurrent neural network for time series modeling, outputting a latent state eigenvector representing the temporal evolution. Finally, this fusion generates five-dimensional channel state data: the original DOP spatial distribution matrix, the polarization time-varying eigenvector, the millimeter-wave scattering frequency-domain eigenvector, the heat conduction time-series eigenvector, and an alignment index to ensure time synchronization of the multi-source features. This channel state data is fully transmitted to the S130 processing module via a high-speed data bus.

[0039] S130 , performing sliding window segmentation on the channel state data to generate a channel state sequence.

[0040] After acquiring the channel state data output by S120, a time mapping relationship for multi-source data is established based on the timestamp alignment index. A sliding processing window with a fixed duration and step interval is set to sequentially capture data segments along the time axis. Four core operations are performed during window data processing: first, a time alignment check is performed to detect the timestamp deviation of each modal feature data within the window. When the deviation exceeds a preset threshold, a spline interpolation algorithm is used for resampling to achieve precise alignment. Second, feature dimension normalization is performed to downsample the polarization spatial matrix to a standard resolution, compress the millimeter-wave feature vector to a low-dimensional space through principal component analysis, and perform numerical normalization mapping on the heat conduction feature vector. Then, spatial feature encoding is performed to apply a convolution kernel operation to the polarization matrix to extract local spatial correlation features and generate a spatial feature map tensor. Finally, serialization and packaging are performed to arrange the processed multi-source features in a time series, forming a structured data packet containing five elements: precise time window start and end markers, a standardized spatial feature map tensor, a reduced-dimensional millimeter-wave frequency-domain feature vector, a normalized heat conduction time series feature vector, and a feature integrity check mark.

[0041] During the sliding window processing, an overlap retention mechanism is employed to ensure temporal continuity: When the window is stepped, the end portion of the data in the previous window is retained as the starting data for the new window. The resulting channel state sequence is a chronologically ordered collection of serialized data packets, each containing spatiotemporally aligned multimodal feature data. This data is output to the S210 adaptive power control module via a high-speed data interface. This sequence serves as the core input for S210 to calculate the transmit power strategy in conjunction with quality of service constraints. It serves as the underlying feature source for historical data samples during the S500 training phase, and dynamically optimizes environmental perception data acquisition parameters through a feedback mechanism in S630.

[0042] Step S200 at least includes steps S210-S230: S210: Acquire a channel state sequence, and perform adaptive power control decision in combination with quality of service constraints.

[0043] The system obtains the channel state sequence generated by step S130 as a basic input source through a high-speed data interface. The channel state sequence consists of multiple data packets arranged in chronological order, and each data packet contains multimodal channel feature data that has been time-space aligned within a specific time window. During specific execution, the system first parses the pre-configured quality of service constraints, which include three core threshold parameters: the minimum communication rate threshold defines the lower limit of the bit rate that the transmission link must maintain, the maximum allowable end-to-end transmission delay threshold limits the upper limit of the total delay of the data packet from the sender to the receiver, and the maximum tolerable bit error rate threshold specifies the acceptable error bit ratio limit for signal demodulation. These threshold parameters are loaded into the decision engine as rigid boundary conditions for power control decisions.

[0044] The channel state sequence is parsed packet by packet, and the standardized spatial feature map tensor contained in each data packet is input into the feature extraction module. This module extracts the channel gain distribution characteristics through convolution kernel operations and generates a quantitative index value that represents the signal propagation path loss. At the same time, the interference intensity parameter is parsed from the millimeter wave frequency domain feature vector. This parameter reflects the superimposed signal energy caused by other transmitting sources in the same frequency band; the noise power spectral density parameter is extracted from the heat conduction time series feature vector. This parameter describes the distribution characteristics of the ambient thermal noise in the frequency domain. The system combines the channel gain, interference intensity, and noise power spectral density parameters parsed in each time window into the current channel quality indicator triple.

[0045] When executing key decision-making judgments, the system compares the current channel quality indicator triplet with the service quality constraint threshold. When the channel gain value exceeds the theoretical threshold required to maintain the minimum communication rate threshold, and the value predicted based on historical transmission delay statistics does not exceed the maximum allowable end-to-end transmission delay threshold, and the theoretical bit error rate calculated based on the bit error rate model is lower than the maximum tolerable bit error rate threshold, the system generates a transmit power reduction instruction. This instruction includes a power adjustment calculation process: based on the difference in channel gain exceeding the theoretical threshold, the corresponding step level is matched in the preset power control step matrix to generate a specific power reduction value. Conversely, when the channel gain is detected to be approaching or below the theoretical threshold, or the transmission delay is predicted to exceed the delay threshold, or the calculated bit error rate exceeds the tolerance threshold, a transmit power increase instruction is generated. The power increment is dynamically matched in the power control step matrix based on the degree of indicator deviation.

[0046] The decision-making mechanism includes an adaptive execution frequency adjustment function. The system continuously monitors the parameter change rate of adjacent data packets in the channel state sequence. When the variance of the channel gain or interference strength changes exceeds a preset fluctuation threshold, a high-frequency decision mode is activated to improve response speed. When the parameter change rate falls below the stability threshold, a low-frequency decision mode is switched to reduce computational overhead. All decision results are encapsulated as a structured instruction set. Each instruction explicitly includes a timestamp, a globally unique identifier of the target transmitter, a power adjustment direction indicator (an up / down enumeration value), the absolute value of the recommended transmit power, and a validity period parameter for the instruction. This structured instruction set is passed as output to the resource allocation module.

[0047] S220 , executing a dynamic resource allocation strategy based on the adaptive power control decision and the resource demand prediction result.

[0048] The system simultaneously receives the adaptive power control decision instruction set output by S210 and the resource demand forecast results generated by the independent forecast module. The resource demand forecast results analyze historical traffic curves, the current number of access devices, and the service growth model to output quantitative predictions of the total spectrum resource demand, total time slot resource demand, and total computing resource demand within the future time window. Spectrum resource demand is expressed as bandwidth demand in megahertz, time slot resource demand is expressed as the number of time slots to be allocated per second, and computing resource demand is converted to standard computing unit equivalents.

[0049] Before executing resource allocation, the system first creates a real-time snapshot of the physical resource pool's status. Regarding spectrum resources, it scans the currently available frequency bands and records each band's center frequency, bandwidth, and interference coupling coefficient for adjacent bands. Regarding time slot resources, it polls the time slot allocation status table to obtain the index of allocable time slot locations. Regarding compute resources, it collects the CPU core availability, remaining memory capacity, and remaining storage space of each edge node. This resource pool status snapshot is then compared with the resource demand forecast to generate a resource gap distribution map.

