A Machine Learning-Based Intelligent Processing and Feature Extraction Method for Antenna Measurement Data

By constructing a standardized measurement data tensor and a machine learning model router, and dynamically matching the processing model, the problems of poor consistency and insufficient reusability in antenna measurement data processing are solved, achieving efficient and reliable feature extraction and result traceability.

CN123116157APending Publication Date: 2026-11-17WEIFANG UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611357886.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing antenna measurement data processing methods suffer from long processing cycles, poor result consistency, difficulty in adapting to feature extraction under multiple scenarios and multiple noise sources, and lack of high interpretability and result traceability.

Method used

By acquiring antenna measurement data, metadata, and calibration data, a standardized measurement data tensor is constructed and a data quality profile is generated. A machine learning model router is used to dynamically match and process the model, performing anomaly identification, noise suppression, missing data reconstruction, and error correction. Multidimensional features are extracted and the processing process is recorded.

Benefits of technology

It achieves efficient and reliable feature extraction in complex scenarios, improves the consistency and interpretability of results, and ensures the traceability and reproducibility of the processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116157A_ABST
    Figure CN123116157A_ABST
Patent Text Reader

Abstract

This invention discloses a machine learning-based intelligent processing and feature extraction method for antenna measurement data, specifically relating to the field of antenna testing and measurement data processing. The invention acquires amplitude, phase, complex field, frequency, angle, polarization, and state data from antenna measurements, maps them to a standardized tensor with unified coordinates, dimensions, and metadata, and generates a quality profile containing missing data, noise, drift, and anomalies. Based on the measurement scenario, data format, and quality profile, it routes and invokes anomaly identification, denoising, missing data reconstruction, and error correction models for collaborative processing. It then extracts frequency domain, spatial domain, radiation pattern, polarization, statistical, and deep representation features, fusing and filtering them through stability evaluation, correlation analysis, and task relevance. Finally, it outputs a structured feature set, confidence level, quality comparison, and traceable records.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of antenna testing and measurement data processing technology, and more specifically, to a machine learning-based intelligent processing and feature extraction method for antenna measurement data. Background Technology

[0002] Antenna measurements typically cover near-field, far-field, and compact-field test scenarios to obtain multi-dimensional core data such as amplitude, phase, and polarization. With the development of broadband, multi-frequency, multi-beam, array antennas, and multi-polarization antennas, a single measurement task may include a large number of frequency points, spatial sampling locations, polarization channels, and equipment status information, resulting in a significant increase in data volume.

[0003] Especially in higher operating frequency bands such as millimeter waves, as the operating wavelength shortens, the antenna measurement results become more sensitive to the probe position, the attitude of the antenna under test, the scanning trajectory, and the phase stability of the measurement system. Even small position deviations, attitude deviations, or equipment drift can result in significant amplitude and phase errors in the measurement data. For high-gain, narrow-beam antennas, the angle sampling accuracy, spatial positioning accuracy, and the integrity of the scanning area also have a significant impact on the calculation of parameters such as main lobe pointing, beamwidth, and sidelobe level, thereby further increasing the difficulty of intelligent processing of measurement data and reliable feature extraction.

[0004] Existing data processing workflows typically rely on manual processes for format conversion, outlier removal, smoothing filtering, missing point interpolation, and radiation pattern parameter calculation. Different measurement systems vary significantly in data fields, coordinate systems, sampling order, and calibration methods, requiring personnel to write scripts or adjust parameters for specific tasks. This results in long processing cycles, poor result consistency, and limited reusability.

[0005] Furthermore, antenna measurement data exhibits significant frequency correlation, spatial correlation, phase continuity, and polarization coupling characteristics. Simple thresholding or fixed filtering methods struggle to simultaneously identify isolated anomalies, regional drift, channel deviation, phase jumps, and environmental disturbances. Conventional interpolation methods are prone to disrupting main lobe, side lobes, nulls, and polarization characteristics when the missing region is large, the beam changes rapidly, or the phase gradient is high.

[0006] While some existing solutions employ machine learning models to process measurement data, they typically only describe how the model performs denoising or classification, without clearly outlining how the data is standardized, how the model is selected based on the measurement scenario, how the quality of the model output is constrained, or how the processed results are transformed into verifiable antenna features. This results in unclear model input / output, untraceable processing links, and insufficient cross-scenario adaptability. Therefore, given the limitations of existing single data processing algorithms in simultaneously handling complex scenarios such as near-field and far-field conditions, and the technical bottlenecks they face including data gaps and the potential for feature distortion due to multi-source noise, a highly interpretable intelligent antenna measurement data processing solution based on a dynamic data quality routing and matching model is needed. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a machine learning-based intelligent processing and feature extraction method for antenna measurement data to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] Acquire antenna measurement data, measurement metadata, and calibration data corresponding to the antenna measurement task to be processed. The antenna measurement data includes at least one of amplitude data, phase data, and complex field data. The measurement metadata includes at least frequency point information, spatial sampling location information, polarization information, measurement equipment status information, and measurement environment information.

