A low-altitude flight intelligent control information system integrating sensor data

Abstract

This invention discloses a low-altitude flight intelligent control information system integrating integrated sensing data, relating to the fields of communication and low-altitude control technology. It includes an integrated sensing base station, a reference signal receiving unit, a cognitive closed-loop engine, and a control decision module. The cognitive closed-loop engine executes a closed-loop process consisting of S1 reference-driven learning, S2 collaborative active identification, and S3 feedback-driven evolution: S1 utilizes reference signals from cooperative targets for directional acquisition and feature annotation, constructing a target physical feature knowledge base; S2, for non-cooperative targets, plans multi-base station collaborative acquisition tasks to maximize information gain, performs hierarchical matching reasoning on multi-dimensional features, and outputs the model, confidence level, and abnormal features; S3 verifies the identification accuracy based on control and response feedback, dynamically updating the knowledge base, feature validity map, and identification model parameters. This invention enables the system to possess continuous learning and self-optimization capabilities, improving the accuracy of identifying non-cooperative targets in complex low-altitude environments.

Description

Technical Field

[0001] This invention relates to the field of communication and low-altitude airspace control technology, and more specifically, to a low-altitude flight intelligent control information system that integrates sensor data. Background Technology

[0002] The widespread use of low-altitude aircraft has placed higher demands on airspace safety management. Integrated communication and sensing technology, capable of simultaneously achieving communication and sensing functions, has become an important technological direction in the field of low-altitude airspace control. However, existing low-altitude flight control systems still face the following technical bottlenecks when dealing with increasingly complex low-altitude environments: First, low-altitude airspace contains a wide variety of flying targets, including various drones, manned aircraft, and non-cooperative targets such as birds, whose characteristics, flight behaviors, and electromagnetic scattering properties vary significantly. Furthermore, targets may undergo hardware modifications, have added payloads, or new models may emerge, making it difficult for pre-defined static feature databases to cover all possibilities.

[0003] Secondly, the low-altitude electromagnetic environment is complex and variable. Meteorological conditions, ground clutter, and multipath effects can all affect the quality of the sensed signal, causing fluctuations in the characteristics of the same target under different conditions. Traditional recognition methods based on fixed thresholds or static template matching are difficult to adapt to such dynamic changes, and their recognition performance is prone to degradation.

[0004] To address the aforementioned issues, existing technologies have proposed several improvements. For example, Chinese Patent CN121218121A discloses a sensor-integrated fusion system and a method for determining indicators. This system calculates sensing performance indicators, such as false alarm rate and false negative rate, by fusing sensing signals and RFID signals. This solution focuses on post-event performance evaluation using the tag signals of cooperative targets; however, its sensing and recognition processes themselves are still based on preset algorithms and features. It does not describe how to utilize cooperative target signals to construct an iteratively updatable feature knowledge base online, nor does it address proactive sensing task planning mechanisms for non-cooperative targets.

[0005] For example, Chinese patent CN120751398A discloses a sensing-integrated beam optimization method, which adaptively decides whether to introduce an independent sensing beam based on the spatial distribution of sensing targets and communication users. This scheme mainly focuses on the dynamic reuse of communication and sensing resources at the beamforming level to improve resource utilization. However, its core revolves around beam coverage of known or preset targets, and it does not delve into how to construct and utilize a continuously evolving cognitive model to achieve refined identification and anomaly judgment of unknown targets. The system lacks the ability to learn and self-optimize from actual control practices.