[0050] The core allocation process implements a three-layer resource coordination mechanism. At the spectrum resource allocation layer, the target transmitter identifier and power adjustment direction in the S210 power control instruction set are parsed. For transmitters performing power reduction, the resulting interference radius reduction is calculated. When the interference radius reduction exceeds the preset reuse activation threshold, a spectrum reuse algorithm is initiated within the coverage area. The frequency reuse factor within the area is recalculated, freeing up previously protected frequency band gaps as available resources. A new frequency band partitioning scheme is generated and allocated to incoming devices. For transmitters performing power increase, the interference enhancement range is calculated based on the power increment. A guard band isolation zone is set within the interference enhancement range, and the resources occupied by this isolation zone are included in the total allocation. At the time slot resource allocation layer, data streams are divided into real-time and non-real-time flows based on a service flow priority classifier. Real-time flows are allocated continuous time slot blocks using an absolute priority preemption mechanism, while non-real-time flows are allocated discrete time slots using a weighted round-robin algorithm with dynamically adjustable weights. Weights are dynamically configured based on the traffic peak distribution predicted by resource demand. At the computing resource allocation layer, a computing task feature vector (including CPU instruction set type requirements, memory bandwidth requirements, and storage IO throughput requirements) is established. Combined with the real-time load matrix of the edge node, the minimum load difference matching algorithm is used to schedule computing tasks to the target node, and an accurate core number quota, memory capacity quota, and storage space quota are allocated to each task.

[0051] Finally, a multidimensional resource allocation plan is generated: the spectrum allocation plan is recorded as a frequency band allocation record table, with each record containing the link identifier, starting frequency, ending frequency, and effective period; the time slot allocation plan is recorded as a time slot allocation mapping table, with each record containing the service flow identifier, frame period number, time slot starting position, and number of occupied time slots; and the computing resource allocation plan is recorded as a computing resource allocation list, with each list containing the task identifier, node identifier, allocated core number list, memory block address, and storage volume label. This plan is output to the joint encoding module via the resource management bus.

[0052] S230 : Jointly encode the adaptive power control decision and the dynamic resource allocation strategy to generate cross-layer optimization parameters.

[0053] The system takes the power control decision instruction set output by S210 and the resource allocation plan generated by S220 as dual input sources and implements cross-layer parameter fusion through a joint encoder. The encoder loads a predefined protocol mapping template, which contains five structured fields: the decision type identifier field uses a 32-bit integer to define the instruction type as a cross-layer collaborative optimization instruction; the power control field uses a TLV structure to encapsulate the target transmitter identifier, power operation code (preset SET_POWER / ADJUST_POWER operation code enumeration value), and power value; the spectrum resource field uses a dynamic array structure to store the contents of the frequency band allocation record table, with each array element containing three parameters: link identifier, start frequency, and end frequency; the time slot resource field uses a bitfield compression format to encode the time slot allocation mapping table, with each record represented by a combination of the starting frame number (16 bits), the starting time slot offset (8 bits), and the number of consecutive time slots (8 bits); and the computation resource field uses a hierarchical description method, with the top layer recording the node identifier and task identifier, and the lower layer linking the core allocation list, memory allocation block descriptor, and storage volume descriptor through pointers.

[0054] The data filling process strictly follows the protocol mapping rules: the target transmitter identifier and power adjustment value are extracted from the power control decision instruction set, converted into a power operation code and power value, and then written into the power control field; the frequency band parameters of each link are extracted from the spectrum allocation plan, arranged in ascending order according to the starting frequency of the frequency band, and then written into the dynamic array of the spectrum resource field; the time slot allocation record is extracted from the time slot allocation mapping table, and written into the time slot resource field after bit field compression; the parameters of each level are extracted from the computing resource allocation list, and the pointer link structure is constructed and written into the computing resource field.

[0055] A protocol control header is added when executing structured encapsulation: the version number field identifies the protocol revision, the timestamp field records the instruction generation timestamp, and the validity period field defines the maximum lifetime of the instruction. A cyclic redundancy check code (CRC32) is generated for the entire data packet using the CRC32 checksum algorithm and appended to the end of the packet to form a complete transmission frame. The final output of the cross-layer optimization parameters is a binary data stream with a strict syntactic structure. Its physical layer contains a protocol control header and a data body. The data body contains layered encapsulation of power control instructions, spectrum allocation parameters, time slot allocation parameters, and computing resource allocation parameters. The checksum provides end-to-end integrity protection. This data stream is distributed to each execution node via the control plane interface.

[0056] Step S300 at least includes steps S310-S330: S310: Obtain cross-layer optimization parameters and decompose them into a computing task subset and a communication task subset.

[0057] The cross-layer optimization parameters generated by S230 are obtained as the core input source. The parameters include the jointly encoded adaptive power control decision and dynamic resource allocation strategy. A structured parsing operation is performed on the cross-layer optimization parameters through a dedicated parameter parsing module. This process first identifies the decision marker bits preset in the joint coding matrix, where the power control identifier uses a high-order byte mask to identify the power adjustment operation type, and the resource allocation identifier uses a low-order byte segment to identify the coding block boundaries of spectrum resources, time slot resources, and computing resources. Based on the bit pattern matching algorithm, the parsing engine splits the original matrix data stream into four independent data channels according to the identifier type: the power control instruction channel, the spectrum allocation parameter channel, the time slot allocation parameter channel, and the computing resource allocation parameter channel. During the computational task subset decomposition phase, the first N-dimensional feature vectors are extracted from the computational resource allocation parameter channel to construct computational load characteristics. These characteristics include a task complexity indicator (mapped from the CPU instruction set type to a preset complexity level) that characterizes computational intensity, a delay sensitivity coefficient (normalized between 0.0 and 1.0) that defines service delay constraints, and a memory demand threshold (absolute value in megabytes) that identifies peak memory demand. During the communication task subset decomposition phase, the M-dimensional feature vectors extracted from the fusion of the spectrum allocation parameter channel and the time slot allocation parameter channel are used to construct transmission demand characteristics. These characteristics include a bandwidth allocation priority (classified by priority weight for real-time and non-real-time flows) that reflects service bandwidth requirements, a link stability index (calculated as a sliding average based on historical packet loss rates) that characterizes wireless channel reliability, and a bit error rate tolerance threshold (floating-point value expressed in scientific notation) that defines the transmission quality baseline. The entire decomposition operation is hardware accelerated through a two-stage pipeline architecture: the first-stage pipeline implements identifier classification by a bit matching circuit, which uses a pre-programmed bit pattern template for parallel matching, identifies the marker bit boundary and triggers the channel separation signal within a single clock cycle; the second-stage pipeline implements feature vector segmentation through a dual-port buffer area, with the front port outputting the computational load feature quantity to the edge computing scheduler, and the back port outputting the transmission demand feature quantity to the link selector, achieving zero-latency parallel output of the two types of feature quantities.

[0058] S320. Schedule edge computing nodes according to the computing task subset, and schedule transmission links according to the communication task subset.

