A real-time wind speed and direction sensing method using adaptive beamforming

Abstract

The application discloses a kind of real-time wind speed and direction sensing method using adaptive beamforming, belong to wind field sensing technical field, the sensing method specific steps are as follows:Ⅰ, initialization adaptive beamforming array, auxiliary environment sensing sensor and computing unit, and the point cloud data of monitoring area is collected in real time, constructs environment semantic map;Ⅱ, based on environment semantic map, generate initial beam weight and pointing strategy, based on the generation result, real-time adjustment each array element weight optimizes echo signal quality;The application significantly improves spatial integrity and time continuity, can actively avoid shelter and interference, key coverage key area, converge faster, higher stability, reduce the influence of abnormal data on overall wind field estimation, significantly improve reliability in complex scene, improve sensing efficiency and effective coverage rate.

Description

Technical Field

[0001] This invention relates to the field of wind field sensing technology, and in particular to a real-time wind speed and direction sensing method using adaptive beamforming. Background Technology

[0002] Wind speed and direction are core physical quantities describing atmospheric motion, and their high spatiotemporal resolution sensing is crucial for engineering safety, environmental monitoring, and the operation of intelligent systems. In complex urban terrain, near-ground boundary layers, low-altitude airspace, and industrial settings, wind fields often exhibit strong non-uniformity, rapid time-varying characteristics, and significant three-dimensional structural features. Traditional wind measurement methods relying on single-point sensors or sparse deployments struggle to simultaneously meet the requirements of real-time performance, continuity, and spatial coverage. In recent years, the development of array sensing technology has led to the gradual maturation of beamforming-based remote sensing wind measurement methods. However, existing methods largely rely on fixed or semi-adaptive beam configurations, primarily adjusting weights based on signal statistical characteristics, lacking effective perception and utilization of wind field disturbances caused by factors such as buildings, vegetation, and moving targets in complex environments. Furthermore, traditional methods typically output discrete measurement point results, making it difficult to directly form a continuous and interpretable three-dimensional wind field description, and limiting their ability to identify and track microscale wind field events. Against this backdrop, inventing a real-time wind speed and direction sensing method utilizing adaptive beamforming becomes particularly important.

[0003] Existing real-time wind speed and direction sensing methods have low spatial integrity and temporal continuity, and cannot actively avoid obstructions and interference, resulting in the inability to focus on key areas. Furthermore, abnormal data has a significant impact on the overall wind field estimation, reducing reliability in complex scenarios and leading to low sensing efficiency and effective coverage. To address this, we propose a real-time wind speed and direction sensing method using adaptive beamforming. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a real-time wind speed and direction sensing method using adaptive beamforming.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A real-time wind speed and direction sensing method utilizing adaptive beamforming, the specific steps of which are as follows:

[0007] Ⅰ. Initialize the adaptive beamforming array, auxiliary environmental perception sensors, and computing units, and collect point cloud data of the monitoring area in real time to construct an environmental semantic map;

[0008] II. Based on the environmental semantic map, generate initial beam weights and pointing strategies, and adjust the weights of each array element in real time based on the generation results to optimize the echo signal quality.

[0009] Ⅲ. Continuously acquire observation information from different spatial locations to form time-series sparse wind field observation data, and reconstruct the three-dimensional wind speed and direction distribution of the entire monitoring area in real time;

[0010] IV. Based on the continuous wind field reconstruction results and the original beamforming output, calculate the joint variation characteristics of signal features in the spatial and temporal dimensions, and identify wind field events;

[0011] V. Based on the current perceived wind field status, perform ultra-short-term wind field evolution simulation, and plan the beam working mode within the future time window based on the simulation results;

[0012] VI. Each node in the sensing network independently executes the real-time wind speed and direction sensing process locally, and periodically uploads model parameter update information, while also integrating the updated parameters.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] This real-time wind speed and direction sensing method utilizes adaptive beamforming to set camera and LiDAR sampling and triggering strategies according to the monitoring scale, automatically switching between static detail and dynamic capture modes. It performs real-time exposure, gain correction, point cloud denoising, and frame quality scores. Subsequently, it generates a geometric baseline based on coarse and fine registration and ground fitting, followed by multimodal semantic segmentation, instantiation, and short-term tracking, and integrates the semantic point cloud into a hierarchical voxel map. Local semantic and geometric features are mapped to beam pointing scores at multiple scales. After balancing coverage, interference, and power constraints, the scores are solved and quantized into hardware control commands. These commands are then sent to the array after simulation verification. Real-time beamforming sampling and adaptive weight updates are performed on the array to generate temporal sparse observations. These sparse observations are then input into a coordinate-type implicit network according to spatiotemporal coordinates for small-batch fine-tuning, and a continuous three-dimensional wind speed and direction field is output. Finally, the spatiotemporal joint discreteness is calculated on the grid points, high-confidence candidates are screened and clustered, the location is instantiated and the trajectory is tracked by data association, and finally event reports and alarms are generated. This significantly improves spatial integrity and temporal continuity, can actively avoid occlusion and interference, focus on covering key areas, converge faster and have higher stability, reduce the impact of abnormal data on the overall wind field estimation, significantly improve the reliability in complex scenarios, and improve perception efficiency and effective coverage. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a flowchart of a real-time wind speed and direction sensing method using adaptive beamforming proposed in this invention. Detailed Implementation

