A method and system for detecting unmanned aerial vehicles based on visible light polarization imaging

Abstract

The present application relates to the technical field of unmanned aerial vehicle detection and identification, and particularly relates to a method and system for detecting unmanned aerial vehicles based on visible light polarization imaging. The method constructs a time-aligned channel state sequence by acquiring multi-source environment perception data, dynamically calculates cross-layer optimization parameters and decomposes them into calculated and transmitted characteristic quantities, schedules edge nodes to cooperatively generate resource scheduling instructions, matches physical resources to construct a virtual instance topology, generates control instructions through a double-channel graph neural network training model for online inference, and finally feeds back to the polarization sensor to realize link calibration. The system integrates polarization imaging and cross-layer optimization technology, significantly improves the accuracy and real-time performance of unmanned aerial vehicle detection in complex environments, reduces the consumption of computing resources, and realizes efficient spectrum reuse and power adaptive adjustment in dynamic environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle detection and identification, and particularly relates to a method and system for detecting unmanned aerial vehicles based on visible light polarization imaging. BACKGROUND

[0002] The widespread application of unmanned aerial vehicles (UAVs) has brought potential security threats, and the demand for efficient and accurate detection of UAVs is increasingly urgent. Existing UAV detection technologies mainly face the following challenges:

[0003] 1. Limitations of single / dual modal perception: Single visible light imaging is easily affected by light changes (glare, backlight), weather conditions (fog, haze), camouflage, and complex background interference, and the detection rate and robustness significantly decrease under low contrast or adverse weather conditions. Traditional infrared imaging is sensitive to temperature and is easily disturbed by solar radiation and heat sources, and has insufficient resolution for detecting small targets (such as small UAVs) at a distance. Although single millimeter wave radar has certain penetration ability and all-weather working characteristics, it has insufficient imaging resolution at close range and is difficult to extract target feature information (such as shape and texture), making it difficult to detect low observable (LowRCS) small UAVs and non-metallic targets, and it is easily disturbed by ground clutter.

[0004] 2. Deficiency in spatio-temporal modeling and dynamic response: Existing methods have insufficient modeling capability for continuous spatio-temporal context information when dealing with high-speed maneuvering targets such as UAVs. Traditional sequence processing methods (such as Kalman filtering and basic RNN) or static image analysis are difficult to effectively capture the complex spatio-temporal evolution rules (such as maneuvering trajectory and attitude change) of target motion and the dynamic interaction with the environment.

[0005] 3. Resource constraints and real-time bottlenecks: Multi-modal perception and complex model calculation bring huge computational overhead. Existing systems often use static resource allocation strategies (such as fixed computation frame rate and preset fusion mode), which cannot adaptively adjust according to the dynamic changes in environmental complexity, target threat level, and system load. SUMMARY

[0006] The present application provides a method and system for detecting unmanned aerial vehicles 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 a complex dynamic environment, and to achieve high-precision real-time response of the unmanned aerial vehicle detection system through spatio-temporal joint modeling and dynamic resource optimization.

[0007] To solve the above technical problems, the present application provides a method for detecting unmanned aerial vehicles based on visible light polarization imaging, comprising:

[0008] The original signals of multi-source environment perception devices are acquired, non-uniformity correction and noise reduction processing are performed, environment perception data containing polarization matrix, point cloud and thermal map are generated through space-time registration, and channel state sequence aligned in time is constructed; the multi-source environment perception device contains visible light polarization imaging sensor array, millimeter wave radar and infrared thermal imager; the visible light polarization imaging sensor synchronously collects original light intensity signals of a target area in four specific polarization directions, and constructs an original data set containing complete polarization state information; the millimeter wave radar acquires target reflection point cloud data according to a preset scanning frequency, and the infrared thermal imager synchronously generates a thermal radiation intensity distribution map;

[0009] The channel state sequence is analyzed to extract channel gain, interference intensity and noise power parameters, power adjustment values are dynamically calculated in combination with quality of service constraints, spectrum multiplexing and time slot allocation are synchronously performed, and cross-layer optimization parameters are jointly encoded;

[0010] The cross-layer optimization parameters are decomposed into calculation load characteristic quantities and transmission demand characteristic quantities, multi-point cooperation and link switching are implemented by scheduling edge nodes, and priority-ordered resource scheduling instructions are output;

[0011] Available units of a physical resource pool are matched to construct a wavelength-graded virtual channel and a secure isolated virtual container template, an instance is deployed through an atomic transaction, and a virtual instance topology subject to topology constraints is generated;

[0012] The virtual instance topology is converted into a normalized feature vector sample set, a double-channel graph neural network is used to train a multi-task model, and power coefficients and resource weight matrix instructions are generated through online inference;

[0013] The process of generating power coefficients and resource weight matrix instructions through online inference includes:

[0014] The virtual instance topology is received, and a topology deviation degree is calculated:

[0015]

[0016] wherein, is the topology deviation degree; is the total number of nodes in the virtual instance topology; is the current state vector; is the historical reference vector; is the Chebyshev norm;

[0017] The deviation degree is output When the model online fine-tuning is triggered, is the topology deviation degree threshold;

[0018] Further, the parameter adjustment instruction is generated:

[0019]

[0020]

[0021] wherein, is the power adjustment amount; is a truncation function; is a weight matrix of power adjustment; is a global feature vector; is a new resource weight allocation scheme; softmax( ): a normalized exponential function; is a weight matrix of resource weight;

[0022] outputting the power adjustment amount and the new resource weight allocation scheme ;

[0023] The power operation code and sparse matrix weight in the instruction are parsed through bit pattern matching, the transmitting end power register and resource pool quota are updated, and the power value is fed back to the polarization sensor calibration link.

[0024] Further, converting the virtual instance topology into a normalized feature vector sample set comprises:

[0025] The system periodically acquires a virtual instance topology historical data set through a distributed storage interface, and the data set is stored in a columnar database in a time series form;

[0026] The virtual instance topology includes a node attribute matrix and an edge connection matrix, the node attribute matrix records a resource type identifier, a real-time occupation rate of a computing unit, an actual value of memory allocation, and a network bandwidth quota, and the edge connection matrix stores a globally unique identifier of a logical link, a sliding window measurement value of transmission delay, and a packet loss rate statistic;

[0027] A topology snapshot is intercepted at a fixed sampling frequency, and a power control coefficient and a resource allocation weight factor fed back at this moment are synchronously associated to form a space-time associated data unit.

[0028] Further, according to the unmanned aerial vehicle detection method, the method further comprises:

[0029] One-hot encoding conversion is performed on the discrete resource type identifier to generate a binary feature vector;

[0030] A minimum-maximum scaling algorithm is used to linearly map the computing unit occupation rate and the memory allocation value to a normalized interval;

[0031] A logarithmic transformation function is applied to the transmission delay measurement value to compress the dynamic range of the data.

[0032] Further, according to the unmanned aerial vehicle detection method, the method further comprises:

[0033] For each virtual node, calculate the arithmetic mean of the transmission delay of all associated logical links and the maximum value of the packet loss rate statistics, generate the edge feature vector;

[0034] The edge feature vector and the node attribute vector are connected at the beginning and end of the feature dimension to generate a topology state vector with unified dimensions.

[0035] Further, training the multi-task model using a dual-channel graph neural network includes:

[0036] The model architecture uses a dual-channel graph neural network design; the first channel integrates a graph attention mechanism to dynamically adjust the neighbor node influence coefficient, and the second channel deploys a multi-layer graph convolution operator to extract deep association features of logical links;

[0037] The outputs of the two channels are subjected to a feature concatenation operation in the fusion layer to concatenate the node feature vector and the edge feature vector into a joint representation vector.

[0038] Further, according to the unmanned aerial vehicle detection method, further comprising:

[0039] A dynamic sample weighting strategy is implemented in the training process; abnormal samples that exceed the quality of service threshold are given three times the sampling weight;

[0040] The loss function contains three optimization objectives; the Huber loss function is used for the delay prediction branch, the cross-entropy loss function is used for the resource allocation branch, and the cosine similarity of adjacent time slice feature vectors is calculated for the topology stability branch.

[0041] Further, according to the unmanned aerial vehicle detection method, further comprising: a sliding time window strategy is applied in the model verification stage; twenty-four hours of continuous data are selected as the training set, and the subsequent six hours of data are selected as the independent verification set;

[0042] When the topology reconstruction error on the verification set increases for three consecutive evaluation periods, an early stop mechanism is triggered to roll back to the historically optimal weight state.

