Low-altitude surveillance radar x-band antenna gain estimation method and system

Abstract

The present application belongs to the technical field of low-altitude radar antenna gain, and discloses a low-altitude monitoring radar X-band antenna gain estimation method and system, which collects micro-parameters such as radiation unit phase distribution and feed network standing wave ratio, establishes the correlation model of the micro-parameters with meso-multiple path signals and macro-scene complexity, introduces a transition zone correction factor iterator, combines temperature compensation, frequency compensation and multi-level loss correction processes, offsets the errors of temperature drift and frequency offset in the 10-12GHz frequency band, improves the accuracy of gain estimation in complex low-altitude environment, constructs a three-dimensional adaptive matrix of RCS, maneuvering acceleration and target type, forms a three-dimensional correction system combined with the complexity dimension of the environment, in the multi-target scene, according to the logic that the smaller the RCS is, the greater the maneuvering acceleration is, and the higher the priority is, the calculation resources are allocated, time division multiplexing architecture is adopted to process targets in batches, and the contradiction between X-band antenna gain response lag and target maneuvering speed is relieved.

Description

Technical Field

[0001] This invention belongs to the field of low-altitude radar antenna gain technology, specifically a method and system for estimating the gain of X-band antennas for low-altitude surveillance radar. Background Technology

[0002] Gain estimation for X-band antennas in low-altitude surveillance radar is fundamentally a multi-field coupling optimization problem involving antenna radiation characteristics, complex low-altitude environments, and dynamic target behavior. The core deficiency of existing technologies is not a matter of surface accuracy or cost, but rather a deep mismatch between the underlying logic and practical application scenarios, as detailed below:

[0003] Existing methods fail to address the cross-scale coupling between microscopic antenna parameters and mesoscopic electromagnetic environment: At the microscopic level, the correlation model between the near-field radiation characteristics and far-field gain of X-band antennas is simplified, neglecting the near-field-far-field transition region shift caused by short wavelengths, while low-altitude detection ranges are mostly located in this transition region, rendering the traditional far-field assumption invalid; At the mesoscopic level, the coupling relationship between multipath effects, clutter interference, and antenna radiation patterns is fragmented, with environmental factors treated as independent correction terms, failing to consider the positive feedback effect of antenna pattern distortion and enhanced environmental interference, such as how the narrow beam of a high-gain X-band antenna exacerbates the direction dependence of multipath interference and gain fluctuations.

[0004] There is a contradiction between the dynamic characteristics of low-altitude targets and the X-band antenna gain response: In terms of time-varying lag, most existing models are static or quasi-static and have not established the correlation between the gain time-varying rate and the scene parameter change rate. The X-band antenna gain has a long lag response to environmental changes, and the fast maneuverability of low-altitude targets can easily lead to significant jumps in detection range. In terms of multi-target adaptation, different low-altitude targets have large differences in radar cross sections and different requirements for antenna gain. Existing unified models have not achieved adaptive matching between target characteristics and gain parameters, resulting in high rates of missed detection and false detection. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for estimating the gain of X-band antennas for low-altitude surveillance radar, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for estimating the gain of an X-band antenna for low-altitude surveillance radar, comprising an acquisition stage, a gain model construction stage, a correction model construction stage, a parameter calibration and optimization stage, an adaptive scheduling stage, a scenario verification stage, and a closed-loop process construction stage.

[0007] Preferably, during the acquisition phase, the phase distribution of the X-band antenna radiating elements and the phase distortion data of the feed network are acquired by scanning equipment. Simultaneously, micro parameters including the amplitude consistency of the radiating elements and the standing wave ratio of the feed network are acquired. A correlation model between the micro parameters and the meso-level environment is established. The influence coefficient between the feed phase distortion and the meso-level multipath signal delay difference is recorded in real time and a dynamic mapping table is formed. At the same time, the health assessment of the diffraction loss at the edge of the radiating elements is introduced. When the loss exceeds a preset threshold, near-field rescanning is automatically triggered.

[0008] By combining dynamically updated data from the geographic information system, the macro-scene is divided into complexity levels according to building density and vegetation coverage. Each level of scene is configured with a corresponding micro-parameter sampling frequency, and a sliding time window matching the sampling frequency is set synchronously. At the same time, the feature fingerprints of typical scenes are pre-stored, so as to realize the adaptive control of the macro-scene on the micro-collection process, environmental data collection cycle and typical scene fingerprint management.

[0009] The specific threshold standards for classifying the complexity of macro-level scenarios are as follows:

[0010] A building density greater than 60% or a vegetation coverage rate greater than 70% is considered a Level 1 scenario (dense urban area), corresponding to a micro-parameter sampling frequency of 2Hz;

[0011] A building density of 20%~60% and a vegetation coverage of 30%~70% constitutes a secondary scenario (mountainous / suburban area), corresponding to a sampling frequency of 1.5Hz;

[0012] A building density of <20% and vegetation coverage of <30% constitutes a Level 3 scene (plains / desert), corresponding to a sampling frequency of 1Hz;

[0013] Building density and vegetation coverage data for each scene level are collected by a Geographic Information System (GIS) with a grid precision of 10m×10m to ensure the accuracy of scene classification.

[0014] The specific parameters of the pre-stored typical scenario feature fingerprint are as follows:

[0015] Urban canyon scene (Level 1 scene): multipath delay difference 0.5~1.2μs, clutter power spectrum -50~-35dBm, electromagnetic interference intensity 45~60dBμV / m;

[0016] Mountainous scene (secondary scene): multipath delay difference 0.2~0.8μs, clutter power spectrum -55~-40dBm, electromagnetic interference intensity 25~40dBμV / m;

[0017] Plains scene (Level 3 scene): Multipath delay difference 0~0.3μs, clutter power spectrum -60~-45dBm, electromagnetic interference intensity 10~20dBμV / m;

[0018] Coastal scenario (new typical scenario): multipath delay difference 0.3~1.0μs, clutter power spectrum -52~-38dBm, electromagnetic interference intensity 30~50dBμV / m; all fingerprint parameters are updated at a frequency of 10 minutes / time to ensure matching with the actual environment.

[0019] Preferably, in the gain model construction stage, the cross-scale parameters from the acquisition stage are combined with the actual gain value derived from the radar echo in reverse, and a transition zone correction factor iterator is established. The least squares method is used to fit the residual between the actual gain and the estimated value. When the detection range is in the transition zone, the correction factor is updated in real time with the rate of change of the multipath power spectrum. For the temperature drift in the X-band 10-12GHz frequency band, a frequency compensation term is added. For every 0.1GHz frequency shift, the feeder loss coefficient is compensated according to the standard of 0.05dB / m.

