A smart antenna coverage area self-adaptive adjustment monitoring method

Abstract

The present application relates to a kind of wisdom antenna coverage area adaptive adjustment monitoring method, to solve the problem of traditional antenna coverage blind area elimination and communication load dynamic adaptation capability deficiency.The core scheme includes: collecting Beidou high-precision positioning data and video stream, constructing the space-time fingerprint atlas of unique space-time signature, realizing visual semantics and space unit binding based on human body detection and tracking, quantifying crowd gathering changes, flow deviation and boundary expansion through semantic feature extraction, generating low-dimensional feature vector expressing communication demand.On this basis, user distribution manifold space mapping is established, and the optimal five-dimensional antenna parameters are obtained through geometric center and similarity matching, to drive antenna adaptive adjustment, while topological fine tuning and online incremental update of static knowledge base are carried out for abnormal manifold evolution.The technology improves the antenna coverage adaptive accuracy and communication demand response efficiency, realizes dynamic generalization and closed-loop optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent antenna adaptive control and multimodal user behavior perception technology, and in particular to a method for adaptive adjustment and monitoring of the coverage area of ​​a smart antenna. Background Technology

[0002] Adaptive adjustment systems for smart antenna coverage areas, as an important means for improving energy efficiency and service quality in modern wireless communication networks, have been widely researched and initially applied. Mainstream solutions in the industry mostly focus on building time-series prediction models using historical user distribution data, or employing deep learning and multi-objective optimization algorithms to achieve parameter adaptation. Typical technical approaches include: recurrent neural networks (such as LSTM) trained based on historical communication traffic and mobile trajectories; prediction models incorporating attention mechanisms or tensor decomposition; large-scale model fine-tuning combined with federated learning; and real-time parameter combination optimization based on multi-strategy genetic algorithms. These methods aim to improve the foresight and fine-grained scheduling capabilities of antenna parameter decisions through feature learning from high-dimensional data.

[0003] In practical engineering deployments, smart antenna parameter adaptive optimization systems generally rely on historical data accumulation, continuously improving their ability to capture user distribution trends in typical scenarios through periodic offline training or incremental learning. Most solutions input high-dimensional features such as real-time user spatial behavior data, wireless channel information, and environmental interference characteristics into complex predictive modeling networks, and then perform parameter search and tuning based on the model output. Theoretically, these methods can achieve good adaptation to static and some low-frequency dynamic scenarios, possessing explicit learning capabilities and fine-grained adjustment. However, with the increasing cost of perceiving multi-source heterogeneous data and the growing complexity of service coverage scenarios, these models generally suffer from significant constraints such as limited generalization ability, delayed response to sudden large-scale changes in user behavior, and high computational resource consumption.

[0004] In practical applications of intelligent antenna adaptive control and multimodal user behavior perception, current mainstream technologies are all based on the "temporal prediction-parameter combination decision-making" paradigm. BeiDou high-precision positioning data technology and computer vision recognition, as important data sources for environmental perception, are often applied to positioning perception and pedestrian target detection, respectively, to improve the spatial coverage capability of a single modality. However, the industry lacks an integrated solution that collaboratively maps BeiDou spatiotemporal fingerprints and visual semantics into a dynamic manifold space of user distribution, using manifold evolution laws for decision-driven decisions. Existing representative systems are mostly used in the following scenarios: communication assurance for large-scale gatherings or sports venues, hotspot scheduling in campuses and parks, and dense multimodal user access in urban public areas. While these technologies can significantly optimize coverage efficiency in known scenarios, they often heavily rely on the quantity and quality of historical data, and have shortcomings in adaptability, interpretability, and ease of deployment in rapidly changing or heterogeneous environments.

[0005] Based on system analysis, the existing technology mainly has the following problems and shortcomings: The generalization ability of dynamic user distribution prediction models is insufficient. Current decision-making relies on static or semi-static models trained in history, which are unable to accurately predict complex dynamic trends such as abnormal crowd gatherings, flow anomalies, or sudden expansion of edge blind spots, thus affecting the real-time performance and effectiveness of parameter adjustments.

[0006] After multi-source sensing data fusion, there are often redundant high-dimensional features that have not been fully distilled, resulting in high model computational complexity, heavy deployment burden on edge devices, and an inability to simultaneously meet the dual constraints of energy saving and low latency.

[0007] The parameter adjustment mechanism is mainly based on "black box" network output, which lacks interpretability and physical traceability of the specific decision-making basis, making it difficult for experts to intervene and optimize operation and maintenance.

[0008] Most model structures are adaptable to historically known scenarios, but they are prone to failure or response lag when encountering sudden environmental changes, boundary mismatches, or new spatial configurations, and global retraining is extremely costly.

[0009] The aforementioned technical bottlenecks indicate that in smart antenna coverage scenarios with highly dynamic demands and strong environmental heterogeneity, there is an urgent need for a new parameter decision-making and feedback mechanism that does not rely heavily on large-scale historical data, can adaptively generalize to unknown dynamic scenarios, and has real-time low computational overhead.

[0010] To address the aforementioned shortcomings, this invention proposes using BeiDou high-precision positioning data to generate regional spatiotemporal fingerprints as prior anchor points. Combined with camera recognition results, semantic distillation is used to extract low-dimensional features strongly correlated with communication needs, constructing a user distribution manifold space. An online incremental manifold update mechanism enables adaptive representation and parameter-driven adjustment. This scheme does not rely on complex black-box time-series prediction models and requires no large-scale labeled data. It fundamentally improves the generalization adaptability and system energy efficiency for complex and unknown scenarios while ensuring real-time response and transparency, promoting the efficient cross-scenario application of smart antenna adaptive technology. Summary of the Invention

[0011] This application provides a smart antenna coverage area adaptive adjustment monitoring method, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0012] This application provides a smart antenna coverage area adaptive adjustment monitoring method, specifically including: S1: Acquire BeiDou high-precision positioning data and camera video stream, perform gridded mapping processing on the BeiDou high-precision positioning data, and generate spatiotemporal fingerprint map units; S2: Based on the spatial coordinate constraints of the spatiotemporal fingerprint map unit, the human detection box and tracking ID in the camera video stream are mapped to the corresponding spatiotemporal fingerprint map unit to generate a spatiotemporal fingerprint association dataset bound to visual semantics. S3: Perform semantic distillation on the visual semantics in the spatiotemporal fingerprint association dataset to extract the rate of change of crowd density, the offset angle of the main moving axis and the expansion speed of the sparse edge region, and generate a low-dimensional communication requirement key semantic feature vector. S4: Construct a two-dimensional orthogonal coordinate system using the key semantic feature vectors of the low-dimensional communication requirements, encode the semantic state of each spatiotemporal fingerprint unit as a dynamic trajectory point, and generate a user distribution manifold space point set; S5: Calculate the geometric centroid position of the user distribution manifold space point set, and perform a matching query in the preset five-dimensional antenna parameter mapping rule cluster based on the geometric centroid position to generate a target five-dimensional antenna parameter combination instruction; S6: Based on the target five-dimensional antenna parameter combination command, drive the antenna actuator to perform physical parameter adjustment actions on the downtilt angle, azimuth angle, elevation, transmission power and polarization mode to generate an adaptively adjusted antenna coverage area.

[0013] S7: Monitor the configuration frequency distribution of the user distribution manifold spatial point set within multiple consecutive periods. If it is determined that the frequency of at least one preset manifold configuration deviates significantly from the historical average, a local manifold topology fine-tuning mechanism is triggered to generate an updated spatiotemporal fingerprint map unit mapping relationship. S8: Replace the corresponding entries in the original static prior knowledge base based on the spatiotemporal fingerprint graph unit mapping relationship to complete the online incremental manifold update.

[0014] The smart antenna coverage area adaptive adjustment monitoring method provided in this application has the following beneficial effects: (1) By constructing a user distribution representation mechanism based on spatiotemporal fingerprinting, this scheme effectively overcomes the technical defects of traditional smart antenna parameter decision-making methods, such as strong dependence on historical data, poor model generalization ability, and high response latency. Existing technologies generally use time-series modeling methods such as LSTM, attention mechanisms, or tensor decomposition for crowd flow prediction, which have problems such as long training cycles, difficulty in adapting to sudden scene changes, and high deployment costs. In contrast, this invention divides the geographic space into grid units with unique spatiotemporal signatures and combines sub-meter-level BeiDou positioning and visual semantic distillation results to establish a static prior knowledge base that integrates the baseline of crowd density, entropy value of movement direction, and distribution of dwell time, thereby realizing a structured and interpretable expression of user spatial behavior patterns. This mechanism does not require large-scale labeled data training, avoids the uncertainty risk brought by black box models, and significantly improves the system's adaptability in unknown or dynamic environments. It is especially suitable for application scenarios with complex crowd flow and strong event-driven characteristics, such as campuses and sports venues.

[0015] (2) The introduction of semantic distillation and manifold space dynamic mapping mechanism transforms antenna parameter adjustment from "parameter search" to "configuration matching," significantly reducing computational complexity and enhancing real-time decision-making. Unlike the high-dimensional parameter iteration process commonly found in multi-objective optimization algorithms or federated learning frameworks, this scheme extracts only semantic features strongly related to communication requirements—such as the rate of change in crowd density, the offset angle of the main moving axis, and the expansion speed of the edge sparse region—and encodes them as dynamic trajectory points in a two-dimensional manifold space, representing "coverage urgency" and "adjustment tolerance," respectively. The collective centroid of the trajectory points of all units directly triggers the lookup table operation of the preset five-dimensional antenna parameter mapping rule cluster, achieving millisecond-level response output. This design not only avoids the complex online optimization solution process but also ensures the transparency and controllability of the decision-making logic, enabling the entire system to operate stably on edge gateway-level devices and meeting the dual constraints of energy saving and low latency control in 5G networks.

[0016] (3) By constructing a closed-loop adaptive feedback system through an online incremental manifold update mechanism, the system can effectively track long-term trend evolution without introducing global retraining. When the frequency of at least one preset manifold configuration deviates continuously from the historical average, the system automatically identifies the abnormal pattern and triggers local topology fine-tuning, correcting only the mapping relationship of fingerprint units near the affected area, thereby avoiding the overall performance degradation problem caused by distribution drift in traditional deep learning models. This mechanism not only preserves the stability of expert experience and offline simulation results, but also gives the system the ability to continuously evolve, enhancing the robustness of long-term deployment. At the same time, since all processing flows are based on lightweight rules and semantic abstraction operations, there is no need to use high-resource-consuming components such as large model fine-tuning, tensor decomposition, or multi-source interference database matching, which greatly reduces hardware investment and operation and maintenance costs.