[0017] Example 1, referring to Figure 1 A real-time wind speed and direction sensing method utilizing adaptive beamforming, the specific steps of which are as follows:

[0018] Initialize the adaptive beamforming array, auxiliary environmental perception sensors, and computing units, and collect point cloud data of the monitoring area in real time to construct an environmental semantic map.

[0019] Specifically, the physical model, serial number, and factory parameter table of each adaptive beamforming array, auxiliary environmental sensing sensor, and computing unit are checked one by one. The operating voltage range and rated current of each device are recorded. A regulated power supply is input to each power supply circuit, followed by a no-load test. Then, a channel-by-channel current rise test is performed with the unit load connected to observe for abnormal voltage drops or overcurrent protection triggering. The system is run for 5-10 minutes under preset operating conditions, and the power supply stability and temperature rise curves are recorded. If the temperature or voltage exceeds the preset allowable values, the faulty device is marked and isolated. The array units are placed in the calibration reference holes on the ground or rack according to the design drawings. Tightening torque and positional deviations are recorded as required. The installation status of each array element is then photographed and recorded. The rangefinder or laser rangefinder measures the three-dimensional coordinates of each array element relative to the reference origin and writes them to the local calibration file. The measured actual coordinates are then compared with the system design coordinates using differential calculation. If the maximum deviation exceeds the allowable threshold, mechanical fine-tuning is initiated. The corrected array element coordinates are then written to the array controller's geometric parameter table and a backup file is generated. Two time synchronization schemes are enabled between the edge computing and sensor nodes: hardware PPS signal and software-based bidirectional timestamp exchange. The control terminal then sends a timestamp request to the sensor node and records the sending and receiving times. The node returns a response and records its local timestamp. Simultaneously, the node clock offset and round-trip delay are calculated based on the interaction results, and the node clock offset is then loaded into the node's local clock adjustment register. The interaction is repeated until the deviation falls within the expected threshold, and the node clock deviation is continuously monitored. If it exceeds the expected threshold, resynchronization is triggered. The calibration transmitter transmits a standard reference signal at a known location. Each array element receives the reference signal multiple times in a static environment and records the amplitude and phase. The ratio or difference between the received amplitude and phase of each array element and the reference value is calculated, and a correction coefficient table for each channel is established. Then, the amplitude compensation and phase compensation registers for each channel are configured in the array controller. At the same time, the reference signal is repeatedly injected, and the consistency of the response of each channel after compensation is checked. Sample data of each receiving channel is sampled for a preset time period by noise sampling. The noise variance and power spectral density estimate of each channel are calculated to determine whether there is abnormal interference. If abnormal interference is found... For frequent interference, use filters or frequency re-fixing strategies to adjust the RF front-end, record the lowest detectable signal level for each channel, and write each noise statistic into a local diagnostic table. Place known geometric calibration boards or markers near the array, and collect data from the camera or LiDAR at multiple angles and distances. Use the calibration data to calculate the intrinsic parameter matrix and distortion coefficients of the camera and laser ranging. For data not used in the calculation, perform backprojection and reprojection error calculations. If the RMS reprojection error is less than the threshold, write it into the intrinsic parameters; otherwise, re-acquire data with diverse angles and distances. Then, write the intrinsic parameters into the sensor driver layer and start the real-time distortion correction pipeline. Simultaneously trigger array reception and LiDAR and camera recording, acquiring a set of simultaneous observations within the same short time window.Based on the initial pose guess of the sensor to the array obtained from on-site measurements, the rigid body transformation is optimized using the least squares method. The optimized extrinsic parameter matrix is ​​then written into the data fusion module. Subsequent sensor data is automatically transformed to the array's unified coordinate system before fusion. A virtual environment is started on the edge computing node, confirming that the required dependent libraries, GPU and TPU drivers, and hardware resources are available and in normal condition. The pre-trained neural implicit wind field model, environment-beam mapping model, and local baseline weight file are loaded into memory from secure storage, along with calibration and noise statistics tables. Checksum hash values ​​are calculated for the loaded model weight file and executable file, and compared with known signatures. If the checksum fails, execution is refused and an alarm is triggered. The public / private key pair, device certificate, and root certificate of the central aggregator required by the edge node are installed into the node's secure storage via a secure channel. Finally, local access control policies are enabled on the node.