[0020] The distance range between the near-field and far-field transition zone of an X-band antenna is calculated using the formula "transition zone distance = 2D² / λ", where D is the effective aperture of the antenna array and λ is the X-band radar wavelength.

[0021] When the X-band frequency is 10 GHz (λ=0.03 m) and the antenna aperture is 1 m, the lower limit of the transition zone distance is approximately 2 × 1² / 0.03 ≈ 66.7 m.

[0022] When the frequency is 12GHz (λ=0.025m) and the antenna aperture is 1m, the lower limit of the transition zone distance = 2×1² / 0.025 = 80m;

[0023] Therefore, in this invention, the near-field and far-field transition zone of the X-band antenna is uniformly defined as 60~80m. When the detection distance is within this range, the real-time update logic of the transition zone correction factor iterator is triggered.

[0024] The mutual coupling loss curves of the radiating elements within the temperature range of -40℃ to 60℃ were obtained by microwave anechoic chamber testing. Combined with the actual far-field gain value derived from radar echoes, a temperature-loss correlation model was established. The ambient temperature collected by the radar's built-in temperature sensor was used as a dynamic input. A multi-level loss correction process, including near-field loss correction and far-field multipath compensation, was incorporated to form a near-field and far-field cooperative gain model adapted to the characteristics of the X-band.

[0025] The specific parameter settings for microwave anechoic chamber testing are as follows:

[0026] The test frequency interval is 0.2GHz, covering the entire 10-12GHz frequency band, with a total of 11 test frequency points: 10GHz, 10.2GHz...12GHz;

[0027] The temperature step is 5℃, covering the range of -40℃ to 60℃, with a total of 21 test temperature points: -40℃, -35℃...60℃;

[0028] During testing, the spacing between radiating elements was fixed at 0.1m, consistent with the actual antenna array layout. The test distance was the far-field distance, ≥2D² / λ, where D is the antenna array aperture, taken as 1m, λ is taken as 0.025m, and the far-field distance is ≥80m.

[0029] Three sets of mutual coupling loss data were collected for each temperature and frequency combination, and the average value was taken as the loss value under that operating condition to ensure the reliability of the loss curve.

[0030] Preferably, in the correction model construction stage, the X-band near-field and far-field cooperative gain model of the receiver gain model construction stage is used to establish a three-dimensional dynamic adaptation correction model. The sliding time window set in the acquisition stage is used to acquire environmental data including multipath delay difference and clutter power spectrum. Based on the lightweight LSTM network, historical data is processed to predict the environmental change trend and generate a pre-corrected gain curve.

[0031] By combining the target RCS, maneuvering acceleration and target type data from the acquisition phase, a three-dimensional adaptation matrix of RCS, maneuvering acceleration and target type is established, and adaptation intervals are defined.

[0032] Based on the electromagnetic interference intensity derived from the macroscopic scene classification during the acquisition stage (Level 1 scene: electromagnetic interference intensity > 40dBμV / m, Level 2 scene: 20-40dBμV / m, Level 3 scene: < 20dBμV / m), environmental complexity is set as the third correction dimension. In complex environments, the time-varying correction weight is set to 0.5 and the target adaptation amplitude is reduced by 30%, while in simple environments, the weight is set to 0.3, forming a three-dimensional correction system of time-varying dimension, target dimension and environmental dimension, realizing dynamic optimization of the basic model in the gain model construction stage.

[0033] Preferably, the parameter calibration and optimization stage receives the measured data from the acquisition stage and the estimated deviation output from the correction model construction stage. The offline stage uses the measured data obtained in the acquisition stage as a basis to construct a linear statistical regression model and adapts it to mountainous and urban scenarios through transfer learning.

[0034] During the online phase, the model parameters are updated using an incremental learning mechanism for each new set of measured gain data derived from radar echoes, based on the sampling frequency of the micro parameters during the acquisition phase.

[0035] K-means clustering is performed on the gain estimation biases of the five consecutive outputs during the correction model construction phase. When the number of biases of a certain class accounts for more than 60% of the five consecutive biases, global optimization of the correction factor is automatically triggered, and the adaptation matrix coefficients of the correction model construction phase are updated synchronously.

[0036] Preferably, the adaptive scheduling phase relies on the three-dimensional dynamic adaptation correction results output by the correction model construction phase and the target priority division rules established in the acquisition phase. According to the priority logic of smaller RCS and greater maneuvering acceleration in the acquisition phase, the micro high-speed UAV is set as priority 1 and the large slow helicopter is set as priority 3. The priority 1 target occupies 50% of the gain computing resources.

[0037] When the number of targets to be detected is greater than 10, a time-division multiplexing architecture is adopted to process targets of different priorities in 3 batches within the time window set in the acquisition phase.

[0038] The system calls upon the typical scene feature fingerprints pre-stored during the data acquisition phase to match the current scene parameters in real time. Once a match is successful, it directly calls upon the model parameters pre-optimized during the model building phase to correct the model.

[0039] Preferably, the scenario verification stage is based on the basic gain model and the three-dimensional correction model of the gain model construction stage, and establishes a joint scenario of complex electromagnetic and dense targets, and records the missed detection data according to the sampling frequency of the acquisition stage.

[0040] For the temperature and loss models in the gain model construction stage and the environmental adaptation logic in the correction model construction stage, a cross-scenario of temperature change and electromagnetic interference is constructed to verify environmental robustness.

[0041] By introducing unknown scenario data and without inputting preset parameters, the synergistic effect of the calibration model in the parameter calibration and optimization stage and the scheduling mechanism in the adaptive scheduling stage is tested. After verification, the gain estimation deviation threshold is output.

[0042] Preferably, the closed-loop process construction stage is based on the model, algorithm and scheduling strategy verified in the previous stage, and adopts a heterogeneous architecture of FPGA and MCU. The FPGA deploys the basic model of the gain model construction stage, the three-dimensional correction algorithm of the correction model construction stage and the scheduling logic of the adaptive scheduling stage; the MCU stores the scene parameters of the acquisition stage and the calibration data of the parameter calibration and optimization stage for offline model management.

[0043] The specific FPGA model is Xilinx Zynq7020, with 55k logic cells, 8.5MB of Block RAM, and an integrated dual-core ARM Cortex-A9 processor with a main frequency of 866MHz. It is responsible for running the basic model, 3D correction algorithm and scheduling logic, and requires ≤80% of the logic resources.

[0044] The specific MCU model is STM32H743VI, with 2MB of Flash capacity, 1MB of RAM capacity, and a main frequency of 480MHz. It is used to store scene parameters and calibration data, and ≥30% of the storage margin needs to be reserved.

[0045] The FPGA communicates with the MCU via an SPI interface with a communication rate of ≥10Mbps and a data transmission delay of ≤1ms; the edge node communicates with the cloud via a 4G / 5G module with a data upload frequency of ≥1 time / hour, ensuring the timeliness of closed-loop optimization.