[0017] In summary, this solution reconstructs the user distribution understanding paradigm, transforming traditional prediction tasks into spatiotemporal semantic manifold evolution analysis, thus achieving an efficient closed loop from perception to decision-making. Its technical approach combines high real-time performance, strong generalization capabilities, and low deployment barriers. It not only overcomes the bottlenecks of existing smart antenna systems in terms of dynamic adaptability and interpretability but also provides a new, highly scalable architectural approach for intelligent wireless resource allocation at the edge. It is particularly suitable for typical application environments requiring rapid response to dynamic changes in the population and accurate matching of communication resource supply, such as large-scale public event security, smart campus management, and emergency communication dispatch. Attached Figure Description

[0018] Figure 1 This is the main flowchart of a smart antenna coverage area adaptive adjustment monitoring method; Figure 2 This is a sub-flowchart of a smart antenna coverage area adaptive adjustment monitoring method; Figure 3 This is another sub-flowchart of a smart antenna coverage area adaptive adjustment monitoring method. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0021] like Figure 1 As shown, this application provides a smart antenna coverage area adaptive adjustment monitoring method, specifically including: S1: Acquire BeiDou high-precision positioning data and camera video stream, perform gridded mapping processing on the BeiDou high-precision positioning data, and generate a spatiotemporal fingerprint map unit with a unique spatiotemporal signature.

[0022] S2: Based on the spatial coordinate constraints of the spatiotemporal fingerprint map unit, the human detection box and tracking ID in the camera video stream are mapped to the corresponding spatiotemporal fingerprint map unit to generate a spatiotemporal fingerprint association dataset bound to visual semantics.

[0023] S3: Perform semantic distillation on the visual semantics in the spatiotemporal fingerprint association dataset to extract the population density change rate, main movement axis offset angle and edge sparse region expansion speed, and generate a low-dimensional communication requirement key semantic feature vector.

[0024] S4: Construct a two-dimensional orthogonal coordinate system using the key semantic feature vectors of the low-dimensional communication requirements, encode the semantic state of each spatiotemporal fingerprint unit into dynamic trajectory points that represent the urgency of coverage and the tolerance for adjustment, and generate a set of user distribution manifold space points.

[0025] S5: Calculate the geometric centroid position of the user distribution manifold space point set, and perform a matching query in the preset five-dimensional antenna parameter mapping rule cluster based on the geometric centroid position to generate a target five-dimensional antenna parameter combination instruction.

[0026] S6: Based on the target five-dimensional antenna parameter combination command, drive the antenna actuator to perform physical parameter adjustment actions on the downtilt angle, azimuth angle, elevation, transmission power and polarization mode to generate an adaptively adjusted antenna coverage area.

[0027] S7: Monitor the configuration frequency distribution of user distribution manifold spatial point set within multiple consecutive periods. If it is determined that the frequency of at least one preset manifold configuration deviates significantly from the historical average, a local manifold topology fine-tuning mechanism is triggered to generate an updated spatiotemporal fingerprint map unit mapping relationship.

[0028] S8: Replace the corresponding entries in the original static prior knowledge base based on the updated spatiotemporal fingerprint map unit mapping relationship, complete the online incremental manifold update, and generate the next cycle parameter decision benchmark with dynamic generalization capability.

[0029] Step S1: Acquire BeiDou high-precision positioning data and camera video stream, perform gridded mapping processing on the BeiDou high-precision positioning data, and generate a spatiotemporal fingerprint map unit with a unique spatiotemporal signature. Specifically, this includes: S1.1: Obtain the original latitude and longitude coordinate sequence reported by the Beidou high-precision positioning data terminal and the original video stream data collected by the camera, and perform timestamp alignment and coordinate system transformation processing on the original latitude and longitude coordinate sequence to generate a standardized positioning point set under a unified spatiotemporal reference.

[0030] This sub-step obtains the raw latitude and longitude coordinate sequence reported by the BeiDou high-precision positioning data terminal and the raw video stream data collected by the camera as input objects. The raw latitude and longitude coordinate sequence undergoes bidirectional matching processing based on the UTC time field reported by the BeiDou terminal and the timestamp of the camera video stream frame to establish a timestamp alignment matrix, ensuring that multi-source data correspond in a unified time domain. The latitude and longitude data in the timestamp alignment matrix undergoes coordinate system transformation processing based on an ellipsoidal model, converting the WGS-84 coordinate system to a local reference system (such as CGCS2000). During the transformation process, the ellipsoidal radius of curvature is calculated and the positioning accuracy is corrected accordingly. The ellipsoidal radius of curvature R can be obtained by the following formula: in, For the semi-major axis of the ellipsoid, Let be the eccentricity of the ellipsoid. The latitude and longitude are represented by the R value. Planar projection processing is performed on the latitude and longitude using the aforementioned R value to generate eastward and northward planar coordinate components under a unified spatial reference. Combined with the camera's inherent extrinsic parameter matrix, the planar coordinate components are spatially registered with the video stream acquisition position vector to correct spatial offsets introduced by the device installation location, thus forming a unified positional reference frame for the set of positioning points. Through a combination of timestamp alignment and spatial registration, the multi-source raw data acquired in the previous step is transformed into a standardized set of positioning points under a unified spatiotemporal reference, achieving spatial consistency and temporal synchronization, providing accurate input conditions for the subsequent adaptive mesh generation in S1.2.

[0031] For example, in an airport outdoor monitoring scenario, the WGS-84 latitude and longitude sequence output by the BeiDou terminal includes a UTC time accuracy of 1ms, a camera video frame rate of 25fps, a frame timestamp accuracy of 40ms, and a bidirectional matching threshold of ±20ms to establish a timestamp alignment matrix. The semi-major axis of the ellipsoid is taken as 6,378,137 meters, the eccentricity of the ellipsoid is taken as 0.08181919, and the latitude range is 31° to 32°. According to the formula, the radius of curvature R of the ellipsoid at 31.5° is approximately 6,378,000 meters. Using this value for planar projection, the eastward and northward components of the positioning point are obtained. After calibration, the camera extrinsic matrix has translation components of (1.2m, -0.5m, 3.0m) and a rotation matrix of three-dimensional orthogonal matrix. Through spatial registration, the east and north components of the planar coordinates are adjusted by 0.8m and -0.3m respectively. The resulting standardized positioning point set has significantly improved matching accuracy with the video acquisition points in the same Euclidean plane coordinate system, and the node error of subsequent grid division is reduced to less than 0.5m.

[0032] S1.2: Based on the geographical coverage of the standardized location point set, the target monitoring area is spatially discretized based on adaptive grid partitioning to generate a set of basic geographic grid units with fixed resolution.

[0033] S1.3: Bind the baseline of human flow density and the entropy value of movement direction of typical time period to each grid cell in the basic geographic grid cell set, and perform spatiotemporal feature injection operation to generate an enhanced geographic grid cell carrying static prior knowledge.

[0034] Based on the set of basic geographic grid cells with fixed resolution generated in step S1.2, the pedestrian flow observation data divided into typical time periods within each grid cell is selected as the input data source, and the pedestrian flow density statistical sequence and the movement direction vector sequence of each individual within a unit time are extracted as the execution objects.

[0035] The periodic mean of the population density statistical sequence is calculated, and the average number of people in each observation period is calculated by sliding the time window. The result is stored as the baseline value of population density of the grid cell in the typical period.

[0036] Probability distribution estimation is performed on the sequence of movement direction vectors, the frequency of occurrence of different direction segments is counted, and the distribution entropy of the movement direction is calculated using the entropy formula, which is: in, Let be the probability value of direction segment i.

[0037] Based on the baseline value of pedestrian density and the entropy value of movement direction, a static prior feature structure is constructed, and the two are used as typical time period feature labels for the grid cell.

[0038] Perform a spatiotemporal feature injection operation to write the static prior feature structure into the grid cell attribute field and retain the index key value associated with the spatial coordinates of the grid cell to generate an enhanced geographic grid cell.

[0039] Through the aforementioned feature injection process, the basic grid cell results from the previous step are transformed into enhanced geographic grid cells carrying static prior knowledge, thus realizing the feature foundation required for the subsequent spatiotemporal signature generation process.

[0040] For example, within a monitored area of ​​a city's commercial district, the basic grid unit resolution is set to 20 meters × 20 meters, with a typical time period from 18:00 to 20:00. During this period, the number of people in a grid unit over 10 consecutive time windows (each window lasting 5 minutes) is 72, 80, 75, 78, 82, 85, 79, 77, 81, and 84, respectively. Calculating the periodic mean yields a baseline pedestrian density of 79.3. Movement directions are statistically analyzed according to eight directional segments: east, northeast, north, northwest, west, southwest, south, and southeast, with probability distributions of 0.25, 0.10, 0.15, 0.05, 0.20, 0.05, 0.10, and 0.10. Calculating the entropy value using the above formula yields an entropy value of approximately 2.52. The baseline of pedestrian density (79.3) and the entropy value of movement direction (2.52) are written as static prior feature labels into the attributes of this grid cell. During subsequent hash encoding, this label will participate in the generation of spatiotemporal signature, effectively improving the uniqueness and generalization adaptability of spatiotemporal fingerprint.

[0041] S1.4: Based on the spatial index information of the enhanced geographic grid unit, a hash encoding mechanism is used to perform joint encryption operations on the grid coordinates and time window features to generate a spatiotemporal signature string with global uniqueness.