[0043] Further, the online inference generates power coefficients and resource weight matrix instructions, including:

[0044] Receive virtual instance topology data stream, convert node attribute matrix and edge connection matrix into standardized feature vectors;

[0045] A lightweight inference engine is used to implement model loading, and graph convolution layers and fully connected layers are merged into a single computing unit through operator fusion optimization.

[0046] Further, according to the unmanned aerial vehicle detection method, further comprising:

[0047] The online inference process adopts a two-stage pipeline architecture; the first stage calculates the dynamic deviation degree of the current topology from the historical benchmark state, and the second stage applies a boundary constraint of -3dB to +3dB to the power adjustment coefficient;

[0048] The parameter adjustment instruction adopts a binary encoding format; the first sixteen bits store the target transmitting node identifier, the middle thirty-two bits record the power adjustment coefficient quantitative value, and the last sixty-four bits sparsely encode the resource allocation weight matrix.

[0049] Further, an unmanned aerial vehicle detection system based on visible light polarization imaging is applied to any of the above methods, comprising:

[0050] An environment perception module is configured to obtain multi-source device original signals, perform non-uniformity correction and noise reduction processing, generate environment perception data containing polarization matrices, point clouds and heat maps through space-time registration, and construct a channel state sequence.

[0051] A cross-layer optimization module is configured to analyze the channel state sequence to extract channel gain, interference strength and noise power parameters, calculate power adjustment values in combination with quality of service constraints, perform spectrum multiplexing and time slot allocation, and generate cross-layer optimization parameters.

[0052] A resource scheduling module is configured to decompose the cross-layer optimization parameters into calculation load characteristics and transmission demand characteristics, schedule edge nodes to implement multi-point cooperation and link switching, and output resource scheduling instructions.

[0053] A virtualization module is configured to match available units of a physical resource pool, construct virtual container templates and virtual channels, deploy instances and generate a virtual instance topology.

[0054] A model inference module is configured to convert the virtual instance topology into a feature vector sample set, train a model using a double-channel graph neural network, and generate power coefficient and resource weight matrix instructions.

[0055] An execution feedback module is configured to analyze the power operation code and sparse matrix weight in the instructions, update the power register and resource pool quota, and feed back the power value to the polarization sensor calibration link.

[0056] The key innovations of the present application include:

[0057] (1) A unified feature representation framework that can effectively fuse visible light polarization imaging (providing material, shape and contour information after suppressing interference) and millimeter wave radar (providing speed, distance and penetration information) is constructed, overcoming the problem of data heterogeneity and realizing deep complementarity and cooperation in information.

[0058] (2) A spatio-temporal feature joint learning architecture is designed, which organically combines models good at capturing long sequence temporal dependencies and models good at extracting spatial global correlations, forms a synergistic effect, realizes end-to-end, integrated modeling and analysis of target spatio-temporal features, and accurately captures the motion trajectory and spatial distribution evolution of dynamic targets.

[0059] (3) An intelligent computing resource dynamic allocation strategy is proposed, which dynamically adjusts the resource input of different data processing links (such as region selection, feature extraction depth, model complexity) according to the real-time environmental complexity perception (such as scene confusion degree, target quantity), target priority evaluation (such as threat level, confidence) and current resource load of the system, while ensuring the core region and target detection accuracy, maximizing the overall processing efficiency.

[0060] The following are its main beneficial effects:

[0061] (1) The core of the present application solves the problem of significant decline in detection performance of traditional single sensor (such as ordinary visible light camera) under complex background (such as cloud layer, strong light, weak light, leaf swing interference) and dynamic scene (high speed of unmanned aerial vehicle, small target, multi-pose). By fusing visible light polarization imaging and millimeter wave radar information, the polarization characteristics can be effectively used to suppress background clutter, enhance target contour and identify material characteristics. Combined with the accurate speed and distance information of millimeter wave radar, complementary advantages are formed. This multi-modal fusion strategy greatly improves the target recognition ability and overall detection confidence of the system under adverse lighting and complex background interference, significantly reduces the miss detection and false detection rate.

[0062] (2) The spatio-temporal joint modeling method of the present application can simultaneously and cooperatively learn the spatial distribution features (such as shape, size, position) of the target and the dynamic trajectory features (such as speed, direction, attitude change) of the target over time. This deep spatio-temporal fusion mechanism enables the system not only to accurately identify the current state of the target, but also to effectively predict the movement trend of the target, realizing stable and continuous tracking of small unmanned aerial vehicle targets moving at high speed and maneuvering.

[0063] (3) In the scene where a large amount of multi-source heterogeneous data (high-resolution polarization image, radar point cloud) needs to be processed and complex spatio-temporal modeling is required, the traditional system often causes processing delay due to limited computing resources, which cannot meet the stringent real-time requirements of unmanned aerial vehicle detection. The resource dynamic optimization mechanism introduced in the present application can intelligently adjust the depth and breadth of the data processing flow according to the current environmental complexity, target threat level and computing resource status. For example, simplify the processing in resource-constrained or low-threat areas, and concentrate resources for detailed analysis when high-threat areas or targets appear. This on-demand allocation strategy ensures that the system can maintain low delay and high frame rate real-time detection and response capability even in complex and variable high-load environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A flowchart of a method for detecting drones based on visible light polarization imaging provided in an embodiment of the present application;

[0065] 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

[0066] 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:

[0067] 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.

[0068] 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.

[0069] 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.

[0070] 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.

[0071] 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.

[0072] 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.

[0073] Step S100 at least includes steps S110-S130:

[0074] S110: Obtain original signals from multi-source environmental perception devices, perform noise suppression and data calibration, and obtain environmental perception data.

[0075] The multi-source environment perception device deployed in the monitoring area includes three types of heterogeneous sensing devices: visible light polarization imaging sensor array, millimeter wave radar and infrared thermal imager. First, the visible light polarization imaging sensor synchronously collects the original light intensity signals of the target area in four specific polarization directions, and constructs the original data set containing complete polarization state information; the millimeter wave radar acquires the target reflection point cloud data according to the preset scanning frequency, and the infrared thermal imager synchronously generates the thermal radiation intensity distribution map. For the polarization imaging data, non-uniformity correction processing is performed: based on the reference image obtained by the calibration board under uniform illumination, the response deviation coefficient of each pixel point of the imaging sensor is calculated, and the original light intensity signal is mapped to the standardized response curve through the linear transformation algorithm, so as to eliminate the response difference of the sensor itself. For the millimeter wave point cloud data, an adaptive distance threshold filtering mechanism is used to eliminate multipath interference noise, which dynamically adjusts the effective detection distance threshold according to the radar scattering cross section, and automatically filters out the discrete noise points beyond the distance threshold. The infrared thermal image data is denoised by time domain median filtering, and the median value of each pixel point in the continuous multiple frames of thermal radiation map is taken, which effectively suppresses the instantaneous thermal noise interference.

[0076] To realize the space-time reference unification of multi-source data, the GPS timing module provides accurate clock source to all perception devices, and establishes a nanosecond-level synchronous time reference. The space registration is performed by using the fixed calibration object coordinate points in three-dimensional space, the millimeter wave point cloud coordinate system is mapped to the polarization imaging coordinate system through affine transformation, and the space alignment of cross-modal data is completed. The final output environment perception data includes four standardized dimensions: four-channel polarization light intensity matrix after space registration, millimeter wave scattering point cloud data set aligned with the polarization imaging coordinate system, thermal radiation intensity distribution map after time domain filtering processing, and synchronous data identification with unified time stamp index. The environment perception data is transmitted to the S120 processing module as the only input source.

[0077] S120, extracts channel feature parameters from the environment perception data, performs time-frequency domain conversion processing, and obtains channel state data.

[0078] Based on the environment perception data output by S110, physical layer channel feature parameters are extracted from three dimensions respectively. In the visible light polarization imaging dimension, first, the light intensity vector and polarization characteristic parameters of each pixel point are calculated: the total light intensity parameter and the difference value of the orthogonal polarization component are derived through the light intensity values of the four polarization directions, and then the polarization degree index and polarization angle parameters representing the polarization characteristics are calculated. The polarization degree index of the full frame pixels of the imaging picture is gridized and counted, generating a polarization degree spatial distribution matrix with consistent resolution as 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, Doppler shift mean value, scattering intensity dispersion, and other feature vectors representing target physical characteristics are calculated. In the infrared thermal image dimension, the temperature gradient change of adjacent pixels along the target motion direction is analyzed to form a feature vector describing the dynamic characteristics of heat conduction.