[0020] Specifically, based on the scale of the monitoring area and the minimum resolvable scale of the target, the camera resolution, exposure time, and frame rate are set, along with the LiDAR rotation rate and single-point repetition rate. Then, two operating modes are initiated: static detail scanning and dynamic capture. During operation, these modes automatically switch based on preset external trigger conditions. Trigger priorities and time windows are configured for each sensor in the edge controller, along with a local caching strategy. After each cycle, the instantaneous sampling quality index is calculated and recorded. If it falls below a preset threshold, the exposure and rate are automatically adjusted, or the cycle is labeled as "low confidence" data. Black level and gain corrections and linearization checks are performed on the camera images. Each camera image is divided into multiple blocks, and the saturation pixel ratio is statistically analyzed to detect... Image local exposure is performed, and image regions with saturation and pixel ratios below a preset threshold are marked as low-weight. Unbelievable points in the LiDAR point cloud with return intensities below a preset threshold or exceeding a preset range are filtered out. Statistical neighborhood analysis is then performed on each point to identify isolated noise points. Secondary filtering is applied to the identified isolated noise points based on normal and curvature. Non-local mean filtering is then applied to each camera image. The corresponding quality score is calculated and recorded for each sample frame. The initial pose relationship between sample frames is obtained based on on-site ranging. Stable geometric features of each point cloud frame are then extracted, and candidate corresponding point pairs are constructed. Iterative optimization based on least squares is used with rigid body transformation parameters as unknowns. Fine registration of each candidate corresponding point pair is repeated until the residual converges. Within a preset range, after registration, the geometric consistency between multiple frames is checked. If a systematic deviation is found, the process reverts to coarse registration or re-acquires sample frames. Based on height stratification or normal consistency, a set of ground candidate points is extracted from the registered point cloud data. The local covariance matrix of the ground candidate point set is calculated, and the normal of the local ground surface is obtained through principal component analysis. The monitoring area is divided into multiple blocks, and one or more local terrain baselines are fitted in each block. The ground height confidence and fitting error of each block are recorded, and blocks with confidence scores below a preset threshold are marked as unresolved areas. Normalization, scale resampling, and data augmentation are performed on each camera image and point cloud data, and the characteristics of the processed camera images and point cloud data are analyzed. Feature information is concatenated and simultaneously input into the corresponding semantic segmentation model. The semantic segmentation model processes each feature information layer by layer through forward propagation, obtains the class probability of each point and pixel, sets an adaptive threshold based on the quality score of the current record, stores semantic labels with class probabilities lower than the adaptive threshold in the map as soft annotations, and projects the segmentation results back to the 3D point cloud to generate a point cloud or semantic block map with semantic labels. After semantic segmentation, connected components within the same semantic category are clustered to generate different instances. Boundary fitting and minimum bounding box calculation are performed on each instance, and a set of short-term Kalman filters are initialized for inter-frame tracking. The unified confidence of the same instance is calculated based on the confidence of LiDAR intensity and visual semantics.Metadata such as category, pose, size, velocity estimation, and confidence level are generated for each instance. Near-field structures and fine obstacles are represented using voxel grids, and large-scale terrain is represented using raster elevation maps. The semantically labeled point cloud of each frame is then projected onto map voxels, and the semantic count and geometric confidence level of each voxel are updated. Based on the updated voxel confidence levels, a weighted fusion strategy is used to update the map values. Voxel values ​​for moving objects that have not been observed for a long time are adjusted according to a time decay rule. Finally, the voxel map undergoes periodic sparsification and hierarchical indexing to generate a complete environmental semantic map.

[0021] It should be noted that the specific formula for calculating the node clock offset is as follows:

[0022] ;

[0023] In the formula, Represents the clock offset of the local node relative to the reference clock; This represents the local timestamp of the control terminal when the request was sent. This represents the local timestamp when the sensor node received the request. Represents the local timestamp when the sensor node sends a response; This represents the local timestamp of the control terminal when it received the response;

[0024] The specific calculation formula for the ratio or difference between the received amplitude and phase of each array element and the reference value is as follows:

[0025] ;

[0026] In the formula, Representing the Amplitude correction factor for each receiving channel; Representing the The actual measured complex envelope of each channel; Representing the Each channel reference complex envelope; Modulus operator for complex numbers; Representing the Phase correction amount for each receiving channel; Phase operator for complex numbers;