[0042] Based on the spatial index information of enhanced geographic grid units, a joint data structure containing grid coordinate values ​​and time window feature parameters is used as the input for encryption operations. The grid coordinate parameters are flattened using a 3D coordinate sequence, arranging latitude, longitude, and elevation components sequentially into a linear vector to reduce dimensional redundancy during subsequent encoding. Periodic normalization mapping is performed on the time window feature parameters, mapping timestamp values ​​to a continuous interval between 0 and 1 to ensure encoding consistency across different sampling periods. The flattened grid coordinate vector and the normalized time window parameters are concatenated to generate a composite key-value vector containing both spatial and temporal information, which serves as the input data for the hash function. A salting strategy is applied, appending grid feature salt values ​​extracted from a static prior knowledge base to the end of the composite key-value vector to enhance collision resistance. Distributed consistent hashing is used to iteratively hash the composite key-value vector, generating intermediate hash values ​​and updating the hash seed according to a preset iteration cycle to ensure the global uniqueness of the output signature. Finally, the hash result obtained from the iterative hashing is formatted and converted to generate a spatiotemporal signature string conforming to the system's identification rules. By using distributed consistent hashing encoding, the enhanced geographic grid unit spatial index information from the previous step is transformed into a globally unique spatiotemporal signature string, achieving unique matching and index retrieval of multi-source heterogeneous data in the spatiotemporal domain.

[0043] For example, in a smart antenna monitoring system, the coordinates of an enhanced geographic grid cell are set as longitude 120.13579, latitude 30.24680, and elevation 45.00000 meters, with a time window feature of 08:30:15 on June 18, 2024. After flattening the 3D coordinate sequence, a linear vector [120.13579, 30.24680, 45.00000] is obtained. After time window normalization, the value is 0.354, resulting in a composite key-value vector [120.13579, 30.24680, 45.00000, 0.354]. The salt value is taken from the average entropy value of the cell's movement direction in the static prior knowledge base, which is 0.872. After concatenation, [120.13579, 30.24680, 45.00000, 0.354, 0.872] is obtained as the hash input. A distributed consistent hashing process with three iterations is adopted. In each round, the seed is updated and recalculated based on the current hash value. The formula is as follows: Where `seed` is the seed value in the iteration, and `key` is the composite key-value vector. After three rounds of iteration, the output hash result is "af93c21b7e8d54f2", which is then formatted to generate the spatiotemporal signature string "TSF-af93c21b7e8d54f2". This signature string is used in the system to uniquely match BeiDou positioning data and camera semantic data, achieving a significant improvement in fast indexing and retrieval and avoiding spatiotemporal domain conflicts.

[0044] S1.5: Write the spatiotemporal signature string as a unique identifier into the corresponding enhanced geographic grid cell attribute field to complete the instantiation of the spatiotemporal fingerprint map cell, so as to output a spatiotemporal fingerprint map cell with a unique spatiotemporal signature for subsequent steps to call.

[0045] The globally unique spatiotemporal signature string generated in the preceding steps is indexed and associated with the attribute fields of the enhanced geographic grid unit, clarifying the storage location and data type definition of the unique identifier. Based on the spatial index key-value call hash dictionary lookup mechanism of the enhanced geographic grid unit, the corresponding grid unit data object is located, and attribute extension operations are performed while maintaining the integrity of the original static prior feature structure. A field write control module is used to load the spatiotemporal signature string into the specified attribute field buffer, and data version control logic is used to add a timestamp of the current operation period to the field to ensure traceability of subsequent differentiated updates. Consistency verification is performed on the written unique identifier, and cross-period conflict detection is called to compare the conflict probability of the current signature string with that in the existing signature database. If a duplicate identifier is detected, a conflict resolution rule chain is triggered, the hash code is recalculated to generate a new signature string after conflict resolution, and the writing process is repeated. For the enhanced geographic grid unit that has been written and verified, the instantiation construction interface is called to transform its state from a normal data object into a spatiotemporal fingerprint map unit entity with a unique spatiotemporal signature, and the entity index is registered in the entity registry for subsequent steps. By using a unique identifier writing and instantiation construction process, the enhanced geographic grid unit generated in the previous step is transformed into a spatiotemporal fingerprint map unit with a unique spatiotemporal signature, thereby achieving accurate referencing effects for subsequent visual semantic binding and spatial constraint matching.

[0046] For example, in a smart antenna coverage area adaptive adjustment monitoring scenario, the spatial index range of the original enhanced geographic grid cell covers the geographic coordinate interval [120.12345, 31.12345] to [120.22345, 31.22345], with a bound population density baseline of 220 people / hour and a movement direction entropy value of 0.68. In the preceding step S1.4, joint hash encoding is performed using the two-dimensional grid coordinates (35, 42) of this grid cell and the time window feature (15:00-15:05), mapping the encoding result to the spatiotemporal signature string "G35x42_T1500". In this step, this string is written to the enhanced geographic grid cell attribute "TemporalSignature" field and appended with the execution cycle number "Cycle20240518_03". During the consistency check, a collision probability lower than the threshold of 0.01 is detected, and no recalculation of hash is required. The instantiated construction interface converts the grid cell into a spatiotemporal fingerprint map cell entity, and registers the index "TS_G35x42_T1500" in the entity registry. In the subsequent S2 step, when the human positioning point mapped by the camera falls within the spatial boundary of this grid cell, this unique identifier can be directly referenced for visual semantic binding operations, ensuring a significant improvement in the spatial matching accuracy of data fusion.

[0047] Step S2: Based on the spatial coordinate constraints of the spatiotemporal fingerprint map unit, map the human detection box and tracking ID in the camera video stream to the corresponding spatiotemporal fingerprint map unit to generate a spatiotemporal fingerprint association dataset bound to visual semantics. Specifically, this includes: S2.1: Obtain the sequence of center pixel coordinates of the human body detection box and the tracking ID identifier after target detection processing in the camera video stream. Perform inverse perspective projection transformation based on the pre-calibrated camera intrinsic parameter matrix and extrinsic parameter rotation and translation matrix to convert the two-dimensional image plane coordinates into a real-time human body positioning point set in the three-dimensional world coordinate system.

[0048] The camera intrinsic parameter matrix is ​​an inherent physical parameter matrix determined by the optical focal length of the camera lens and the pixel size of the imaging sensor. It is used to establish the geometric mapping relationship between the two-dimensional image pixel coordinate system and the camera's own coordinate system. The extrinsic parameter rotation and translation matrix is ​​a pose parameter matrix characterizing the camera's installation posture in the three-dimensional world coordinate system. It includes a rotation component describing the shooting angle and a translation component describing the installation position. Both the camera intrinsic parameter matrix and the extrinsic parameter rotation and translation matrix are pre-obtained based on the camera's factory physical specifications calibration and on-site spatial pose calibration.

[0049] The input is the camera video stream after target detection processing, containing the sequence of center pixel coordinates of the human detection bounding box and its corresponding tracking ID identifier. Based on the calibrated camera intrinsic and extrinsic rotation and translation matrices, a 3D reconstruction projection mapping model is constructed. The intrinsic matrix is ​​inverted on the center pixel coordinate sequence to reflect the normalized coordinate relationship of the camera's imaging plane. The normalized plane coordinates are combined with the extrinsic matrix through matrix multiplication to achieve an inverse perspective projection transformation from the image plane to the world coordinate system. For the normalized 3D coordinates obtained from the inverse transformation, combined with the known camera installation height and tilt angle, affine transformation parameter adjustment is performed to eliminate coordinate offsets caused by lens distortion. During processing, the 2D pixel coordinates (u,v) are converted to world coordinates (X,Y,Z) using the following formula: ( , , () represents the real-time coordinates of the human body's location point in the three-dimensional world coordinate system after transformation. It is the inverse of the extrinsic rotation and translation matrix. is the inverse of the camera intrinsic parameter matrix, (u, v) represents the pixel coordinates of the center of the human detection box on the two-dimensional image plane, s represents the scale factor, and t is the camera extrinsic parameter translation vector.

[0050] Through the aforementioned multi-level matrix operations, the input two-dimensional pixel coordinate sequence is transformed into a real-time human positioning point set in a three-dimensional world coordinate system, while retaining the tracking ID to maintain the temporal uniqueness of the target, achieving a precise mapping from visual detection results to physical spatial positioning. Using inverse perspective projection transformation, the results of the previous step are converted into three-dimensional world coordinate data suitable for subsequent spatial inclusion determination, achieving accurate matching between the user's location and the grid cells of the antenna coverage area.

[0051] For example, a camera mounted 6 meters high on the side of a city road has a resolution of 1920×1080 pixels, a horizontal viewing angle of 90°, and a vertical viewing angle of 60°. The intrinsic parameter matrix elements are fx = 1050, fy = 1048, cx = 960, and cy = 540. The extrinsic parameter matrix includes an installation tilt angle of -15° and translation amounts of (0.5m, 0.0m, 6.0m). The center pixel coordinates of the target detection box are (1200, 400), corresponding to a tracking ID of 21. After performing the intrinsic parameter inverse matrix operation, the normalized coordinates are (0.2286, -0.1296). Multiplying the normalized coordinates by the extrinsic parameter rotation matrix, and combining this with the installation height and translation vector, yields the world coordinates (8.2m, -4.6m, 0.0m). Through distortion compensation and affine correction, the final positioning point is maintained at (8.1m, -4.5m, 0.0m). The 3D positioning result and tracking ID are output simultaneously, providing input data for subsequent spatial inclusion determination. It has been verified that the positioning deviation is controlled within 0.15m under different lighting conditions, which significantly improves the spatial mapping accuracy of dynamic human body position.

[0052] S2.2: Based on the geographic latitude and longitude information of the real-time human positioning point set and the spatial boundary constraints of the spatiotemporal fingerprint map unit with a unique spatiotemporal signature generated in the previous steps, perform spatial inclusion determination processing of the point within the polygon to establish the target spatiotemporal fingerprint map unit index to which each real-time human positioning point belongs.

[0053] Based on the geographic latitude and longitude data of the real-time human location point set generated in step S2.1, and the spatial boundary data of the spatiotemporal fingerprint map unit with a unique spatiotemporal signature output in step S1.5, the input conditions for spatial inclusion determination are established.

[0054] The latitude and longitude coordinates of each real-time human positioning point are subjected to coordinate precision quantization processing to ensure that the resolution of the input values ​​during the judgment process is not lower than the preset meter-level threshold.

[0055] A closed polygon index is constructed using the spatial boundary vertex sequence in the spatiotemporal fingerprint unit attributes, and a sorting rule is used to ensure the consistency of vertex order to avoid judgment errors.

[0056] The latitude and longitude coordinates of human body positioning points and the sequence of polygon vertices are uniformly converted into Cartesian plane coordinate system to eliminate the influence of geographic coordinate system curvature on inclusion determination.

[0057] The spatial inclusion determination is performed by calling the ray intersection method. For each location point, a horizontal ray is emitted from that point, and the number of intersections with the polygon boundary is counted.