[0046] The deviation threshold of edge nodes is verified according to the scenario verification stage. The real-time estimated gain is compared with the actual value. When the threshold is exceeded, local parameter fine-tuning is triggered.

[0047] The cloud periodically aggregates multi-node deviation data, uses gradient descent to optimize the gain model, constructs a phase correction factor library, and pushes it to edge nodes.

[0048] The present invention also provides a low-altitude surveillance radar X-band antenna gain estimation system, based on the above method, including a data acquisition module, a basic model construction module, a dynamic correction module, a parameter calibration module, an adaptive scheduling module, a scene verification module, and an engineering closed-loop module;

[0049] The data acquisition module includes a scanning device, a geographic information system module, and a temperature sensor; the scanning device acquires microscopic parameters of the X-band antenna, establishes a correlation model between the microscopic parameters and the mesoscopic environment, and evaluates the diffraction loss at the edge of the radiating element to trigger a rescan; the geographic information system module classifies the macroscopic scene into complexity levels according to building density and vegetation coverage.

[0050] The basic model building module is connected to the data acquisition module. Based on the acquired cross-scale parameters, and with the actual gain value backed by radar echo as the benchmark, a transition zone correction factor iterator is established. A frequency compensation term is added for the temperature drift in the X-band 10-12GHz frequency band. A temperature-loss correlation model is established by combining temperature sensor data and microwave anechoic chamber test curves. Multi-level loss correction is incorporated to form a near-field and far-field collaborative basic gain model.

[0051] The dynamic correction module, connected to the basic model building module, collects environmental data, predicts environmental changes and generates pre-corrected gain curves through a lightweight LSTM network; it establishes a three-dimensional adaptation matrix by combining the target RCS, maneuver acceleration and target type, and divides the environmental complexity dimension according to the electromagnetic interference intensity, forming a three-dimensional correction system of time-varying dimension, target dimension and environmental dimension.

[0052] The parameter calibration module is connected to the dynamic correction module. It offline establishes a linear statistical regression model based on measured data and adapts it to multiple scenarios through transfer learning. Online, it adds measured data according to the sampling frequency, updates parameters using incremental learning, and clusters the continuous gain estimation bias to trigger correction factor optimization.

[0053] The adaptive scheduling module is connected to the dynamic correction module and the data acquisition module. Based on the three-dimensional dynamic adaptation correction results output by the dynamic correction module and the target priority division rules established by the data acquisition module, the target priority is divided according to the logic that the smaller the RCS and the greater the maneuvering acceleration, the higher the priority. The time-division multiplexing architecture is used to process multiple targets in batches and call the typical scene feature fingerprint matching parameters.

[0054] The scenario verification module is connected to the basic model construction module, dynamic correction module, and parameter calibration module. It establishes complex electromagnetic and dense target joint scenarios and temperature change and electromagnetic interference cross scenarios, introduces unknown scenario data, verifies the basic model, correction model and calibration-scheduling synergy effect, and outputs the gain estimation deviation threshold.

[0055] The engineering closed-loop module, connected to the aforementioned modules, adopts a heterogeneous architecture of FPGA and MCU. The FPGA deploys the basic model and correction algorithm, while the MCU stores the parameters. Edge nodes compare and estimate the gain with the actual value in real time and make fine adjustments. The cloud aggregates data to optimize the correction factor library, forming a closed-loop process.

[0056] The beneficial effects of this invention are as follows:

[0057] 1. This invention establishes a correlation model between microscopic parameters such as the phase distribution of radiating elements and the standing wave ratio of the feed network, and the mesoscopic multipath signal and the complexity of the macroscopic scene. At the same time, it introduces a transition zone correction factor iterator, combined with temperature compensation, frequency compensation and multi-level loss correction process, to offset the errors of temperature drift and frequency offset in the 10-12GHz band. It solves the problem of transition zone estimation deviation, which significantly improves the accuracy of gain estimation in complex low-altitude environments and is more in line with actual detection needs.

[0058] 2. This invention uses a lightweight LSTM network to process environmental data, predict environmental changes, and generate pre-corrected gain curves. At the same time, it constructs a three-dimensional adaptation matrix of RCS, maneuvering acceleration, and target type, and forms a three-dimensional correction system by combining the environmental complexity dimension. In multi-target scenarios, computing resources are allocated according to the logic that the smaller the RCS and the greater the maneuvering acceleration, the higher the priority. The time-division multiplexing architecture is used to process targets in batches, which effectively alleviates the contradiction between the lag in X-band antenna gain response and the fast maneuvering of targets, and reduces the missed detection and false detection of key objects such as micro high-speed targets.

[0059] 3. This invention designs an offline and online dual-mode parameter calibration scheme: in the offline stage, it adapts to mountainous and urban scenarios through transfer learning, while in the online stage, it updates model parameters in real time through incremental learning; at the same time, it monitors deviations through K-means clustering and triggers correction factor optimization; the engineering deployment adopts a heterogeneous architecture of FPGA and MCU, and the edge nodes can fine-tune parameters based on the deviation threshold. The cloud periodically summarizes data to optimize the correction factor library, forming a closed loop throughout the process; thus, the system can still operate stably under complex scenarios such as temperature changes and electromagnetic interference, reducing long-term operation and maintenance costs and better meeting the engineering application needs of low-altitude surveillance radar. Attached Figure Description

[0060] Figure 1 This is a flowchart of the low-altitude surveillance radar X-band antenna gain estimation method of the present invention;

[0061] Figure 2 This is a flowchart illustrating the core process of building and refining the gain model in this invention.

[0062] Figure 3 This is the core flowchart for closed-loop optimization and scheduling verification of the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] like Figures 1 to 3 As shown in the figure, this invention provides a method for estimating the gain of an X-band antenna for low-altitude surveillance radar, including an acquisition stage, a gain model construction stage, a correction model construction stage, a parameter calibration and optimization stage, an adaptive scheduling stage, a scenario verification stage, and a closed-loop process construction stage. The specific implementation of each stage is as follows:

[0065] The acquisition phase involves using a 0.1 mm near-field scanning device (vector network analyzer paired with a high-precision near-field scanning frame) to collect data on the phase distribution of the X-band antenna radiating elements and the phase distortion of the feed network. Simultaneously, it collects micro-parameters such as the amplitude consistency of the radiating elements and the standing wave ratio of the feed network, establishes a correlation model between the micro-parameters and the meso-level environment, and records the influence coefficient between the feed phase distortion and the meso-level multipath signal delay difference in real time. For every 0.5° increase in phase distortion, the delay difference fluctuation threshold is increased by 10%, and a dynamic mapping table is formed. At the same time, a health assessment of the diffraction loss at the edge of the radiating elements is introduced, and near-field rescanning is automatically triggered when the loss exceeds 5%.