[0027] The specific calculation formula for statistical neighborhood analysis is as follows:

[0028] ;

[0029] In the formula, Representing the The average neighborhood distance metric for each point; This represents the number of neighboring points used for statistics; The Euclidean distance function represents the distance between two points; Representing the Three-dimensional coordinate vectors of points; Representing the Three-dimensional coordinate vectors of points;

[0030] The specific calculation formula for fine registration is as follows:

[0031] ;

[0032] In the formula, Represents the total residual of registration; Represents the number of corresponding point pairs to be registered; Represents the rotation matrix from the source coordinate system to the target coordinate system; Representing the The coordinates of each corresponding point in the source point cloud; Represents the translation vector; Representing the There are corresponding points on the target.

[0033] Based on the environmental semantic map, initial beam weights and pointing strategies are generated. Based on the generation results, the weights of each array element are adjusted in real time to optimize the echo signal quality.

[0034] Specifically, the environmental semantic map is projected onto a local planar grid centered on the array. The frequency of semantic occurrence of each voxel is counted, and an initial feature description is formed by combining the geometric features of the unit. Multi-scale features of each block are extracted through predefined local filters, and the multi-scale features are stitched together to generate corresponding block feature tensors. The feature tensors of each block are matched with the directional response template of the array to obtain local scores for multiple beam pointing and elevation angle pairs. The local scores of the entire monitoring area are then accumulated to form a global score map of candidate directions. Non-maximum suppression is applied to the global score map, and multiple peak directions are selected as the initial candidate pointing set. The importance of the corresponding channel for each candidate direction is calculated. Based on target coverage utility, interference and occlusion penalties, and resource and power constraints, a set of target utility functions is established. Then, according to semantic categories and preset security policies, corresponding weights are assigned to each candidate direction. The target utility function is then formalized into a beam weight optimization objective function, and the initial candidate weights that satisfy the target utility function are obtained through fast approximate solution. The actual array RF and phased array unit hardware... To address the limitations, amplitude and phase separation is performed on each initial candidate weight. A least-squares or nearest-neighbor quantization strategy is used to map the amplitude and phase of each channel to the corresponding hardware-supported grid points. Simultaneously, quantization errors are recorded and the quantization loss of the entire array is calculated. If the quantization loss exceeds a preset threshold, local re-optimization is performed on each channel, and the quantization loss and local optimization are repeated until the quantization loss falls below the preset threshold. The quantized amplitude and phase are then encoded into specific control commands. An array factor model is used to rapidly simulate each control command, calculating the peak value in the main lobe direction, the amplitude of the main sidelobes, and the distribution of receive and transmit gains in the semantic region. The simulation results are used to verify whether a preset threshold is met. Control commands that do not meet the threshold are corrected and adjusted, and the simulation is repeated until the threshold is met. Then, a confidence score for each control command is calculated based on simulation metrics and corresponding semantic importance. The control commands are sorted from highest to lowest confidence score, and the control commands with confidence scores higher than the preset threshold are output. The control command with the highest confidence score is selected as the primary control scheme, and the rest are backup control schemes.

[0035] Specifically, each edge controller calculates the total transmit power of the current control scheme. If the total transmit power exceeds the preset total power limit of the corresponding node, all channels are scaled up proportionally. Then, the scaled control schemes are converted into the control command format of the RF front-end. The actual register values ​​fed back by the front-end after transmission are checked. If any channel transmission fails or a register returns an abnormal value, a single channel is triggered to fall back to safe mode and recorded as a transmission anomaly. Down-conversion and ADC sampling are performed on each channel according to the configured sampling rate and window length to obtain the baseband sampling stream of the corresponding channel. Then, local DC removal, low-pass filtering, and windowing are performed on each baseband sampling stream. The short-time Fourier transform of each frame of baseband sampling stream is calculated, and the baseband sampling stream of each channel is time-aligned according to a unified time base. Each sampling frame is marked with a corresponding timestamp and frame number. For each beam pointing, from the current control... The system reads the control commands of each array element under the specified direction and converts them into corresponding complex weight coefficients. Then, through linear weighted summation, it performs complex weight synthesis on a sample-by-sample basis at each time point within the frame, generating complex value outputs for each beam at different times. Subsequently, it calculates the complex value output energy, phase stability, and spectral width of each beam. At the same time, it normalizes each beam using the local noise baseline during non-working periods to generate corresponding signal-to-noise ratio and confidence values. The beam indicators, signal-to-noise ratio, and confidence values ​​are used as quality indicators. Based on the quality indicators of each beam and the predefined optimization objectives, it uses the output of the current frame and the local reference to calculate the incremental update of each complex weight in the current cycle. After the current cycle ends, it updates the complex weights of each channel. Then, based on the signal-to-noise ratio distribution and semantic priority of the beams in multiple frames, it selects the direction to sweep in the next time window or maintains the current direction, while switching the pointing strategy.