[0058] If the number of intersections satisfies the odd number determination rule, the location point is assigned to the corresponding spatiotemporal fingerprint spectrum unit; if it is even, the inclusion detection of neighboring units is performed to establish a unique attribution index.

[0059] Based on the spatial inclusion determination results, each real-time human positioning point is bound to a unique target spatiotemporal fingerprint map unit index, realizing the mapping and transformation from location data to spatiotemporal unit index, which serves as the input key value for S2.3 multi-source fusion mounting.

[0060] By processing spatial inclusion determination, the real-time human positioning point set from the previous step is transformed into a unique target spatiotemporal fingerprint map unit index corresponding to each point, thereby achieving the spatial binding effect between positioning information and spatiotemporal fingerprint units.

[0061] For example, within a 500m × 500m monitoring area, the spatiotemporal fingerprinting unit boundary is a rectangular polygon composed of four vertices. The latitude and longitude of the vertices are 30.0000°N, 120.0000°E to 30.0045°N, 120.0045°E, respectively, and the latitude and longitude of the human body positioning point are 30.0020°N, 120.0020°E. Converting the latitude and longitude of the vertices and positioning point to a Cartesian coordinate system with the lower left corner vertex of the area as the origin, the positioning point coordinates are obtained as (222m, 222m). Using the ray intersection method, a ray is emitted from this point along the positive X-axis. A single intersection with the polygon boundary is detected, which meets the odd-number determination rule. Therefore, the positioning point belongs to the rectangular unit index ID=05. For another location point with latitude and longitude of 30.0050°N and 120.0050°E, the transformed coordinates are (555m, 555m). With 0 ray intersections, it is determined to be outside the polygon. Neighbor detection is performed, and it is ultimately assigned to cell ID=06. In this scenario, the index binding results after spatial inclusion determination are ID=05 and ID=06, effectively ensuring accurate matching in subsequent multi-source data fusion. This achieves a one-to-one correspondence between human location data and spatiotemporal cell indices within the target area, and significantly improves the accuracy of subsequent semantic fusion and parameter decision-making.

[0062] S2.3: Using the target spatiotemporal fingerprint map unit index as the association key value, the corresponding human body detection box attribute information and tracking ID identifier are injected into the attribute field of the spatiotemporal fingerprint map unit, and a multi-source data fusion mounting operation is performed to generate spatiotemporal fingerprint association data entries bound to visual semantics.

[0063] The human body detection box attribute information refers to the set of numerical features extracted in real time from the camera video stream through target detection, which describes the visual state of each human target in the picture; the human body detection box attribute information comes from deep learning inference on video frames in the camera video stream, and while recognizing the human body and generating its pixel-level bounding box, the visual semantic data calculated includes the center coordinates, width and height of the box, detection confidence and optional pose key points.

[0064] S2.4: Perform time window sliding aggregation processing on the spatiotemporal fingerprint association data entries bound to visual semantics, and filter instantaneous false detection noise based on the continuity characteristics of the tracking ID identifier to generate a spatiotemporal fingerprint association dataset with temporal consistency.

[0065] S2.5: Based on the spatiotemporal fingerprint association dataset with temporal consistency, perform a data structure serialization and encapsulation operation to package spatial coordinates, spatiotemporal signatures, population size, and movement trajectory features into standardized intermediate data objects, so as to output a spatiotemporal fingerprint association dataset bound to visual semantics for subsequent semantic distillation operations.

[0066] The spatial coordinates refer to the unique spatial region identifier of the spatiotemporal fingerprint unit to which each real-time human positioning point belongs, which is bound by the spatial inclusion determination process; the spatiotemporal signature is the encoding derived from the spatiotemporal fingerprint unit itself, used to uniquely identify the temporal and spatial attributes of the unit; the number of people is obtained by aggregating the number of human detection box attribute information and tracking ID identifiers injected into the attribute field of the same spatiotemporal fingerprint unit; the movement trajectory feature is the sequence pattern reflecting the position change of the target in continuous spatiotemporal units, which is extracted by sliding aggregation processing in the time window based on the continuity of the tracking ID identifier.

[0067] like Figure 2 As shown, step S3: Perform semantic distillation on the visual semantics in the spatiotemporal fingerprint association dataset to extract the rate of change in crowd density, the offset angle of the main movement axis, and the expansion speed of the sparse edge region, generating a low-dimensional semantic feature vector for key communication requirements. Specifically, this includes: S3.1: Based on the historical period human detection box number sequence and the current period human detection box number sequence in the spatiotemporal fingerprint association dataset, perform sliding window difference operation to eliminate instantaneous noise interference and quantify the density fluctuation amplitude per unit time, thereby generating the instantaneous change in crowd density; then perform normalization mapping processing on the instantaneous change in crowd density to obtain a standardized crowd density change rate.

[0068] In this embodiment, the sequence of human detection boxes in the historical period and the sequence of human detection boxes in the current period are both sequence data extracted from the spatiotemporal fingerprint association dataset in chronological order, representing the change in the number of human targets within the same spatiotemporal fingerprint map unit. They are generated by statistically analyzing the number of people in the historical time window and the latest time window respectively and arranging them in chronological order in the continuous time window sliding aggregation process of the human detection box attribute information and tracking ID identifier injected into the attribute field of the unit.

[0069] The normalization mapping process is specifically performed based on the Min-Max linear normalization function to eliminate instantaneous noise interference and quantify the density fluctuation amplitude per unit time. In other embodiments, it can also be performed based on the Z-Score normalization function or the Sigmoid nonlinear saturation function to adapt to the data distribution characteristics of different dynamic ranges and improve the robustness of the normalized population density change rate.

[0070] S3.2: Using the tracking ID coordinate trajectory data of continuous frames in the spatiotemporal fingerprint association dataset, perform principal component analysis dimensionality reduction processing to extract the feature vector representing the overall movement trend of the crowd as the main movement direction reference; then calculate the angle between the main movement direction reference and the preset grid normal direction of the spatiotemporal fingerprint map unit to obtain the main movement axis offset angle reflecting the degree of deviation of the crowd flow.

[0071] The tracking ID coordinate trajectory data refers to the spatial coordinate sequence corresponding to each tracking ID in a continuous time frame, extracted from the spatiotemporal fingerprint association dataset based on the continuity of the tracking ID identifier. The tracking ID coordinate trajectory data is derived from the human detection box attribute information (such as the center coordinate of the box) injected into the spatiotemporal fingerprint map unit attribute field, which is grouped by tracking ID identifier and then processed by time window sliding aggregation to filter noise, forming temporal position data reflecting the target's movement path.

[0072] Based on the tracking ID coordinate trajectory data of consecutive frames in the spatiotemporal fingerprint association dataset, the world coordinate sequence of each tracking ID within a specified time window is selected as the analysis input to ensure that the trajectory data has undergone spatial coordinate system unification and temporal consistency screening. The trajectory data is centered by subtracting the mean from the trajectory point coordinates to eliminate displacement deviation, forming a zero-mean trajectory matrix as the input matrix for principal component analysis. The covariance matrix of this trajectory matrix is ​​calculated to quantify the joint change trend of each coordinate component, and eigenvalue decomposition is performed to extract the corresponding eigenvectors. Based on the magnitude of the eigenvalues, the eigenvector corresponding to the largest eigenvalue is selected as the main movement direction reference representing the overall movement trend of the population, and it is normalized to a unit direction vector to eliminate the influence of scale on angle calculation. The preset grid normal direction vector of the spatiotemporal fingerprint atlas unit is obtained, and vector angle calculation is performed. The main movement axis offset angle is calculated using the following formula: Where V is the reference vector for the main movement direction, and N is the grid normal direction vector. The dot product of vector V and vector N is represented, where |V| and |N| represent the magnitudes of the corresponding vectors, respectively. The calculated offset angle is then bound to the corresponding spatiotemporal signature to form intermediate-state feature data reflecting the degree of deviation in crowd flow.

[0073] By using principal component analysis for dimensionality reduction and vector angle calculation, the spatiotemporal fingerprint correlation data trajectory information from the previous step is transformed into main movement axis offset angle data that reflects the degree of deviation of the overall flow trend of the population, thereby achieving accurate quantification of the flow direction deviation in communication coverage requirements.

[0074] For example, for a set of tracking IDs within the coverage area, 120 consecutive frames of trajectory data are sampled at a frequency of 10 Hz, resulting in a world coordinate sequence of 12 sampling points per trajectory, in meters. After centering the trajectory sequence, a zero-mean trajectory matrix is ​​formed. Its covariance matrix is ​​calculated, and eigenvalue decomposition is performed, yielding a maximum eigenvalue of 4.56 and a corresponding eigenvector of [0.707, 0.707]. This vector is normalized and used as vector V, with the preset grid normal direction N vector set to [0, 1]. The dot product V·N = 0.707 is calculated using the formula, with vector magnitudes |V|=1 and |N|=1. Substituting these values ​​into the formula, θ ≈ 45° is obtained. This offset angle is marked as a medium-level flow deviation in subsequent feature fusion, resulting in a significantly improved antenna beam adjustment safety margin width. Verification results show that even in scenarios with this flow deviation, the system can maintain coverage continuity without increasing transmit power.

[0075] S3.3: For the distribution set of human detection boxes located in the boundary region of the spatiotemporal fingerprint map unit in the spatiotemporal fingerprint association dataset, perform convex hull contour extraction and area evolution monitoring to identify the morphological changes of the sparse region at the coverage edge; calculate the contour expansion rate per unit time based on the morphological changes of the sparse region, thereby generating the edge sparse region expansion speed that represents the growth trend of the coverage blind area.

[0076] The human detection box distribution set refers to the set of all human detection box spatial locations located in the boundary region of the spatiotemporal fingerprint map unit, selected from the spatiotemporal fingerprint association dataset. The human detection box distribution set is derived from the human detection boxes that have completed spatial inclusion determination and are bound to each spatiotemporal fingerprint map unit. The human detection boxes are then subjected to secondary screening based on their spatial coordinates and relative positional relationship with the unit boundary, resulting in a specific subset for analyzing edge coverage patterns.