[0010] Based on the measurement metadata, the antenna measurement data is subjected to format mapping, coordinate alignment, dimension unification and sampling sequence rearrangement to construct a standardized measurement data tensor, and a data quality profile is generated based on the standardized measurement data tensor;

[0011] Based on the measurement metadata and the data quality profile, the measurement scenario category, data form category, and data defect category are determined. The measurement scenario category, the data form category, data size, sampling density, quality labels generated by the data quality profile, and target analysis tasks are input into the model router. The model router outputs the model combination, model execution order, and corresponding model parameter configuration to select a target processing model that matches the measurement scenario category, the data form category, and the data defect category from a pre-established machine learning model library.

[0012] The target processing model is used to perform at least two of the following processes on the standardized measurement data tensor: outlier identification, noise suppression, missing data reconstruction, and measurement error correction, to obtain corrected measurement data;

[0013] Based on the corrected measurement data, a feature extraction data domain corresponding to the target analysis task is determined. When the corrected measurement data is near-field measurement data and the target analysis task includes far-field performance analysis, a near-field-to-far-field transformation is performed on the corrected measurement data to obtain far-field radiation pattern data. Based on the corrected measurement data and the far-field radiation pattern data, at least two types of features are extracted from frequency domain features, spatial domain features, radiation pattern features, polarization features, statistical features, and model representation features to form a candidate feature set.

[0014] Perform feature fusion, stability evaluation, and redundancy screening on the candidate feature set to obtain the target feature set;

[0015] Output the target feature set, the confidence level corresponding to each target feature, and a traceable processing record for recording the data source, model version, processing parameters, and quality changes of each processing step.

[0016] Preferably, the antenna measurement data is organized into data packets according to the measurement task. The data packets include the original measurement matrix, effective sampling mask, channel identifier, timestamp, coordinate system identifier, measurement device identifier, antenna attitude information, probe information, cable information, and calibration file index. The format mapping includes mapping the data fields output by different measurement devices to a unified field dictionary and saving the mapping relationship between the original fields and the unified fields.

[0017] Preferably, the generation of the data quality profile includes: identifying missing sampling points, duplicate sampling points, saturated sampling points, abrupt sampling points, and phase jump points; calculating the integrity index, noise index, drift index, continuity index, and channel consistency index for each frequency point, each spatial sampling position, each polarization channel, and each measurement channel; and generating quality labels, defect location masks, and processing priorities based on each index.

[0018] Preferably, the machine learning model library includes anomaly detection models, denoising models, missing data reconstruction models, error correction models, and feature encoding models.

[0019] Preferably, when the target processing model includes an anomaly identification model and a denoising model, the anomaly identification model learns a multidimensional data distribution based on normal measurement data and anomaly labels sampling points, local regions, or channels that deviate from the multidimensional data distribution; the denoising model takes a standardized measurement data tensor containing noise defects and the anomaly labels as input, and outputs denoised measurement data that retains the main lobe, side lobes, nulls, phase gradient, and polarization change characteristics.

[0020] Preferably, when the target processing model includes a missing data reconstruction model, the missing data reconstruction model generates reconstructed data based on the frequency correlation, spatial correlation, polarization correlation, and phase continuity around the missing location, and generates a reconstruction confidence score for each reconstruction location. When the reconstruction confidence score is lower than a preset threshold determined by the physical consistency check, the processing system sets the feature output channel of the corresponding spatial location to zero and isolates it to avoid polluting the global radiation pattern feature calculation. At the same time, it generates a supplementary measurement suggestion instruction containing specific frequency points and coordinate ranges and pushes it into the manual scheduling queue, and controls the workstation to suspend the processing of the current problem grid, and jumps to the next independent frequency point or independent polarization channel to continue to execute the feature extraction task, so as to ensure the non-blocking operation of the overall data processing link.

[0021] Preferably, when the target processing model includes an error correction model, the error correction model identifies and corrects systematic errors caused by probe response, cable loss, channel gain difference, phase zero drift, temperature change and positioning deviation based on calibration data, equipment status information and measurement environment information; the error correction model outputs the corrected measurement data and the estimation results of various errors.

[0022] Preferably, the candidate feature set includes at least the following: gain features, gain characteristics, gain variation features within the operating frequency band, half-power beamwidth features, main lobe pointing features, side lobe level features, front-to-back ratio features, null position features, phase center features, group delay features, axial ratio features, cross-polarization features, pattern symmetry features, spatial energy distribution features, and model characterization features for antenna type identification, production consistency discrimination, or fault mode identification.

