A method and system for recognizing flight intention of a flight target based on common sense data

Abstract

The application provides a flight target flight intention recognition method and system based on common sense data, the method comprising: acquiring a historical trajectory point set of a flight target and real-time common sense data, the real-time common sense data comprising a real-time speed sequence, a wireless signal strength, a wireless signal frequency band and a current position coordinate of the flight target; calculating a behavior threat score according to the real-time speed sequence and the historical trajectory point set; calculating a signal threat score according to the wireless signal strength and the wireless signal frequency band; calculating a geofencing risk score according to the current position coordinate; constructing an observation state set of the flight target according to the behavior threat score, the signal threat score, the geofencing risk score and the real-time common sense data and inputting the observation state set into a preset hidden Markov model, so that the hidden Markov model generates a flight intention of the flight target at a current time through probabilistic reasoning, and the accuracy of flight intention recognition of the flight target is improved.

Description

Technical Field

[0001] This application relates to the field of low-altitude safety technology for power grids, and in particular to a method and system for recognizing the flight intention of a flying target based on sensor data. Background Technology

[0002] Traditional low-altitude safety monitoring of power grids primarily relies on a combination of radar, video surveillance, and manual inspections. However, civilian drones, characterized by their low altitude, slow speed, and small radar cross-section, are prone to escaping the effective detection range of traditional radars in complex electromagnetic environments or densely built-up areas. This makes them difficult to locate and track accurately, posing real safety risks such as collisions and deliberate damage to power transmission lines, substations, and other power facilities. Furthermore, while dedicated low-altitude radars offer strong detection capabilities, they suffer from high transmission power, limited deployment in residential areas, and insufficient vertical coverage, making it difficult to meet the wide-area, continuous sensing needs of the power grid. Manual inspections and ordinary video surveillance, on the other hand, suffer from inefficiency, are greatly affected by environmental and lighting conditions, and experience communication delays leading to analysis lags, failing to meet the high standards required for real-time, accurate behavioral analysis and intent determination of low-altitude moving targets.

[0003] 5G-Advanced (5G-A) introduces revolutionary technologies such as Integrated Sensing and Communication (ISAC), enabling high-precision sensing of target distance, speed, angle, and trajectory using communication signals themselves. Leveraging its high bandwidth, low latency, and high reliability, it provides a new technological foundation for real-time dynamic data acquisition and transmission of low-altitude targets for power grids. However, current applications based on 5G-A sensing data are mostly focused on target detection and trajectory tracking. Facing the increasing demands for low-altitude security for power grids, there is still a lack of an intelligent method that can deeply integrate multi-dimensional sensing data, quantitatively assess behavioral anomalies, and ultimately dynamically infer the target's potential flight intentions. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a method and system for recognizing the flight intention of a flight target based on synesthesia data, thereby improving the accuracy of recognizing the flight intention of a flight target.

[0005] In a first aspect, embodiments of this application provide a method for recognizing the flight intention of a flight target based on synesthetic data, including: Acquire the historical trajectory point set of the flight target and real-time sensing data within a preset time period. The real-time sensing data includes the real-time speed sequence, wireless signal strength, wireless signal frequency band, and current position coordinates of the flight target. The behavioral threat score of the flight target is calculated based on the real-time velocity sequence and the historical trajectory point set. The signal threat score of the flying target is calculated based on the wireless signal strength and the wireless signal frequency band. The geofencing risk score of the flight target is calculated based on the current location coordinates; Based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensor data, a set of observation states for the flight target is constructed. The set of observed states is input into a preset hidden Markov model, so that the hidden Markov model generates the second hidden state of the flight target at the current moment through probabilistic deduction based on the current set of observed states and the first hidden state of the flight target at the previous moment, and then uses the second hidden state as the flight intention of the flight target at the current moment.

[0006] This application provides a method for identifying the flight intent of a flight target based on synesthetic data. First, it acquires multi-dimensional real-time data and historical trajectories of the flight target, and calculates threat scores from three independent dimensions: behavior, signal, and geofencing. This achieves a multi-faceted, quantitative assessment of the abnormal state of the flight target. Subsequently, this embodiment innovatively fuses these multi-dimensional quantitative scores with the original synesthetic data to construct an observation state set that comprehensively reflects the instantaneous state of the target. Finally, the observation state is input into a preset Hidden Markov Model (HMM). Utilizing the probabilistic inference capability of the HMM, combined with the target's previous intent state, the most probable flight intent at the current moment is inferred. Compared to traditional methods relying solely on single radar data or simple threshold alarms, the technical solution provided in this application, through deep fusion of multi-source heterogeneous data and dynamic inference based on probabilistic models, can more comprehensively and accurately understand the behavioral patterns and potential threats of flight targets. This effectively reduces false alarms and missed alarms caused by incomplete data or rigid rules, significantly improving the accuracy and intelligence level of identifying low-altitude security threats to the power grid, and providing a reliable decision-making basis for proactive defense and precise countermeasures of the power grid.