[0077] Based on the distribution set of human detection boxes located in the boundary regions of the spatiotemporal fingerprint spectral units in the spatiotemporal fingerprint association dataset, a spatial index of the pixel coordinate sequence of human detection boxes in the boundary regions is constructed. This sequence is then converted into a corresponding set of geographic coordinates using a geometric mapping function for subsequent morphological analysis. Convex hull extraction is performed on this set of geographic coordinates. Andrew's monotone chain 2D convex hull algorithm is used to construct a minimum convex polygon contour containing all boundary detection points in a two-dimensional plane. The contour vertex sequence is then sorted clockwise to meet the input requirements for area evolution monitoring. Time series calculations are performed on the area change of this convex hull contour over continuous sampling periods. A polygon area calculation method based on the Shoelace formula is introduced, with the expression: in and Let x and y represent the x and y coordinates of the i-th vertex of the convex hull, respectively. and Let represent the coordinates of the next vertex immediately adjacent to the i-th vertex. When i=n, and The values ​​are respectively taken as the coordinates of the first vertex. and This is used to achieve polygon closure, where n represents the total number of vertices of the convex hull contour, and A represents the area of ​​the convex hull polygon. The area results from consecutive periods are used to construct an area evolution sequence, and a first-order difference operation is performed on this sequence to obtain the area change per unit time.

[0078] Dividing the change in area per unit time by the change in the radius of the circumcircle of the current convex hull profile yields the profile expansion rate, which is calculated using the following formula: Where v is the expansion rate. The change in area per unit time. This represents the change in the radius of the circumscribed circle per unit time. The sign of the expansion rate is used to determine the trend of sparse region morphology changes. When v > 0, it indicates sparse region expansion; when v < 0, it indicates sparse region contraction. Through this processing method, the spatial distribution results of the human detection boxes from the previous step are transformed into the expansion rate of the edge sparse region, representing the growth trend of the coverage blind zone, thus quantifying the dynamic changes of the coverage edge. For example, in a smart antenna monitoring area, the BeiDou+ positioning system captured 48 detection points within the boundary area, with geographical coordinates ranging from 120.123°E to 120.125°E and 30.123°N to 30.126°N. The convex hull areas of the boundary detection box sequence obtained by the camera in the most recent three periods were 12.8 m², 14.5 m², and 15.1 m², respectively. Performing a first-order difference on the area sequence yields an area change AΔ1 of approximately 1.35 m² per unit time, and a circumscribed circle radius change rΔ1 of 0.25 m. This is calculated using the formula... The calculated spread rate is 5.4 m² / m, indicating a continuous expansion trend of the sparse region. Applying this result to the parameter decision module, and incorporating the sparse region expansion rate as part of the boundary adjustment tolerance calculation, can significantly improve the edge coverage integrity within the antenna beam refresh cycle.

[0079] S3.4: Obtain the population density change rate, main movement axis offset angle and edge sparse region expansion speed generated in the above steps, and perform multi-source heterogeneous feature alignment and weighted fusion processing to construct a composite feature structure containing three-dimensional information of density dynamics, flow direction deviation and boundary expansion; by performing vector space encoding processing on the composite feature structure, a low-dimensional communication requirement key semantic feature vector is finally generated.

[0080] like Figure 3 As shown, step S4: Construct a two-dimensional orthogonal coordinate system using the low-dimensional communication demand key semantic feature vector, and encode the semantic state of each spatiotemporal fingerprint unit into dynamic trajectory points representing coverage urgency and adjustment tolerance, generating a user distribution manifold space point set. Here, the semantic state refers to the association between the user's communication demand characteristics and network coverage status at a specific spatiotemporal location; it is obtained by extracting the population density change rate, main movement axis offset angle, and edge sparse region expansion speed from the low-dimensional communication demand key semantic feature vector, and mapping them into dynamic trajectory points representing coverage urgency and adjustment tolerance.

[0081] Step S4 specifically includes: S4.1: Based on the population density change rate contained in the key semantic feature vector of the low-dimensional communication demand, perform normalized weighted mapping processing to generate a scalar value representing the coverage urgency of the current area signal load pressure.

[0082] In this embodiment, the normalized weighted mapping process is specifically constructed based on the Sigmoid function, which uses its smooth nonlinear characteristics to robustly map the rate of change of the input to the [0,1] interval. In other embodiments, a linear piecewise function can also be used to achieve fast calculation with low complexity, or a competitive weight mechanism can be constructed based on the Softmax function to adapt to the dynamic and robust requirements of signal load pressure representation in different scenarios.

[0083] S4.2: Based on the main moving axis offset angle and edge sparse region expansion speed contained in the key semantic feature vector of the low-dimensional communication requirements, perform multi-dimensional fuzzy logic fusion operation to generate an adjustment tolerance scalar value that characterizes the antenna beam adjustment safety boundary.

[0084] Based on the main moving axis offset angle and the edge sparse region expansion rate contained in the key semantic feature vector of the low-dimensional communication requirements, a multi-dimensional fuzzy logic fusion method is used to collaboratively process the two types of inputs to generate an adjustment tolerance scalar value characterizing the safety boundary of antenna beam adjustment. Membership function mapping is performed on the main moving axis offset angle to convert continuous offset angle values ​​into coverage risk level membership degrees. These risk level membership degrees establish a three-segment gradient curve (low, medium, and high risk) based on the consistency between the deviation direction and the beam coverage axis. Membership function mapping is also performed on the edge sparse region expansion rate to convert the outward expansion rate per unit time into boundary instability membership degrees. These instability membership degrees establish a three-segment gradient curve (stable, fluctuating, and drastic) based on the boundary change rate and the growth trend of the coverage blind zone. A two-dimensional fuzzy rule matrix is ​​established, with the main moving axis offset angle risk level membership degree as the first input dimension and the edge sparse region expansion rate instability membership degree as the second input dimension. A corresponding adjustment tolerance fuzzy output set is set according to the linkage risk level of each combination. A fuzzy aggregation operation is performed on the matrix output set to merge the membership values ​​of all rules into a unified fuzzy set representing the comprehensive adjustment risk. The fuzzy set of comprehensive adjustment risk is then defuzzified using the centroid method, and the adjustment tolerance scalar value is calculated using the following formula: in, To adjust the tolerance scalar value, To defuzzify the output domain variables, To comprehensively adjust the membership function of risk, fuzzy logic fusion and defuzzification are used to transform the main movement axis offset angle and edge sparse region expansion speed in the low-dimensional communication requirement key semantic feature vector from the previous step into quantitative indicators reflecting the antenna beam adjustment safety boundary. This achieves the technical effect of dynamically evaluating the antenna adjustment fault tolerance range based on real-time user distribution.

[0085] For example, in an urban commercial area scenario where an outdoor base station covers a radius of 800 meters, the real-time offset angle of the main mobile axis is 15°, and the expansion speed of the sparse edge area is measured to be 1.2 m / s. The offset angle membership function is defined as a piecewise trapezoidal curve with low risk [0°, 10°], medium risk [10°, 20°], and high risk [20°, 30°]. The 15° corresponds to a membership degree of 0.7 for medium risk and 0.3 for low risk. The expansion speed membership function is defined as a piecewise trapezoidal curve with a membership degree of [0, 0.5], [0.5, 1.5], and [1.5, 3.0], with the 1.2 m / s corresponding to a membership degree of 0.8 for fluctuation and 0.2 for stability. The two-dimensional rule matrix is ​​set to output a medium safety boundary when the medium risk and fluctuation states are combined, corresponding to a membership degree distribution in the range of [0.4, 0.6]. After fuzzy aggregation, a comprehensive adjustment risk membership function is obtained. Numerical integration yields a formula where the numerator is 0.5 × the integral zone value, and the denominator is the total integral zone value. The final adjustment tolerance scalar value is calculated to be 0.52. This value is used as the input for the ordinate component to construct a two-dimensional orthogonal coordinate system. Combined with the coverage urgency value on the abscissa, a dynamic trajectory point set is generated. Verification shows that the adjustment tolerance curve is stable within a continuous period, and beam adjustment does not trigger coverage blind spots, achieving a significant improvement in system adjustment safety.

[0086] S4.3: Using the coverage urgency scalar value as the abscissa component and the adjustment tolerance scalar value as the ordinate component, a two-dimensional orthogonal feature coordinate system is constructed to establish the geometric projection reference surface of the user distribution state in the manifold space.

[0087] Based on the dual scalar input conditions of the coverage urgency scalar value and the adjustment tolerance scalar value, the two-dimensional coordinate construction processing module is invoked to load the coverage urgency scalar value into the horizontal axis component register area and the adjustment tolerance scalar value into the vertical axis component register area.

[0088] Numerical domain consistency calibration is performed on the horizontal and vertical coordinate components. An interval normalization mechanism is used to uniformly map the horizontal and vertical components to the closed interval [0,1] to ensure that the indicators of different dimensions are comparable and computationally stable in the same geometric space.

[0089] Orthogonal axis elements are generated using a structured coordinate builder. By setting the unit length of the basis vectors of the horizontal and vertical axes to be equal, and defining the physical correspondence between the origin position and the zero-coverage urgency and zero-adjustment tolerance, a standard two-dimensional orthogonal feature coordinate system is formed.

[0090] A spatial projection transformation matrix is ​​introduced during the coordinate system generation process to perform affine transformation on coordinate points with different combinations of coverage urgency and adjustment tolerance, so as to eliminate the geometric distortion effect caused by the fluctuation of eigenvector values.

[0091] The horizontal and vertical components are combined into coordinate pairs through geometric projection, and their corresponding geometric positions are established in the orthogonal characteristic coordinate system to generate a reference surface for subsequent manifold state projection.

[0092] By using the coordinate system and geometric position establishment method described above, the key semantic feature vectors of low-dimensional communication requirements from the previous step are transformed into a geometric projection reference surface with precise spatial correspondence, thereby achieving a rigorous mapping between user distribution status and two-dimensional spatial position, and providing a standardized spatial reference for dynamic trajectory point instantiation.