[0066] Formula for calculating diffraction loss at the edge of a radiating element: This represents the diffraction loss at the edge of the radiating element, expressed in % . This represents the X-band radar wavelength, measured in meters (m). The range of values ​​is 10-12 GHz, corresponding to wavelengths of 0.025-0.03 m. This represents the distance from the edge of the radiating element to the near-field scanning surface, in meters (m), measured by a 0.1 mm near-field scanning device. This represents the phase distortion angle of the radiating element, expressed in degrees, and is calculated from the phase distribution data collected by the scanning equipment.

[0067] Triggering condition: When When this happens, near-field rescan is automatically triggered.

[0068] Combining dynamically updated data from a Geographic Information System (GIS), the macro-scenes are divided into complexity levels based on building density and vegetation coverage. Each level of scene is configured with a corresponding micro-parameter sampling frequency, with 2Hz for densely populated urban areas and 1Hz for plains areas.

[0069] A sliding time window matching the sampling frequency is set synchronously, and feature fingerprints of typical scenarios are pre-stored to achieve adaptive control of macro-scenes on micro-acquisition processes, environmental data acquisition cycles, and fingerprint management of typical scenarios.

[0070] Examples of specific derived rules for macroscopic scene classification and electromagnetic interference intensity are as follows:

[0071] After classifying the macro-level scene complexity by combining dynamically updated building density and vegetation coverage data from a geographic information system, the electromagnetic interference intensity under each scene is simultaneously collected using the radar's built-in electromagnetic interference detection module, with the unit being dBμV / m. A mapping relationship between scene complexity and electromagnetic interference intensity is then established.

[0072] Level 1 scenario (building density > 60% or vegetation coverage > 70%): Electromagnetic interference intensity default > 40dBμV / m (complex environment); Level 2 scenario (20% ≤ building density ≤ 60% and 30% ≤ vegetation coverage ≤ 70%): Electromagnetic interference intensity default 20-40dBμV / m (moderate environment);

[0073] Level 3 scenario (building density < 20% and vegetation coverage < 30%): Electromagnetic interference intensity default < 20dBμV / m (simple environment);

[0074] This mapping relationship serves as the specific basis for determining the environmental complexity dimension during the model construction phase.

[0075] In the gain model construction stage, the microscopic radiation parameters, mesoscopic environmental parameters, and macroscopic scene classification data from the acquisition stage are combined. The actual gain value derived from the radar echo (calculated by back-calculation using the radar receiving channel calibration coefficient and the known RCS of the target) is used as a benchmark to establish a transition zone correction factor iterator. The least squares method is used to fit the residual between the actual gain and the estimated value. When the detection range is in the transition zone, the correction factor is updated in real time with the rate of change of the multipath power spectrum. For every 2dB increase in the rate of change, the correction factor is increased by 0.3. For the temperature drift (±0.2GHz, caused by the heat dissipation of the host) in the X-band 10-12GHz frequency band, a frequency compensation term is added. For every 0.1GHz frequency shift, the feeder loss coefficient is compensated according to the standard of 0.05dB / m.

[0076] Formula for inverse calculation of actual radar gain: This indicates the actual gain of the radar antenna, in dB. This represents the measured power of the radar receiving channel, in dBm, and is collected by the radar receiving module. This indicates the detection distance between the radar and the target, in meters, and is measured by the radar ranging module. The X-band radar wavelength is represented in meters, and the formula for calculating diffraction loss at the edge of the same radiating element is also included. This represents the total loss of the radar system, in dB, including feeder loss, clutter loss, etc., with a preset value of 0.05 dB / m × feeder length. This represents the measured RCS of the target, in meters. 2 The data is collected by the scanning equipment during the acquisition phase.

[0077] Formula relating the rate of change of multipath power spectrum to correction factor: This represents the transition zone correction factor after real-time updates; This represents the baseline value of the correction factor, obtained by fitting the residuals using the least squares method; the default value is 1.0. This represents the multipath power spectrum change rate, in dB, extracted by the radar clutter suppression module, with a value range of 0-10 dB.

[0078] Logical explanation: For every 2dB increase in the multipath power spectrum change rate, the correction factor is increased by 0.3.

[0079] Formulas for frequency offset and feeder loss compensation: The value represents the feeder loss compensation after frequency shift, in dB; 0.05 is the feeder loss coefficient for every 0.1 GHz frequency shift, in dB / (m·0.1 GHz). This indicates the feeder length in meters (m), a parameter inherent to radar hardware, such as 5m. This represents the actual frequency offset in GHz, collected by the radar frequency detection module, with a value range of ±0.2 GHz.

[0080] Example: When the feeder length is 5m and the frequency offset is 0.2GHz, the compensation value is... =0.05×5×(0.2 / 0.1)=0.5dB.

[0081] By using microwave anechoic chamber testing (high and low temperature chamber temperature control, data collected every 5℃), the mutual coupling loss curves of radiating elements within the temperature range of -40℃ to 60℃ were obtained. Combined with the actual far-field gain value derived from radar echoes, a temperature-loss correlation model was established. The ambient temperature collected by the radar's built-in temperature sensor was used as a dynamic input, and a multi-level loss correction process of near-field loss correction and far-field multipath compensation was incorporated to form a near-field and far-field cooperative gain model adapted to the characteristics of the X-band.

[0082] In the modified model construction stage, the X-band near-field and far-field collaborative gain model of the receiver gain model construction stage is used to establish a three-dimensional dynamic adaptation correction model. The 50ms sliding time window set in the acquisition stage (matching the sampling frequency of micro parameters) is used to collect environmental data such as multipath delay difference and clutter power spectrum (extracted by radar clutter suppression module). Based on a lightweight LSTM network with a parameter scale of <100,000, historical data is processed to predict the environmental change trend in the next 100ms and generate a pre-corrected gain curve. The lightweight LSTM network contains 2 hidden layers and uses ReLU activation function and Adam optimizer.

[0083] Core structural parameters of the LSTM model:

[0084] Input layer dimension: 2-dimensional, corresponding to two features of the input environment data: multipath delay difference and clutter power spectrum;

[0085] Hidden layer configuration: 2 hidden layers, each containing 128 LSTM neurons, with fully connected layers between them, and a dropout layer (dropout rate 0.2) added after each layer to suppress overfitting;

[0086] Output layer dimension: 1-dimensional, outputting a continuous gain correction sequence within the next 100ms, outputting 10 data points per 10ms;

[0087] Network connection relationship: Input layer → Hidden layer 1 (LSTM+ReLU activation) → Dropout layer 1 → Hidden layer 2 (LSTM+ReLU activation) → Dropout layer 2 → Fully connected output layer (Linear activation).