[0036] Example 2, refer to Figure 1 A real-time wind speed and direction sensing method utilizing adaptive beamforming, the specific steps of which are as follows:

[0037] Continuously acquire observation information from different spatial locations to form time-series sparse wind field observation data, and reconstruct the three-dimensional wind speed and direction distribution of the entire monitoring area in real time.

[0038] Specifically, the latest sparse wind field observation data is retrieved in chronological order and represented using four-dimensional spatiotemporal coordinates. Then, corresponding observation entries are generated for each sparse wind field observation data set, and these entries are stored in an online cache. Stratified sampling is performed based on the spatiotemporal distribution and confidence level of the sparse wind field observation data. The sampled data are normalized, and corresponding weight values ​​are calculated based on the confidence level of each observation sample. A coordinate-based implicit network structure is used as the wind field reconstruction model, with four-dimensional spatiotemporal coordinates as input and wind speed and direction components as outputs. A prior model weighted wind field reconstruction model from offline pre-training is then loaded as initial parameters. The pre-processed observation samples are input into the wind field reconstruction model, which processes each observation sample layer by layer through forward propagation and outputs predicted wind speed and direction values ​​for the corresponding monitoring area. Simultaneously, based on the current monitoring area... The model calculates the total loss between the predicted and actual wind speed and direction values, and uses the automatic differentiation framework of each edge device to calculate the gradient value of the wind field reconstruction model parameters corresponding to the current total loss. Based on the gradient value, the model parameters are adjusted. The wind field reconstruction model parameters are trained and updated repeatedly until the total loss converges to the preset fault tolerance range. Then, the performance of the trained wind field reconstruction model is verified using observation samples that were not involved in the training. If the performance does not meet the preset expectation, the model parameters are rolled back to the previous version and updated again. The latest sparse wind field observation data is input into the verified wind field reconstruction model. The wind field reconstruction model outputs grid sampling of the continuous wind speed vector field through forward propagation. Then, the continuous wind speed vector field is projected into the corresponding wind speed magnitude and wind direction angle, and smoothed by the prediction results of each block. At the same time, the prediction results are stitched together to form the corresponding continuous three-dimensional wind speed and direction distribution field.

[0039] It should be noted that the short-time average power, phase statistics, and center frequency and spectral width of the Doppler spectrum of the synthesized output stream of each beam are calculated and recorded. Then, the instantaneous Doppler displacement of the synthesized output stream of each beam is calculated using the phase difference method or continuous phase tracking to obtain the radial velocity component of the wind speed. At the same time, the corresponding confidence value is calculated. If the confidence value is higher than the preset threshold, the observation is considered a valid observation. Each valid observation is packaged into an observation unit according to the prescribed structure. The similarity of each observation unit is calculated, and observation units with similarity higher than the preset allowable threshold are merged. Then, lightweight coding is used to write the observation units as sparse wind field observation data into the edge database.

[0040] Furthermore, in this embodiment, the observation unit specifically includes a timestamp, beam identifier, spatial pointing, projection point, power, phase, Doppler, confidence level, and semantic label;

[0041] Stratified sampling prioritizes observations from the most recent time, high confidence levels, and semantically important regions.

[0042] Based on the continuous wind field reconstruction results and the original beamforming output, the joint variation characteristics of signal features in the spatial and temporal dimensions are calculated, and wind field events are identified.

[0043] Specifically, the continuously reconstructed 3D wind field and the local observations obtained from beamforming are mapped to the same voxel index, and the local spatial dispersion of each voxel position is calculated and used as the spatial feature of each voxel. Based on a preset time interval, a short-term history buffer is maintained for each voxel, and the temporal changes of this short-term history buffer are calculated and used as the temporal features of each voxel. Then, the spatial and temporal features of each voxel are synchronized and aligned. The spatial dispersion and temporal changes at each voxel are normalized, and the spatiotemporal joint dispersion of the corresponding voxel is synthesized by weighted summation. Voxels with spatiotemporal joint dispersion below a preset threshold are then removed, and a candidate voxel set is established based on the remaining voxels. Based on the candidate voxels, the voxels in the candidate voxel set are further classified into... A set of voxels is selected, and the confidence values ​​of each wind field time are calculated using a linear discriminator. Candidate event labels for each gradient are output. Based on spatial connectivity, each candidate voxel is clustered into multiple instance clusters. The geometric location, spatial range, and initial intensity description of each cluster are calculated. Corresponding time identifiers are added to each instance cluster. Based on the spatial distance, instance shape differences, and feature similarity among the instance clusters, the Hungarian algorithm is used to calculate the matching degree between the current time and the existing trajectory of each newly identified instance cluster and the trajectory of the previous time. If the matching degree is higher than a preset threshold, the match is considered successful, and the position of the trajectory is updated. Otherwise, the match fails, and the number of failures is recorded. If the number of failures exceeds a preset threshold, the matching stops, and an event report is output.