[0006] Therefore, how to enable low-altitude flight control systems to have continuous learning capabilities, build feature knowledge bases online using cooperative targets, achieve proactive collaborative perception and deep reasoning for non-cooperative targets, and obtain feedback from each control practice to iteratively optimize their own models, thereby maintaining long-term stable recognition performance in dynamic and complex low-altitude environments, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a low-altitude flight intelligent control information system that integrates sensor data to address the problems of insufficient identification capability, limited environmental adaptability, and rigid system cognitive model that makes continuous optimization difficult when facing non-cooperative, unknown, or modified flight targets in existing low-altitude flight control systems.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A low-altitude flight intelligent control information system integrating sensor data includes: The integrated sensing base station is configured to transmit and receive radio frequency signals; The reference signal receiving unit is configured to receive a reference signal carrying identification information transmitted by the cooperative flight target; The cognitive closed-loop engine is connected to both the integrated sensing base station and the reference signal receiving unit; and The control and decision-making module is connected to the cognitive closed-loop engine; The cognitive closed-loop engine is characterized in that it is configured to perform the following steps: S1. In response to the cooperative target information reported by the reference signal receiving unit, the integrated sensing base station is scheduled to collect directional signals from the corresponding target, extract physical layer feature vectors from the collected signals, bind the feature vectors with the identity information, and associate them with the base station waveform parameters, beam pointing parameters and the relative geometric relationship between the target and the base station at the time of collection, and store them as tagged feature samples in the target physical feature knowledge base. S2. When a non-cooperative target is detected, with the goal of reducing the uncertainty about the target's identity, and constrained by the transmit power and task timing of each integrated sensing base station, an optimization function is constructed and solved to generate a multi-dimensional signal acquisition task sequence consisting of multiple sub-tasks. Multiple integrated sensing base stations are scheduled to collaboratively perform acquisition according to the task sequence, and a multi-dimensional physical layer feature vector of the non-cooperative target is extracted from the acquired multi-dimensional signal. Then, the multi-dimensional feature vector is matched hierarchically with the target physical feature knowledge base. The first layer performs a coarse classification of physical categories based on at least one of the following features: the target's speed of motion and the mean radar cross section. The second layer loads the corresponding matching model based on the coarse classification results, performs fine-grained model matching on the micro-Doppler spectrum features and radio frequency fingerprint features of the target, and outputs the candidate model and the corresponding matching probability as confidence. The third layer calculates the residual vector of the refined model matching, compares the absolute value of the residual of each feature dimension with the dynamic threshold set for that dimension based on the statistical distribution of historical samples of the same model in the knowledge base, and identifies feature dimensions that exceed the threshold as abnormal features. The output includes the identification conclusions, including the model number, confidence level, and abnormal features. S3. Receive control and handling feedback for the non-cooperative target, and use the feedback to verify the accuracy of the identification conclusion output by S2; based on the verification result, update the sample weights in the target physical feature knowledge base or supplement new samples, and update the feature validity map at the same time, and adjust the parameters of the classifier or anomaly detection model used in the hierarchical matching; wherein, the feature validity map is a dynamic data structure that records the discrimination weights of different physical layer features under different acquisition conditions for different physical category targets, and is used to prioritize the planning of feature dimensions with discrimination weights greater than a preset threshold and their corresponding acquisition conditions when generating a multi-dimensional signal acquisition task sequence in S2; The control and management decision module generates control and management instructions based on the identification conclusions.

[0009] Furthermore, the physical layer feature vector includes at least two of the following: radar cross section time series, micro-Doppler spectrum features, and radio frequency fingerprint features.

[0010] Furthermore, each record in the target physical feature knowledge base is associated with and stored as the target's physical layer feature vector, the waveform parameters and beam pointing parameters of the integrated sensing base station when the vector is acquired, and the azimuth and elevation angles between the target and the base station.

[0011] Furthermore, in S2, a multi-dimensional signal acquisition task sequence is generated. Specifically, the task planner in the cognitive closed-loop engine uses the information entropy of reducing the uncertainty of the target identity as the objective function, and uses the available transmission power of each base station and the maximum number of parallel tasks as constraints. It uses a distributed optimization algorithm to solve the problem and outputs a task sequence consisting of multiple sub-tasks. Each sub-task specifies the identifier of the scheduled base station, the signal waveform, the beam direction, and the time window.

[0012] Furthermore, in the hierarchical matching, the first layer performs coarse classification of physical categories based on a statistical learning rule set, which includes judgment rules set according to radar cross section threshold and velocity threshold.

[0013] Furthermore, in the hierarchical matching, the dynamic threshold is set as the interval boundary formed by the mean plus or minus K times the standard deviation, based on the mean and standard deviation of the historical sample numerical distribution of each feature dimension of each known model target in the knowledge base, where K is a preset constant.

[0014] Furthermore, the update rule for the feature validity map is as follows: when an identification is verified as accurate, the discrimination weight of the feature on which the identification depends under the corresponding acquisition conditions is increased; when an identification is verified as incorrect, the discrimination weight of the feature on which the identification depends under the corresponding acquisition conditions is decreased.

[0015] Furthermore, the update of the feature validity map adopts a time decay weighted average algorithm. The new score is the weighted sum of the historical score and the current identification confidence, where the weight of the historical score is the forgetting factor and the weight of the current identification confidence is 1 minus the forgetting factor.

[0016] Furthermore, in S3, the parameters of the classifier or anomaly detection model used for hierarchical matching are adjusted as follows: the multidimensional physical layer feature vector collected this time and the verified real model are used as new samples and added to the training set of the classifier for incremental learning; the standardized residual vector calculated this time and its anomaly label are added to the training set of the anomaly detection model for updating.

[0017] Furthermore, during the system initialization phase, the cognitive closed-loop engine continuously executes S1, using the reference signals emitted by the cooperative flight target to construct the initial data set of the target physical feature knowledge base.