[0023] Preferably, the feature fusion, stability evaluation, and redundancy screening include: normalizing candidate features of different dimensions; comparing the feature consistency of repeated measurement results or perturbation enhancement results for the same measurement task; deleting redundant features or reducing the output priority of redundant features based on the correlation between features, the contribution of features to the target analysis task, and feature confidence; and associating the retained candidate features according to frequency, spatial location, polarization mode, and task label to form the target feature set.

[0024] Preferably, the method further includes model training and updating steps: constructing training samples based on historical antenna measurement data, simulation data, standard component measurement data, and manually confirmed defect labels; performing coordinate perturbation, amplitude and phase perturbation, noise injection, sampling missing simulation, and environmental drift simulation on the training samples to generate enhanced samples; training the model in the machine learning model library using the training samples and the enhanced samples; after the target processing model is put into use, generating feedback samples based on the manual review results, supplementary test results, and subsequent task results, and incrementally updating the corresponding model when the update conditions are met, while retaining the model version and performance evaluation records before and after the update.

[0025] The technical effects and advantages of this invention are as follows:

[0026] This invention integrates raw data from different coordinate systems and sampling methods into a single access framework by unifying data packets and standardizing measurement tensors. Building upon this, it no longer relies on a single fixed algorithm but dynamically matches processing strategies based on data quality profiling and model routing mechanisms. Simultaneously, it introduces multi-dimensional correlation constraints to verify the output results, thereby suppressing side effects such as over-smoothing, main lobe shift, and side lobe distortion while improving the reliability of corrections in complex scenarios. After processing, in addition to conventional physical parameters, the system simultaneously extracts deep representational features and adds stability evaluations and confidence level annotations, providing richer quantitative references for performance evaluation and production consistency analysis. Throughout the entire process, data sources, model versions, and parameter adjustments are fully recorded, ensuring the reproducibility and auditability of the results. It also allows for closed-loop incremental updates triggered by manual review, creating a closed loop between engineering traceability and actual correction feedback. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0028] Figure 2 This is a schematic diagram of the data flow for intelligent processing and feature extraction in this invention. Detailed Implementation

[0029] 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.

[0030] Example 1

[0031] Please see Figure 1As shown, this invention provides a machine learning-based intelligent processing and feature extraction method for antenna measurement data. The method in this embodiment includes steps S100 to S600. Each step can be executed by an antenna measurement workstation, a data processing server, an edge computing device, or a cloud computing platform, or it can be executed collaboratively by multiple devices.

[0032] Step S100: Acquire antenna measurement data, measurement metadata, and calibration data. Antenna measurement data may include amplitude, phase, in-phase component, quadrature component, complex electric field, received power, gain, or pre-generated radiation pattern data. Measurement metadata includes measurement task identifier, antenna identifier, measurement time, measurement system identifier, measurement mode, coordinate system, frequency list, sampling angle, scanning position, polarization, probe attitude, antenna attitude under test, sampling interval, integration time, temperature, humidity, device self-test status, and operation log. Calibration data includes probe calibration files, cable loss curves, channel gain and phase calibration results, reference antenna measurement results, and device null data.

[0033] Step S200: Construct a standardized measurement data tensor and generate a data quality profile. The processing system reads data format descriptions from different data sources, maps the original fields to a unified field dictionary, performs coordinate transformation based on coordinate system identifiers and attitude information, unifies dimensions according to power units, electric field units, angle units, and frequency units, and sorts the sampled data according to frequency dimension, spatial dimension, polarization dimension, and channel dimension. The standardized measurement data tensor can adopt a multi-dimensional structure of "task × frequency × spatial location × polarization × channel," and stores a valid flag, source flag, and timestamp for each sampling unit.

[0034] Step S300 involves model routing, anomaly identification, denoising, missing data reconstruction, and error correction. The processing system first determines whether the current task exhibits high noise, local anomalies, continuous missing data, channel bias, phase drift, or environmental drift based on the data quality profile. Then, it selects the target processing model based on the measurement scenario and defect type. Different models can run serially or in parallel, and the model output is written back to correct the measurement data tensor after a physical consistency check.

[0035] Step S400: Perform performance feature extraction on the intelligently processed measurement data. The processing system first determines the data domain corresponding to feature extraction based on the measurement scenario and target analysis task. For tasks that have already obtained far-field measurement data, antenna performance features are directly extracted from the corrected far-field data; for near-field measurement data, when the target analysis task includes far-field performance analysis, near-field-far-field transformation is performed on the near-field measurement data after anomaly identification, denoising, missing data reconstruction, and error correction to obtain the corresponding far-field radiation pattern data.

[0036] Based on the corrected measurement data and the far-field pattern data obtained from near-field and far-field transformation, performance characteristics with clear antenna physical significance, such as gain, gain variation with frequency, main lobe peak value, main lobe pointing, half-power beamwidth, sidelobe level, front-to-back ratio, null position, axial ratio, cross-polarization level, and pattern symmetry, are extracted.