[0093] For example, in a communication coverage optimization scenario in a certain urban area, the coverage urgency scalar value is normalized to 0.85, and the adjustment tolerance scalar value is normalized to 0.35. The coverage urgency scalar value is input into the horizontal axis component register area, and the adjustment tolerance scalar value is input into the vertical axis component register area. After interval normalization calibration, both components are simultaneously mapped to the [0,1] domain. In the coordinate system generation stage, the unit length of the horizontal and vertical axis basis vectors is set to the equivalent value of a 10-meter physical coverage radius, and the origin corresponds to the state where there is no user clustering and no adjustment is required. The constructed two-dimensional orthogonal feature coordinate system uses the affine transformation matrix [1,0;0,1] to represent the standard reference plane without scaling or rotation. The horizontal and vertical components are combined into a coordinate pair (0.85,0.35) and projected onto a specific position in the reference plane. This position is located at a length of 8.5 units on the horizontal axis and 3.5 units on the vertical axis, corresponding to the user distribution state with high coverage urgency and medium to low adjustment tolerance. Performance verification shows that this mapped position can accurately guide the antenna beam adjustment direction and safety margin during the subsequent dynamic trajectory point instantiation process, thereby significantly improving the real-time performance and matching degree of parameter decision-making.

[0094] S4.4: For each spatiotemporal fingerprint map unit, the coverage urgency scalar value and the adjustment tolerance scalar value are encapsulated into a coordinate pair, and a dynamic trajectory point instantiation operation is performed to generate a single-point manifold state entity bound to a specific spatiotemporal signature.

[0095] Based on the coverage urgency scalar value and adjustment tolerance scalar value generated in the preceding steps, the unique spatiotemporal signature attribute field of the spatiotemporal fingerprint graph unit is called as the binding identifier to establish a one-to-one correspondence with the semantic state of the unit.

[0096] Numerical precision unification processing is performed on the scalar values ​​of coverage urgency and adjustment tolerance, using floating-point formatting operations and limiting the number of decimal places to ensure component consistency in the two-dimensional orthogonal feature coordinate system.

[0097] The coordinate encapsulation function is used to combine the processed coverage urgency scalar value and the adjustment tolerance scalar value into a coordinate pair structure, clarifying that the horizontal component corresponds to the signal load pressure and the vertical component corresponds to the beam adjustment safety boundary.

[0098] A unique spatiotemporal signature index is embedded in the coordinate pair structure, and a hash linked list storage strategy is adopted to achieve fast retrieval and instance location. The structure is then encapsulated as a dynamic trajectory point data entity.

[0099] The instantiation and memory allocation process is performed on the dynamic trajectory point data entity, a globally unique reference handle is allocated and written to the state manager, forming a single-point manifold state entity bound to a specific spatiotemporal signature.

[0100] Through the above encapsulation and instantiation process, the scalar value result of the previous step is transformed into spatially computable coordinate point data, realizing the real-time presentation of the user distribution state in the manifold space and the basis for subsequent topological operations.

[0101] For example, in a smart antenna coverage area monitoring scenario, the low-dimensional communication demand key semantic feature vector generated by BeiDou positioning and camera recognition includes a coverage urgency scalar value of 0.82 and an adjustment tolerance scalar value of 0.46. After performing precision unification processing on both, the floating-point format is ensured to be consistent. The two are encapsulated into a coordinate pair using the following formula: (0.82, 0.46). This coordinate pair is bound to the spatiotemporal signature string "TSF_20230915_142300_GX09", and a dynamic trajectory point entity is generated using a hash linked list structure with a handle index of HTP_000134. After the state manager completes the entity allocation, the horizontal axis of this single-point manifold state in the two-dimensional orthogonal feature coordinate system reflects high load pressure, and the vertical axis reflects the medium beam adjustment safety boundary. In the subsequent S4.5 set assembly stage, the system directly calls the handle HTP_000134 to participate in the configuration calculation of the user distributed manifold space point set, thereby significantly improving the response capability to the coverage requirements of this region in the configuration geometric centroid calculation.

[0102] S4.5: Aggregate the single-point manifold state entities generated by all spatiotemporal fingerprint map units, and perform set topology assembly processing to generate a complete set of user distribution manifold spatial points that describe the evolution trend of user spatial distribution at the current moment.

[0103] Step S5: Calculate the geometric centroid position of the user distribution manifold space point set, and perform a matching query based on the geometric centroid position in a preset five-dimensional antenna parameter mapping rule cluster to generate a target five-dimensional antenna parameter combination instruction. Specifically, this includes: S5.1: Perform weighted centroid calculation based on the user distribution manifold space point set to obtain the manifold space geometric centroid coordinate vector representing the overall communication demand trend of the current coverage area.

[0104] S5.2: Construct a five-dimensional antenna parameter mapping rule cluster library using expert experience calibration data and offline simulation results to generate a static prior mapping index table containing the feature fingerprints of typical manifold configurations and the corresponding optimal antenna parameter combinations.

[0105] The expert experience calibration data and offline simulation output results are used as input conditions for constructing mapping rules. The input conditions include typical user distribution manifold spatial coordinate patterns, historical coverage adjustment records, and corresponding five-dimensional antenna parameter execution status. Based on the verified coverage scenarios and parameter adjustment cases in the expert experience calibration data, the pattern feature extraction operation is performed to map the two-dimensional coordinate combination of coverage urgency and adjustment tolerance to the five-dimensional parameter value domain. Based on the communication performance evaluation indicators of different manifold configurations in the offline simulation results, the performance priority sorting process is performed to determine the optimal five-dimensional antenna parameter combination entry for each configuration. A multi-dimensional feature fingerprint encoding mechanism is adopted to encapsulate the geometric features, dynamic evolution trends, and coverage performance indicators of each typical manifold configuration into a unique feature fingerprint vector. Through static prior index generation processing, a mapping table from feature fingerprints to five-dimensional antenna parameter combinations is established, and this mapping table is stored as a rule cluster library for subsequent matching queries. During the construction process, in order to meet the fast retrieval requirements of the rule index, a hash index optimization process is performed to convert the feature fingerprint mapping relationship into an index structure that can respond to matching requests in O(1) time complexity. By using the above processing method, the geometric centroid input from the previous step is transformed into a static prior mapping index table that can be matched, thereby achieving the expected technical effect of locking the optimal five-dimensional antenna parameter combination with low latency and high precision.

[0106] For example, in the scenario of adaptive adjustment of smart antenna coverage in a city square, the expert-calibrated data includes four types of user distribution manifold configurations: uniform coverage, single hotspot clustering, dual hotspot distribution, and edge expansion. The coverage urgency value for each manifold configuration ranges from 0 to 1, and the adjustment tolerance value ranges from 0 to 0.8. In the offline simulation results, the signal coverage quality, interference suppression level, and energy consumption performance under each configuration are calculated based on the propagation model. The five-dimensional antenna parameter combination with the highest comprehensive score for coverage quality and energy consumption is selected. For example, the single hotspot clustering configuration corresponds to a downtilt angle of 12 degrees, an azimuth angle of 75 degrees, an elevation of 18 meters, a medium-to-high level transmit power, and a vertical polarization. When performing feature fingerprint encoding, the coverage urgency and adjustment tolerance are paired to form a two-dimensional coordinate pair, which, together with the simulation evaluation index, forms a feature vector containing five consecutive fields. This feature vector is then converted into a 64-bit index key value using a hash function. During the mapping table generation phase, the feature fingerprint key values ​​of the four configuration types and the corresponding five-dimensional antenna parameter combination entries are written into the rule cluster library. A three-layer hash bucket structure is established during the retrieval optimization process to accommodate high-concurrency matching requests. If the coverage urgency calculated by the geometric centroid in actual operation is 0.65 and the adjustment tolerance is 0.32, then the index entry for the single-hotspot clustered configuration is locked through feature fingerprint encoding and hash retrieval. The five-dimensional antenna parameter combination associated with this configuration is directly output, achieving fast and high-precision parameter decision response.

[0107] S5.3: Perform multidimensional Euclidean distance metric calculation on the coordinate vector of the spatial geometric centroid of the manifold to obtain the similarity matching score sequence between the coordinate vector of the spatial geometric centroid of the manifold and the feature fingerprints of each typical manifold configuration in the static prior mapping index table.

[0108] S5.4: Perform maximum likelihood estimation filtering based on the similarity matching score sequence to lock the target typical manifold configuration feature fingerprint and its associated optimal antenna parameter combination entries that have the highest matching degree with the coordinate vector of the geometric centroid of the manifold space.

[0109] Based on the input conditions of the similarity matching score sequence, a dataset containing the matching relationships between the coordinate vector of the geometric centroid of the manifold space and the feature fingerprints of each typical manifold configuration in the static prior mapping index table is selected as the execution object. Probability distribution estimation processing is performed on this dataset to construct a normalized probability density function for the matching scores to eliminate inconsistencies in the scoring scales of fingerprints from different configurations. The probability density function is used to calculate the likelihood values ​​of the feature fingerprints of each typical manifold configuration, mapping the matching scores as independent variables to the likelihood values, and determining the peak position of the likelihood values ​​using the maximum likelihood estimation criterion. The configuration feature fingerprints at the peak positions are indexed and located, and their corresponding five-dimensional antenna parameter combination entries in the mapping rule cluster library are extracted. Consistency verification processing is performed to verify whether the matching features of the five-dimensional antenna parameter combination entries with the coordinate vector of the geometric centroid of the manifold space under preset rule weights satisfy the optimal conditions. Through the maximum likelihood estimation filtering method, the similarity matching score sequence from the previous step is transformed into a lock on the feature fingerprints of the target typical manifold configuration and their associated optimal antenna parameter combination entries, achieving precise selection of the antenna parameter combinations.

[0110] For example, in a scenario where the coverage area of ​​a smart antenna is dynamically adjusted, the geometric centroid coordinate vector of the manifold space is (0.72, 0.35). The static prior mapping index table contains 15 typical manifold configuration feature fingerprints and their corresponding five-dimensional antenna parameter combinations. Similarity matching score sequences obtained through multidimensional Euclidean distance metric are 0.88, 0.64, 0.91, and 0.57, respectively. After normalizing the score sequences to the [0,1] interval, a probability density function is constructed, and the likelihood value of each configuration fingerprint is calculated using the maximum likelihood estimation method. in, To score In configuration fingerprint Given the given conditions, n represents the total number of samples. Numerical optimization is used to search for the location with the highest likelihood value. The corresponding configuration fingerprint is the third type, with a matching score of 0.91. The associated five-dimensional antenna parameter combination includes a downtilt angle of 10.5°, an azimuth angle of 132°, an elevation of 15.2 meters, a transmit power level of 3, and vertical polarization. Consistency verification of this entry shows that, under the regular weight matrix, the matching features between this combination and the geometric centroid coordinate vector of the manifold space satisfy the preset optimal conditions. The output of this embodiment is to lock the feature fingerprint of the third typical manifold configuration and its optimal antenna parameter combination, which can be directly encapsulated into control commands in subsequent steps to achieve precise dynamic adjustment of the coverage area.