[0018] The technical effects and advantages of this invention are as follows: Compared to existing technologies, this invention establishes a benchmark-driven learning mechanism, enabling the autonomous construction and continuous enrichment of the feature knowledge base. When the benchmark signal receiving unit receives the benchmark signal transmitted by a cooperative target, the cognitive closed-loop engine immediately schedules the integrated sensing base station to collect directional signals from the target, extracting physical layer feature vectors such as radar cross-section time series, micro-Doppler spectrum features, and radio frequency fingerprint features from the echo data. These feature vectors are bound to the identity information in the benchmark signal and associated with the base station waveform parameters, beam pointing parameters, and the relative geometric relationship between the target and the base station at the time of acquisition, stored as tagged feature samples in the target physical feature knowledge base. Through this mechanism, this invention transforms commonly existing cooperative targets in the airspace into continuously updated training data sources, allowing the knowledge base to continuously accumulate feature samples covering different models, flight states, and environmental conditions as the system operates, providing an evolving cognitive foundation for subsequent non-cooperative target identification.

[0019] Compared to existing technologies, this invention establishes a collaborative active identification mechanism, enabling knowledge-guided perception and deep reasoning for non-cooperative targets. When a non-cooperative target is detected, the cognitive closed-loop engine first queries a feature validity map, which records the identification and discrimination weights of different physical layer features under different acquisition conditions for targets of different physical categories. Based on this, the task planning module aims to maximize the information gain for target identity while satisfying base station resource constraints, planning the optimal task sequence for multi-base station collaborative acquisition, and actively selecting the most discriminative feature dimensions and the most favorable observation geometry for acquisition. After acquiring multi-dimensional feature data, the feature matching module performs a three-layer progressive reasoning: the first layer performs coarse classification of physical categories based on motion speed and the mean radar cross section; the second layer loads the corresponding matching model based on the coarse classification results, performs refined model matching on micro-Doppler spectrum features and radio frequency fingerprint features, and outputs candidate models and confidence levels; the third layer calculates the matching residual vector, compares the residuals of each feature dimension with a dynamic threshold set based on historical statistical distribution, and identifies feature dimensions exceeding the threshold as abnormal features. Through the above mechanism, the present invention focuses sensing resources on acquiring information that can best eliminate target uncertainty, while achieving accurate identification of target model and discovery of abnormal features.

[0020] Compared to existing technologies, this invention establishes a feedback-driven evolution mechanism, enabling the self-iteration and continuous optimization of the system's cognitive model. When a control command is executed for a non-cooperative target, the system receives feedback containing the target's actual model number. The feedback learning module compares the current identification conclusion with the actual situation to verify accuracy. Based on the verification result, the system performs at least one of the following update operations: storing the collected feature samples as new samples in the target physical feature knowledge base or adjusting the weights of relevant samples; updating the feature validity map, increasing the discriminative weight of the used features under the corresponding collection conditions if the identification is accurate, and decreasing the corresponding weight if the identification is incorrect; adding the collected feature vectors and the verified actual model number as new samples to the training set of the classifier and anomaly detection model for incremental learning. Through this mechanism, the result of each identification task becomes a driving force for system evolution. The updated knowledge base provides richer reference samples for subsequent stages, the updated feature validity map guides subsequent task planning to make better perceptual decisions, and the updated model makes subsequent matching inference more accurate. Attached Figure Description

[0021] Figure 1 This is a linear structure diagram of the system hardware and data flow of the present invention.

[0022] Figure 2 This is a three-stage linear flowchart of the cognitive closed-loop engine of the present invention.

[0023] Figure 3 This is a hierarchical matching branch diagram for the collaborative identification stage of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1 As attached Figures 1 to 3 The implementation details of the intelligent control and management information system for low-altitude flight, which integrates sensor data, are as follows: I. Overall System Deployment This system is deployed in airspace requiring intelligent control of low-altitude aircraft, such as airport airspace protection zones, urban low-altitude logistics routes, security areas for major events, or protection zones for critical infrastructure. The vertical range of the controlled airspace is set from the ground to a height of 1000 meters, covering the typical operating altitude of low-altitude UAVs.

[0026] The system hardware consists of three parts: A grid-like distributed integrated sensing base station, with the distance between base stations set at 1 to 5 kilometers according to radio frequency coverage requirements, ensures seamless coverage of the controlled airspace; Each reference signal receiving unit is deployed at a base station site or a separate location to receive identification signals actively transmitted by the cooperative aircraft. And a server cluster located in the control center, which runs the core processing engine. All components are interconnected via high-bandwidth, low-latency wired or wireless communication networks.