[0059] Edge computing node scheduling is performed based on the computational load characteristics output by S310. The scheduler first accesses the edge computing node resource registry to obtain real-time resource status. The registry continuously collects three key indicators of each node through distributed probes: the number of available computing units (indicating the number of remaining CPU cores), the graphics processor load rate (the 3D rendering resource utilization expressed as a percentage), and the memory utilization ratio (the ratio of allocated memory to total capacity). Node fitness matching calculations are then performed: the task complexity identifier in the computational load characteristics is input into the node processing capacity scoring function. This function generates a basic score based on the compatibility of the complexity level and the node hardware configuration, superimposes the negative weighted values ​​of the graphics processor load rate and the memory utilization ratio, and finally outputs a node fitness matrix quantized to 0.100 points. The dynamic allocation strategy incorporates two core mechanisms: When the latency sensitivity coefficient exceeds a preset threshold (typically 0.8), a multi-point coordination mechanism is activated. Using a task decomposition algorithm, a single task is split into K logically contiguous parallel subtasks. L edge nodes with a physical distance of less than 200 meters are selected from a node geographic topology database to form a coordination group, and subtask packets are distributed via multicast. When the memory demand threshold exceeds 80% of a node's available memory, a vertical expansion mechanism is triggered, invoking a pre-configured memory pool sharing interface to map the excess memory demand to the virtual address space of the shared memory pool. Dynamically attaching physical memory blocks is then implemented via the memory controller. During the scheduling of communication task subsets, the link selector loads the transmission demand characteristics output by S310 as a decision-making basis. A set of candidate links is initialized based on bandwidth allocation priorities. A weighted evaluation module then calculates the degree of compatibility between the link stability index and the bit error rate tolerance threshold, assigning a weight of 0.6 to the link stability index and a weight of 0.4 to the bit error rate tolerance threshold. The link with the highest overall score is then selected as the primary transmission path. When the stability index of the millimeter wave link falls below the threshold (the empirical value is 35), the seamless switching mechanism is triggered, automatically forwarding the service flow to the Sub6GHz backup link, and recalculating the optimal frequency band combination through the frequency band reallocation module. This process is completed within 50 milliseconds and maintains service flow continuity.

[0060] S330: Prioritize the response status of edge computing nodes and transmission links, and output resource scheduling instructions.

[0061] The edge node resource response message and transmission link status report scheduled by S320 are received as dual inputs. The resource response message contains three core fields: task processing delay estimate (a microsecond-level processing time prediction value), node resource reservation status code (a status code indicating the success or failure of CPU / GPU / memory resource reservation), and cache queue length (an integer value representing the number of tasks currently to be processed). The link status report contains the actual allocated bandwidth value (the measured bandwidth in megahertz), the channel quality indicator value (a 0100 scale value converted based on the signal-to-noise ratio), and the transmission delay measurement value (a millisecond-level measurement result of the end-to-end one-way transmission delay). The priority sorting operation is implemented through a three-level pipeline evaluation architecture: the first-level evaluator calculates the node response efficiency ratio, whose input is the product of the estimated task processing delay and the node resource reservation status code, where the status code value is 1.0 for the success state and 0.1 for the failure state. The product result is normalized and converted into a computing resource priority score in the range of 0100; the second-level evaluator calculates the link transmission efficiency ratio, whose input is the weighted sum of the actual allocated bandwidth value and the channel quality indicator value, with the bandwidth value participating in the calculation with a weight of 0.7 and the channel quality participating in the calculation with a weight of 0.3. The weighted sum result is logarithmically transformed and outputs a communication resource priority score in the range of 0100; the third-level evaluator performs joint sorting, using a two-layer sorting algorithm to first sort in descending order by the computing resource priority score, and then in ascending order by the communication resource priority score if the scores are the same, to generate a globally unique resource priority sequence. A three-level encapsulation structure is used when outputting resource scheduling instructions: the instruction header contains a 32-bit precision global timestamp (synchronized with the system master clock) and a 64-bit transaction identifier (including the task identifier and version number); the instruction body records the sorted node identifier sequence (a list of node IDs arranged in descending order of priority) and the link identifier sequence (a list of link IDs arranged in descending order of score), and each identifier is attached with a weight coefficient to indicate the degree of priority difference; an 8-bit resource synchronization flag is attached to the end of the instruction. When the 0th bit of the flag is set, the resource matching operation of S410 is triggered. The 1st bit indicates whether cross-domain resource coordination is required, and the remaining bits are reserved for extended functions.

[0062] Step S400 at least includes steps S410-S430: S410: Obtain resource scheduling instructions and match available resource units in a physical resource pool.

[0063] Receive the resource scheduling instruction output from step S330, which includes a globally prioritized subset of computing tasks and a subset of communication tasks, along with their resource requirement attributes. These resource requirement attributes include the number of CPU cores, GPU memory capacity, and memory bandwidth threshold parameters required for computing-intensive tasks, as well as the upper limit on end-to-end transmission latency, minimum guaranteed bandwidth, and bit error rate tolerance threshold defined for communication-intensive tasks. When performing a matching operation, first parse the task type identifier in the instruction (the computing-intensive identifier points to the floating-point operation-intensive or matrix operation-intensive subclass, and the communication-intensive identifier distinguishes between real-time streaming and non-real-time batch transmission types). Access the resource status database of the physical resource pool, which continuously synchronizes the real-time physical resource parameters of the edge computing node: including CPU core availability (the percentage of idle cores collected via kernel performance counters), GPU memory occupancy (memory allocation status obtained via the device driver interface), remaining memory capacity (the physical memory available value in megabytes), the bandwidth idle rate of the transmission link (the proportion of available bandwidth converted from the port queue depth), and the port availability status (the link protocol layer handshake signal detection result).

[0064] The resource matching engine performs a two-tiered screening logic: for a subset of computing tasks, it generates a computing power requirement constraint (requiring that the target node's single-core clock speed multiplied by the number of available cores be greater than or equal to a minimum computing power threshold) and a latency constraint (the sum of the node processing latency and the network transmission latency be less than a preset threshold). For a subset of communication tasks, it generates a bandwidth redundancy constraint (the link's available bandwidth be greater than 120% of the task's required bandwidth) and a reliability constraint (the link's bit error rate be less than 1e5 and support forward error correction). A two-way verification mechanism ensures matching accuracy: In the forward verification phase, resource screening criteria are generated based on the constraint expressions, and candidate resource units that meet all criteria are retrieved from the physical resource pool. In the reverse verification phase, the service level agreement commitments of the candidate resource units are extracted (including the 99.99% computing task completion rate promised by the computing node and the millisecond-level delay jitter limit guaranteed by the communication link), verifying whether these metrics meet the fault tolerance and real-time requirements of the task's quality of service constraints. The final output list of available resource units is stored in a structured data format. Each record contains a globally unique identifier for the resource unit (a 128-bit hash value generated based on the naming rules of the resource pool registry), a resource type classification code (a binary classification code for computing resources / communication resources), physical location coordinates (a triplet of latitude, longitude, and elevation), and real-time performance indicators (the computing node includes the current CPU load rate and memory fragmentation rate, and the communication link includes the current round-trip delay and packet loss rate). This list serves as the input data basis for step S420.

[0065] S420: Construct a virtualized resource template according to available resource units and execute resource instantiation deployment.

[0066] The list of available resource units output from step S410 is parsed and divided into computing resource groups and communication resource groups based on the resource type classification code. The virtualized resource template construction process is implemented at two levels: at the computing resource level, a virtual computing container template is created based on the heterogeneous characteristics of edge computing nodes. Specifically, key parameters of the computing resource units are extracted, including the number of CPU cores (the number of allocable logical cores), GPU model (model identifier supporting the CUDA computing architecture), and memory capacity (contiguous physical memory block size). Lightweight virtual machine instances are then configured at the hypervisor level through the kernel-level virtualization module. The configuration process includes three core operations: first, an independent security domain isolation environment is allocated to each virtual machine, which implements data isolation through a memory encryption engine and IO access control lists; second, a time slot scheduling strategy is configured to bind the virtual machine time slice to the physical CPU clock cycle, and a time slicing round-robin algorithm is used to ensure multi-task parallelism; third, a pass-through mode is enabled for GPU resources, bypassing the virtualization layer and directly mapping the physical GPU device to the virtual machine device space to ensure the computational efficiency of the polarization imaging processing task; at the same time, a dynamic quota adjustment mechanism is configured for memory resources, which automatically expands the memory allocation limit based on the memory pressure indicators collected by the virtual machine monitor to adapt to sudden memory load demands.