[0079] Further, the above features are processed in time and frequency domains: the polarization degree spatial distribution matrix is processed by sliding window Fourier transform, the frequency energy distribution of the polarization degree parameters of each pixel point is calculated within a continuous multi-frame time window, and the low-frequency energy proportion is extracted as a feature vector representing time-varying characteristics; the millimeter wave Doppler shift data is processed by wavelet packet decomposition algorithm to extract the energy entropy parameters of the target sub-band at a specific decomposition level, generating a frequency domain feature vector; the infrared heat conduction feature vector is input into a recurrent neural network for time series modeling, and an implicit state feature vector representing the time evolution law is output. Finally, five-dimensional channel state data is generated: including the original polarization degree spatial distribution matrix, the polarization time-varying feature vector, the millimeter wave scattering frequency domain feature vector, the heat conduction time series feature vector, and the alignment index identifier to ensure the time synchronization of multi-source features. The channel state data is transmitted to S130 processing module through high-speed data bus.

[0080] S130, the channel state data is segmented by sliding window to generate channel state sequences.

[0081] After obtaining the channel state data output by S120, a time mapping relationship of the multi-source data is established according to the timestamp alignment index. A sliding processing window with a fixed time length and a step interval is set, and data segments are sequentially intercepted along the time axis. Four core operations are performed in the window data processing process: first, time alignment verification is performed to detect the timestamp deviation of each modal feature data in the window, and when the deviation exceeds the preset threshold, a spline interpolation algorithm is used for resampling to achieve accurate alignment; second, feature dimension normalization processing is implemented to downsample the polarization degree space matrix to a standard resolution, the millimeter wave feature vector is compressed to a low-dimensional space through principal component analysis, and the thermal conductivity feature vector is numerically normalized and mapped; then, spatial feature coding is performed, and a convolution kernel operation is applied to the polarization degree matrix to extract local spatial correlation features and generate a spatial feature map tensor; finally, serialization and packaging are performed, and the processed multi-source features are arranged in time sequence to form a structured data packet containing five elements: accurate time window start and end markers, standardized spatial feature map tensor, downsampled millimeter wave frequency domain feature vector, normalized thermal conductivity time series feature vector, and feature integrity verification identifier.

[0082] An overlap reservation mechanism is used in the sliding window processing process to ensure time domain continuity: when the window steps, the data at the end of the previous window is reserved as the starting data of the new window. The finally generated channel state sequence is a set of serialized data packets arranged in time sequence, each data packet containing multi-modal feature data aligned in time and space, which is output to the S210 adaptive power control module through a high-speed data interface. This sequence is used as the core input in S210 to calculate the transmission power strategy in combination with the quality of service constraints, and in the S500 training stage, it is used as the source of the underlying features of the historical data samples, and dynamically optimizes the environmental perception data acquisition parameters through the feedback mechanism of S630.

[0083] Step S200 includes at least steps S210-S230:

[0084] S210, obtaining a channel state sequence, combining service quality constraint conditions, and performing adaptive power control decision.

[0085] The system obtains the channel state sequence generated by step S130 as the basic input source through a high-speed data interface. The channel state sequence is composed of multiple data packets arranged in time sequence, each data packet containing multi-modal channel feature data processed in time and space. Specifically, the system first parses the preconfigured service quality constraint conditions, which include three core threshold parameters: a minimum communication rate threshold defining the lower limit of the bit rate that the transmission link must maintain, a maximum allowable end-to-end transmission delay threshold limiting the total delay of the data packet from the sending end to the receiving end, and a maximum tolerable error rate threshold specifying the acceptable error bit ratio limit of signal demodulation. These threshold parameters are loaded into the decision engine as rigid boundary conditions for power control decision.

[0086] The channel state sequence is processed packet by packet, and the standardized spatial feature map tensor contained in each data packet is input into a feature extraction module. The module extracts channel gain distribution features through convolution kernel operation to generate quantized indicator values representing signal propagation path loss. At the same time, an interference intensity parameter is parsed from the millimeter wave frequency domain feature vector, which reflects the superimposed signal energy caused by other transmitters in the same frequency band; a noise power spectral density parameter is extracted from the thermal conduction time sequence feature vector, which describes the distribution characteristics of environmental 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 a current channel quality indicator triple.

[0087] When executing critical decision judgments, the system jointly compares the current channel quality indicator triple with the service quality constraint threshold. When the channel gain value is higher than the theoretical threshold required to maintain the minimum communication rate threshold, and the predicted value of the historical transmission delay does not exceed the maximum allowed 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. The instruction includes a power adjustment amount calculation process: according to the difference between the channel gain and 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 approach or be lower than the theoretical threshold, or the predicted transmission delay is likely to break through the delay threshold, or the calculated bit error rate exceeds the tolerance threshold, a transmit power increase instruction is generated, and the power increment is dynamically matched in the power control step matrix according to the indicator deviation.

[0088] The decision mechanism includes an adaptive execution frequency adjustment function, and the system continuously monitors the parameter change rate of adjacent data packets in the channel state sequence. When the variance of channel gain or interference intensity exceeds the preset fluctuation threshold, the high-frequency decision mode is activated to improve the response speed; when the parameter change rate is below the stability determination threshold, the low-frequency decision mode is switched to reduce the operation overhead. All decision results are packaged into a structured instruction set, and each instruction explicitly includes a timestamp identifier, a target transmitter global unique identifier, a power adjustment direction indicator (an enumerated value of rise / fall), a recommended transmit power absolute value, and a valid period parameter of the instruction. The structured instruction set is transmitted as output to the resource allocation module.

[0089] S220, based on the adaptive power control decision and resource demand prediction results, execute a dynamic resource allocation strategy.

[0090] The system synchronously receives the adaptive power control decision instruction set output from S210 and the resource demand prediction result generated by the independent prediction module. The resource demand prediction result outputs quantitative prediction values of the total amount of spectrum resource demand, the total amount of time slot resource demand and the total amount of computing resource demand in a future time window by analyzing the historical traffic volume curve, the current access device quantity and the traffic growth model. The spectrum resource demand is expressed by the bandwidth demand value in megahertz, the time slot resource demand is expressed by the number of time slots to be allocated per second, and the computing resource demand is converted into the equivalent value of standard computing units.

[0091] Before performing resource allocation, the system first establishes a real-time state snapshot of the physical resource pool: in the spectrum resource dimension, the currently available frequency bands are scanned and the center frequency, bandwidth value and adjacent frequency band interference coupling coefficient of each frequency band are recorded; in the time slot resource dimension, the allocable time slot position index is obtained by polling the time slot allocation state table; in the computing resource dimension, the CPU core availability, memory remaining capacity and storage space margin of each edge node are collected. The resource pool state snapshot is matched with the resource demand prediction result to generate a resource gap distribution diagram.

[0092] The core allocation process implements a three-layer resource coordination mechanism: in the spectrum resource allocation layer, the target transmitter identifier and power adjustment direction in the S210 power control instruction set are analyzed. For the transmitter performing power reduction, the interference radius shrinkage value caused by power reduction is calculated, and when the interference radius shrinkage exceeds the preset multiplexing activation threshold, the spectrum multiplexing algorithm is started in its coverage area: the frequency reuse factor in the area is recalculated, the original required protection frequency band gap is released as available resource, a new frequency band division scheme is generated and allocated to the access device to be accessed. For the transmitter performing power boost, the interference enhancement range is calculated according to the power increment, and a protection band isolation zone is set in the interference enhancement range, which occupies resources into the total amount of resources allocated. In the time slot resource allocation layer, data streams are divided into real-time streams and non-real-time streams based on the traffic flow priority classifier, real-time streams are allocated continuous time slot blocks using the absolute priority preemption mechanism, and non-real-time streams are allocated discrete time slots using the weighted round robin algorithm with dynamically adjustable weights. The weight value is dynamically configured according to the traffic peak value distribution in the resource demand prediction. In the computing resource allocation layer, a computing task feature vector (including CPU instruction set type demand, memory bandwidth demand and storage IO throughput demand) is established, combined with the real-time load matrix of the edge node, and the minimum load difference matching algorithm is used to schedule the computing task to the target node, and an accurate core number quota, memory capacity quota and storage space quota are allocated to each task.