[0111] S5.5: Parse and encapsulate the control instructions according to the optimal antenna parameter combination entry to generate a target five-dimensional antenna parameter combination instruction containing downtilt angle value, azimuth angle value, elevation value, transmit power level and polarization mode identifier.

[0112] Step S6: Based on the target five-dimensional antenna parameter combination command, drive the antenna actuator to perform physical parameter adjustments on the downtilt angle, azimuth angle, elevation, transmit power, and polarization mode to generate an adaptively adjusted antenna coverage area. Specifically, this includes: S6.1: Obtain the downtilt adjustment component and azimuth adjustment component in the target five-dimensional antenna parameter combination command, and perform pulse width modulation processing on the antenna servo motor driver based on mechanical transmission control processing to generate a mechanical displacement control signal with specific rotation accuracy, thereby obtaining the downtilt mechanical execution state and azimuth mechanical execution state.

[0113] The system acquires the downtilt and azimuth adjustment components from the target five-dimensional antenna parameter combination command generated by S5.5, analyzes their values ​​and units, and converts them into the standardized angle format used by the antenna servo drive control system. It then calls the instruction interface of the mechanical execution unit, inputting the standardized downtilt and azimuth values ​​as input to the core processing module of the mechanical transmission control. In the mechanical transmission control, based on the transmission ratio parameters, moment of inertia coefficient, and friction torque correction value of the current antenna mechanical transmission pair, it performs a mapping calculation between the control quantity and the target rotation angle to generate the desired motor rotation step size matrix. Finally, it uses pulse width modulation (PWM)... A Modulation-Pulse Width Modulation (PWM) signal generator constructs a PWM control signal sequence containing duty cycle, period, and phase information based on the components of the step matrix. During the PWM control signal output process, the phase synchronization mechanism of the motor driver ensures that the output channels corresponding to the downtilt adjustment component and the azimuth adjustment component are triggered synchronously to avoid phase deviation during angle adjustment. The PWM control signal is driven to the corresponding servo motors via a power amplifier module, and the actual rotation step size is sampled and fed back in real time by the motor encoder to complete the closed-loop output of the downtilt and azimuth mechanical execution states. Through mechanical transmission control processing and the PWM drive mechanism, the angle components in the target five-dimensional antenna parameter combination command from the previous step are converted into a precise and controllable mechanical displacement execution state, realizing precise adjustment of the antenna array in the two degrees of freedom of downtilt and azimuth.

[0114] For example, in a smart antenna adaptive adjustment scenario, the target five-dimensional antenna parameter combination command includes a downtilt adjustment component of 3.5° and an azimuth adjustment component of 12.0°. During execution, the formulas for converting 3.5° and 12.0° into radian values ​​used by the servo control system are as follows: and The resulting downtilt angle is approximately 0.0611 radians, and the azimuth angle is approximately 0.2094 radians. In the mechanical transmission control processing, assuming a transmission ratio of 1:50, a moment of inertia coefficient of 0.0025 kg·m², and a friction torque correction value of 0.001 N·m, the step matrix components required to calculate the target rotation angle are... and The PWM control signal period is set to 20 ms, with duty cycles of 55% and 60% respectively, and the phase difference is controlled within ±0.5° for synchronous output. During servo motor execution, the encoder provides real-time feedback that the step size error is within ±0.02 steps, ensuring complete consistency between the tilt angle and azimuth angle adjustment positions and the target values. The output mechanical execution state meets the high-precision antenna beam pointing pre-adjustment requirements.

[0115] S6.2: Based on the downtilt angle mechanical execution state and the azimuth angle mechanical execution state, a closed-loop correction process is performed on the physical pointing of the antenna array using a high-precision angle sensor feedback loop to eliminate mechanical transmission gap error, thereby obtaining the calibrated antenna beam pointing vector.

[0116] S6.3: Obtain the transmit power adjustment component in the target five-dimensional antenna parameter combination command, and dynamically adjust the bias voltage of the RF front-end power amplifier module based on digital predistortion technology and automatic gain control processing to match the current channel loss characteristics, thereby obtaining the target RF radiated power level.

[0117] The transmit power adjustment component from the target five-dimensional antenna parameter combination command is obtained. The bias voltage control interface of the RF front-end power amplifier module is invoked to perform initial bias sampling to establish the current operating point voltage value. Based on the initial bias voltage value and the transmit power adjustment component, digital predistortion coefficients are calculated. An inverse transfer function model is established to inversely compensate for the nonlinear distortion characteristics of the signal. These inverse transfer function coefficients are injected into the digital predistortion processing unit to pre-adjust the amplitude and phase of the baseband signal to be transmitted, ensuring that the waveform after power amplification approximates the ideal state. Automatic gain control is performed based on the real-time power detection result of the predistortion output signal. The gain adjustment amount is dynamically calculated to correct the amplification coefficient of the power amplifier module to match the current channel loss characteristics. This gain adjustment amount is calculated as a ratio of the difference between the target power and the detected power, using the following formula: in, The target radio frequency radiated power level, The current detected power level is given by ΔG, which represents the required gain adjustment. The calculated ΔG is converted into a bias voltage fine-tuning command, and the voltage is dynamically corrected in a closed-loop manner through the RF power amplifier bias regulator to ensure the output power remains stable at the set value. Through the above-mentioned combined digital predistortion and automatic gain adjustment process, the target power setting from the previous step is transformed into the stable bias voltage and gain state of the RF power amplifier, achieving accurate control of the RF radiated power level and reducing channel distortion.

[0118] For example, in a smart antenna system, the target five-dimensional antenna parameter combination command specifies a transmit power level of 45dBm, corresponding to a target RF radiated power level setting of 31.6W, while the currently detected power level is 28.4W. After acquiring the transmit power adjustment component, the system performs predistortion coefficient calculation, obtaining inverse transfer function coefficients of [-0.003, 0.015, -0.002], which are then injected into the digital predistortion unit to correct the baseband signal. In the power detection stage, the automatic gain control process calculates ΔG according to the formula: The result was 0.101, indicating a need to increase the gain by approximately 10.1%. This gain adjustment was converted into a bias voltage adjustment value of 0.28V and applied to the power amplifier module through a bias regulator. During the process, the power detection point showed that the adjusted output power stabilized at 31.6W, and the channel power spectrum remained flat, verifying a significant reduction in signal distortion and a uniform distribution of received signal strength within the coverage area, demonstrating high communication quality.

[0119] S6.4: Based on the calibrated antenna beam pointing vector and the target radio frequency radiated power level, and combined with the polarization mode adjustment component in the target five-dimensional antenna parameter combination command, the phase phase and amplitude distribution of each radiating element are weighted and reconstructed using phased array feed network phase synthesis technology to change the electromagnetic wave polarization characteristics, thereby obtaining the reconstructed antenna radiation pattern.

[0120] Based on the calibrated antenna beam pointing vector and the target radio frequency radiated power level, the polarization mode adjustment component in the target five-dimensional antenna parameter combination command is obtained as the initial polarization configuration input.

[0121] The polarization mode is analyzed by adjusting the polarization mode of the input to establish the physical definition parameters of the target electromagnetic wave polarization form, including the polarization angle, polarization type (linear polarization, circular polarization, elliptical polarization) and its rotation direction.

[0122] The calibrated antenna beam pointing vector and the target RF radiated power level are jointly mapped to the radiating element matrix index space of the phased array feed network to determine the set of radiating elements participating in the polarization reconstruction operation and their initial phase and amplitude values.

[0123] Based on phase synthesis technology using a phased array feed network, phase difference and amplitude weighting calculations are performed on the driving signals of the radiating elements. By adjusting the phase offset and amplitude ratio of each element, the superimposed field forms a wavefront in space that conforms to the target polarization characteristics. The following phase weighting calculation formula is used to achieve unified adjustment of the phase of each element: in, Let be the phase value of the i-th radiating element. As a reference phase, This is the global phase offset. This is the phase weighting factor.

[0124] For amplitude adjustment, amplitude normalization and proportional weighting are performed, and the amplitude is configured using the following formula: in, Let i be the amplitude value of the i-th radiating element. For reference range, This is the amplitude gain factor.

[0125] The aforementioned phase and amplitude distributions are applied to the excitation signal generation module of the phased array feed network to form a spatial superposition state of composite polarization wavefronts.

[0126] Based on the spatial superposition results of the wavefront, a reconstructed antenna radiation pattern that conforms to the target polarization characteristics is output, thereby achieving synchronous adjustment of beam shape and polarization mode.

[0127] By using phase synthesis and amplitude weighting processing through a phased array feed network, the beam pointing and power configuration results from the previous step are transformed into an antenna state that has the target polarization characteristics and satisfies the design radiation pattern in space, thereby achieving adaptive polarization adjustment and coverage area optimization.

[0128] S6.5: Based on the reconstructed antenna radiation pattern, the signal strength distribution within the service area is mapped and extrapolated in real time using a spatial electromagnetic field propagation model to verify the coverage blind spot elimination effect and interference suppression level, thereby generating an adaptively adjusted antenna coverage area.

[0129] Step S7: Monitor the configuration frequency distribution of the user distribution manifold spatial point set over multiple consecutive periods. If it is determined that the frequency of at least one preset manifold-like configuration significantly deviates from the historical average, a local manifold topology fine-tuning mechanism is triggered to generate an updated spatiotemporal fingerprint spectral unit mapping relationship. The preset manifold-like configuration refers to typical categories with different spatial distribution patterns, such as uniform or clustered, identified through clustering morphological analysis of the user distribution manifold spatial point set. The configuration frequency distribution is a statistical result of the number of occurrences of various manifold configurations within a preset observation period, obtained by statistical processing using a sliding time window based on the manifold configuration label sequence. Step S7 specifically includes: S7.1: Obtain the user distribution manifold spatial point set within multiple consecutive periods, and perform clustering morphological analysis on the user distribution manifold spatial point set to generate a manifold configuration label sequence representing the current population distribution topology.

[0130] S7.2: Based on the manifold configuration label sequence, the frequency of occurrence of various manifold configurations within a preset observation period is calculated using sliding time window statistical processing to generate a real-time manifold configuration frequency distribution vector.