[0037] For near-field measurement data, near-field features such as scan integrity, local amplitude and phase distribution, spatial energy distribution, and phase gradient can be directly extracted; at the same time, feature encoding models can be used to extract model characterization features that reflect the overall distribution pattern of measurement data from different antennas.

[0038] Step S500 involves feature fusion, stability evaluation, and redundancy screening. The processing system normalizes the dimensions of candidate features, compares the consistency of features from repeated measurements, data enhanced by different perturbations, or adjacent frequency points, and ranks the features according to the target task. Features that are highly relevant but contribute little to the target analysis task can be merged or deleted, while low-confidence features can be retained but have their output priority reduced.

[0039] Step S600: Output structured features, confidence scores, and traceable processing records. Structured features can be saved by task, frequency, spatial region, polarization mode, and feature category. Traceable processing records include at least the following: input data summary, data quality profile, model routing results, model version, model parameters, outlier locations, reconstruction locations, correction amounts, processing time, and quality changes before and after processing.

[0040] Example 2

[0041] In one specific implementation, the processing system creates a task-level data packet for each measurement task. The task-level data packet consists of a raw data area, a metadata area, a calibration data index area, a quality marker area, and a processing record area. The raw data area only stores the original sampled values ​​and the original file index, and is not overwritten during subsequent processing, thus ensuring the traceability of the original data.

[0042] For planar near-field scan data, the spatial position dimension can be represented by both horizontal and vertical coordinates; for cylindrical near-field data, the spatial position dimension can be represented by axial position and azimuth angle; for spherical near-field or far-field data, the spatial position dimension can be represented by azimuth angle and elevation angle. The processing system converts various position descriptions into a unified coordinate object while retaining the original coordinate values.

[0043] The generation of the quality profile consists of two parts: rule detection and learning detection. Rule detection is used to detect missing files, abnormal fields, duplicate coordinates, unit conflicts, out-of-range errors, illegal phase values, and device status errors; learning detection is used to detect data distribution shifts, local texture anomalies, cross-channel inconsistencies, and slow drift. The results of both are combined into a quality label.

[0044] The quality score of a single sampling unit can be represented as a weighted combination of integrity, anomaly level, noise level, drift level, and channel consistency. For example:

[0045]

[0046] Where Q is the quality score, I is the integrity index, A is the anomaly level, N is the noise level, D is the drift level, and C is the channel consistency index. to The weights are determined based on the measurement task.

[0047] Example 3

[0048] The model library can include unsupervised anomaly detection models, convolutional denoising autoencoders, time-frequency joint denoising networks, conditional generative reconstruction models, graph neural network reconstruction models, regression correction models, feature encoding models, and ensemble discriminant models. Each model has applicable scenario labels, input dimension requirements, sampling density requirements, supported defect types, training data versions, and performance metrics.

[0049] The inputs to the model router include measurement method, data dimension, number of frequency points, spatial sampling pattern, number of polarizations, number of channels, data size, quality label, missing proportion, outlier proportion, noise level, and target analysis task. The model router can be implemented using decision trees, gradient boosting models, or lightweight neural networks. The model router outputs model combination, execution order, window size, number of iterations, confidence threshold, and physical consistency check rules. The model router is implemented using a decision tree algorithm. Its core branch rules include: when the input feature "quality label - noise level" > 0.5 and the missing proportion < 5%, the router outputs "Model combination: time-frequency joint denoising network; Execution order: single-step execution; Parameter configuration: smooth window". When the missing proportion When the "Measurement Scene Category" is "Cylinder Near Field", the router outputs the following model combination: a denoising model plus a graph neural network to reconstruct the model; the execution order is: denoising first, then reconstruction; the parameter configuration is: reconstruction confidence threshold. .

[0050] For example, for tasks with a few isolated abrupt changes in the far-field single-frequency radiation pattern, the model router can choose between an anomaly identification model and a local reconstruction model; for tasks with random noise and several missing scan lines in the broadband planar near-field data, the model router can first call the denoising model, and then call the reconstruction model that utilizes frequency correlation and spatial correlation; for tasks with overall phase drift under different temperature conditions, the model router can prioritize calling the error correction model.

[0051] To avoid the processing model producing results that do not conform to the antenna's physical characteristics, the model output needs to undergo a physical consistency check. The physical consistency check includes the following quantitative judgment rules:

[0052] Phase gradient rationality: Calculate the phase difference between adjacent spatial sampling points. The system calculates the second-order phase difference. To determine whether phase winding has occurred: if detected And local second-order phase difference Less than the preset phase change threshold (e.g.) If the trend is smooth and indeed caused by winding, then phase unwinding recovery is performed; if and (Indicating a genuine abnormal jump rather than tangling), the physical consistency check of the sampling point is deemed to have failed; where, phase tangling refers to the phenomenon where the measured phase jumps periodically when the actual continuous phase crosses the boundary of the preset phase representation interval, since the phase output by the measurement system is usually limited to the preset phase representation interval; phase detangling is a process of identifying the periodic jumps and compensating for the corresponding integer periodic phase amount to restore the phase change to continuity;

[0053] Range of main lobe position variation: Compare the deviation of the pointing angle of the main lobe peak in the radiation pattern before and after reconstruction. ,like If so, it is determined to be unsuccessful;

[0054] Frequency response continuity: Calculate the absolute value of the difference in gain between adjacent frequency points. If it exceeds 3dB, it is judged as failing.