[0027] II. Integrated Sensor Base Station The integrated sensing and communication base station is equipped with a phased array antenna array with digital beamforming capabilities, enabling simultaneous communication coverage and radar detection of low-altitude targets. The base station operates in the radio frequency band, ranging from 4000MHz to 8000MHz, suitable for low-altitude UAV detection and communication. The base station transmitter can dynamically select signal waveforms from a pre-set waveform library, which includes at least linear frequency modulated continuous wave, pulse Doppler, and orthogonal frequency division multiplexing waveforms to accommodate different detection distances and resolution requirements.

[0028] The base station receiver has a dynamic range of no less than 80dB. Its digital signal processing unit samples and digitizes the received radio frequency signal at a sampling rate of [missing information]. The obtained raw IQ data is then uploaded to the cognitive closed-loop engine. The base station supports simultaneous operation of multiple beams, enabling the detection and tracking of multiple surrounding targets.

[0029] III. Reference Signal Receiving Unit The reference signal receiving unit listens to and decodes broadcast signals containing identification information actively transmitted by low-altitude cooperative aircraft, including registered consumer drones, industrial drones, electric vertical takeoff and landing (eVTOL) aircraft, and manned general aviation aircraft. Its scanned frequency set includes: 1090MHz ADS-B signals for receiving broadcast automatic dependent surveillance information from manned aircraft and some large drones; and 2400MHz and 5800MHz UAV remote identification signals for receiving UAV broadcast identification information compliant with national standards.

[0030] The reference signal receiving unit decodes and verifies the received signal. The verification methods include cyclic redundancy check comparison and identity whitelist comparison to ensure that the received identity information is authentic and reliable.

[0031] After successful verification, the unit generates a first data packet with a uniform format, which includes a timestamp. Identity ID, latitude and longitude ,speed ,course and signal-to-noise ratio And send it to the cognitive closed-loop engine in real time.

[0032] IV. Cognitive Closed-Loop Engine The cognitive closed-loop engine is deployed on a server cluster in the control center, with 128GB of physical memory and 10TB of storage capacity. The engine is implemented in a modular software architecture, with core modules including: signal processing, task planning, feature matching, knowledge base management, and feedback learning.

[0033] The engine executes a closed-loop process in three stages to achieve autonomous learning, intelligent identification, and continuous optimization of low-altitude flying targets: (I) S1 stage: Benchmark-driven learning This phase utilizes cooperative objectives from routine flights within the airspace to build an initial structured feature knowledge base for the system.

[0034] Cognitive closed-loop engine at all times The first data packet reported by the reference signal receiving unit was received. Then, first determine the base station closest to the target. Calculate the target relative to azimuth Pitch angle and straight-line distance Azimuth The range of values ​​is to Pitch angle The range of values ​​is to .

[0035] Subsequently, the engine generates the first configuration instruction. And sent to the base station First configuration command Includes the following parameters: Beam pointing Transmission power Waveform type, center frequency ,bandwidth and collection duration .

[0036] Transmit power Determined by the following formula: ; in, The minimum echo signal-to-noise ratio threshold required by the system is set to 10dB to ensure that the echo signal quality is sufficient to extract fine features. Path loss is calculated using the free space model: ; For the speed of light, take m / s; The noise figure of the receiving system is set to 5 dB.

[0037] bandwidth Based on the desired distance resolution dR, a distance resolution dR of 1m is set to meet the requirement of differentiating the size of the drone: ; Collection time Based on the desired speed resolution dv setting, a speed resolution dv of 0.1 m / s can accurately capture the drone's hovering, low-speed flight, and other states. ; in Wavelength, derived from center frequency Calculated.

[0038] base station According to the first configuration instruction The parameters are transmitted as signals and the echoes are received synchronously to obtain the raw IQ data. And then it is sent back to the cognitive closed-loop engine.

[0039] Signal processing module Perform a two-dimensional Fourier transform to generate a range-Doppler matrix, from which the radar cross section time series of the target unit is extracted. , This represents the sequence length. The radar cross section (RCS) reflects the electromagnetic wave reflection characteristics of a target; different UAV models have different RCS characteristics.

[0040] The signal processing module performs a short-time Fourier transform on the echo to obtain the micro-Doppler spectrum and extracts the preceding signal. The frequency of the main harmonic Amplitude and bandwidth , The value is 5.

[0041] The micro-Doppler spectrum, generated by rotor rotation, is a key feature for identifying UAV type and model. If communication signal components are detected, the feature vectors of their carrier frequency offset (df) and phase noise are extracted. These radio frequency fingerprint features can be used to distinguish different individuals of the same model.