[0067] At the communication resource level, a virtual channel template is created based on the physical characteristics of the transmission link. Specifically, the core parameters of the communication resource unit are extracted, including available bandwidth (guaranteed minimum bandwidth value), transmission delay (physical layer signal propagation delay), and port protocol (an identifier that supports Ethernet / Fibre Channel protocols), and a virtual switching path is configured on the data plane through a software-defined network controller. The configuration uses a segmented routing strategy to implement differentiated services: exclusive wavelength channels are allocated for high-priority communication tasks (such as drone control instructions), and a wavelength selection switch is used to lock a specific optical wavelength to establish an end-to-end optical path; statistical multiplexing channels are deployed for low-priority tasks (such as historical data return), and time slot resources are dynamically allocated in the time-division multiplexing frame structure; and a quality of service identifier is injected into the border gateway protocol routing announcement message to enable core network equipment to identify the service level of the virtual channel.

[0068] After the template is built, resource instantiation and deployment are executed: the infrastructure management interface (following the RESTful API specification) is called to deploy the virtual computing container template to the selected edge computing node physical resources. This process writes the virtual machine configuration descriptor to the host machine's virtualization control register through the device driver layer; the flow table rules are sent to the selected transmission link switching device using the standard flow table format of OpenFlow protocol version 1.5. The flow table entries clearly include input port matching rules, service quality identifier re-marking actions, and output port forwarding instructions. The deployment process uses an atomic transaction mechanism to ensure consistency: a two-phase commit protocol is used to coordinate the deployment operations of multiple resource units. If the instantiation of any resource unit fails (such as a virtual machine startup timeout or flow table delivery error), an automatic rollback mechanism is triggered to release the allocated resources (including destroying the created virtual machine instance and revoking the delivered flow table entries) until all associated resource units return a deployment success response. The final generated virtual resource instance set contains three core elements: virtual machine instance ID (a globally unique identifier composed of the host identifier and the virtual machine serial number), virtual channel ID (a 128-bit identifier composed of the source and destination node IDs and the service level code), and resource binding relationship (a table that records the mapping relationship between virtual machine instances and host computing nodes, and virtual channels and physical links).

[0069] S430: Perform logical connection mapping on the resource instantiation deployment result to generate a virtual instance topology.

[0070] The set of virtual resource instances output from step S420 is obtained, and the physical location coordinates of the virtual machine instance (used to calculate geographic distance), network address information (IPv6 address and VLAN tag combination), and endpoint identifiers of the virtual channel (source port MAC address and destination port IP address pair) are extracted from it. The logical connection mapping operation is performed in three consecutive stages: During the computing node interconnection relationship construction phase, based on the network address information of the virtual machine instance, the network topology parameters between each instance (including the number of transmission path hops, the transmission delay of each hop, and the path redundancy index) are obtained through the enhanced internal gateway routing protocol. In view of the task characteristics of the polarization imaging processing pipeline: a point-to-point direct connection channel is established for the image preprocessing node and the feature recognition node, and a static routing table is configured at the IP layer to bypass the core router to achieve direct communication; in view of the data distribution requirements of the multimodal fusion task, a multicast communication tree is constructed for the spectral analysis node and the polarization clustering node, and a shortest path tree with the spectral analysis node as the root is constructed through the protocol-independent multicast protocol. The connection relationship construction adopts an improved minimum spanning tree algorithm to optimize the communication cost: with the virtual machine instance as the vertex and the geographical distance multiplied by the link unit price as the weight edge, the minimum communication cost topology structure is generated through the Prim algorithm, and the equal-cost multipath rule is automatically injected to achieve load balancing.

[0071] During the communication link load rule mapping phase, the service quality level identifier (QoS) encoded in the virtual channel ID is parsed (categorized as high reliability and high real-time). High-reliability channels (such as the drone control command transmission channel) are mapped to a dual-fiber redundant link architecture, switching to the backup path within 50 milliseconds if the primary fiber fails. A strict priority queue scheduling policy is configured for real-time channels (such as polarized video streams), assigning the highest priority queue to the switch egress queue and setting a minimum bandwidth guarantee. Traffic shaping rules are also incorporated into the mapping process: Based on the bandwidth allocation weight coefficient (normalized value from 0.0 to 1.0) in the dynamic resource allocation policy of step S220, token bucket parameters are set for each virtual channel (token generation rate = weight coefficient × total physical link bandwidth, bucket depth = maximum allowed burst traffic × weight coefficient).

[0072] During the topology description framework generation phase, the aforementioned node connectivity and link load rules are integrated into a directed graph structure model: With virtual machine instances as vertices, vertex attributes include resource type identifiers (computing node classification codes) and computing capacity values ​​(peak computing power in TFLOPS). With virtual channels as directed edges, edge attributes include transmission protocol types (TCP / UDP / RDMA protocol identifiers), bandwidth reservation values ​​(minimum guaranteed bandwidth in Mbps), and maximum allowable latency (end-to-end transmission latency upper limit). Constraint verification is performed through a topology verification module: First, the sum of the latency of all critical paths (such as the path from the preprocessing node to the feature recognition node) is calculated to ensure that it does not exceed the end-to-end latency budget defined in the S210 power control decision. Second, the wavelength allocation non-conflict condition is verified, checking that there are no duplicate wavelength allocation records on the same fiber link and that the optical layer resource allocation rules required by the S320 transmission link scheduling are met. The final output is a virtual instance topology file that complies with the GraphML 1.2 standard. This file uses an XML syntax structure to record complete node configuration parameters (including virtual machine startup command line parameters), link parameters (including token bucket configuration details), and routing policy table (static routing and equal-cost multi-path entries). This file serves as the topology state input for step S510.

[0073] Step 500 at least includes steps S510-S530: S510: Obtain historical operation data of the virtual instance topology and construct a topology state sample set.

[0074] The system periodically obtains the virtual instance topology historical data set generated by step S430 through the distributed storage interface, and the data set is stored in a columnar database in the form of a time series. Each virtual instance topology contains two core data structures: the node attribute matrix records the resource type identifier, real-time occupancy of the computing unit, actual memory allocation value and network bandwidth quota configuration parameters of each virtual node in the form of a two-dimensional table; the edge connection matrix stores the globally unique identifier of the logical link between nodes, the transmission delay sliding window measurement value and the packet loss rate statistics. During the data extraction process, the system establishes a timestamp alignment index mechanism: the topology snapshot is captured at a fixed sampling frequency, and the power control coefficient and resource allocation weight factor fed back by the S600 resource control module at that moment are synchronously associated to form a spatiotemporal associated data unit.