[0055] If the check fails, the processing system lowers the model output confidence level, reverts to more conservative processing parameters, or marks the corresponding area as requiring manual review.

[0056] Example 4

[0057] Anomaly detection can be performed at the sampling point level, local window level, scan line level, and channel level. The processing system segments the standardized measurement data tensor into a sliding window and constructs input features such as amplitude, phase gradient, local spectrum, neighborhood difference, cross-frequency consistency, and cross-polarity consistency. The unsupervised anomaly detection model learns the normal data distribution using normal samples and outputs anomaly scores.

[0058] When identifying anomalies in phase data, the processing system prioritizes calculations in the phase expansion or complex domain to avoid misjudgments caused by phase jumps at positive and negative boundaries. For in-phase and quadrature components, a complex representation can be constructed first, followed by calculation of the complex neighborhood difference and phase gradient.

[0059] Denoising models can employ one-dimensional convolutional networks to process frequency sequences, two-dimensional convolutional networks to process planar scan data, and spherical convolutional or graph neural networks to process irregular angle sampling data. During training, high-quality measurement data is used as target samples, and noisy samples are generated by adding white noise, impulse noise, correlated noise, baseline drift, and channel bias.

[0060] The denoising loss not only constrains the numerical difference between the output and the target sample, but also constrains the main lobe peak, side lobe region energy, phase gradient, and polarization relationship. During actual inference, the denoising model simultaneously receives a data quality mask, enabling the model to focus on processing low-quality regions while preserving as much of the original value as possible in high-quality regions.

[0061] For abnormal sampling points identified by the model, the processing system does not delete them directly. Instead, it first records the abnormality type, abnormality score, and neighborhood status, and then decides whether to use neighborhood replacement, local model reconstruction, or retain the original value and add risk markers based on the abnormality range.

[0062] Example 5

[0063] The missing data reconstruction model takes a missing mask, valid sampled data, frequency coordinates, spatial coordinates, polarization identifiers, and device status as input. For isolated missing points, the model can reconstruct the data using local neighborhoods; for continuous missing regions, the model simultaneously utilizes information from adjacent frequency points, adjacent spatial regions, and other polarization channels for conditional reconstruction.

[0064] In planar near-field scenes, the reconstruction model can treat the scanned grid as a two-dimensional image, with the real and imaginary parts serving as input channels respectively; in irregular scanned scenes, the sampling locations are constructed as a graph structure. The initial feature vector of node V contains the magnitude and phase values ​​of that point. The weights of edge E are calculated based on the spatial Euclidean distance and frequency difference. The graph neural network layer uses a graph convolutional network, and its feature update formula is:

[0065]

[0066] in, This is represented as an adjacency matrix with self-loops added. Represented as a degree matrix, For the first The learnable weight matrix of the layer, It is a ReLU activation function; For the first The node feature matrix of the layer, its initial state It includes two-dimensional features of the amplitude and phase values ​​of each sampling node;

[0067] The adjacency matrix The edge in weight The Gaussian kernel function is used to calculate the result based on the spatial Euclidean distance and frequency difference. The specific calculation formula is as follows:

[0068]

[0069] in , They are nodes and spatial coordinates, , These are the corresponding test frequencies. and These are the scaling constants for space and frequency, respectively.

[0070] The model is trained end-to-end by minimizing the mean squared error (MSE) loss function at known sampling points.

[0071] Each reconstruction location outputs a reconstruction value and a reconstruction confidence score. The reconstruction confidence score can be determined based on model uncertainty, the effective sampling ratio of the neighborhood, the differences in outputs from multiple models, and the results of physical consistency checks. For locations with low confidence, the processing system does not use them as normal measured values, but instead reduces their weight in the target feature calculation and generates supplementary testing suggestions.

[0072] The error correction model is used to handle probe response error, cable loss, channel gain difference, phase zero drift, temperature drift, positioning deviation, and turntable hysteresis. The processing system reads calibration data and environmental data, corrects directly calculable errors according to deterministic calibration rules, and then estimates residual errors using a machine learning regression model.