[0093] The final multi-dimensional resource allocation scheme is generated: the spectrum allocation scheme is recorded as a frequency band allocation record table, each record containing a link identifier, a start frequency, an end frequency, and a valid period; the time slot allocation scheme is recorded as a time slot allocation mapping table, each record containing a service flow identifier, a frame period number, a time slot start position, and a number of occupied time slots; the computing resource allocation scheme is recorded as a computing resource allocation list, each list containing a task identifier, a node identifier, a list of allocated core numbers, a memory block address, and a storage volume label. The scheme is output to the joint encoding module through the resource management bus.

[0094] S230, jointly encoding the adaptive power control decision and the dynamic resource allocation strategy to generate cross-layer optimization parameters.

[0095] The system obtains the power control decision instruction set output in S210 and the resource allocation scheme generated in S220 as double input sources, and implements cross-layer parameter fusion through a joint encoder. The encoder loads a predefined protocol mapping template, which contains five types of structured fields: a decision type identifier field uses a 32-bit integer number to define the instruction type as a cross-layer cooperative optimization instruction; a power control field encapsulates the target transmitting end identifier, the power operation code (preset SET_POWER / ADJUST_POWER operation code enumeration value), and the power value using the TLV structure; the spectrum resource field stores the frequency band allocation record table content using a dynamic array structure, each array element containing three groups of parameters: link identifier, start frequency, and end frequency; the time slot resource field encodes the time slot allocation mapping table using a bit field compression format, each record represented by a combination of start frame number (16 bits), start time slot offset (8 bits), and consecutive time slot number (8 bits); the computing resource field uses a hierarchical description method, with the top layer recording the node identifier and the task identifier, and the lower layer linking the core allocation list, the memory allocation block descriptor, and the storage volume descriptor through pointers.

[0096] The data filling process strictly implements the protocol mapping rules: the target transmitting end identifier and the power adjustment value are extracted from the power control decision instruction set, converted into the power operation code and the power value, and written into the power control field; the frequency band parameters of each link are extracted from the spectrum allocation scheme, arranged in ascending order of frequency band start frequency, and written into the dynamic array of the spectrum resource field; the time slot allocation records are extracted from the time slot allocation mapping table, compressed by bit field, and written into the time slot resource field; the hierarchical parameters are extracted from the computing resource allocation list, and a pointer linking structure is constructed and written into the computing resource field.

[0097] Adding protocol control header when performing structured encapsulation: the version number field identifies the protocol revision version, the timestamp field records the instruction generation timestamp, and the validity period field defines the longest survival period of the instruction. The CRC32 check algorithm is used to generate the cyclic redundancy check code of the entire data packet, which is attached to the tail of the data packet to form a complete transmission frame. The final output of the cross-layer optimization parameter is a binary data stream with a strict syntax structure, which includes a protocol control header and a data body in the physical layer. The data body is divided into power control instructions, spectrum allocation parameters, time slot allocation parameters, and computing resource allocation parameters, which are encapsulated in layers. The check code provides end-to-end integrity protection. The data stream is distributed to each execution node through the control plane interface.

[0098] Step S300 includes at least steps S310-S330:

[0099] S310, obtaining cross-layer optimization parameters, and decomposing into a computing task subset and a communication task subset.

[0100] The cross-layer optimization parameter generated by S230 is obtained as a core input source, which contains jointly encoded adaptive power control decisions and dynamic resource allocation strategies. A structured parsing operation is performed on the cross-layer optimization parameter by a dedicated parameter parsing module. The operation first identifies the preset decision marker bits in the jointly encoded matrix. The power control identifier uses the high byte mask to identify the power adjustment operation type, and the resource allocation identifier identifies the encoding block boundaries of the spectrum resource, time slot resource, and computing resource through the low byte segment. Based on the bit pattern matching algorithm, the parsing engine splits the original matrix data stream into four independent data channels, namely the power control instruction channel, the spectrum allocation parameter channel, the time slot allocation parameter channel, and the computing resource allocation parameter channel, according to the identifier type. In the computing task subset decomposition stage, the first N-dimensional feature vector is extracted from the computing resource allocation parameter channel to construct the computing load feature quantity, which includes the task complexity identifier representing the computing intensity (mapped from the CPU instruction set type to the preset complexity level), the time delay sensitivity coefficient defining the service time delay constraint (a normalized value with a value range of 0.0 to 1.0), and the memory demand threshold identifying the memory demand peak (an absolute value in megabytes). In the communication task subset decomposition stage, the last M-dimensional feature vector is extracted from the fusion of the spectrum allocation parameter channel and the time slot allocation parameter channel to construct the transmission demand feature quantity, which includes the bandwidth allocation priority reflecting the service bandwidth demand (priority weight classified by real-time flow / non-real-time flow), the link stability index representing the wireless channel reliability (sliding average value calculated based on the historical packet loss rate), and the error code rate tolerance threshold defining the transmission quality bottom line (floating point value represented by scientific notation). The entire decomposition operation is implemented by a two-stage pipeline architecture for hardware acceleration: the first-stage pipeline classifies the identifiers by a bit matching circuit, which performs parallel matching using a pre-recorded bit pattern template to identify the marker bit boundaries and trigger the channel separation signal within a single clock cycle; the second-stage pipeline implements feature vector segmentation through a dual-port buffer area, the front port outputs the computing load feature quantity to the edge computing scheduler, and the back port outputs the transmission demand feature quantity to the link selector, achieving zero-delay parallel output of the two types of feature quantities.

[0101] S320, scheduling the edge computing node according to the computing task subset and scheduling the transmission link according to the communication task subset.

[0102] The edge computing node scheduling is performed based on the calculation load feature quantity output by S310. The scheduler first accesses the edge computing node resource registration table to obtain the real-time resource state, and the registration table continuously collects the number of available computing units (identifying the number of remaining CPU cores), the graphics processor load rate (the three-dimensional rendering resource occupancy rate expressed in percentage), and the memory occupancy ratio (the ratio of allocated memory to total capacity) of each node as three key indicators. Then the node adaptation degree matching calculation is performed: the task complexity identifier in the calculation load feature quantity is input into the node processing capacity scoring function, which generates a basic score based on the compatibility of complexity level and node hardware configuration, superimposes the negative weighted values of graphics processor load rate and memory occupancy ratio, and finally outputs the node adaptation degree matrix quantified as 0100 points. The dynamic allocation strategy contains two core mechanisms: when the delay sensitivity coefficient exceeds the preset threshold (typical value is 0.8), the multi-point cooperative mechanism is activated, the single task is split into K logical continuous parallel sub-tasks by the task decomposition algorithm, L edge nodes with a physical distance of less than 200 meters are selected according to the node geographical topology database to form a cooperative group, and the sub-task package is distributed in a multicast manner; when the memory demand threshold exceeds 80% of the available memory of the node, the vertical expansion mechanism is triggered, the pre-set memory pool sharing interface is called, the excess memory demand is mapped to the virtual address space of the shared memory pool, and the dynamic mounting of the physical memory block is implemented through the memory controller. In the communication task subset scheduling process, the link selector loads the transmission demand feature quantity output by S310 as the basis for decision making: based on the bandwidth allocation priority, the candidate link set is initialized, the matching degree of the link stability index and the error code rate tolerance threshold is calculated through the weighted evaluation module, the link stability index is given a weight of 0.6, and the error code rate tolerance threshold is given a weight of 0.4, and the link with the highest comprehensive score is output as the main transmission path. When the stability index of the millimeter wave link is lower than the threshold value (empirical value is 35), the seamless switching mechanism is triggered, the business flow is automatically forwarded to the Sub6GHz backup link, and the optimal frequency band combination is recalculated through the frequency band redistribution module. The process is completed within 50 milliseconds and maintains the continuity of the business flow.

[0103] S330, the response state of the edge computing node and the transmission link is prioritized, and a resource scheduling instruction is output.