[0075] During structured preprocessing of the raw topology data, feature engineering transformations are implemented. For discrete resource type identifiers in the node attribute matrix, a one-hot encoding converter is used to generate binary feature vectors, ensuring that each resource type corresponds to a unique activation bit pattern. For continuous numerical features, including computational unit occupancy and memory allocation values, a min-max scaling algorithm is used to linearly map them to a normalized range between zero and one, eliminating the impact of dimensional differences on subsequent analysis. Furthermore, the transmission delay measurements in the edge connection matrix are processed, and a logarithmic transformation function is applied to compress the data dynamic range to address the problem of non-convergence in model training caused by long-tail distributions. The key feature fusion stage utilizes a neighborhood aggregation strategy. For each virtual node, the arithmetic mean of the transmission delay and the statistical maximum of the packet loss rate of all its associated logical links are calculated to form an edge feature vector representing the connection quality. This edge feature vector is then concatenated with the node's own attribute vector along the feature dimension to generate a unified topology state vector. The final output topology state sample set consists of a sequence of timestamp-labeled feature vectors. Data masking is performed before being written to the training database, and sensitive configuration parameters are hidden through field masking, providing a standardized input source for network optimization model training.

[0076] S520: Training a network optimization model based on the topology state sample set and the service quality indicator.

[0077] The system loads the topology state sample set constructed by S510 and performs model training in combination with the service quality constraints defined in S200. The service quality indicators include the end-to-end transmission delay threshold, the task completion rate baseline value, and the system energy consumption upper limit. These indicators serve as supervisory signals to guide the optimization direction of the model. The model architecture adopts a dual-channel graph neural network design: the first channel integrates the graph attention mechanism, dynamically adjusts the influence coefficient of neighbor nodes through learnable weight parameters, and captures the asymmetric dependency relationship between virtual nodes; the second channel deploys multi-layer graph convolution operators, iteratively performs feature propagation on the edge connection matrix, and extracts the deep correlation features of the logical link. The outputs of the two channels implement feature cascade operations in the fusion layer, splicing the node feature vector and the edge feature vector into a joint representation vector.

[0078] The training process implements a dynamic sample weighting strategy: when loading batch data, outliers exceeding the service quality threshold are assigned a triple sampling weight. This enhances the model's ability to detect critical conditions, such as when delays exceed the specified threshold or task failures are detected. The loss function design incorporates three optimization objectives: the delay prediction branch uses the Huber loss function to measure the deviation between the predicted transmission delay and the actual measured value. This function automatically switches to a linear loss mode to prevent gradient explosion when the error is large. The resource allocation branch uses a cross-entropy loss function to compare the resource allocation weights output by the model with the actual policy feedback from the S600. The topology stability branch calculates the cosine similarity of the feature vectors of adjacent time slices as a regularization constraint to suppress unnecessary oscillations in the topology structure. A sliding time window strategy is used during the model validation phase: 24 consecutive hours of topology state data are used as the training set, and the following six hours of data are used as an independent validation set. If the topology reconstruction error on the validation set shows an increasing trend for three consecutive evaluation cycles, an early stopping mechanism is triggered, and the model automatically rolls back to the historical optimal weight state. The resulting network optimization model consists of a graph neural network encoder and a multi-task prediction head. The encoder outputs a latent space feature vector with a dimension of 256. The multi-task prediction head, through a fully connected layer, outputs power adjustment coefficients and a resource reallocation weight matrix in parallel. Model parameters are synchronized to all edge nodes via a distributed parameter server to ensure global inference consistency.

[0079] S530: Input the real-time virtual instance topology into the network optimization model to generate parameter adjustment instructions.

[0080] The system receives the virtual instance topology data stream output in real time by the S430 module and performs topology data conversion operations: the node attribute matrix and edge connection matrix of the current topology are converted into standardized feature vectors in strict accordance with the structured processing rules defined by S510. Real-time verification is implemented before model inference: if it is detected that the difference between the topology generation timestamp and the current system time exceeds 500 milliseconds, the expired topology is discarded and a regeneration request is triggered. The model loading link is implemented using a lightweight inference engine: the network optimization model trained by S520 is converted into an intermediate representation format of the computational graph, and the graph convolution layer and the fully connected layer are merged into a single computing unit through operator fusion optimization technology to reduce memory access overhead during inference.

[0081] The online inference process utilizes a two-stage pipeline architecture. In the first stage, after the graph neural network encoder outputs the latent space feature vector, the topology stability monitoring submodule is immediately activated to calculate the dynamic deviation of the current topology from the historical baseline state. In the second stage, at the output of the multi-task prediction head, bounds are imposed on the power adjustment coefficient, limiting its value to a physically feasible range of -3dB to +3dB. The resource allocation weight matrix is ​​also probabilistically normalized to ensure that the sum of all resource allocation weights equals unity. The parameter adjustment instruction generation stage incorporates a security verification mechanism: the predicted power adjustment coefficient is fed into the S200's quality of service constraint verifier. If simulation results indicate that the task completion rate falls below a preset threshold, the backup strategy generator is activated to invoke the historically optimal parameter configuration. The final instruction output is encoded in binary format: the first 16 bits store the globally unique identifier of the target transmitting node; the middle 32 bits record the quantized value of the power adjustment coefficient with a quantization accuracy of 0.1dB; and the last 64 bits use sparse matrix coding to store the resource allocation weight matrix. The instruction is transmitted to the S610 resource execution module via a high-speed data bus for parsing and implementation, forming a closed-loop control chain from topology perception to parameter regulation.

[0082] In another embodiment: S510: Obtain historical operating data of the virtual instance topology and construct a topology state sample set.

[0083] The virtual instance topology history dataset generated by S430 is periodically obtained through the distributed storage interface and stored in a columnar database in the form of time series. Each virtual instance topology contains: Node attribute matrix: records resource type identifiers , computing unit real-time occupancy (scope ), actual value of memory allocation and network bandwidth quotas; Edge connection matrix: stores the globally unique identifier of the logical link and the transmission delay sliding window measurement value and packet loss rate statistics ; Establish a timestamp alignment index mechanism: capture a topology snapshot at a 10Hz sampling frequency and synchronously associate it with the power control coefficient fed back by S600 at that moment and resource allocation weight factors , forming a spatiotemporal correlation data unit. Implementing a structured preprocessing process: The expression for node feature engineering transformation is:

[0084] in:

[0085] node The normalized attribute vector of ; :node Resource type identifier (discrete enumeration value: 0 = polarization imaging unit, 1 = millimeter wave radar unit); : Calculate unit occupancy; : Actual value of memory allocation; : Minimum / maximum computing occupancy of historical data sets; : Minimum / maximum memory value of historical data set; : vector concatenation operator; : One-hot encoding function; Output: Node The normalized attribute vector of

[0086] Furthermore, the delay characteristics are transformed:

[0087] in, : Logical link The transmission delay measurement value of : Delay after logarithmic transformation; Input: original delay measured by S120 in real time, output: smoothed delay characteristics; Furthermore, neighborhood feature aggregation is constructed: For nodes Computation: Generate edge eigenvectors:

[0088] in, is the mean delay of the associated link; is the maximum packet loss rate; Topology state vector generation:

[0089] in, For nodes The topological state vector of : node feature vector; : Neighborhood feature aggregation; Output: Node The topological state vector ; Final sample set ,in is the timestamp, is the total number of nodes, is the number of samples. Indicates the timestamp When the node in the virtual instance topology The topological state vector of . This sample set is directly used as the input of the S520 model.

[0090] S520: Train a network optimization model based on the topology state sample set and service indicators.