[0073] The training samples for the calibration model can consist of a standard antenna, a reference device, and repeated measurement data under different environmental conditions. Model inputs include frequency, channel, device status, temperature, humidity, scan direction, and historical calibration values. Outputs are amplitude corrections, phase corrections, or coordinate corrections. To ensure interpretability, the processing system stores both deterministic calibration values ​​and residual calibration values ​​estimated by the model.

[0074] Example 6

[0075] like Figure 2 As shown, the corrected measurement data enters the multi-domain feature extraction layer. Frequency domain features can include the mean gain within the operating frequency band, gain fluctuation, resonant point location, bandwidth, phase slope, group delay, frequency response smoothness, and correlation between adjacent frequency points.

[0076] Spatial domain and pattern characteristics can include main lobe peak value, main lobe pointing, half-power beamwidth, first null position, maximum sidelobe level, integrated sidelobe level, front-to-back ratio, pattern symmetry, beam skew, null depth, null position, and regional energy distribution. For two-dimensional or three-dimensional patterns, beam solid angle, curvature near the peak direction, and angular spacing between multiple beams can also be calculated.

[0077] Polarization characteristics can include axial ratio, cross-polarization level, polarization isolation, polarization elliptic parameters, main polarization direction, polarization variation with frequency, and correlation between different polarization channels.

[0078] Statistical features may include mean, variance, skewness, kurtosis, local gradient distribution, spectral entropy, spatial entropy, proportion of outlier regions, and proportion of reconstructed regions. Model representation features are automatically extracted from corrected measurement data by the feature encoding model and can be used for antenna model identification, production consistency determination, fault mode identification, or measurement task clustering.

[0079] To ensure the interpretability of features, the processing system establishes a correlation between the model's representation features and the original frequency points, spatial regions, and polarization channels. Attention weights, feature sensitivity analysis, or local occlusion tests can be used to determine which data regions the model's representation features primarily originate from, and the correlation results are written into a traceable processing log.

[0080] Furthermore, the processing system can generate a structured performance description of the antenna under test based on the aforementioned physical performance characteristics and model characterization characteristics. The structured performance description includes at least one of the following: gain level and its variation trend within the operating frequency band, beamwidth and its variation trend with frequency, main lobe pointing and stability, side lobe level, front-to-back ratio, null characteristics, and polarization performance.

[0081] When the target analysis task is antenna type identification, the physical performance characteristics and model characterization characteristics are input into the pre-trained antenna type identification model, and the preset antenna category to which the antenna under test belongs and the corresponding category confidence level are output according to the feature matching relationship between the model and the training samples of the preset antenna category.

[0082] When the target analysis task is to describe performance characteristics, the extracted gain, beamwidth, main lobe pointing, side lobe level and polarization parameters are judged in intervals or analyzed in trends according to the pre-set performance evaluation rules, and structured labels are generated to describe the performance characteristics of the antenna under test.

[0083] Example 7

[0084] Candidate features are first normalized according to their feature type. For features with definite physical units, both the original and normalized values ​​are retained; for features representing the model, only the normalized values, encoding dimensions, and corresponding model versions are saved.

[0085] Stability evaluation can utilize repeated measurements of the same task, or it can apply small noise perturbations, coordinate perturbations, frequency sampling, or sampling point occlusion to the current data, comparing the changes in features before and after the perturbation. Features that change little and remain consistent in repeated measurements are assigned higher stability scores.

[0086] Task contribution can be determined through feature importance, ablation experiments, changes in classification accuracy, changes in regression error, or expert rules. Correlation analysis is used to identify highly correlated feature groups, and the processing system can retain features that are stable, have clear physical meaning, or are computationally inexpensive.

[0087] Feature confidence The following calculations were performed using the joint mapping between the exponential decay model and the data quality profile:

[0088]

[0089] Among them, attenuation coefficient Based on the thermal noise floor calibration of the current measurement system (value range at room temperature is...), This nonlinear mapping can effectively suppress (Quality Score) lower than To prevent the confidence level from decaying too quickly, the system avoids the accidental deletion of valid features. When the confidence level of a feature falls below the output threshold, the processing system either does not output the feature or marks it as a reference value.

[0090] Example 8

[0091] The training samples consist of historical measured data, standard part data, simulation data, and manually labeled data. Historical measured data is used to reflect real equipment noise and environmental changes; standard part data is used to establish performance benchmarks; simulation data is used to supplement difficult-to-obtain fault samples and missing samples; and manually labeled data is used to provide labels for anomaly types, defect ranges, and processing results.

[0092] Before training, the processing system divides the data into training, validation, and test sets according to antenna type, measurement method, frequency band, sampling density, and data source to avoid adjacent data from the same measurement task entering the training and test sets simultaneously. For broadband tasks, stratified sampling by frequency band can also be used to ensure coverage of different frequency bands.

[0093] Data augmentation includes amplitude perturbation, phase perturbation, noise injection, random missing data, continuous region missing data, scan line missing data, coordinate offset, channel gain bias, and temperature drift simulation. Augmentation parameters are determined based on the statistical distribution of historical quality profiles to make the augmented samples closely approximate actual defects.