[0044] Based on the current perceived wind field status, a short-term wind field evolution simulation is performed, and the beam operation mode within the future time window is planned based on the simulation results.

[0045] Each node in the sensing network independently executes the real-time wind speed and direction sensing process locally, and periodically uploads model parameter update information while integrating the updated parameters.

Claims

1. A real-time wind speed and direction sensing method utilizing adaptive beamforming, characterized in that, The specific steps of this sensing method are as follows: Ⅰ. Initialize the adaptive beamforming array, auxiliary environmental perception sensors, and computing units, and collect point cloud data of the monitoring area in real time to construct an environmental semantic map; II. Based on the environmental semantic map, generate initial beam weights and pointing strategies, and adjust the weights of each array element in real time based on the generation results to optimize the echo signal quality. Ⅲ. Continuously acquire observation information from different spatial locations to form time-series sparse wind field observation data, and reconstruct the three-dimensional wind speed and direction distribution of the entire monitoring area in real time; The specific steps for real-time reconstruction of the three-dimensional wind speed and direction distribution across the entire monitoring area are as follows: S4.1: Extract the latest sparse wind field observation data in chronological order and represent the sparse wind field observation data in four-dimensional spatiotemporal coordinates. Then, generate corresponding observation entries for each sparse wind field observation data and put each observation entry into the online cache. Perform stratified sampling based on the spatiotemporal distribution and confidence level of the sparse wind field observation data. S4.2: Normalize the observation samples obtained by sampling, calculate the corresponding weight value based on the confidence of each observation sample, adopt a coordinate-based implicit network structure as the wind field reconstruction model, take four-dimensional spatiotemporal coordinates as input and wind speed and wind direction components as output, and then load the prior model weight wind field reconstruction model from offline pre-training as the initial parameters. S4.3: Input the preprocessed observation samples into the wind field reconstruction model. The wind field reconstruction model processes each observation sample layer by layer through forward propagation and outputs the predicted wind speed and direction values ​​in the corresponding monitoring area. At the same time, based on the actual wind speed and direction values ​​in the current monitoring area, the total loss between the predicted value and the actual value is calculated. S4.4: Calculate the gradient values ​​of the wind field reconstruction model parameters corresponding to the current total loss using the automatic differentiation framework of each edge device, and adjust the model parameters based on the gradient values. Repeat the training and updating of the wind field reconstruction model parameters multiple times until the total loss converges to the preset fault tolerance range. Then, use observation samples that did not participate in the training to verify the performance of the trained wind field reconstruction model. If the performance does not meet the preset expectations, roll back the model parameters to the previous version and update them again. S4.5: Input the latest sparse wind field observation data into the validated wind field reconstruction model. The wind field reconstruction model outputs grid sampling of the continuous wind speed vector field through forward propagation. Then, the continuous wind speed vector field is projected into the corresponding wind speed magnitude and wind direction angle, and smoothed by the prediction results of each block. At the same time, the prediction results are stitched together to form the corresponding continuous three-dimensional wind speed and wind direction distribution field. IV. Based on the continuous wind field reconstruction results and the original beamforming output, calculate the joint variation characteristics of signal features in the spatial and temporal dimensions, and identify wind field events; V. Based on the current perceived wind field status, perform ultra-short-term wind field evolution simulation, and plan the beam working mode within the future time window based on the simulation results; VI. Each node in the sensing network independently executes the real-time wind speed and direction sensing process locally, and periodically uploads model parameter update information, while also integrating the updated parameters.