[0007] Furthermore, the step of calculating the behavioral threat score of the flight target based on the real-time velocity sequence and the historical trajectory point set includes: Based on the real-time velocity sequence, calculate the velocity change value of the flight target at each moment within the preset time period, and determine the maximum velocity change value from each velocity change value; The velocity mutation value of the flight target is calculated based on the maximum velocity change value and the preset normalized velocity. The probability of the flight target staying in each preset area is determined by analyzing the historical trajectory point set. Then, the trajectory entropy of the flight target is calculated based on each of the staying probabilities and the weighting factors corresponding to each preset area. The behavioral threat score of the flight target is calculated based on the velocity mutation value and the trajectory entropy.

[0008] This application further defines the calculation method for behavioral threat scoring. By analyzing real-time speed sequences, the "maximum speed change value" is extracted to quantify abnormal speed changes, effectively capturing suspicious behaviors such as sudden acceleration, sharp turns, or evasive maneuvers by the flight target. Simultaneously, by analyzing historical trajectory point sets to calculate "trajectory entropy," the randomness and unpredictability of flight paths can be quantified, thereby identifying unplanned flight path behaviors such as loitering and detours. Combining speed change values ​​reflecting instantaneous maneuvering anomalies with trajectory entropy reflecting long-term behavioral patterns constitutes the behavioral threat score, making the assessment of flight behavior no longer a single, instantaneous judgment, but a comprehensive evaluation that considers both short-term dynamic anomalies and long-term pattern anomalies. This assessment method can effectively distinguish between normal inspections (fixed flight path, stable speed) and potential threat behaviors (chaotic trajectory, speed changes), enhancing the depth and reliability of threat perception at the behavioral level and providing more robust and granular behavioral feature input for subsequent intent determination.

[0009] In one possible implementation, calculating the signal threat score of the flight target based on the wireless signal strength and the wireless signal frequency band includes: The first signal threat index of the flight target is determined by comparing the wireless signal strength with a preset signal strength threshold. Determine whether the wireless signal frequency band is within a preset list of illegal frequency bands, and determine the second signal threat index of the flight target based on the determination result; The signal threat score of the flight target is calculated based on the first signal threat index and the second signal threat index.

[0010] This application specifies the calculation method for signal threat scoring, integrating two key indicators—Radio Signal Strength Indicator (RSSI) and signal frequency band—for threat determination. By comparing real-time signal strength with a preset threshold, abnormal signal enhancement (potentially indicating relaying or close-range manipulation) or abnormal signal weakening (potentially an attempt to evade detection) can be identified. Simultaneously, by checking whether the signal frequency band belongs to a preset list of illegal frequency bands, "black flight" drones using unauthorized frequency bands for communication can be directly identified. These two indicators are fused through logical judgment to generate multi-level signal threat scores, achieving refined hierarchical assessment of abnormal communication signals. Compared to traditional methods that rely solely on signal strength or only perform frequency band blacklist / whitelist filtering, this embodiment's dual-judgment mechanism significantly improves the accuracy and robustness of identifying illegal or malicious drone signals, effectively combating simple signal camouflage or power modulation, and providing reliable signal-dimensional evidence for subsequent intent identification.

[0011] Furthermore, the step of calculating the geofencing risk score of the flight target based on the current location coordinates includes: Calculate the Euclidean distances between the flight target and each boundary coordinate point in the preset electronic fence boundary based on the current position coordinates, and determine the shortest Euclidean distance. The shortest Euclidean distance is input into a preset inverse proportional function to calculate the geofencing risk score of the flight target.

[0012] This application defines a method for calculating geofence risk scores. The risk score is calculated by substituting the shortest Euclidean distance from the current location of a flying target to the boundary of the geofence into an inverse proportional function. The advantage of this method lies in its intuitive and dynamically continuous evaluation results: when the target is infinitely far from the no-fly zone, the score approaches zero, indicating negligible risk; as the target approaches or even touches the geofence boundary, the score increases sharply and even approaches infinity (in practice, a very small constant is usually introduced to avoid division by zero), thus clearly reflecting the security principle of "the closer the distance, the higher the risk" in a quantitative form. This risk scoring model based on continuous distance provides better early warning of potential intrusion risks than a simple "enter / not enter" Boolean judgment and quantifies and classifies the risks at different distances. This enables the security system to achieve risk-sensitive adaptive responses, improving the accuracy of subsequent identification of the flying target's flight intent.