[0104] The edge node resource response message and the transmission link state report scheduled from S320 are received as two-way input. The resource response message includes three core fields: task processing delay estimation (processing time prediction value with microsecond level precision), node resource reservation state code (state code identifying whether CPU / GPU / memory resource reservation is successful), and cache queue length (integer value of the number of tasks to be processed); the link state report includes actual allocated bandwidth value (measured bandwidth in megahertz), channel quality indication value (0100 scale value converted from signal-to-noise ratio), and transmission delay measurement value (millisecond level measurement result of end-to-end one-way transmission delay). The priority sorting operation is realized through a three-stage pipeline evaluation architecture: the first-stage evaluator calculates the node response efficiency ratio, which is the product of the task processing delay estimation and the node resource reservation state code, where the state code success state takes the value of 1.0 and the failure state takes the value of 0.1. The product result is normalized and converted into a 0100 range calculation resource priority score; the second-stage evaluator calculates the link transmission efficiency ratio, which is the weighted sum of the actual allocated bandwidth value and the channel quality indication 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 converted to output a 0100 range communication resource priority score; the third-stage evaluator performs joint sorting, using a double-layer sorting algorithm to arrange in descending order of the calculation resource priority score, and in ascending order of the communication resource priority score in the case of the same score, to generate a globally unique resource priority sequence. The output resource scheduling instruction adopts a three-stage encapsulation structure: the instruction header includes a 32-bit precision global timestamp (synchronized with the system master clock) and a 64-bit transaction identifier (including task identification and version number); the instruction body records the sorted node identifier sequence (node ID list arranged in descending order of priority) and the link identifier sequence (link ID list arranged in descending order of score), with each identifier attached with a weight coefficient to identify the priority difference; the instruction tail is attached with an 8-bit resource synchronization flag bit, which is set when the first 0 bit is set to trigger the resource matching operation of S410, and the second 1 bit identifies whether cross-domain resource coordination is needed, and the remaining bits are reserved for extension functions.

[0105] Step S400 includes at least steps S410-S430:

[0106] S410, acquire the resource scheduling instruction, and match the available resource units of the physical resource pool.

[0107] Receiving the resource scheduling instruction output from step S330, which contains a globally prioritized subset of computing tasks and a subset of communication tasks and their resource requirement attributes. The resource requirement attributes cover the number of CPU cores required for compute-intensive tasks, GPU memory capacity, memory bandwidth threshold parameters, and the upper limit of end-to-end transmission delay defined for communication-intensive tasks, the minimum bandwidth guarantee value, and the error code rate tolerance threshold. When performing the matching operation, first parse the task type identifier in the instruction (the compute-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 state database of the physical resource pool, which continuously synchronizes the real-time physical resource parameters of the edge computing node: including CPU core availability rate (the idle core proportion collected by the kernel performance counter), GPU memory occupancy rate (the memory allocation state obtained through the device driver interface), memory remaining capacity (physical memory available value in megabytes), and transmission link bandwidth idle rate (available bandwidth proportion converted from port queue depth), and port availability state (link protocol layer handshake signal detection result).

[0108] Performing a double-layer screening logic through the resource matching engine: for the subset of computing tasks, generating an algorithm requirement constraint expression (requiring the single-core frequency of the target node multiplied by the available core number to be greater than or equal to the minimum algorithm threshold) and a delay constraint expression (the sum of node processing delay and network transmission delay is less than the preset threshold value); for the subset of communication tasks, generating a bandwidth redundancy constraint expression (the available bandwidth of the link is greater than 120% of the task demand bandwidth value) and a reliability constraint expression (the link error code rate is lower than 1e5 and supports forward error correction mechanism). A bidirectional verification mechanism is used to ensure the matching accuracy: in the forward verification stage, resource screening conditions are generated according to the constraint expressions, and candidate resource units that meet all the conditions are retrieved from the physical resource pool; in the reverse verification stage, the service level agreement commitment indicators of the candidate resource units (including the 99.99% computing task completion rate promised by the computing node and the upper limit of millisecond-level delay jitter guaranteed by the communication link) are extracted, and it is verified whether the indicators match the fault tolerance requirements and real-time requirements in the service quality constraint conditions of the tasks. The finally output available resource unit list is stored in a structured data format, each record contains a globally unique identifier of 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 two-classification code of computing resource / communication resource), physical location coordinates (a three-tuple of longitude, latitude, and elevation), and real-time performance indicators (the computing node includes current CPU load rate and memory fragmentation rate, and the communication link includes current round-trip time and packet loss rate), which serves as the input data basis for step S420.

[0109] S420, constructing a virtualized resource template from the available resource units and performing resource instantiation deployment.

[0110] The available resource unit list output by the parsing step S410 is classified into a computing resource group and a communication resource group according to the resource type classification code. The virtualization resource template construction process is implemented in two levels: in the computing resource level, a virtual computing container template is created according to the heterogeneous characteristics of the edge computing node. The specific implementation is as follows: the key parameters of the computing resource unit are extracted, including the number of CPU cores (the number of assignable logical cores), the GPU model (the model identifier supporting the CUDA computing architecture), and the memory capacity (the size of the continuous physical memory block), and a lightweight virtual machine instance is configured in the Hypervisor layer through the kernel-level virtualization module. The configuration process includes three core operations: first, an independent secure domain isolation environment is allocated to each virtual machine, which realizes data isolation through a memory encryption engine and an IO access control list; second, a time slice scheduling strategy is configured to bind the virtual machine time slice and the physical CPU clock cycle, and the time slicing rotation algorithm is used to guarantee the parallelism of multiple tasks; third, a pass-through mode is enabled for the GPU resource, which directly maps the physical GPU device to the virtual machine device space through the virtualization layer, ensuring the computing efficiency of the polarization imaging processing task; and a dynamic quota adjustment mechanism is configured for the memory resource, which automatically expands the upper limit of memory allocation according to the memory pressure index collected by the virtual machine monitor, adapting to the sudden memory load demand.

[0111] In the communication resource level, a virtual channel template is created according to the physical characteristics of the transmission link. The specific implementation is as follows: the core parameters of the communication resource unit are extracted, including the available bandwidth (the minimum guaranteed bandwidth value), the transmission delay (the physical layer signal propagation delay), and the port protocol (the identifier supporting the Ethernet / fiber channel protocol), and a virtual switching path is configured in the data plane through a software-defined network controller. The configuration adopts a segment routing strategy to implement differentiated services: an exclusive wavelength channel is allocated for high-priority communication tasks (such as unmanned aerial vehicle control instructions), a specific optical wavelength is locked through a wavelength selection switch to establish an end-to-end optical path; a statistical multiplexing channel is deployed for low-priority tasks (such as historical data backhaul) to dynamically allocate time slot resources in a time division multiplexing frame structure; and a quality of service identifier is injected into the border gateway protocol route advertisement message to enable the core network device to identify the service level of the virtual channel.

[0112] After the template is built, a resource instantiation deployment operation is performed: an infrastructure management interface (complying with RESTful API specifications) is called to deploy the virtual computing container template to the selected edge computing node physical resources, which writes a virtual machine configuration descriptor to the host's virtualization control registers through the device driver layer; flow table rules are issued to the selected transport link switching device in the standard flow table format of OpenFlow protocol version 1.5, and the flow table entries explicitly contain input port matching rules, quality of service 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, and when any resource unit instantiation fails (such as virtual machine startup timeout or flow table issue error), an automatic rollback mechanism is triggered to release the allocated resources (including destroying the created virtual machine instance and revoking the issued 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 (globally unique identifier composed of host identifier and virtual machine sequence number), virtual channel ID (128-bit identifier composed of source and destination node ID and service level code), and resource binding relationship (mapping relationship table recording the mapping relationship between virtual machine instance and host computing node, and virtual channel and physical link).

[0113] S430, logical connection mapping is performed on the resource instantiation deployment result to generate a virtual instance topology.

[0114] The virtual resource instance set output by the obtaining step S420 is obtained, from which the physical location coordinates of the virtual machine instance (used for calculating the geographic distance), network address information (IPv6 address combined with VLAN tag), and endpoint identifier of the virtual channel (source port MAC address and destination port IP address pair) are extracted. The logical connection mapping operation is implemented in three consecutive stages:

[0115] In the computing node interconnection relationship construction stage, based on the network address information of the virtual machine instance, the network topology parameters (including transmission path hop count, transmission delay of each hop, and path redundancy index) between instances are obtained through an enhanced interior gateway routing protocol. According to the task characteristics of the polarization imaging processing pipeline: a point-to-point direct 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 realize direct communication; a multicast communication tree is constructed for the spectral analysis node and the polarization clustering node to meet the data distribution requirements of the multi-modal fusion task, and a shortest path tree with the spectral analysis node as the root is constructed through a protocol-independent multicast protocol. The connection relationship construction uses an improved minimum spanning tree algorithm to optimize the communication cost: the virtual machine instance is taken as the vertex, and the geographical distance multiplied by the link unit price is taken as the weighted edge, the Prim algorithm is used to generate the minimum communication cost topology structure, and the equivalent multipath rule is automatically injected to realize load balancing.