[0042] After normalizing the above features, the signal processing module concatenates them into a first physical layer feature vector. .

[0043] The knowledge base management module will use the first physical layer feature vector With the first data packet The identity identifier (ID) is bound to the data to form labeled training samples, which are then stored together with the data acquisition condition metadata in the association table of the target physical feature knowledge base. middle.

[0044] The metadata for the acquisition conditions includes: transmit power. ,frequency ,bandwidth Waveform parameters, beam pointing ,distance timestamp Environmental index This contextual information allows subsequent matching to be performed under similar conditions, improving recognition accuracy.

[0045] Environmental Index From the temperature at the time of collection ,humidity air pressure calculate: ; Among them, weight These are preset constants. Environmental factors can affect signal propagation and scattering characteristics; recording environmental indices helps to understand the reasons for these characteristic changes.

[0046] Through continuous execution of Phase S1, the system accumulates samples covering multiple models of low-altitude aircraft under different flight states and weather conditions, constructing the initial core of the target physical characteristic knowledge base. The multiple target models include DJI series drones, Autel series drones, various industrial-grade drones, and light sport aircraft such as Cessna, with flight states including hovering, cruise, climb, descent, and maneuvering.

[0047] (II) S2 Phase: Collaborative Active Identification This phase initiates a collaborative sensing and identification process for non-cooperative targets that are not emitting reference signals, such as unidentified aircraft or potential bird interference. This is equivalent to the system using existing knowledge to actively explore and perform deep reasoning about unknown targets.

[0048] When multiple integrated sensing base stations are at time Simultaneously, a first data packet not associated with any data packet was reported. target point At that time, the task planning module is triggered. Target point trace Includes the following parameters: position coordinates velocity components and the mean radar cross section This information, obtained through the base station's standard detection patterns, indicates the presence of a suspicious target.

[0049] The task planning module first queries the Feature Performance Matrix (EM). The Feature Performance Matrix (EM) is a three-dimensional matrix, with each dimension representing the number of target categories. Number of feature types Number of combinations of collection conditions Target categories include rotary-wing drones, fixed-wing drones, birds, and manned aircraft; Feature types include radar cross section sequence, micro-Doppler frequency, and radio frequency fingerprint; The combination of acquisition conditions is obtained by discretizing the waveform, bandwidth, polarization mode, and observation angle. Matrix elements The value indicates the collection conditions. Lower features To distinguish categories The contribution value is initially set to 0.5.

[0050] The task planning module aims to maximize information gain (IG) while satisfying the resource constraints of each base station, and plans collaborative sensing tasks accordingly. The information gain (IG) is calculated as follows: ; Among them, prior entropy , The target belongs to the category The prior probability is set based on historical statistics; Let be the expected posterior entropy. Maximizing information gain means choosing the acquisition scheme that best eliminates the uncertainty of the target.

[0051] The task planning module uses an auction algorithm to solve for the optimal task sequence: broadcasting a set of subtasks to each integrated sensor base station; each base station calculates its bid price based on its available transmission power, maximum number of parallel tasks, and geometric relationship with the target; The task planning module allocates tasks based on the bid price, forming a second configuration instruction set. Second configuration instruction set It contains multiple subtasks, each of which specifies the identifier of the scheduled base station, the signal waveform to be used, the beam direction, and the time window for signal acquisition.

[0052] The task planning module references the Feature Performance Matrix (EM) during planning, prioritizing features with high discriminative power and their corresponding optimal observation geometry. For low-flying UAVs, base stations with small pitch angles are preferred to obtain clearer micro-Doppler features; for high-speed maneuvering targets, the duration of each acquisition session is shortened to avoid feature blurring caused by target movement.

[0053] Taking two subtasks as an example, the second configuration instruction set Include: First subinstruction Send to base station The parameters are: beam pointing beamwidth Pick The waveform is pulse Doppler, and the power is... Duration Narrow beams are used to acquire RCS sequences with high angular resolution.

[0054] Second sub-instruction Send to base station The parameters are: beam pointing beamwidth Pick The waveform is a linear frequency modulated continuous wave with power... Duration The wide beam ensures that the target remains covered during maneuvering and is used for continuous acquisition of the micro-Doppler spectrum.

[0055] base station and Based on the sub-instructions received, each component performs data acquisition and transmits back the raw IQ data. and The signal processing module processes the returned data in parallel and extracts feature vectors. and .

[0056] The feature fusion strategy is as follows: for similar features, a weighted average is applied based on the signal-to-noise ratio to improve feature quality; for complementary features, they are directly concatenated to form a more comprehensive feature description. After fusion, the second physical layer feature vector set is obtained. .