[0088] Specific steps for training an LSTM model:

[0089] Step 1: Dataset Construction and Preprocessing - Collect historical environmental data for three typical low-altitude scenarios (dense urban areas, mountains, and plains). Collect 10,000 samples for each scenario. Each sample contains environmental data for 20 consecutive sliding time windows + gain correction labels for 1 target time window. Perform min-max normalization on the input features. Multipath delay difference (range 0.1-2μs) and clutter power spectrum (range -60~-20dBm) are uniformly normalized to the [0, 1] interval.

[0090] Step 2: Dataset partitioning – Divide the dataset into training, validation, and test sets in a 7:2:1 ratio. The training set contains balanced samples from various scenarios, the validation set is used to monitor overfitting, and the test set uses data from unknown scenarios (coastal ports).

[0091] Step 3: Training parameter configuration - The initial learning rate is set to 0.001, and a step-wise learning rate decay is used, which decays to 0.9 of the current value every 10 rounds. The number of training rounds is 50, the batch size is 32, the loss function is the mean squared error (MSE), and the optimizer Adam has β1=0.9, β2=0.999, and ε=1e-8.

[0092] Step 4: Training process monitoring - When the validation set loss does not decrease for 5 consecutive rounds, the early stopping mechanism is triggered to save the current optimal model weights; after training, the model prediction error is calculated using the test set, requiring the pre-corrected gain curve to have a good fit of ≥92% with the actual gain change trend.

[0093] After training, the model performance needs to be validated using a test set. Specific evaluation metrics are as follows:

[0094] The mean square error (MSE) must be ≤0.005, the mean absolute error (MAE) must be ≤0.05dB, and the fit between the pre-corrected gain curve and the actual gain change trend must be ≥92%. At the same time, considering the real-time requirements of low-altitude radar, the model inference latency must be ≤10ms. This latency index is based on the Xilinx Zynq7020 FPGA hardware platform test to ensure that the radar real-time processing requirements are met after deployment.

[0095] The adaptation logic between the model and the low-altitude radar scenario:

[0096] Input data correlation: The multipath delay difference and clutter power spectrum of the LSTM input are directly taken from the sliding time window data of the acquisition stage and synchronized in real time with the output of the X-band radar clutter suppression module to ensure the timeliness of the input data;

[0097] Output data fusion: The pre-corrected gain curve (10 consecutive correction values) output by LSTM is fused with the time-varying correction weight of the three-dimensional correction system in a "point-by-point product" manner. That is, the final time-varying correction value at a single time point = pre-corrected gain value × time-varying correction weight of the corresponding scene, 0.5 for complex environment, 0.4 for medium environment, and 0.3 for simple environment.

[0098] Scene adaptation enhancement: During training, samples in complex scenes such as densely populated urban areas are assigned a weight of 0.5, mountain scenes 0.3, and plain scenes 0.2. At the same time, ±5% random noise is added to simulate electromagnetic interference fluctuations, thereby improving the model's robustness to complex low-altitude electromagnetic environments.

[0099] Combining target RCS, maneuvering acceleration, and target type data from the acquisition phase, a three-dimensional adaptation matrix for RCS, maneuvering acceleration, and target type is established, and adaptation intervals are defined. Gain is enhanced by 1.2 dB for micro-UAVs and suppressed by 0.6 dB for high-speed helicopters with triggering amplitude-limiting filtering. Maneuvering acceleration is calculated from the radar echo Doppler frequency shift, and target type is distinguished by echo polarization characteristics and pulse width.

[0100] The matching rules for target type, RCS, maneuvering acceleration, and gain correction are as follows:

[0101] The micro high-speed UAV has an RCS < 0.05m² and a maneuvering acceleration > 3m / s². The basic gain correction is +1.2dB. In complex environments, the adaptation range is reduced by 30% to +0.84dB, while in simple environments, there is no reduction and it is still +1.2dB.

[0102] For conventional civilian drones, the corresponding RCS is 0.05~0.5m², the maneuvering acceleration is 1~3m / s², the basic gain correction is +0.8dB, the adaptation range is reduced by 30% in complex environments to +0.56dB, and it remains at +0.8dB in simple environments.

[0103] For small fixed-wing aircraft with RCS of 0.5~8m² and maneuvering acceleration of 1~3m / s², the basic gain correction is +0.3dB. In complex environments, the adaptation range is reduced by 30% to +0.21dB, while in simple environments, it remains at +0.3dB.

[0104] For large helicopters with RCS > 8m² and maneuvering acceleration < 1m / s², the basic gain correction is -0.6dB. In complex environments, the adaptation range is reduced by 30% to -0.42dB, while in simple environments it remains at -0.6dB.

[0105] For low-speed balloons with an RCS of 5~20m² and a maneuvering acceleration of <0.5m / s², the basic gain correction is -0.3dB. In complex environments, the adaptation range is reduced by 30% to -0.21dB, while in simple environments it remains at -0.3dB.

[0106] Electromagnetic interference intensity is derived from the macroscopic scene classification during the acquisition phase. Electromagnetic interference intensity >40dBμV / m is considered complex, 20-40dBμV / m is considered moderate, and <20dBμV / m is considered simple. Environmental complexity is set as the third correction dimension. In complex environments, the time-varying correction weight is set to 0.5 and the target adaptation amplitude is reduced by 30%. In simple environments, the weight is set to 0.3, forming a three-dimensional correction system of time-varying dimension, target dimension and environmental dimension, realizing dynamic optimization of the basic model in the gain model construction phase.

[0107] After determining the environmental complexity based on the mapping relationship between scene complexity and electromagnetic interference intensity, the following correction rules are applied:

[0108] Complex environment (electromagnetic interference intensity > 40 dBμV / m): Time-varying correction weight = 0.5, target adaptation amplitude reduced by 30%;

[0109] In moderate environments (electromagnetic interference intensity 20-40 dBμV / m): time-varying correction weight = 0.4, target adaptation amplitude reduced by 15%;

[0110] Simple environment (electromagnetic interference intensity < 20 dBμV / m): Time-varying correction weight = 0.3, target adaptation amplitude reduced by 0%;

[0111] A three-dimensional correction system based on time-varying factors, objectives, and environment is formed.

[0112] To improve the parameter accuracy of the 3D dynamic adaptation correction model in the correction model construction stage, the parameter calibration and optimization stage receives the measured data from the acquisition stage and the estimated deviation output from the correction model construction stage. In the offline stage, based on 50 sets of measured data (covering day and night periods) of plain scene obtained in the acquisition stage, a linear statistical regression model is constructed. Through transfer learning, it is adapted to mountain and urban scenes, and 10 sets of unique interference parameters are added to each scene.