[0094] Model evaluation metrics include anomaly detection accuracy, false negative rate, false positive rate, signal-to-noise ratio change before and after denoising, reconstruction error, main lobe position deviation, side lobe level deviation, phase continuity, feature stability, and processing time. Only when the model's metrics on the test set and standard component data meet the preset requirements will it be added to the official model library.

[0095] During the online usage phase, the system collects manual review results, supplementary test results, and subsequent performance analysis results to form feedback samples. Incremental training is triggered when the number of feedback samples reaches a threshold, the data distribution changes significantly, or the model performance continues to decline. Regression testing is performed before the new model is deployed, and the old model is retained to support result reproduction and version rollback.

[0096] Example 9

[0097] In a far-field measurement example in the X-band (8-12 GHz), the measurement task included five test frequencies, horizontal polarization, and vertical polarization patterns, including 10 GHz. Impulse interference was present at some angular locations in the raw data, with the phase data at 10.5 GHz showing an overall shift (approximately 30 degrees), and two adjacent angular sampling points missing.

[0098] The processing system first aligns the angle sequences of different frequency points and generates a missing mask. The quality profile identifies pulse interference regions, overall phase shift, and missing locations. The model router selects an anomaly detection model, a local reconstruction model, and a phase drift correction model.

[0099] The anomaly identification model combines amplitude abrupt changes, phase gradients, and consistency of adjacent frequency points to mark pulse interference points; the local reconstruction model uses adjacent angle and adjacent frequency point data to recover missing positions; and the phase drift correction model corrects the overall offset based on the reference channel and historical calibration results.

[0100] Subsequently, the main lobe peak value, beamwidth, maximum sidelobe level, front-to-back ratio, null location, and cross-polarization level are extracted. The processing system generates a confidence score for each feature based on the correction region ratio, model confidence, and repeatability consistency, and outputs the anomaly location and correction process.

[0101] Example 10

[0102] In a planar near-field measurement example, the measurement data consists of complex near-field scan matrices at multiple frequency points (sampling interval is...). Two scan lines showed continuous gaps, and some areas were affected by phase noise caused by equipment vibration. The reconstruction confidence threshold was set to 0.85.

[0103] The processing system constructs data channels from the real and imaginary parts, and forms a three-dimensional data tensor based on frequency, lateral position, and vertical position. Quality profiling identifies regions with continuous missing data and high phase noise. The model router selects a two-dimensional convolutional denoising model and a frequency-space joint reconstruction model.

[0104] The denoising model uses local spatial texture and adjacent frequency information to suppress phase noise, while the reconstruction model uses data from both sides of the missing scan line and adjacent frequency data to reconstruct the missing region. The model output undergoes a near-field and far-field transformation consistency check, comparing the changes in far-field main lobe position, side lobe level, and energy before and after processing.

[0105] When the physical consistency check of the reconstructed region passes, the system continues to extract near-field amplitude and phase distribution features and far-field radiation pattern features; when the check fails, the system retains the original missing markers and generates scan line retest suggestions.

[0106] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent processing and feature extraction of antenna measurement data based on machine learning, characterized in that, include: Acquire antenna measurement data, measurement metadata, and calibration data corresponding to the antenna measurement task to be processed. The antenna measurement data includes at least one of amplitude data, phase data, and complex field data. The measurement metadata includes at least frequency point information, spatial sampling location information, polarization information, measurement equipment status information, and measurement environment information. Based on the measurement metadata, the antenna measurement data is subjected to format mapping, coordinate alignment, dimension unification and sampling sequence rearrangement to construct a standardized measurement data tensor, and a data quality profile is generated based on the standardized measurement data tensor; Based on the measurement metadata and the data quality profile, the measurement scenario category, data form category, and data defect category are determined. The measurement scenario category, the data form category, data size, sampling density, quality labels generated by the data quality profile, and target analysis tasks are input into the model router. The model router outputs the model combination, model execution order, and corresponding model parameter configuration to select a target processing model that matches the measurement scenario category, the data form category, and the data defect category from a pre-established machine learning model library. The target processing model is used to perform at least two of the following processes on the standardized measurement data tensor: outlier identification, noise suppression, missing data reconstruction, and measurement error correction, to obtain corrected measurement data; Based on the corrected measurement data, a feature extraction data domain corresponding to the target analysis task is determined. When the corrected measurement data is near-field measurement data and the target analysis task includes far-field performance analysis, a near-field-to-far-field transformation is performed on the corrected measurement data to obtain far-field radiation pattern data. Based on the corrected measurement data and the far-field radiation pattern data, at least two types of features are extracted from frequency domain features, spatial domain features, radiation pattern features, polarization features, statistical features, and model representation features to form a candidate feature set. Perform feature fusion, stability evaluation, and redundancy screening on the candidate feature set to obtain the target feature set; Output the target feature set, the confidence level corresponding to each target feature, and a traceable processing record for recording the data source, model version, processing parameters, and quality changes of each processing step. 2.The method of claim 1, wherein: The antenna measurement data is organized into data packets according to the measurement task. The data packets include the original measurement matrix, effective sampling mask, channel identifier, timestamp, coordinate system identifier, measurement device identifier, antenna attitude information, probe information, cable information, and calibration file index. The format mapping includes mapping the data fields output by different measurement devices to a unified field dictionary and saving the mapping relationship between the original fields and the unified fields.