2. The real-time wind speed and direction sensing method using adaptive beamforming according to claim 1, characterized in that, The specific steps for constructing an environmental semantic map by collecting point cloud data of the monitoring area in real time, as described in Step I, are as follows: S1.1: Based on the scale of the monitoring area and the minimum resolvable scale of the target, set the camera resolution, exposure time and frame rate, as well as the rotation rate and single-point repetition rate of the LiDAR. Then start two operating modes: static detail scanning and dynamic capture. At the same time, the mode will automatically switch according to the preset external trigger conditions during operation. S1.2: Configure trigger priority and time window for each sensor in the edge controller, and configure local caching strategy. Then, calculate and record the instantaneous sampling quality index after each cycle. If it is lower than the preset threshold, automatically adjust the exposure and rate or mark the cycle as "low confidence" data. S1.3: Perform black level and gain correction and linearization verification on camera images, divide each camera image into multiple blocks, and count the saturation pixel ratio to detect local exposure of the image. Image areas with saturation and pixel ratio below the preset threshold are marked as low weight. Unbelievable points in the LiDAR point cloud with return intensity below the preset threshold or exceeding the preset range are screened out. Then, statistical neighborhood analysis is performed on each point to identify isolated noise points. Secondary filtering is performed on the identified isolated noise points based on normal and curvature. Then, non-local mean filtering is applied to each camera image. Finally, the corresponding quality score is calculated and recorded for each sample frame. S1.4: Based on the on-site ranging, the initial pose relationship between each sample frame is obtained. Then, the stable geometric features of each point cloud frame are extracted, and candidate corresponding point pairs are constructed. The least squares-based iterative optimization is adopted with the rigid body transformation parameters as unknowns. The fine registration of each candidate corresponding point pair is repeated until the residual converges to the preset range. After registration, the geometric consistency between multiple frames is checked. If a systematic deviation is found, it is back to coarse registration or the sample frames are reacquired. S1.5: Based on the height stratification or normal consistency, extract the set of ground candidate points from the registered point cloud data, calculate the local covariance matrix of the ground candidate point set, and obtain the local ground normal through principal component analysis. Divide the monitoring area into multiple blocks, and fit one or more local terrain baselines in each block. Record the ground height confidence and fitting error of each block, and mark the blocks with confidence scores below the preset threshold as unresolved areas. S1.6: Normalize, scale resample, and data augment the images and point cloud data of each camera respectively, and stitch together the feature information of the processed images and point cloud data. Then, input them into the corresponding semantic segmentation model. The semantic segmentation model processes each feature information layer by layer through forward propagation and obtains the class probability of each point and pixel. Based on the quality score of the current record, an adaptive threshold is set. The semantic labels with class probabilities lower than the adaptive threshold are stored in the map in the form of soft annotations. The segmentation results are projected back to the 3D point cloud to generate a point cloud or semantic block map with semantic labels. S1.7: After semantic segmentation is completed, cluster the connected components within the same semantic category to generate different instances. Perform boundary fitting and minimum bounding box calculation on each instance, and initialize a set of short-term Kalman filters for inter-frame tracking. Calculate the unified confidence of the same instance based on LiDAR intensity and visual semantic confidence, and generate metadata such as category, pose, size, velocity estimation and confidence for each instance. S1.8: Use voxel grids to represent near-field structures and fine obstacles, use raster elevation maps to represent large-scale terrain, then project the semantically labeled point cloud of each frame into map voxels, and update the semantic count and geometric confidence of the voxel. Based on the updated voxel confidence, use a weighted fusion strategy to update map values, adjust the voxel values ​​of moving objects that have not been observed for a long time according to the time decay rule, and then perform periodic sparsification and hierarchical indexing on the voxel map to generate a complete environmental semantic map.

3. The real-time wind speed and direction sensing method using adaptive beamforming according to claim 2, characterized in that, The specific steps for generating the initial beam weights and pointing strategy described in step II are as follows: S2.1: Project the environmental semantic map onto a local planar grid centered on the array, count the number of occurrences of semantics for each voxel, and combine the geometric features of the cell to form an initial feature description. Extract multi-scale features of each block through a predefined local filter, and stitch the multi-scale features to generate the corresponding block feature tensor. S2.2: Match the feature tensor of each block with the directional response template of the array to obtain the local scores of multiple beam pointing pitch angle pairs. Then, accumulate the local scores of the entire monitoring area to form a global score map of candidate directions. Perform non-maximum suppression on the global score map, select multiple peak directions as the initial candidate pointing set, and calculate the channel importance of each candidate direction. S2.3: Based on target coverage utility, interference and occlusion penalties, and resource and power constraints, a set of target utility functions are established. Then, according to semantic categories and preset security policies, corresponding weights are assigned to each candidate direction. The target utility function is then formalized into a beam weight optimization objective function, and the initial candidate weights that satisfy the target utility function are obtained through fast approximate solution. S2.4: Due to the hardware limitations of the actual array RF and phase control units, the amplitude and phase of each initial candidate weight are separated. The least square or nearest neighbor quantization strategy is used to map the amplitude and phase of each channel to the corresponding hardware-supported grid point. At the same time, the quantization error is recorded and the quantization loss of the entire array is calculated. S2.5: If the quantization loss is higher than the preset threshold, local re-optimization is performed on each channel, and the quantization loss and local optimization are repeatedly calculated until the quantization loss is lower than the preset threshold. The quantized amplitude and phase are then encoded into specific control commands. S2.6: Use the array factor model to quickly simulate each control command, calculate the peak value in the main lobe direction, the amplitude of the main side lobes, and the distribution of receive and transmit gains in the semantic region. Check whether the preset threshold is met based on the simulation results. For control commands that do not meet the threshold, make corrections and adjustments, and re-simulate until the threshold is met. Then, calculate the credibility score of each control command based on the simulation indicators and the corresponding semantic importance. Sort each control command according to the credibility score from high to low, and output the control commands with credibility scores higher than the preset threshold. At the same time, the control command with the highest credibility score is used as the main control scheme, and the rest are backup control schemes.