[0091] load Sample set and integration of service quality constraints defined by S200 .in, The minimum task completion rate that the system needs to ensure; The maximum energy consumption allowed by the system. The network optimization model uses the LSTM-Transformer parallel architecture: Time series feature extraction:

[0092] in, is the time series feature vector of node n at time t; :time node The state vector of : Always hide the status; : LSTM parameters (including input gate , Forget Gate , output gate weight); Based on trainable parameters Long short-term memory (LSTM) neural networks; Output: Time series features (Dimension 128).

[0093] Spatial feature extraction:

[0094] in, is the spatial feature; : Time sliding window parameter; is the moment in the time sliding window The node topology state vector.

[0095] Output: Spatial features (Dimension 128).

[0096] Furthermore, feature fusion is performed:

[0097] in, For nodes At the moment The fusion feature vector of : Learnable weight parameter (initial value 0.5); Output: fused features (dimension 128); Multi-objective loss function:

[0098] in, : Measured end-to-end delay; : Delay prediction value; : real resource weight; is the Huber loss function, which is used for the regression task of delay prediction; CE( ) is the cross entropy loss function, which is used for the classification task of resource weight allocation; The resource weight vector predicted by the model; is the fused feature vector of node n at time t; is the fused feature vector of node n at time t−1; is the Chebyshev norm, which is used to calculate the maximum deviation between eigenvectors; is the weight coefficient; Output: Multi-objective loss value .

[0099] Optimization through backpropagation Parameters, the model is fully updated every 24 hours.

[0100] S530: Input the real-time virtual instance topology into the network optimization model to generate parameter adjustment instructions.

[0101] Receive the virtual instance topology output by S430 in real time : Topological deviation calculation:

[0102] in, is the topological deviation, which indicates the difference between the current topological state and the historical benchmark; is the total number of nodes in the virtual instance topology; : current state vector; : historical benchmark vector; Output: Deviation ,when Trigger model online fine-tuning. is the topology deviation threshold.

[0103] Furthermore, the parameter adjustment instruction generates:

[0104]

[0105] in, is the power adjustment amount; is the truncation function; is the weight matrix for power adjustment; is the global eigenvector; A new resource weight allocation scheme; (global eigenvector); : normalized exponential function; is the weight matrix of resource weights; Output: Power adjustment and a new resource weight allocation scheme .

[0106] Instruction encoding and transmission: binary instruction packet structure: The first 16 bits: target node ID (corresponding to the S110 device number); Middle 32 bits: quantized power coefficient (accuracy 0.1dB); Last 64 bits: Sparse coding (non-zero value index + value); The instructions are transmitted to the S620 execution unit via a low-latency channel.

[0107] This module achieves three core technological breakthroughs: (1) Dynamic feature closed loop. Fusion of polarization imaging resource data Millimeter wave delay data , Real-time detection of S130 scattering characteristic anomalies ( Trigger model update); (2) Spatiotemporal parallel modeling. Capturing the temporal characteristics of infrared heat conduction (LSTM memory gating mechanism) and modeling the spatial distribution of polarization (Transformer multi-head attention); (3) Resource-energy consumption collaborative optimization. Power adjustment instructions Realize dynamic adjustment of energy consumption and resource weight Optimize resource allocation efficiency.

[0108] Step S600 at least includes steps S610-S630: S610: Analyze the power control coefficient and resource allocation weight in the parameter adjustment instruction.

[0109] The system receives the parameter adjustment instruction generated in step S530 via the high-speed instruction bus. This instruction is stored in a fixed-length data structure using binary encoding. When performing the parsing operation, the instruction decoding engine first identifies the version identifier in the protocol control header, verifies the integrity of the cyclic redundancy check code, and then strips the protocol header. The core parsing process is implemented through a triple mapping mechanism: 1. Field positioning mechanism: Locate the starting offset of the power control field and the starting offset of the resource allocation field according to the instruction template index table pre-burned in the parsing module.

[0110] 2. Coefficient Extraction Mechanism: Within the power control field, the first 16 bits store the target transmitting node identifier, the middle 16 bits are the power operation code, and the last 32 bits store the quantized value of the power adjustment coefficient. The system uses a shift register to separate these three sets of parameters and inputs the quantized value into a floating-point converter to convert it into a signed floating-point power control coefficient.

[0111] 3. Weight Parsing Mechanism: The resource allocation field uses sparse matrix encoding to store a four-dimensional resource weight vector (compute / storage / bandwidth / time slot resources). The parser calls the sparse matrix decoder to reconstruct the complete vector based on a preset dimension mapping table (dimension 0 corresponds to CPU weight, dimension 1 corresponds to memory weight, dimension 2 corresponds to bandwidth weight, and dimension 3 corresponds to time slot weight). Each weight value is normalized to the range [0, 1] by dividing an 8-bit fixed-point number by 256, ensuring that the sum of all weight components is strictly equal to 1.

[0112] The parsing process implements dual validation: the power control coefficients are immediately fed into a bound constraint checker after being restored, and the resource allocation weight vectors are verified for normalization using a cumulative checker. The final output is a structured parsing result: a power control coefficient package (containing the target node ID, operation type, and coefficient value) and a resource allocation weight package (containing a four-dimensional weight vector), which are transmitted to the S620 execution module via a shared memory area.

[0113] S620: Update the transmitting end power according to the power control coefficient, and update the resource pool quota according to the resource allocation weight.

[0114] The system executes two physical parameter update processes in parallel, both of which are implemented at the hardware level through the device driver layer interface: The transmitter power update process is as follows: 1. Baseline value acquisition: The current baseline power value of the target transmitting node is read through the power control interface. This interface calls the device driver's ioctl command POWER_GET to obtain the integer power value in 0.1dBm units from the wireless network card register.

[0115] 2. New value calculation: Perform differentiated operations based on the power operation type analyzed by S610: When the operation code is ADJUST_POWER: new power value = reference power value + power control coefficient; When the operation code is SET_POWER: new power value = power control coefficient; The calculated results are input into the saturation processor (limited to the ±20dBm dynamic range supported by the physical device).

[0116] 3. Hardware Write: Construct a power reconfiguration instruction packet (containing the target device MAC address, new power value, and effective timestamp) and write it to the network card power control register using the device driver's ioctl command POWER_SET. The visible light communication transmitter (such as the polarization imaging device described in S110) updates the LED drive current via the I²C bus; the RF transmitter updates the amplifier bias voltage via the RF front-end control bus.

[0117] The resource pool quota update process is as follows: 1. State snapshot collection: Access the physical resource pool management module (i.e., the resource status database of S410) to obtain the real-time resource quadruple: the number of available CPU cores (by reading the idle core counter in / proc / cpuinfo), the remaining memory capacity (by calling the meminfo system call), the free storage block list (by querying the block device allocation table), and the available time slot bitmap (extracted from the S220 time slot allocation mapping table).

[0118] 2. Quota recalculation: CPU quota: total number of cores × CPU weight → number of logical cores allocated to the target VM; Memory quota: total memory × memory weight → target container's memory limit; Storage quota: total number of free storage blocks × storage weight → block device allocation for the target volume; Timeslot quota: total timeslots × timeslot weight → timeslot allocation bitmap of target service flow; The calculation process uses an integer truncation strategy to ensure that resource allocation does not overflow (e.g., 12 cores × 0.33 weight → 3 cores are allocated).