[0116] In the communication link load rule mapping phase, the quality of service level identifier encoded by the virtual channel ID is parsed (divided into high reliability level and high real-time level). The high reliability channel (such as the unmanned aerial vehicle control instruction transmission channel) is mapped to the dual-fiber redundant link architecture, and when the main optical fiber path fails, it is switched to the standby path within 50 milliseconds; the real-time channel (such as the polarization video stream) is configured with a strict priority queue scheduling strategy, and the highest priority queue is allocated in the switch export queue and the minimum bandwidth guarantee value is set. At the same time, traffic shaping rules are injected in the mapping process: according to the bandwidth allocation weight coefficient (normalized value from 0.0 to 1.0) in the step S220 dynamic resource allocation strategy, the token bucket parameters of each virtual channel are set (token generation rate = weight coefficient x total bandwidth of physical link, bucket depth = maximum allowed burst traffic x weight coefficient).

[0117] In the topology description framework generation phase, the aforementioned node connection relationship and link load rule are integrated into a directed graph structure model: the virtual machine instance is taken as a vertex, and the vertex attribute includes a resource type identifier (computing node classification code) and a computing capacity value (peak computing power in TFLOPS); the virtual channel is taken as a directed edge, and the edge attribute includes a transmission protocol type (TCP / UDP / RDMA protocol identifier), a bandwidth reservation value (minimum guaranteed bandwidth in Mbps), and a maximum allowed delay (upper limit value of end-to-end transmission delay). Constraint verification is performed through a topology verification module: first, the total delay 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 delay budget value defined in the S210 power control decision; second, the wavelength allocation conflict-free condition is verified, and the same optical fiber link is checked for repeated wavelength allocation records to meet the optical layer resource allocation rules of the S320 transmission link scheduling requirement. Finally, a virtual instance topology file conforming to the GraphML 1.2 standard is output, which records the complete node configuration parameters (including virtual machine startup command line parameters), link parameters (including token bucket configuration details), and routing strategy table (static routing and equal cost multi-path entries) in XML syntax structure. The file serves as the topology state input of step S510.

[0118] Step 500 at least includes steps S510-S530:

[0119] S510, obtain the historical running data of the virtual instance topology, and construct a topology state sample set.

[0120] The system periodically acquires the virtual instance topology history dataset generated by step S430 through a distributed storage interface, which is stored in a columnar database in time series form. Each virtual instance topology contains two core data structures: the node attribute matrix records the resource type identifier, the real-time occupancy rate of the computing unit, the actual value of memory allocation, and the 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 sliding window measurement value of the transmission delay, and the packet loss rate statistics. During data extraction, the system establishes a timestamp alignment indexing mechanism: taking snapshots of the topology at a fixed sampling frequency, synchronously associating the power control coefficients and resource allocation weight factors fed back by the S600 resource control module at that moment, and forming a spatiotemporally correlated data unit.

[0121] Feature engineering conversion is implemented when performing structured preprocessing on the original topology data: for discrete resource type identifiers in the node attribute matrix, a binary feature vector is generated through a one-hot encoding converter, so that each resource type corresponds to a unique active bit pattern; for continuous numerical features, including the computing unit occupancy rate and the memory allocation value, the min-max scaling algorithm is used to linearly map them to the normalized interval of zero to one, eliminating the interference of dimension differences on subsequent analysis; at the same time, the transmission delay measurement value in the edge connection matrix is processed, and the log transformation function is applied to compress the dynamic range of the data, solving the problem of model training divergence caused by long-tail distribution. In the key feature fusion stage, the neighborhood aggregation strategy is adopted: for each virtual node, the arithmetic mean of the transmission delay of all associated logical links and the statistical maximum of the packet loss rate are calculated to form an edge feature vector representing the connection quality; the edge feature vector and the node attribute vector are concatenated at the beginning and end of the feature dimension to generate a topology state vector with uniform dimensions. The final output topology state sample set consists of a sequence of timestamp-labeled feature vectors, and a data desensitization operation is performed before writing to the training database, hiding sensitive configuration parameters through field masking technology to provide a standardized input source for network optimization model training.

[0122] S520, based on the topology state sample set and the quality of service indicators, training a network optimization model.

[0123] The system loads S510 the constructed topology state sample set and performs model training in combination with the service quality constraint conditions defined in S200. The service quality indicators include an end-to-end transmission delay threshold, a task completion rate benchmark value, and a system energy consumption upper limit, which 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 a graph attention mechanism to dynamically adjust the influence coefficient of neighboring nodes through learnable weight parameters, capturing asymmetric dependency relationships between virtual nodes; the second channel deploys multi-layer graph convolution operators to iteratively perform feature propagation on the edge connection matrix, extracting deep association features of logical links. The outputs of the two channels are subjected to feature concatenation operations in the fusion layer to splice the node feature vectors and edge feature vectors into joint representation vectors.

[0124] The training process implements a dynamic sample weighting strategy: when loading batch data, abnormal samples that exceed the service quality indicator threshold are assigned three times the sampling weight, for example, when detecting records with excessive delay or task failure events, the model's perception of critical states is enhanced. The loss function design includes three optimization objectives: in the delay prediction branch, the Huber loss function is used to measure the deviation between the predicted transmission delay and the actual measured value, which automatically switches to a linear loss mode when the error is large to avoid gradient explosion; the resource allocation branch compares the resource allocation weight output by the model with the actual strategy feedback from S600 through the cross-entropy loss function; the topology stability branch calculates the cosine similarity of adjacent time slice feature vectors as a regularization constraint term to suppress unnecessary oscillation of the topology structure. In the model verification phase, a sliding time window strategy is applied: twenty-four hours of topology state data are selected as the training set, and the subsequent six hours of data are used as an independent validation set. When the topology reconstruction error on the validation set shows an upward trend for three consecutive evaluation periods, the early stopping mechanism is triggered and automatically rolled back to the historically optimal weight state. The final generated network optimization model includes a graph neural network encoder and a multi-task prediction head: the encoder outputs a two-hundred-and-fifty-six-dimensional hidden space feature vector; the multi-task prediction head outputs power adjustment coefficients and resource reallocation weight matrices in parallel through fully connected layers. Model parameters are updated to all edge nodes through a distributed parameter server synchronization mechanism to ensure global reasoning consistency.

[0125] S530, inputting the real-time virtual instance topology into the network optimization model to generate parameter adjustment instructions.

[0126] The system receives the virtual instance topology data stream output by the S430 module in real time, performs a topology data conversion operation: converts the node attribute matrix and edge connection matrix of the current topology into a standardized feature vector, strictly following the structured processing rules defined in S510. Before model reasoning, real-time verification is implemented: if 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 stage is implemented using a lightweight inference engine: the network optimization model trained in S520 is converted into a computation graph intermediate representation format, and the graph convolution layer and the fully connected layer are merged into a single computation unit through operator fusion optimization technology, reducing memory access overhead during inference.

[0127] The online inference process adopts a two-stage pipeline architecture: the first stage activates the topology stability monitoring submodule immediately after the graph neural network encoder outputs the hidden space feature vector, and calculates the dynamic deviation of the current topology from the historical baseline state; the second stage applies boundary constraint conditions to the power adjustment coefficient in the multi-task prediction head output stage, limiting its value to the physically feasible interval of -3dB to +3dB, and simultaneously implementing probability normalization processing on the resource allocation weight matrix to ensure that the sum of all resource allocation weights is equal to one. The parameter adjustment instruction generation stage includes a safety verification mechanism: input the predicted power adjustment coefficient into the service quality constraint verifier of S200, and when the simulation results show that the task completion rate is below the preset threshold, enable the backup strategy generator to call the historical optimal parameter configuration. The final output instruction uses binary encoding format: the first sixteen bits store the globally unique identifier of the target transmitting node; the middle thirty-two bits record the quantized value of the power adjustment coefficient, with a quantization precision of 0.1dB; the last sixty-four bits store the resource allocation weight matrix using sparse matrix encoding technology. The instruction is transmitted to the S610 resource execution module for analysis and implementation through the high-speed data bus, forming a closed-loop control chain from topology perception to parameter regulation.