4. The real-time wind speed and direction sensing method using adaptive beamforming according to claim 3, characterized in that, The specific steps for adjusting the weights of each array element in real time to optimize the echo signal quality, as described in Step II, are as follows: S3.1: Each edge controller calculates the total transmit power of the current control scheme. If the total transmit power exceeds the preset total power limit of the corresponding node, all channels are scaled up by the same ratio. Then, the scaled control schemes are converted into the control command format of the RF front end. The actual register value fed back by the front end after transmission is checked. If any channel fails to transmit or the register returns an abnormal value, a single channel is triggered to fall back to safe mode and recorded as a transmission abnormality. S3.2: Perform down-conversion and ADC sampling on each channel according to the configured sampling rate and window length to obtain the baseband sampling stream of the corresponding channel. Then, perform local DC removal, low-pass filtering and windowing on each baseband sampling stream, calculate the short-time Fourier transform of each frame of baseband sampling stream, and align the baseband sampling stream of each channel according to a unified time base. Mark each sampling frame with the corresponding timestamp and frame number. S3.3: For each beam pointing, read the control command of each array element under that pointing from the current control scheme, convert it into the corresponding complex weight coefficient, and then perform complex weight synthesis of each beam sample by sample at time point within the frame through linear weighted summation to generate the complex value output of each beam at different times. Then calculate the complex value output energy, phase stability and spectral width of each beam. At the same time, use the local noise baseline during non-working periods to normalize each beam to generate the corresponding signal-to-noise ratio and confidence value, and use the beam indicators, signal-to-noise ratio and confidence value as quality indicators. S3.4: Based on the quality indicators of each beam and the predefined optimization objectives, use the output of the current frame and the local reference to calculate the incremental update of each complex weight in the current cycle, and update the complex weight of each channel after the current cycle ends. Then, based on the signal-to-noise ratio distribution and semantic priority of the multi-frame beams, select the direction to sweep in the next time window or maintain the current pointing, and switch the pointing strategy at the same time.

5. The real-time wind speed and direction sensing method using adaptive beamforming according to claim 1, characterized in that, The specific steps for calculating the joint variation characteristics of signal features in the spatial and temporal dimensions and identifying wind field events, as described in step IV, are as follows: S5.1: Map the continuously reconstructed three-dimensional wind field and the local observations obtained from beamforming to the same voxel index, calculate the local spatial dispersion of each voxel position, and use it as the spatial feature of each voxel. Based on the preset time interval, maintain a set of short-term history buffers for each voxel, calculate the time changes of each short-term history buffer, and use them as the time features of each voxel. Then, synchronize and align the spatial features and time features of each voxel. S5.2: Normalize the spatial dispersion and time variation of each voxel, and synthesize the spatiotemporal joint dispersion of the corresponding voxel by weighted summation. Then, filter out the corresponding voxels whose spatiotemporal joint dispersion is lower than the preset threshold, and establish a candidate voxel set based on the remaining voxels. S5.3: Based on the candidate voxel set of each voxel in the candidate voxel set, calculate the confidence value of each wind field time through a linear discriminator, and output the candidate event labels of each gradient. According to spatial connectivity, cluster each candidate voxel into multiple instance clusters, calculate the geometric location, spatial range and initial intensity description of each cluster, and add corresponding time labels to each instance cluster. S5.4: Based on the spatial distance, instance shape differences, and feature similarity of each instance cluster, the Hungarian algorithm is used to calculate the matching degree between the current time and the existing trajectory of each newly identified instance cluster at the previous time. If the matching degree is higher than the preset threshold, the matching is considered successful, and the position of the trajectory is updated. Otherwise, the matching fails, and the number of failures is recorded. If the number of failures is higher than the preset threshold, the matching is stopped, and an event report is output.

Images