[0119] 3. Policy implementation: Send quota update instructions to the resource pool manager: CPU core binding: write the cpu.cfs_quota_us parameter through the cgroups subsystem; Memory limit setting: configured through the memory.limit_in_bytes control group file; Storage volume expansion: Call the LVM tool to execute the lvresize command to adjust the logical volume size; Timeslot reallocation: Update the OFDMA frame structure through the timeslot allocation mapping table in S220; The update operation uses an atomic commit protocol: the write cache transaction mechanism is enabled at the device driver layer, and a completion notification is sent to the S630 module only when the power register is written successfully and the resource quota update returns an ACK signal.

[0120] S630: Feedback the updated transmitter power and resource pool quota to the environment perception data acquisition module.

[0121] The system constructs a structured feedback data packet and establishes a closed-loop path with the S110 perception module: 1. Data packet structure: Power feedback item: encapsulates the actual effective power value of the target transmitting node (the verification value is read back from the device register) and the global timestamp (synchronized with the sliding window time base of S130); Quota feedback item: records the four-dimensional resource allocation results (number of CPU cores allocated, memory upper limit, storage volume size, time slot bitmap hash value) and resource pool version identifier; The data format uses TLV encoding (Type-Length-Value), and the type identifier strictly matches the parsing template of the S110 preprocessing module.

[0122] 2. Transmission channel establishment: The physical layer uses the PCIe high-speed bus to transmit power feedback items (low latency requirements); The application layer transmits quota feedback items (metadata required) through the gRPC framework; The channel configuration parameters inherit the high-speed data interface definition of S130 (same data frame size and flow control parameters).

[0123] 3. Perception layer fusion processing: After the environmental perception data acquisition module (S110) receives the feedback packet: The power value is used to calibrate the received signal strength benchmark: the polarization imaging sensor adjusts the ADC reference voltage based on the optical transmission power, and the millimeter wave radar corrects the path loss compensation factor based on the RF power; Quota information is used to allocate data preprocessing resources: the image correction algorithm dynamically adjusts the thread pool size based on the number of available CPU cores; the point cloud filtering task sets the point cloud cache queue depth based on the memory limit; Finally, in the feature extraction stage of S120, the polarization degree matrix (including the light intensity value after power calibration) and millimeter wave scattering characteristics (clustering accuracy based on resource constraints) in the channel state sequence both carry updated system state information, providing closed-loop input for the S520 network optimization model.

[0124] Example 2: Figure 2 FIG. 1 shows a structural block diagram of a drone detection system based on visible light polarization imaging according to an embodiment of the present invention. Figure 2 As shown, the structure may include: The environmental perception module 10 is configured to acquire raw signals from multiple source devices, perform non-uniformity correction and noise reduction, and generate environmental perception data consisting of a polarization matrix, point cloud, and heat map through spatiotemporal registration to construct a channel state sequence. The environmental perception module initiates a workflow responsible for collecting raw signal data from polarization imaging sensors and other multi-source devices (such as conventional optical imaging devices and possible auxiliary sensors). To ensure the accuracy of subsequent analysis, the module performs critical non-uniformity correction operations to eliminate sensor response differences and performs noise reduction to improve signal quality. Subsequently, through precise spatiotemporal registration technology, it fuses and aligns the polarization information (polarization matrix), three-dimensional spatial information (point cloud), and thermal distribution information (heat map) from different devices to generate unified environmental perception data. Ultimately, it constructs a channel state sequence reflecting the current channel conditions, laying a solid data foundation for the entire detection process.

[0125] The cross-layer optimization module 20 is configured to parse the channel state sequence to extract channel gain, interference intensity, and noise power parameters. It then calculates power adjustment values ​​based on quality of service constraints, performs spectrum reuse and time slot allocation, and generates cross-layer optimization parameters. The core task of the cross-layer optimization module is to parse the channel state sequence transmitted by the environment perception module. Through in-depth analysis, the module accurately extracts key wireless communication parameters, including channel gain, environmental interference intensity, and background noise power. These parameters, along with system-preset quality of service constraints (such as data transmission delay requirements and recognition accuracy thresholds), are input into the optimization algorithm. Based on these parameters, the module calculates the optimal power adjustment value to balance energy consumption and performance requirements. It also intelligently implements spectrum resource reuse and dynamic allocation of communication time slots, generating a cross-layer optimization parameter set that incorporates the optimization decisions. This module's role is to dynamically allocate resources in complex wireless environments, ensuring the reliable and efficient transmission of sensing data and providing a stable information link for detection tasks.

[0126] The resource scheduling module 30 is configured to decompose the cross-layer optimization parameters into computing load characteristics and transmission demand characteristics, schedule edge nodes to implement multi-point collaboration and link switching, and output resource scheduling instructions. The resource scheduling module is responsible for converting the abstract parameters output by the cross-layer optimization module into specific resource allocation actions. It decomposes the cross-layer optimization parameters into two types of key characteristics: one is the computing load characteristics that need to be borne by the edge computing nodes, and the other is the transmission demand characteristics that have clear requirements for data transmission bandwidth and delay. Based on this, the module coordinates the capabilities of multiple edge computing nodes in the network, directs them to implement multi-point collaborative computing to share the load, and decisively executes link switching operations when the communication link status changes, ensuring the real-time availability and optimal matching of computing and transmission resources, and finally outputs executable specific resource scheduling instructions to direct the flow and use of physical resources.

[0127] Virtualization module 40 is configured to match available units in the physical resource pool, construct virtual container templates and virtual channels, deploy instances, and generate a virtual instance topology. The module's core function is to intelligently match available compute, storage, and network units in the physical resource pool according to resource scheduling instructions. Based on the matching results, it constructs lightweight virtual container templates to encapsulate the operating environment and creates efficient virtual channels to meet data transmission requirements. The module then deploys the processing instances required by the detection algorithm on demand within this virtualized infrastructure, ultimately generating a virtual instance topology diagram depicting these virtual instances and their interconnections. Through abstraction and pooled management, this module significantly improves the system's utilization efficiency and deployment flexibility for heterogeneous hardware resources.

[0128] The model inference module 50 is configured to convert the virtual instance topology into a set of feature vector samples, employ a dual-channel graph neural network training model, and generate power coefficient and resource weight matrix instructions. This model inference module becomes the core driver of the detection task. It converts the virtual instance topology generated by the virtualization module, representing the configuration of computing and communication resources, into a set of normalized feature vector samples that can be processed by a machine learning model (specifically, a dual-channel graph neural network). Using these sample sets, the module drives a pre-trained dual-channel graph neural network to perform inference operations. This network can deeply explore the complex relationship between topology, detection performance, and resource efficiency, ultimately outputting two key decision instructions: a power coefficient for fine-tuning the transmit power of each sensor and communication node, and a resource weight matrix instruction for dynamically allocating quotas for different resources (such as CPU, memory, and bandwidth) within the resource pool. This module implements an intelligent decision-making closed loop based on environmental conditions and resource configuration.