[0113] During the online phase, the sampling frequency of micro parameters is based on the acquisition phase. For each new set of measured gain data based on radar echo back-inference, the model parameters are updated using an incremental learning mechanism. The new data and historical data are weighted at 7:3.

[0114] K-means clustering is performed on the five consecutive gain estimation biases (actual value - estimated value) output during the correction model construction phase. When the number of biases of a certain type (e.g., positive bias > 0.5dB) accounts for more than 60% of the five consecutive biases, global optimization of the correction factor is automatically triggered. For example, the multipath correction factor is increased by 0.2, and the adaptation matrix coefficients of the correction model construction phase are updated synchronously to achieve continuous adaptation of model parameters to the actual scenario.

[0115] K-means clustering sets the number of clusters to 3, corresponding to negative bias, no bias, and positive bias scenarios, respectively. The initial cluster center values ​​are -0.8dB, 0dB, and +0.8dB, respectively. There are two termination conditions for clustering iteration, and iteration can be stopped if either one is met: first, the offset of the cluster center in two consecutive iterations is <0.01dB; second, the number of iterations reaches 10. By specifying the clustering parameters, the stability and controllability of the clustering results are ensured.

[0116] The adaptive scheduling phase relies on the three-dimensional dynamic adaptation correction results output by the correction model construction phase and the target priority division rules established in the acquisition phase. According to the priority logic of smaller RCS and greater maneuvering acceleration in the acquisition phase, the micro high-speed UAV is set as priority 1 and the large slow helicopter is set as priority 3. The priority 1 target occupies 50% of the gain computing resources.

[0117] Based on the rule that higher priority is given to those with smaller RCS and greater maneuvering acceleration, specific priority criteria are determined as follows:

[0118] Priority 1: Micro high-speed targets (RCS < 0.05m² and maneuvering acceleration > 3m / s², such as micro high-speed UAVs) are exclusively allocated 50% of the gain computing resources;

[0119] Priority 2: Medium-sized, medium-speed targets (0.05m²≤RCS≤8m² and 1m / s²≤maneuvering acceleration≤3m / s², such as conventional civilian drones and small fixed-wing aircraft), allocate 30% gain computing resources;

[0120] Priority 3: Large, slow-moving targets (RCS > 8m² and maneuvering acceleration < 1m / s², such as large helicopters) are allocated 20% of the gain computing resources.

[0121] When the number of detected targets is greater than 10, a time-division multiplexing architecture is adopted. Within the 50ms time window set in the acquisition phase, targets of different priorities are processed in 3 batches to avoid computational congestion.

[0122] The time-division multiplexing architecture uses a fixed time slot allocation rule for its 50ms time window. Priority 1 targets (micro high-speed UAVs) occupy 25ms time slots, accounting for 50% of the total time window; priority 2 targets (conventional civilian UAVs, small fixed-wing aircraft) occupy 15ms time slots, accounting for 30% of the total time window; and priority 3 targets (large helicopters, low-speed balloons) occupy 10ms time slots, accounting for 20% of the total time window. Time slot switching adopts a seamless switching mechanism with a switching latency of ≤1ms to avoid target data loss. If the number of high-priority targets is <3, the remaining time slots will be dynamically allocated to low-priority targets to ensure that the computing resource utilization rate is ≥90% and improve resource utilization efficiency.

[0123] The system calls upon the typical scene feature fingerprints pre-stored during the acquisition phase, which are stored in a hash table. These fingerprints include multipath delay differences of 0.5 to 1.2 μs in urban canyons and power spectra of mountain clutter. The system matches the current scene parameters in real time. Once a match is successful, the system directly calls upon the model parameters pre-optimized during the model building phase to correct the model parameters, ensuring the continuity of gain estimation when switching scenes.

[0124] The scenario verification phase, specifically, establishes a complex electromagnetic and dense target joint scenario based on the basic gain model and the three-dimensional correction model constructed during the gain model construction phase. This scenario simulates a multipath delay difference of 0.5–1.2 μs in a core urban area and an RCS of 0.01–0.1 m for a swarm of 20 drones. 2 Record the missed detection data for 48 hours according to the sampling frequency of the collection phase;

[0125] For the temperature and loss model in the gain model construction stage and the environmental adaptation logic in the correction model construction stage, a cross-scenario of temperature change and electromagnetic interference was constructed, including temperature cycling from -30℃ to 50℃ and electromagnetic interference of 50dBm, to verify the environmental robustness.

[0126] Introduce third-party unknown scenario data, such as salt spray and multipath environment in coastal ports, without inputting preset parameters, and test the synergistic effect of calibration model in parameter calibration and optimization stage and scheduling mechanism in adaptive scheduling stage. The gain estimation deviation is required to be <0.8dB. After verification, the gain estimation deviation threshold is output to ensure the applicability of the technical solution in the first five stages in all scenarios.

[0127] The closed-loop process construction phase is based on the models, algorithms, and scheduling strategies verified in the preceding phases. It adopts a heterogeneous architecture of FPGA and MCU. The FPGA (Xilinx Zynq series, integrating ARM Cortex-A9) deploys the basic model of the gain model construction phase, the three-dimensional correction algorithm of the correction model construction phase, and the scheduling logic of the adaptive scheduling phase. The MCU (STM32H7 series) stores the scene parameters of the acquisition phase and the calibration data of the parameter calibration and optimization phase for offline model management.

[0128] The edge node is verified according to the deviation threshold (±0.5dB) in the scenario verification stage. Every 10 detection cycles, the real-time estimated gain is compared with the actual value. When the threshold is exceeded, local parameter fine-tuning is triggered. The correction factor in the model construction stage is adjusted through the proportional-integral controller.

[0129] The proportional integral formula for fine-tuning edge node parameters: This indicates the amount of fine-tuning of the correction factor in the modified model; This represents the scaling factor, which defaults to 0.2 and is calibrated by the deviation threshold during the scenario verification phase. This represents the integral coefficient, which defaults to 0.1 and is calibrated by the deviation threshold during the scenario verification phase. This represents the real-time measured gain, in dB, and is calculated using the same formula as the actual radar gain. This represents the real-time estimated gain, in dB, output by the modified model. This represents the integration time, measured in seconds, and is taken as 10 detection cycles, such as 10 × 0.05 s = 0.5 s;

[0130] Triggering condition: When When the deviation threshold of 0.5dB (scenario verification stage) is reached, the formula is triggered to calculate the fine-tuning amount.

[0131] The coefficient calibration of the PI controller is based on the gain estimation deviation threshold (default ±0.5dB) output during the scenario verification phase. The specific calibration formula is as follows:

[0132] proportionality coefficient proportionality coefficient =0.2 × (Actual Deviation Threshold / 0.5), Integral Coefficient =0.1×(actual deviation threshold / 0.5); For example, when the deviation threshold determined after scenario verification is ±0.8dB, =0.2×(0.8 / 0.5)=0.32, =0.1×(0.8 / 0.5)=0.16.