[0131] S7.3: Perform a deviation measurement operation on the real-time manifold configuration frequency distribution vector and the historical manifold configuration frequency mean benchmark stored in the static prior knowledge base to generate a set of abnormal manifold configurations that significantly deviate from the threshold.

[0132] S7.4: Based on the spatiotemporal fingerprint unit index contained in the abnormal manifold configuration set, extract the affected local spatiotemporal fingerprint map unit subset, and perform mapping weight recalibration processing on the local spatiotemporal fingerprint map unit subset based on the latest low-dimensional communication requirement key semantic feature vector to generate local manifold topology fine-tuning instructions to be updated.

[0133] S7.5: Based on the local manifold topology fine-tuning instruction, perform incremental replacement operations on the mapping relationships of corresponding entries in the original static prior knowledge base to generate updated spatiotemporal fingerprint spectrum unit mapping relationships with dynamic generalization capabilities. The static prior knowledge base is a benchmark database established during initialization to store historical manifold configuration frequency averages and various mapping relationships such as spatiotemporal fingerprint spectrum unit mapping relationships and five-dimensional antenna parameter mapping rule clusters. The static prior knowledge base originates from an initial knowledge set constructed after offline analysis and rule extraction of historical operating data or expert experience, and is continuously optimized through an incremental update mechanism during system online operation.

[0134] Step S8: Replace the corresponding entries in the original static prior knowledge base based on the updated spatiotemporal fingerprint graph unit mapping relationship, complete the online incremental manifold update, and generate the next cycle parameter decision benchmark with dynamic generalization capability. Specifically, this includes: S8.1: Obtain the updated spatiotemporal fingerprint map unit mapping relationship and the original static prior knowledge base generated by the previous steps, and perform indexing and positioning processing on the original static prior knowledge base to identify the old version mapping entries of fingerprint units near the affected area corresponding to the triggering local manifold topology fine-tuning mechanism.

[0135] S8.2: Based on the identified old mapping entries and the updated spatiotemporal fingerprint map unit mapping relationship, perform difference comparison and consistency verification processing to generate a mapping data packet to be written, which includes a unique spatiotemporal signature, a correction value for the human flow density baseline, an offset value for the movement direction entropy value, and an update value for the dwell time distribution characteristics.

[0136] S8.3: Use the mapping data packet to be written to perform atomic replacement operations on the corresponding entries in the original static prior knowledge base, so as to eliminate the conflict between the old and new data versions and generate an incrementally updated version of the static prior knowledge base with the latest spatiotemporal semantic features.

[0137] S8.4: Based on the incrementally updated static prior knowledge base, perform the linkage recalibration process of the global mapping rule cluster to synchronously adjust the affected rule weights in the preset five-dimensional antenna parameter mapping rule cluster, and generate a dynamic generalized parameter decision benchmark that adapts to the current user distribution trend.

[0138] S8.5: Based on the dynamic generalization parameter decision benchmark, perform the next cycle parameter decision ready state flag setting operation to activate the parameter decision module's adaptive response capability to new scenarios, complete the online incremental manifold update closed loop, and output the next cycle parameter decision benchmark with dynamic generalization capability.

[0139] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0140] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0141] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for adaptively adjusting monitoring of smart antenna coverage area, characterized in that, Specifically, it includes: S1: Acquire BeiDou high-precision positioning data and camera video stream, perform gridded mapping processing on the BeiDou high-precision positioning data, and generate spatiotemporal fingerprint map units; S2: Based on the spatial coordinate constraints of the spatiotemporal fingerprint map unit, the human detection box and tracking ID in the camera video stream are mapped to the corresponding spatiotemporal fingerprint map unit to generate a spatiotemporal fingerprint association dataset bound to visual semantics. S3: Perform semantic distillation on the visual semantics in the spatiotemporal fingerprint association dataset to extract the rate of change of crowd density, the offset angle of the main moving axis and the expansion speed of the sparse edge region, and generate a low-dimensional communication requirement key semantic feature vector. S4: Construct a two-dimensional orthogonal coordinate system using the key semantic feature vectors of the low-dimensional communication requirements, encode the semantic state of each spatiotemporal fingerprint unit as a dynamic trajectory point, and generate a user distribution manifold space point set; S5: Calculate the geometric centroid position of the user distribution manifold space point set, and perform a matching query in the preset five-dimensional antenna parameter mapping rule cluster based on the geometric centroid position to generate a target five-dimensional antenna parameter combination instruction; S6: Based on the target five-dimensional antenna parameter combination command, drive the antenna actuator to perform physical parameter adjustment actions on the downtilt angle, azimuth angle, elevation, transmission power and polarization mode to generate an adaptively adjusted antenna coverage area.

2. The method for adaptive adjustment and monitoring of smart antenna coverage area according to claim 1, characterized in that, The camera video stream contains the sequence of center pixel coordinates of the human detection box and its corresponding tracking ID identifier.

3. The method for adaptive adjustment and monitoring of smart antenna coverage area according to claim 2, characterized in that, Step S2 specifically includes: The sequence of center pixel coordinates and tracking ID identifier of the human body detection box after target detection processing in the camera video stream are obtained. Based on the pre-calibrated camera intrinsic parameter matrix and extrinsic parameter rotation and translation matrix, inverse perspective projection transformation is performed to obtain a real-time human body positioning point set. Based on the geographic latitude and longitude information of the real-time human positioning point set and the spatial boundary constraints of the spatiotemporal fingerprint map unit, the spatial inclusion determination process of the point within the polygon is performed to establish the target spatiotemporal fingerprint map unit index to which each real-time human positioning point belongs. Using the target spatiotemporal fingerprint map unit index as the association key, the corresponding human body detection box attribute information and tracking ID identifier are injected into the attribute field of the spatiotemporal fingerprint map unit, and a multi-source data fusion mounting operation is performed to generate a spatiotemporal fingerprint association data entry bound to the visual semantics. Perform time window sliding aggregation processing on the spatiotemporal fingerprint association data entries bound to visual semantics, filter instantaneous false detection noise based on the continuity characteristics of the tracking ID identifier, and generate a spatiotemporal fingerprint association dataset; Based on the spatiotemporal fingerprint association dataset, a data structure serialization and encapsulation operation is performed to package spatial coordinates, spatiotemporal signatures, population size, and movement trajectory features into intermediate data objects, and output the spatiotemporal fingerprint association dataset.

4. The method for adaptive adjustment and monitoring of smart antenna coverage area according to claim 1, characterized in that, Step S3 specifically includes: Based on the historical period human detection box number sequence and the current period human detection box number sequence in the spatiotemporal fingerprint association dataset, a sliding window difference operation is performed to generate the instantaneous change in crowd density, and the instantaneous change in crowd density is normalized and mapped to obtain a standardized rate of change in crowd density. Using the tracking ID coordinate trajectory data of continuous frames in the spatiotemporal fingerprint association dataset, principal component analysis is performed to reduce the dimension, extract the feature vector representing the overall movement trend of the population as the main movement direction reference, and calculate the angle between the main movement direction reference and the preset grid normal direction of the spatiotemporal fingerprint map unit to obtain the main movement axis offset angle. For the distribution set of human detection boxes located in the boundary region of the spatiotemporal fingerprint map unit in the spatiotemporal fingerprint association dataset, convex hull contour extraction and area evolution monitoring are performed to identify the morphological changes of the sparse region covering the edge, and the contour expansion rate per unit time is calculated based on the morphological changes of the sparse region to generate the edge sparse region expansion speed. The rate of change of the crowd density, the offset angle of the main moving axis and the expansion speed of the sparse edge region are obtained. Multi-source heterogeneous feature alignment and weighted fusion processing are performed to construct a composite feature structure. The low-dimensional communication requirement key semantic feature vector is generated by performing vector space encoding processing on the composite feature structure.

5. The smart antenna coverage area adaptive adjustment monitoring method according to claim 4, characterized in that, The normalization mapping process is specifically executed based on the Min-Max linear normalization function.

6. The method for adaptive adjustment and monitoring of smart antenna coverage area according to claim 4, characterized in that, The composite feature structure contains three-dimensional information on density dynamics, flow direction deviation, and boundary extension.

7. The method for adaptive adjustment and monitoring of smart antenna coverage area according to claim 4, characterized in that, Step S4 specifically includes: Based on the rate of change of crowd agglomeration in the key semantic feature vector of the low-dimensional communication demand, a normalized weighted mapping process is performed to generate a scalar value of coverage urgency. Based on the main moving axis offset angle and edge sparse region expansion speed in the key semantic feature vector of the low-dimensional communication requirements, multi-dimensional fuzzy logic fusion operation is performed to generate an adjustment tolerance scalar value. Using the coverage urgency scalar value as the abscissa component and the adjustment tolerance scalar value as the ordinate component, a two-dimensional orthogonal feature coordinate system is constructed, and a geometric projection reference plane is established. For each spatiotemporal fingerprint map unit, the coverage urgency scalar value and the adjustment tolerance scalar value are encapsulated into a coordinate pair, and a dynamic trajectory point instantiation operation is performed to generate a single-point manifold state entity. Aggregate all the single-point manifold state entities generated by the spatiotemporal fingerprint map units, perform set topology assembly processing, and generate the user distribution manifold spatial point set.

8. The method for adaptive adjustment and monitoring of smart antenna coverage area according to claim 7, characterized in that, The normalized weighted mapping process is specifically executed based on the Sigmoid function.

9. The method for adaptive adjustment and monitoring of smart antenna coverage area according to claim 1, characterized in that, After step S6, the following is included: S7: Monitor the configuration frequency distribution of the user distribution manifold spatial point set within multiple consecutive periods. If it is determined that the frequency of at least one preset manifold configuration deviates significantly from the historical average, a local manifold topology fine-tuning mechanism is triggered to generate an updated spatiotemporal fingerprint map unit mapping relationship. S8: Replace the corresponding entries in the original static prior knowledge base based on the spatiotemporal fingerprint graph unit mapping relationship to complete the online incremental manifold update.

10. The method for adaptive adjustment and monitoring of smart antenna coverage area according to claim 1, characterized in that, The target five-dimensional antenna parameter combination command includes downtilt angle value, azimuth angle value, elevation value, transmit power level, and polarization mode identifier.

Images