[0057] The feature matching module matches the feature vector set of the second physical layer. Perform hierarchical matching and inference: The first layer is a coarse classification of physical categories. The feature matching module applies a rule set based on statistical learning. Perform a coarse classification. Rule set. Predetermined by statistical learning of historical samples in a target physical feature knowledge base. For example, if the mean radar cross-section... Less than -20dBsm, speed If the velocity is less than 20 m / s and the vertical acceleration is approximately zero, the target's physical category is determined to be "small rotary-wing drone". This layer quickly eliminates obviously irrelevant categories, narrowing down the scope for subsequent fine-tuning.

[0058] The second layer is refined model matching. The feature matching module matches the physical categories output from the first layer. Load the corresponding dedicated classifier Dedicated classifier A support vector machine model is used, utilizing all the target physical feature knowledge bases. The sample training is used to distinguish different models within the same category, such as DJI Mavic 3 and DJI Air 2S.

[0059] The second physical layer feature vector set Input-specific classifier The model probability distribution is obtained. Take the model with the highest probability. The highest probability value is used as the confidence level to represent the most likely outcome. .

[0060] The third layer is anomaly feature analysis. The feature matching module retrieves the model number from the target physical feature knowledge base. For samples collected under similar conditions, calculate their characteristic mean. With covariance matrix Calculate the set of feature vectors for the second physical layer. fusion vector Mahalanobis distance: ; Mahalanobis distance comprehensively measures the deviation of the feature vector from the standard model, and calculates the standardized residuals of each dimension: ; in For the first The mean of the dimensional features, The standard deviation is the standard deviation. The standardized residuals reveal which specific feature dimension exhibits an anomaly.

[0061] Dynamic threshold Based on historical statistical distribution: For each known model target in the target physical feature knowledge base, statistically analyze the numerical distribution of each feature dimension in historical samples and calculate the mean. with standard deviation ,set up , The preset constant is set to 3.

[0062] If the absolute value of the standardized residual in a certain dimension Greater than If the feature in that dimension is an anomaly, then the anomaly list will be output. For example, if the micro-Doppler frequency characteristics of a drone are abnormal, it may indicate that the drone has had different propellers replaced or a load has been added.

[0063] The feature matching module ultimately generates the identification results. Identification results Includes: Target Temporary Identifier (TID), Physical Category ,model Confidence level Exception list and the identification results Sending the data to the control and decision-making module is equivalent to the system providing an "identity assessment" and "abnormal alert" for the target.

[0064] (III) S3 Stage: Feedback-Driven Evolution This phase utilizes feedback information from control and management processes to drive iterative optimization of system knowledge, models, and strategies, enabling the system to learn from each practice.

[0065] Once a control instruction is executed targeting a non-cooperative target, such as law enforcement officers forcing a drone to land and verifying its identity, the system receives feedback on the control and handling results regarding the target's TID (Target Information ID). Feedback on the results of control and disposal Includes verified target model number And descriptions of modifications, such as "Upon investigation, the drone is a ** brand, but a throwing device has been added."

[0066] The feedback learning module will identify the results. Feedback on control and disposal results Compare: If the model With the actual model Consistent, and the exception list If the identification matches the modification description, the identification is considered valid; otherwise, the identification is considered to be flawed.

[0067] Based on the validation results, the feedback learning module performs update operations at the following three levels: The first level involves updating the target physical feature knowledge base. If the identification is successful, the knowledge base management module increases the weight of relevant samples in the knowledge base. : ; in The learning rate is set to 0.1, indicating that successfully identified samples have higher priority in future matching. The fused vector collected in this instance is also used... With the actual model New samples are added to the target physical feature knowledge base to cover more sample variations. If there are biases in the identification, the sample weights are adjusted or supplementary samples are added and corrected according to the actual situation.

[0068] The second level involves updating the feature performance matrix. The feedback learning module updates the feature performance matrix EM with the data corresponding to the current acquisition conditions. and key features The scoring uses a time-decay weighted average algorithm: ; in The forgetting factor, with a value ranging from 0.9 to 0.99, controls the weight of historical experience and new experience; For this utility, take the confidence level. If the identification is accurate, the score of the feature used under the corresponding conditions will be increased; if the identification is incorrect, the corresponding score will be decreased.

[0069] The third step involves updating the classifier and anomaly detection model. This involves fusing the collected vectors... Compared with the verified real model As a new sample, it is added to a dedicated classifier. Incremental learning is performed on the training set to adapt the classifier to new model variants. The standardized residual vector calculated in this study is then used... The abnormal labels are added to the training set of the anomaly detection model for updating. The anomaly detection model uses a single-class support vector machine, which enables it to identify more types of anomaly patterns.