[0133] After the coefficient calibration is completed, it needs to be verified by 100 sets of actual measurement data to ensure that the gain deviation after fine-tuning is ≤ 80% of the deviation threshold, so as to ensure that the fine-tuning effect is stable and reliable.

[0134] The cloud aggregates multi-node deviation data weekly, and uses the gradient descent method to optimize the gain model and build a phase correction factor library. For example, the multipath correction factor in coastal areas is reduced by 0.15. The data is then pushed to edge nodes via the MQTT protocol, forming a closed loop of the entire process of collection, modeling, correction, calibration, scheduling, verification, and optimization to meet the needs of engineering applications.

[0135] This invention also provides a low-altitude surveillance radar X-band antenna gain estimation system, based on the above method, including a data acquisition module, a basic model construction module, a dynamic correction module, a parameter calibration module, an adaptive scheduling module, a scene verification module, and an engineering closed-loop module;

[0136] The data acquisition module includes a 0.1 mm near-field scanning device, a geographic information system module, and a temperature sensor. The scanning device collects microscopic parameters of the X-band antenna, establishes a correlation model between the microscopic parameters and the mesoscopic environment, and evaluates the diffraction loss at the edge of the radiating element to trigger a rescan. The geographic information system module divides the macroscopic scene into complexity levels according to building density and vegetation coverage, and configures a dedicated microscopic parameter sampling frequency for each level.

[0137] The basic model building module is connected to the data acquisition module. Based on the acquired cross-scale parameters, and with the actual gain value backed by radar echo as the benchmark, a transition zone correction factor iterator is built. A frequency compensation term is added for the temperature drift in the X-band 10-12GHz frequency band. A temperature-loss correlation model is established by combining temperature sensor data and microwave anechoic chamber test curves. Multi-level loss correction is incorporated to form a near-field-far-field collaborative basic gain model.

[0138] The dynamic correction module, connected to the basic model building module, collects environmental data in a 50ms sliding time window, predicts environmental changes in the next 100ms through a lightweight LSTM network and generates a pre-corrected gain curve; it establishes a three-dimensional adaptation matrix of RCS, maneuver acceleration and target type by combining target RCS, maneuver acceleration and target type, and divides the environmental complexity dimension according to the electromagnetic interference intensity, forming a three-dimensional correction system of time-varying dimension, target dimension and environmental dimension.

[0139] The parameter calibration module is connected to the dynamic correction module. It establishes a linear statistical regression model offline based on measured data of plain scene and adapts it to multiple scenes through transfer learning. Online, it adds measured data according to the sampling frequency, updates parameters by incremental learning, and clusters the continuous gain estimation bias to trigger the optimization of correction factor.

[0140] The adaptive scheduling module is connected to the dynamic correction module and the data acquisition module. Based on the three-dimensional dynamic adaptation correction results output by the dynamic correction module and the target priority division rules established by the data acquisition module, the target priority is divided according to the logic that the smaller the RCS and the greater the maneuvering acceleration, the higher the priority. The time-division multiplexing architecture is used to process multiple targets in batches and call the typical scene feature fingerprint matching parameters.

[0141] The scenario verification module is connected to the basic model construction module, dynamic correction module, and parameter calibration module. It establishes complex electromagnetic and dense target joint scenarios and temperature change and electromagnetic interference cross scenarios, introduces unknown scenario data, verifies the basic model, correction model and calibration-scheduling synergy effect, and outputs the gain estimation deviation threshold.

[0142] The engineering closed-loop module, connected to the aforementioned modules, adopts a heterogeneous architecture of FPGA and MCU. The FPGA deploys the basic model and correction algorithm, while the MCU stores the parameters. Edge nodes compare and estimate the gain with the actual value in real time and make fine adjustments. The cloud aggregates data to optimize the correction factor library, forming a closed-loop process.

[0143] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0144] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for estimating the gain of an X-band antenna for low-altitude surveillance radar, characterized in that, It includes the acquisition phase, gain model construction phase, correction model construction phase, parameter calibration and optimization phase, adaptive scheduling phase, scenario verification phase, and closed-loop process construction phase; Acquisition phase: Collect microscopic parameters of X-band antenna, establish a correlation model between microscopic and mesoscopic levels, and classify scene complexity according to building density and vegetation coverage. Gain model construction stage: Based on the actual gain value, establish a transition region correction factor iterator, supplement temperature and frequency compensation terms, incorporate loss correction, and form a basic gain model; Model building phase: Collect environmental data, predict environmental changes through LSTM network, establish a three-dimensional adaptation matrix of RCS, maneuvering acceleration and target type, add environmental complexity dimension, and form a three-dimensional correction system; Parameter calibration and optimization phase: Establish regression model to adapt to the scenario when offline, update model parameters when online, and optimize correction factors when deviation exceeds the standard; Adaptive scheduling phase: Prioritize targets based on RCS and maneuver acceleration parameters, and use scenario fingerprints to match scenario parameters; Scenario verification phase: Establish joint and cross-scenario tests, and introduce unknown scenario data for verification; Closed-loop process construction phase: The model algorithm is deployed using a heterogeneous architecture of FPGA and MCU to establish a closed loop between the edge and the cloud; The gain model construction stage combines the cross-scale parameters of the acquisition stage, and uses the actual gain value derived from the radar echo as a benchmark to establish a transition zone correction factor iterator, fits the residual between the actual gain and the estimated value, and updates the correction factor in real time with the multipath power spectrum change rate when the detection range is in the transition zone; a frequency compensation term is added for the temperature drift of the X-band 10-12GHz frequency band, and feeder loss is compensated according to the frequency band offset. The mutual coupling loss curves of the radiating elements within the temperature range were obtained by microwave anechoic chamber testing. Combined with the actual far-field gain value derived from the radar echo, a temperature-loss correlation model was established. The ambient temperature collected by the radar's built-in temperature sensor was used as a dynamic input. A multi-level loss correction process, including near-field loss correction and far-field multipath compensation, was incorporated to form a near-field and far-field collaborative basic gain model. The modified model construction stage receives the X-band near-field and far-field collaborative gain model in the gain model construction stage, establishes a three-dimensional dynamic adaptation correction model, collects environmental data through a sliding time window, processes historical data based on LSTM network, predicts environmental change trends, and generates a pre-corrected gain curve. By combining the target RCS, maneuvering acceleration and target type data from the acquisition phase, a three-dimensional adaptation matrix of RCS, maneuvering acceleration and target type is established, and adaptation intervals are defined. Based on the electromagnetic interference intensity derived from the macro-scenario classification, environmental complexity is set as the third correction dimension. The complex environment and the simple environment are respectively configured with corresponding time-varying correction weights and target adaptation amplitudes, forming a three-dimensional correction system of time-varying dimension, target dimension and environmental dimension.