[0129] The execution feedback module 60 is configured to parse the power opcode and sparse matrix weights in the instruction, update the power register and resource pool quota, and feed back the power value to the polarization sensor calibration link. This execution feedback module is responsible for implementing the intelligent decisions of the model inference module and forming a closed loop. It accurately parses the received instruction, identifying the power opcode (indicating the specific power adjustment action) and the resource weight matrix (indicating the direction of resource quota adjustment). Based on this information, the module updates the power register status within the system and adjusts the transmit power level of each sensor and communication node in real time. Simultaneously, it also updates the allocation quota of various resource types in the physical resource pool to ensure that computing and transmission resources are supplied on demand. Most importantly, this module feeds back the updated actual power value to the polarization sensor calibration link for real-time correction of the signal acquisition accuracy of the polarization imaging sensor, thus forming a feedback loop from execution to perception, continuously improving the detection stability and accuracy of the entire system in dynamic environments.

Claims

1. A method for detecting drones based on visible light polarization imaging, characterized in that: include: The system acquires raw signals from multi-source environmental perception devices, performs non-uniformity correction and noise reduction processing, generates environmental perception data including a polarization matrix, point cloud, and thermal map through spatiotemporal registration, and constructs a time-aligned channel state sequence. The multi-source environmental perception device includes a visible light polarization imaging sensor array, a millimeter-wave radar, and an infrared thermal imager. The visible light polarization imaging sensor synchronously collects raw light intensity signals from the target area in four specific polarization directions to construct a raw data set containing complete polarization state information. The millimeter-wave radar acquires target reflection point cloud data at a preset scanning frequency, and the infrared thermal imager synchronously generates a thermal radiation intensity distribution map. Analyze the channel state sequence to extract channel gain, interference strength, and noise power parameters, dynamically calculate the power adjustment value based on quality of service constraints, synchronously perform spectrum reuse and time slot allocation, and jointly encode to generate cross-layer optimization parameters; Decompose cross-layer optimization parameters into calculated load characteristics and transmission demand characteristics, schedule edge nodes to implement multi-point coordination and link switching, and output prioritized resource scheduling instructions; Match available units in the physical resource pool, build secure and isolated virtual container templates and wavelength-graded virtual channels, deploy instances through atomic transactions, and generate virtual instance topologies with topology constraint verification. The virtual instance topology is converted into a normalized feature vector sample set, a dual-channel graph neural network is used to train a multi-task model, and online inference is used to generate power coefficient and resource weight matrix instructions. The power opcode and sparse matrix weight in the instruction are parsed by bit pattern matching, the transmitter power register and resource pool quota are updated, and the power value is fed back to the polarization sensor calibration link.

2. The drone detection method according to claim 1, characterized in that: The process of generating power coefficient and resource weight matrix instructions by online reasoning includes: Receive the virtual instance topology and calculate the topology deviation: ； in, is the topological deviation; is the total number of nodes in the virtual instance topology; is the current state vector; is the historical benchmark vector; is the Chebyshev norm; Output deviation ,when When triggering the online fine-tuning of the model, is the topology deviation threshold; Furthermore, the parameter adjustment instruction generates: ； in, is the power adjustment amount; is the truncation function; is the weight matrix for power adjustment; is the global eigenvector; A new resource weight allocation scheme; : normalized exponential function; is the weight matrix of resource weights; Output power adjustment and a new resource weight allocation scheme .

3. The drone detection method according to claim 2, characterized in that: Also includes: Perform one-hot encoding conversion on discrete resource type identifiers to generate binary feature vectors; The computing unit occupancy rate and memory allocation value are linearly mapped to the normalized interval using the minimum and maximum scaling algorithm; Applying a logarithmic transformation function to the transmission delay measurements compresses the data dynamic range.

4. The drone detection method according to claim 3, characterized in that: Also includes: For each virtual node, calculate the arithmetic mean of the transmission delay and the maximum statistical value of the packet loss rate of all its associated logical links to generate the edge feature vector; The edge feature vector and the node attribute vector are connected end to end in the feature dimension to generate a topological state vector with unified dimension.

5. The drone detection method according to claim 1, characterized in that: Using dual-channel graph neural networks to train multi-task models includes: The model architecture adopts a dual-channel graph neural network design; the first channel integrates a graph attention mechanism to dynamically adjust the influence coefficients of neighboring nodes, and the second channel deploys multi-layer graph convolution operators to extract deep correlation features of logical links; The outputs of the two channels are subjected to feature concatenation in the fusion layer, concatenating the node feature vector and the edge feature vector into a joint representation vector.

6. The drone detection method according to claim 5, characterized in that: Also includes: The training process implements a dynamic sample weighting strategy; Assign triple sampling weight to abnormal samples that exceed the service quality index threshold; The loss function contains three optimization objectives: the delay prediction branch adopts the Huber loss function, the resource allocation branch adopts the cross entropy loss function, and the topological stability branch calculates the cosine similarity of the feature vectors of adjacent time slices.

7. The drone detection method according to claim 6, characterized in that: Also includes: The sliding time window strategy is applied in the model validation phase; Twenty-four consecutive hours of data were selected as the training set, and the subsequent six hours of data were used as an independent validation set; When the topology reconstruction error on the validation set increases for three consecutive evaluation cycles, the early stopping mechanism is triggered to roll back to the historical optimal weight state.

8. The drone detection method according to claim 1, characterized in that: Instructions for generating power coefficients and resource weight matrices for online inference include: Receive virtual instance topology data stream and convert node attribute matrix and edge connection matrix into normalized feature vector; A lightweight inference engine is used to load the model, and the graph convolution layer and the fully connected layer are merged into a single computing unit through operator fusion optimization.

9. The drone detection method according to claim 8, characterized in that: Also includes: The online inference process adopts a two-stage pipeline architecture; The first level calculates the dynamic deviation of the current topology from the historical benchmark state, and the second level imposes a boundary constraint of -3dB to +3dB on the power adjustment coefficient; The parameter adjustment instruction adopts a binary coding format; the first sixteen bits store the target transmitting node identifier, the middle thirty-two bits record the quantized value of the power adjustment coefficient, and the last sixty-four bits are sparsely coded to store the resource allocation weight matrix.

10. A UAV detection system based on visible light polarization imaging, applied to the method according to any one of claims 1 to 9, characterized in that: include: An environmental perception module is configured to acquire raw signals from multiple source devices, perform non-uniformity correction and noise reduction, generate environmental perception data including polarization matrix, point cloud, and heat map through spatiotemporal registration, and construct a channel state sequence; a cross-layer optimization module configured to parse the channel state sequence to extract channel gain, interference strength, and noise power parameters, calculate power adjustment values ​​based on quality of service constraints, perform spectrum reuse and time slot allocation, and generate cross-layer optimization parameters; A resource scheduling module is configured to decompose cross-layer optimization parameters into calculation load characteristic quantities and transmission demand characteristic quantities, schedule edge nodes to implement multi-point coordination and link switching, and output resource scheduling instructions; A virtualization module is configured to match available units in the physical resource pool, build virtual container templates and virtual channels, deploy instances, and generate virtual instance topology; A model inference module, configured to convert the virtual instance topology into a set of feature vector samples, train the model using a dual-channel graph neural network, and generate power coefficient and resource weight matrix instructions; The execution feedback module is configured to parse the power operation code and sparse matrix weight in the instruction, update the power register and resource pool quota, and feed back the power value to the polarization sensor calibration link.

Images