[0070] Through synchronous updates at these three levels, the system completes the process of extracting experience and optimizing itself from a specific practice. The updated knowledge base provides richer reference samples for the subsequent S2 stage, the updated feature performance matrix guides subsequent task planning to make smarter perceptual decisions, and the updated model makes subsequent matching inference more accurate.

[0071] V. Control and Decision Module The control and decision-making module receives the identification results output by the cognitive closed-loop engine. It makes decisions based on the built-in low-altitude airspace management rules and generates corresponding control instructions.

[0072] The control decision module first calculates the threat score. Threat score Trajectory). Threat assessment function Taking the following factors into consideration: Inherent risks of the target model For example, large drones are riskier than micro drones; Confidence When the confidence level is too low, a conservative strategy should be adopted. Exception list The presence of unusual features may indicate that the target has been modified or has malicious intent; Distance from the boundary of the no-fly zone The closer to the no-fly zone, the higher the risk; The target's current trajectory, and whether it is flying towards a sensitive area. Inherent risks associated with the target model. The assignment of values ​​references the drone classification standards and historical flight data published by the Civil Aviation Administration of China. Threat Score The output range is 0 to 100.

[0073] The control decision module is based on threat scores. The system queries the contingency plan database and generates control instructions that comply with airspace management regulations. The contingency plan database includes: like If so, a log recording instruction is generated, and the target information is stored in the blacklist database for future reference. like Then, a command is generated to send a radio warning to the target, and the warning information is sent to the target through Automatic Dependent Surveillance-Broadcast or Remote Identification Protocol for Unmanned Aerial Vehicles. like This generates a command to perform directional radio frequency interference on the target, with interference parameters... According to the target model The corresponding common communication and navigation frequency bands and preset interference effectiveness settings include interference methods such as blocking image transmission links or navigation signals; like If so, it generates an instruction authorizing the use of higher-level countermeasures, such as notifying law enforcement officers to rush to the scene or activating electromagnetic capture equipment.

[0074] Control commands are issued to the corresponding integrated sensing base stations or dedicated countermeasure equipment for execution, and the execution results are recorded for subsequent evaluation.

[0075] VI. System Initialization and Closed-Loop Operation During the system initialization phase, the cognitive closed-loop engine continuously executes Phase S1, utilizing the reference signals continuously emitted by cooperative flight targets that are normally present in the airspace to drive the acquisition and feature annotation of signals from various low-altitude aircraft, constructing the initial dataset for the target physical feature knowledge base. When the number of samples for the main categories in the target physical feature knowledge base reaches more than 100, it indicates that the system has sufficient prior knowledge, and Phase S2 begins normalized operation.

[0076] In daily operation, Phase S1 runs continuously, constantly enriching the knowledge base with newly emerging cooperative targets; whenever a non-cooperative target is detected, Phase S2 is triggered to perform collaborative identification; after each control and response, Phase S3 uses feedback information to optimize the system. The identification results output by Phase S2... The control and management decision-making module generates control and management instructions; the control and management results generated after the control and management instructions are executed are fed back. The S3 phase is triggered; the optimization results of the target physical feature knowledge base, feature performance matrix, classifier, and anomaly detection model in the S3 phase are directly fed into the task planning and matching reasoning process in the subsequent S2 phase.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-altitude flight intelligent control information system integrating sensor data, characterized in that, include: The integrated sensing base station is configured to transmit and receive radio frequency signals; The reference signal receiving unit is configured to receive a reference signal carrying identification information transmitted by the cooperative flight target; The cognitive closed-loop engine is connected to the integrated sensing base station and the reference signal receiving unit, respectively. as well as The control and decision-making module is connected to the cognitive closed-loop engine; The cognitive closed-loop engine is characterized in that it is configured to perform the following steps: S1. In response to the cooperative target information reported by the reference signal receiving unit, the integrated sensing base station is scheduled to collect directional signals from the corresponding target, extract physical layer feature vectors from the collected signals, bind the feature vectors with the identity information, and associate them with the base station waveform parameters, beam pointing parameters and the relative geometric relationship between the target and the base station at the time of collection, and store them as tagged feature samples in the target physical feature knowledge base. S2. When a non-cooperative target is detected, with the goal of reducing the uncertainty about the target's identity, and constrained by the transmit power and task timing of each integrated sensing base station, an optimization function is constructed and solved to generate a multi-dimensional signal acquisition task sequence consisting of multiple sub-tasks. Multiple integrated sensing base stations are scheduled to collaboratively perform acquisition according to the task sequence, and a multi-dimensional physical layer feature vector of the non-cooperative target is extracted from the acquired multi-dimensional signal. Then, the multi-dimensional feature vector is matched hierarchically with the target physical feature knowledge base. The first layer performs a coarse classification of physical categories based on at least one of the following features: the target's speed of motion and the mean radar cross section. The second layer loads the corresponding matching model based on the coarse classification results, performs fine-grained model matching on the micro-Doppler spectrum features and radio frequency fingerprint features of the target, and outputs the candidate model and the corresponding matching probability as confidence. The third layer calculates the residual vector of the refined model matching, compares the absolute value of the residual of each feature dimension with the dynamic threshold set for that dimension based on the statistical distribution of historical samples of the same model in the knowledge base, and identifies feature dimensions that exceed the threshold as abnormal features. The output includes the identification conclusions, including the model number, confidence level, and abnormal features. S3. Receive control and handling feedback for the non-cooperative target, and use the feedback to verify the accuracy of the identification conclusion output by S2; based on the verification result, update the sample weights in the target physical feature knowledge base or supplement new samples, and update the feature validity map at the same time, and adjust the parameters of the classifier or anomaly detection model used in the hierarchical matching; wherein, the feature validity map is a dynamic data structure that records the discrimination weights of different physical layer features under different acquisition conditions for different physical category targets, and is used to prioritize the planning of feature dimensions with discrimination weights greater than a preset threshold and their corresponding acquisition conditions when generating a multi-dimensional signal acquisition task sequence in S2; The control and management decision module generates control and management instructions based on the identification conclusions.