2. The method for estimating the gain of an X-band antenna for low-altitude surveillance radar according to claim 1, characterized in that, The acquisition phase involves collecting phase distribution data of X-band antenna radiating elements and phase distortion data of the feed network using a scanning device. Simultaneously, micro-parameters, including amplitude consistency of radiating elements and standing wave ratio of the feed network, are acquired. A correlation model between micro-parameters and the meso-level environment is established. The influence coefficient between feed phase distortion and meso-level multipath signal delay difference is recorded in real time to form a dynamic mapping table. At the same time, a health assessment of radiating element edge diffraction loss is introduced. When the loss exceeds a preset threshold, near-field rescanning is automatically triggered. By combining dynamically updated data from the geographic information system, the macro-scenes are divided into complexity levels according to building density and vegetation coverage. Each level of scene is configured with corresponding micro-parameter sampling frequency and sliding time window, and the feature fingerprints of typical scenes are pre-stored.

3. The method for estimating the gain of an X-band antenna for low-altitude surveillance radar according to claim 2, characterized in that, The parameter calibration and optimization stage receives the measured data from the acquisition stage and the estimated deviation from the output of the correction model construction stage. The offline stage constructs a linear statistical regression model and adapts it to multiple scenarios through transfer learning. During the online phase, the model parameters are updated using an incremental learning mechanism based on the sampling frequency of micro-parameters. K-means clustering is performed on the gain estimation biases of multiple consecutive iterations. When the proportion of a certain type of bias reaches a preset threshold, global optimization of the correction factor is automatically triggered, and the coefficients of the adaptation matrix are updated synchronously.

4. The method for estimating the gain of an X-band antenna for low-altitude surveillance radar according to claim 3, characterized in that, The adaptive scheduling phase relies on the three-dimensional dynamic adaptation correction results output by the correction model construction phase and the target priority division rules established in the acquisition phase to divide the target priority according to the logic that the smaller the RCS and the greater the maneuvering acceleration, the higher the priority. When the number of detected targets reaches a preset number, a time-division multiplexing architecture is adopted to process targets of different priorities in batches within the time window set during the acquisition phase. The system calls upon the typical scene feature fingerprints pre-stored during the data acquisition phase to match the current scene parameters in real time. Once a match is successful, it directly calls upon the model parameters pre-optimized during the model building phase to correct the model.

5. The method for estimating the gain of an X-band antenna for low-altitude surveillance radar according to claim 4, characterized in that, The scenario verification phase refers to the basic gain model and the three-dimensional correction model in the gain model construction phase, which are used to establish a complex electromagnetic and dense target joint scenario, and record relevant test data according to the sampling frequency of the acquisition phase. For the temperature-loss correlation model in the gain model construction stage and the environmental adaptation logic in the correction model construction stage, a cross-scenario of temperature change and electromagnetic interference is constructed to verify environmental robustness. By introducing unknown scenario data and without inputting preset parameters, the synergistic effect of the calibration model in the parameter calibration and optimization stage and the scheduling mechanism in the adaptive scheduling stage is tested. After verification, the gain estimation deviation threshold is output.

6. The method for estimating the gain of an X-band antenna for low-altitude surveillance radar according to claim 5, characterized in that, The closed-loop process construction phase is based on the models, algorithms, and scheduling strategies verified in the preceding phases. It adopts a heterogeneous architecture of FPGA and MCU. The FPGA deploys the basic gain model in the gain model construction phase, the three-dimensional correction algorithm in the correction model construction phase, and the scheduling logic in the adaptive scheduling phase. The MCU stores the scene parameters in the acquisition phase and the calibration data in the parameter calibration and optimization phase for offline model management. The deviation threshold of edge nodes is verified according to the scenario verification stage. The real-time estimated gain is compared with the actual value. When the threshold is exceeded, local parameter fine-tuning is triggered. The cloud periodically aggregates multi-node deviation data, uses gradient descent to optimize the gain model, constructs a phase correction factor library, and pushes it to edge nodes.

7. A gain estimation system for an X-band antenna of a low-altitude surveillance radar, based on the method of claim 6, characterized in that, It includes a data acquisition module, a basic model building module, a dynamic correction module, a parameter calibration module, an adaptive scheduling module, a scenario verification module, and an engineering closed-loop module; The data acquisition module includes a scanning device, a geographic information system module, and a temperature sensor; it collects microscopic parameters of the X-band antenna and establishes a correlation model between the microscopic parameters and the mesoscopic environment; the geographic information system module classifies the macroscopic scene into complexity levels according to building density and vegetation coverage. The basic model building module is connected to the data acquisition module. Based on the acquired cross-scale parameters, a transition zone correction factor iterator is established, a frequency compensation term is added, and a temperature-loss correlation model is established by combining temperature sensor data and microwave anechoic chamber test curves. Multi-level loss correction is incorporated to form a near-field and far-field collaborative basic gain model. The dynamic correction module, connected to the basic model building module, collects environmental data, predicts environmental changes through an LSTM network, and generates a pre-corrected gain curve. It combines the target RCS, maneuvering acceleration, and target type to establish a three-dimensional adaptation matrix, divides the environmental complexity dimension according to the electromagnetic interference intensity, and forms a three-dimensional correction system. The parameter calibration module is connected to the dynamic correction module. It establishes a linear statistical regression model offline based on measured data and adapts it to multiple scenarios through transfer learning. It uses incremental learning to update parameters and clusters the gain estimation bias to trigger the optimization of the correction factor. The adaptive scheduling module, connected to the dynamic correction module and the data acquisition module, relies on the three-dimensional dynamic adaptation correction results and target priority division rules to divide the target priority according to the logic that the smaller the RCS and the greater the maneuvering acceleration, the higher the priority, and calls the typical scene feature fingerprint matching parameters. The scenario verification module is connected to the basic model building module, dynamic correction module, and parameter calibration module to establish complex electromagnetic and dense target joint scenarios and temperature change and electromagnetic interference cross scenarios. It introduces unknown scenario data to verify the basic gain model, correction model, and the collaborative effect of calibration and scheduling. The engineering closed-loop module connects with the data acquisition module, basic model construction module, dynamic correction module, parameter calibration module, adaptive scheduling module, and scenario verification module. It adopts a heterogeneous architecture of FPGA and MCU. Edge nodes compare and estimate the gain with the actual value in real time and make fine adjustments. The cloud aggregates data to optimize the correction factor library, forming a closed-loop optimization between the edge and the cloud.

Images