[0128] In another embodiment:

[0129] S510, obtains historical running data of the virtual instance topology and constructs a topology state sample set.

[0130] The virtual instance topology historical data set generated by S430 is periodically obtained through a distributed storage interface, and the data set is stored in a columnar database in time series form. Each virtual instance topology includes:

[0131] Node attribute matrix: records resource type identifier , real-time occupancy rate of computing unit (range ), actual value of memory allocation , and network bandwidth quota;

[0132] Edge connection matrix: stores globally unique identifiers of logical links, transmission latency sliding window measurement values and packet loss rate statistics ;

[0133] Establish timestamp alignment index mechanism: take topology snapshot at 10Hz sampling frequency, synchronize associated power control coefficient feedback by S600 at this moment and resource allocation weight factor , form space-time correlation data unit. Implement structured preprocessing process:

[0134] The expression of node feature engineering conversion is:

[0135]

[0136] Among them:

[0137] The normalized attribute vector of node ; : resource type identifier of node (discrete enumeration value: 0=polar imaging unit, 1=millimeter wave radar unit); : calculation unit occupancy rate; : actual value of memory allocation; : minimum / maximum calculation occupancy rate of historical data set; : minimum / maximum memory value of historical data set; : vector splicing operator; : one-hot encoding function;

[0138] Output: normalized attribute vector of node

[0139] Further, delay feature transformation is performed:

[0140]

[0141] Among them, : transmission delay measurement value of logical link ; : log-transformed delay;

[0142] Input: original delay measured by S120 in real time, output: smoothed delay feature;

[0143] Further, neighborhood feature aggregation is constructed:

[0144] For node , calculate: generate edge feature vector:

[0145] ​in, is the mean delay of the associated link; is the maximum packet loss rate;

[0146] Topology state vector generation:

[0147]

[0148] in, For nodes The topological state vector of : node feature vector; : Neighborhood feature aggregation;

[0149] Output: Node The topological state vector ;

[0150] 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.

[0151] S520: Train a network optimization model based on the topology state sample set and service indicators.

[0152] 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

[0153] The network optimization model uses the LSTM-Transformer parallel architecture:

[0154] Time series feature extraction:

[0155]

[0156] 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) ; a long short-term memory (LSTM) neural network based on trainable parameters ;

[0157] Output: temporal feature (dimension 128).

[0158] Spatial feature extraction:

[0159]

[0160] wherein, is the spatial feature; is the time sliding window parameter; is the node topology state vector at time in the time sliding window.

[0161] Output: spatial feature (dimension 128).

[0162] Further, feature fusion is performed:

[0163]

[0164] wherein, is the fusion feature vector of node at time ; is a learnable weight parameter (initial value 0.5);

[0165] Output: fusion feature (dimension 128);

[0166] Multi-objective loss function:

[0167]

[0168] wherein, is the measured end-to-end delay; is the delay prediction value; is the true resource weight; is a Huber loss function for the regression task of delay prediction; CE( ) is a cross-entropy loss function for the classification task of resource weight allocation; is the resource weight vector predicted by the model; is the fusion feature vector of node n at time t; is the fusion feature vector of node n at time t−1; is the Chebyshev norm used to calculate the maximum deviation between feature vectors; is the weight coefficient;

[0169] Output: Multi-objective loss loss value .

[0170] Optimization by backpropagation Parameters, update the model in full every 24 hours.

[0171] S530, input the real-time virtual instance topology into the network optimization model to generate parameter adjustment instructions.

[0172] Receive S430 the real-time output of the virtual instance topology :

[0173] Topology deviation degree calculation:

[0174]

[0175] Where, is the topology deviation degree, indicating the difference between the current topology state and the historical benchmark; is the total number of nodes in the virtual instance topology; : current state vector; : historical benchmark vector;

[0176] Output: deviation degree When trigger the online fine-tuning of the model. is the topology deviation degree threshold.

[0177] Further, the parameter adjustment instruction generation:

[0178]

[0179]

[0180] Where, is the power adjustment amount; is the truncation function; is the weight matrix of power adjustment; is the global feature vector; is the new resource weight allocation scheme; (global feature vector); : normalized exponential function; is the weight matrix of resource weight;

[0181] Output: power adjustment amount and the new resource weight allocation scheme .

[0182] Instruction encoding and transmission: binary instruction package structure:

[0183] Front 16 bits: target node ID (corresponding to S110 device number);

[0184] Middle 32 bits: quantized power coefficient (accuracy 0.1 dB);

[0185] Back 64 bits: Sparse coding of (non-zero value index + value) of;

[0186] The instruction is transmitted to the S620 execution unit through a low-delay channel.

[0187] This module realizes three core technology breakthroughs:

[0188] (1) Dynamic feature closed loop. Fusion of polarization imaging resource data and millimeter wave time delay data ,

[0189] Real-time detection of S130 scattering feature anomalies (trigger model update when );

[0190] (2) Space-time parallel modeling. Capture infrared heat conduction time sequence characteristics (LSTM memory gate mechanism), model polarization space distribution (Transformer multi-head attention);

[0191] (3) Resource-energy consumption collaborative optimization. Power adjustment instruction Realize dynamic adjustment of energy consumption, resource weight Optimize resource allocation efficiency.

[0192] Step S600 includes at least steps S610-S630:

[0193] S610, analyze the power control coefficient and resource allocation weight in the parameter adjustment instruction.

[0194] The system receives the parameter adjustment instruction generated by step S530 through the high-speed instruction bus, which is stored in a fixed-length data structure in binary encoding format. 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 realized through a triple mapping mechanism:

[0195] 1. Field positioning mechanism: according to the instruction template index table pre-burned in the parsing module, the starting offset of the power control field and the starting offset of the resource allocation field are located.

[0196] 2. Coefficient extraction mechanism: In the power control field, the first 16 bits store the target transmitter node identifier, the middle 16 bits are the power operation code, and the last 32 bits store the power adjustment coefficient quantization value. The system separates the three groups of parameters through a shift register, and inputs the quantization value into a floating-point converter to restore it to a signed floating-point power control coefficient.

[0197] 3. Weight resolution mechanism: The resource allocation field uses sparse matrix encoding to store a four-dimensional resource weight vector (compute / memory / bandwidth / slot resources). The resolver calls the sparse matrix decoder to reconstruct the complete vector according to the 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 slot weight). Each weight value is restored to a normalized value in the interval [0, 1] by dividing it by 256 through an 8-bit fixed-point number operation, ensuring that the sum of all weight components is strictly equal to 1.

[0198] Double verification is implemented during the resolution process: the power control coefficient is immediately input into the boundary constraint checker after being restored; the resource allocation weight vector is verified for normalization by the cumulative checker. The final output is a structured resolution result: a power control coefficient package (including target node ID, operation type, coefficient value) and a resource allocation weight package (including a four-dimensional weight vector), which are transmitted to the S620 execution module through the memory sharing area.

[0199] S620, updates the transmitter power according to the power control coefficient and updates the resource pool quota according to the resource allocation weight.

[0200] The system executes two physical parameter update processes in parallel, both of which are implemented through device driver layer interfaces to perform hardware-level operations:

[0201] The transmitter power update process is as follows:

[0202] 1. Reference value acquisition: read the current reference power value of the target transmitter node through the power control interface. This interface calls the ioctl command word POWER_GET of the device driver program to obtain the integer power value from the wireless network card register in units of 0.1 dBm.

[0203] 2. New value calculation: perform differential operations according to the power operation type resolved by S610:

[0204] When the operation code is ADJUST_POWER: new power value = reference power value + power control coefficient;

[0205] When the operation code is SET_POWER: new power value = power control coefficient;

[0206] The calculation result is input into the saturation processor (limited to the ±20 dBm dynamic range supported by the physical device).

[0207] 3. Hardware write: construct a power reconfiguration instruction package (containing target device MAC address, new power value, effective timestamp), write it into the network card power control register through the ioctl command word POWER_SET of the device driver. The visible light communication transmitter (polar imaging device as described in S110) updates the LED drive current through the I2C bus; the radio frequency transmitter updates the amplifier bias voltage through the radio frequency front-end control bus.