2. The low-altitude flight intelligent control information system integrating sensor data as described in claim 1, characterized in that, The physical layer feature vector includes at least two of the following: radar cross section time series, micro-Doppler spectrum features, and radio frequency fingerprint features.

3. The low-altitude flight intelligent control information system integrating sensor data as described in claim 1, characterized in that, Each record in the target physical feature knowledge base is associated with and stored as the target's physical layer feature vector, the waveform parameters and beam pointing parameters of the integrated sensing base station when the vector is acquired, and the azimuth and elevation angles between the target and the base station.

4. The low-altitude flight intelligent control information system integrating sensor data as described in claim 1, characterized in that, In S2, a multi-dimensional signal acquisition task sequence is generated. Specifically, the task planner in the cognitive closed-loop engine takes reducing the information entropy of uncertainty about the target identity as the objective function, and uses the available transmission power of each base station and the maximum number of parallel tasks as constraints. It uses a distributed optimization algorithm to solve the problem and outputs a task sequence consisting of multiple sub-tasks. Each sub-task specifies the identifier of the scheduled base station, the signal waveform, the beam direction, and the time window.

5. The low-altitude flight intelligent control information system integrating sensor data according to claim 1, characterized in that, In the hierarchical matching, the first layer performs coarse classification of physical categories based on a statistical learning rule set, which includes judgment rules set according to radar cross section threshold and velocity threshold.

6. The low-altitude flight intelligent control information system integrating sensor data according to claim 1, characterized in that, In the hierarchical matching, the dynamic threshold is set as the interval boundary formed by the mean plus or minus K times the standard deviation, based on the mean and standard deviation of the historical sample values ​​of each feature dimension of each known model target in the knowledge base, where K is a preset constant.

7. The low-altitude flight intelligent control information system integrating sensor data according to claim 1, characterized in that, The update rule for the feature validity map is as follows: when an identification is verified as accurate, the discrimination weight of the feature on which the identification depends under the corresponding acquisition conditions is increased; when an identification is verified as incorrect, the discrimination weight of the feature on which the identification depends under the corresponding acquisition conditions is decreased.

8. The low-altitude flight intelligent control information system integrating sensor data according to claim 7, characterized in that, The feature validity map is updated using a time decay weighted average algorithm. The new score is the weighted sum of the historical score and the current identification confidence, where the weight of the historical score is the forgetting factor and the weight of the current identification confidence is 1 minus the forgetting factor.

9. The low-altitude flight intelligent control information system integrating sensor data according to claim 1, characterized in that, In step S3, the parameters of the classifier or anomaly detection model used for hierarchical matching are adjusted as follows: the multi-dimensional physical layer feature vector collected this time and the verified real model are used as new samples and added to the training set of the classifier for incremental learning; the standardized residual vector calculated this time and its anomaly label are added to the training set of the anomaly detection model for updating.

10. The low-altitude flight intelligent control information system integrating sensor data according to claim 1, characterized in that, During the system initialization phase, the cognitive closed-loop engine continuously executes S1, using the reference signals emitted by the cooperative flight target to construct the initial data set of the target physical feature knowledge base.

Images