3. The method for intelligent processing and feature extraction of antenna measurement data based on machine learning according to claim 1, characterized in that: The generation of the data quality profile includes: identifying missing sampling points, duplicate sampling points, saturated sampling points, abrupt sampling points, and phase jump points; calculating the integrity index, noise index, drift index, continuity index, and channel consistency index for each frequency point, each spatial sampling position, each polarization channel, and each measurement channel; and generating quality labels, defect location masks, and processing priorities based on each index.

4. The method for intelligent processing and feature extraction of antenna measurement data based on machine learning according to claim 1, characterized in that: The machine learning model library includes anomaly detection models, denoising models, missing data reconstruction models, error correction models, and feature encoding models.

5. The method for intelligent processing and feature extraction of antenna measurement data based on machine learning according to claim 4, characterized in that: When the target processing model includes an anomaly detection model and a denoising model, the anomaly detection model learns a multidimensional data distribution based on normal measurement data and anomaly labels sampling points, local regions or channels that deviate from the multidimensional data distribution. The denoising model takes a standardized measurement data tensor containing noise defects and the anomaly label as input, and outputs denoised measurement data that retains the main lobe, side lobes, nulls, phase gradient, and polarization change characteristics.

6. The method for intelligent processing and feature extraction of antenna measurement data based on machine learning according to claim 4, characterized in that: When the target processing model includes a missing data reconstruction model, the missing data reconstruction model generates reconstructed data based on the frequency correlation, spatial correlation, polarization correlation, and phase continuity around the missing location, and generates a reconstruction confidence score for each reconstruction location. When the reconstruction confidence score is lower than a preset threshold determined by the physical consistency check, the processing system sets the feature output channel of the corresponding spatial location to zero and isolates it to avoid polluting the global radiation pattern feature calculation. At the same time, it generates a supplementary measurement suggestion instruction containing specific frequency points and coordinate ranges and pushes it into the manual scheduling queue, and controls the workstation to suspend the processing of the current problem grid and step to the next independent frequency point or independent polarization channel to continue the feature extraction task.

7. The method for intelligent processing and feature extraction of antenna measurement data based on machine learning according to claim 4, characterized in that: When the target processing model includes an error correction model, the error correction model identifies and corrects systematic errors caused by probe response, cable loss, channel gain difference, phase zero drift, temperature change and positioning deviation based on calibration data, equipment status information and measurement environment information; the error correction model outputs the corrected measurement data and the estimation results of various errors.

8. The method for intelligent processing and feature extraction of antenna measurement data based on machine learning according to claim 1, characterized in that: The candidate feature set includes at least the following: gain features, gain characteristics, gain variation features within the operating frequency band, half-power beamwidth features, main lobe pointing features, side lobe level features, front-to-back ratio features, null position features, phase center features, group delay features, axial ratio features, cross-polarization features, pattern symmetry features, spatial energy distribution features, and model characterization features used for antenna type identification, production consistency discrimination, or fault mode identification.

9. The method for intelligent processing and feature extraction of antenna measurement data based on machine learning according to claim 1, characterized in that: The feature fusion, stability evaluation, and redundancy screening include: normalizing candidate features of different dimensions; comparing the feature consistency of repeated measurement results or perturbation enhancement results for the same measurement task; deleting redundant features or reducing the output priority of redundant features based on the correlation between features, the contribution of features to the target analysis task, and feature confidence; and associating the retained candidate features by frequency, spatial location, polarization mode, and task label to form the target feature set.

10. The method for intelligent processing and feature extraction of antenna measurement data based on machine learning according to claim 1, characterized in that: It also includes model training and update steps: constructing training samples based on historical antenna measurement data, simulation data, standard component measurement data, and manually confirmed defect labels; Coordinate perturbation, amplitude and phase perturbation, noise injection, sampling missing simulation, and environment drift simulation are performed on the training samples to generate enhanced samples; The model in the machine learning model library is trained using the training samples and the augmented samples; after the target processing model is put into use, feedback samples are generated based on the results of manual review, supplementary testing and subsequent tasks, and the corresponding model is incrementally updated when the update conditions are met, while retaining the model version and performance evaluation records before and after the update.