[0208] The resource pool quota update process is as follows:

[0209] 1. State snapshot collection: access the physical resource pool management module (i.e. the resource state database of S410), obtain real-time resource quadruples: available CPU core number (read / proc / cpuinfo idle core counter), remaining memory capacity (call meminfo system call), free storage block linked list (query block device allocation table), and available time slot bitmap (extract from S220 time slot allocation mapping table).

[0210] 2. Quota recalculation:

[0211] CPU quota: total core number x CPU weight -> logical core number allocated to the target virtual machine;

[0212] Memory quota: total memory x memory weight -> target container memory upper limit value;

[0213] Storage quota: total number of free storage blocks x storage weight -> block device allocation amount of target volume;

[0214] Time slot quota: total number of time slots x time slot weight -> time slot allocation bitmap of target service flow;

[0215] The calculation process adopts an integer truncation strategy to ensure that resource allocation does not overflow (e.g. 12 cores x 0.33 weight -> allocate 3 cores).

[0216] 3. Strategy implementation: send quota update instructions to the resource pool manager:

[0217] CPU core binding: write cpu.cfs_quota_us parameter through cgroups subsystem;

[0218] Memory upper limit setting: configure through memory.limit_in_bytes control group file;

[0219] Storage volume expansion: call LVM tool to execute lvresize command to adjust logical volume size;

[0220] Time slot reallocation: update OFDMA frame structure through S220 time slot allocation mapping table;

[0221] The update operation adopts an atomic commit protocol: a write cache transaction mechanism is enabled at the device driver layer, and only when the power register write is successful and the resource quota update returns an ACK signal, a completion notification is sent to the S630 module.

[0222] S630, the updated transmit power and resource pool quota are fed back to the environment perception data acquisition module.

[0223] The system constructs a structured feedback data packet and establishes a closed-loop path with the S110 perception module:

[0224] 1. Data packet construction:

[0225] Power feedback item: encapsulates the actual effective power value of the target transmitting node (verified value read back from the device register) and the global timestamp (synchronized with the sliding window time reference of S130);

[0226] Quota feedback item: records the four-dimensional resource allocation result (CPU core allocation number, memory upper limit value, storage volume size, time slot bitmap hash value) and resource pool version identifier;

[0227] The data format adopts TLV encoding (Type-Length-Value), and the type identifier strictly matches the parsing template of the S110 preprocessing module.

[0228] 2. Transmission channel establishment:

[0229] The physical layer uses the PCIe high-speed bus to transmit the power feedback item (low delay requirement);

[0230] The application layer transmits the quota feedback item through the gRPC framework (needs to carry metadata);

[0231] The channel configuration parameters inherit the high-speed data interface definition of S130 (same data frame size and flow control parameters).

[0232] 3. Perception layer fusion processing:

[0233] After the environment perception data acquisition module (S110) receives the feedback packet:

[0234] The power value is used to calibrate the received signal strength reference: the polarization imaging sensor adjusts the ADC reference voltage according to the light transmission power, and the millimeter wave radar corrects the path loss compensation factor according to the radio frequency power;

[0235] The quota information acts on data preprocessing resource allocation: the image correction algorithm dynamically adjusts the thread pool size according to the available CPU core number; the point cloud filtering task sets the point cloud cache queue depth according to the memory upper limit value;

[0236] At the feature extraction stage of S120, the polarization degree matrix (containing the power calibrated light intensity values) in the channel state sequence and the millimeter wave scattering features (based on the clustering accuracy of resource constraints) both carry updated system state information, providing closed-loop input for the network optimization model of S520.

[0237] Embodiment two: Figure 2 A structural block diagram of a UAV detection system based on visible light polarization imaging according to an embodiment of the present application is shown. As shown, the structure can include: Figure 2

[0238] The environment perception module 10 is configured to acquire multi-source device original signals, perform non-uniformity correction and noise reduction processing, generate environment perception data containing polarization matrices, point clouds, and heat maps through space-time registration, and construct a channel state sequence. The environment perception module starts the workflow, and this module is responsible for acquiring original signal data from polarization imaging sensors and other multi-source devices (such as ordinary optical imaging devices and possible auxiliary sensors). In order to ensure the accuracy of subsequent analysis, this module performs key non-uniformity correction operations to eliminate sensor response differences and performs noise reduction processing to improve signal quality. Subsequently, it aligns and fuses polarization information (polarization matrices), three-dimensional spatial information (point clouds), and heat distribution information (heat maps) from different devices through precise space-time registration technology, generates unified environment perception data, and finally constructs a channel state sequence reflecting the current channel conditions, laying a solid data foundation for the entire detection process.

[0239] The cross-layer optimization module 20 is configured to analyze the channel state sequence to extract channel gain, interference strength, and noise power parameters, calculate power adjustment values in combination with quality of service constraints, perform spectrum multiplexing and time slot allocation, and generate cross-layer optimization parameters. The core task of the cross-layer optimization module is to analyze the channel state sequence delivered by the environment perception module. Through in-depth analysis, this module can accurately extract key wireless communication parameters, including channel gain, interference strength present in the environment, and background noise power. These parameters are input into the optimization algorithm together with the system's preset quality of service constraints (such as data transmission delay requirements, identification accuracy thresholds, etc.). The module calculates the optimal power adjustment value to balance energy consumption and performance requirements, intelligently performs spectrum resource multiplexing and allocation, and dynamically allocates communication time slots, generating a set of cross-layer optimization parameters containing optimization decisions. The role of this module is to dynamically allocate resources in complex wireless environments to ensure reliable and efficient transmission of perception data, providing stable information link protection for detection tasks.

[0240] ​The resource scheduling module 30 is configured to decompose the cross-layer optimization parameters into a computing load feature quantity and a transmission demand feature quantity, schedule the edge nodes to implement multi-point cooperation 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 feature quantities: one is the computing load feature quantity that needs to be borne by the edge computing node, and the other is the transmission demand feature quantity that has 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 cooperative computing to share the load, and decisively executes link switching operations when the communication link state changes, ensuring the real-time availability and optimal matching of computing and transmission resources. Finally, it outputs executable specific resource scheduling instructions to guide the flow and usage of physical resources.

[0241] The 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 core function of the module is to intelligently match the available computing, storage, and network units in the physical resource pool according to the requirements of the resource scheduling instructions. It constructs lightweight virtual container templates based on the matching results to encapsulate the running environment, and creates efficient virtual channels to meet the data transmission requirements. Then, the module deploys processing instances required by the detection algorithm on these virtualized infrastructures as needed, and finally generates a virtual instance topology structure diagram describing these virtual instances and their interconnection relationships. Through abstraction and pool management, the module significantly improves the utilization efficiency and deployment flexibility of heterogeneous hardware resources in the system.

[0242] The model inference module 50 is configured to convert the virtual instance topology into a feature vector sample set, train a model using a dual-channel graph neural network, and generate a power coefficient and a resource weight matrix instruction. The model inference module becomes the core driver of the detection task. It converts the virtual instance topology structure generated by the virtualization module, which represents the configuration of computing and communication resources, into a standardized feature vector sample set that can be processed by a machine learning model (especially a dual-channel graph neural network). Using these sample sets, the module drives the pre-trained dual-channel graph neural network to perform inference operations. This network can deeply mine the complex relationships between the topology structure and detection performance and resource efficiency, and finally output two key decision instructions: one is the power coefficient used to finely regulate the transmission power of each sensor and communication node, and the other is the resource weight matrix instruction used to guide the dynamic allocation of different resource (such as CPU, memory, bandwidth) quotas in the resource pool. The module realizes an intelligent decision-making closed loop based on environmental state and resource configuration.

[0243] The execution feedback module 60 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. The execution feedback module is responsible for implementing the intelligent decision of the model inference module and forming a closed loop. It accurately parses the received instruction, identifies the power operation code (indicating specific power adjustment action) and resource weight matrix (indicating resource quota adjustment direction) therein. The module updates the power register state in the system according to the above, and adjusts the transmission power level of each sensor and communication node in real time. At the same time, it also updates the allocation quota of each type of resource in the physical resource pool, ensuring that the calculation and transmission resources are supplied on demand. What is particularly important is that the 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, thereby forming a feedback loop from execution to perception, continuously improving the detection stability and accuracy of the entire system in a dynamic environment.

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 feature vector sample set, 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