[0013] In one possible implementation, constructing the observation state set of the flight target based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensor data includes: Based on a preset first weighting coefficient, the behavioral threat score, the signal threat score, and the geofence risk score are weighted and summed to obtain a multimodal feature fusion threat score; Based on the real-time velocity sequence in the real-time sensing data, the velocity change vector of the flying target at the current moment is calculated; The flight altitude of the flight target at the current moment is determined based on the current position coordinates in the real-time sensing data; Based on the wireless signal strength and frequency band in the real-time sensing data, the signal characteristics of the flying target at the current moment are determined; By combining the multimodal feature fusion threat score, the velocity mutation vector, the flight altitude, and the signal features, an observation state set of the flight target is constructed.

[0014] This application's embodiments specifically define the construction method of the observation state set. Instead of simply inputting threat scores from various dimensions directly into the model, it employs an innovative two-stage fusion strategy. In the first stage, threat scores from the behavioral, signal, and geographical dimensions are fused into a comprehensive "multimodal feature fusion threat score" through weighted summation. This score provides a macroscopic quantification of the overall anomaly level of the target. In the second stage, this comprehensive score is combined with more physically meaningful micro-features (such as velocity mutation vectors, flight altitude, and signal features) directly extracted from the original synesthetic data to jointly constitute the observation state set. This construction method combines high-level semantic information (fusion score) and low-level physical features (original data features), providing the Hidden Markov Model with extremely rich and comprehensive observation input. The high-level score provides a rapid judgment of threat level, while the low-level features retain detailed information distinguishing different intent categories, enabling the model to perform probability matching and intent inference more accurately, significantly improving the accuracy of intent classification.

[0015] Furthermore, the flight target flight intention recognition method also includes: before calculating the multimodal feature fusion threat score, calculating the cumulative trajectory offset distance of the flight target based on the historical trajectory point set and the preset reference inspection route trajectory sequence, and weighting and summing the behavioral threat score, the signal threat score, the geofence risk score and the cumulative trajectory offset distance according to the preset second weight coefficient to obtain the multimodal feature fusion threat score.

[0016] This embodiment additionally incorporates the key feature of "cumulative trajectory deviation distance" into the calculation of the multimodal feature fusion threat score. Whether the trajectory deviates from the preset legitimate inspection route is one of the most direct and powerful pieces of evidence for judging the legitimacy of the flight target's intent. This embodiment calculates the trajectory deviation distance through a weighted dynamic time warping algorithm, which can effectively align and compare the real-time trajectory with the reference trajectory in the time dimension. Even if there are differences in speed and pauses, the overall deviation can be accurately calculated. This quantified trajectory anomaly index is then weighted and fused with the threat scores of the other three dimensions, so that the final "multimodal feature fusion threat score" includes information from the four core dimensions of behavior, signal, geography, and trajectory. This improvement makes feature fusion more comprehensive, especially enhancing the detection capability of covert threats such as "disguised normal flight but actually deviating from the flight path." An aircraft exhibiting normal speed and normal signal but with a severely deviated trajectory will also have its threat score increased due to the contribution of the trajectory dimension, thus being identified as an anomaly by the system earlier and more accurately. This compensates for the deficiencies of single or a few dimensional features, further improving the reliability and comprehensiveness of the entire intent recognition system.

[0017] In one possible implementation, the step of inputting the set of observed states into a preset Hidden Markov Model (HMM) so that the HMM generates the second hidden state of the flight target at the current moment through probabilistic deduction based on the current set of observed states and the first hidden state of the flight target at the previous moment, includes: The set of observed states is input into a preset hidden Markov model so that the hidden Markov model determines the observation probability of various hidden states of the flight target at the current moment based on the current set of observed states and the preset observation probability matrix. The observation probability matrix is ​​obtained by modeling based on a Gaussian mixture model. Based on the first hidden state, the prior probabilities corresponding to various hidden states of the flight target at the current moment are calculated through the current state transition probability matrix, wherein the current state transition probability matrix is ​​dynamically updated based on the relative distance between the flight target and the nearest power facility; The second hidden state of the flight target at the current moment is determined based on each of the observed probabilities and each of the prior probabilities.

[0018] This application's embodiments specifically disclose the internal mechanism of Hidden Markov Model (HMM) intention reasoning, clarifying the intention reasoning process as a combination of two probabilistic pieces of information: first, the "observation probability" calculated based on the current set of observed states and the observation probability matrix (modeled by a Gaussian mixture model), which reflects the likelihood of the target being in each hidden intention state under the various observed features; second, the "prior probability" calculated based on the intention state at the previous moment and the dynamically updated state transition probability matrix, which reflects the trend of the target transitioning to each intention state after considering historical states and the current environmental context (reflected in the dynamic transition probability). Finally, the most likely intention state is determined by combining these two probabilities. The advantages of this mechanism are: first, it utilizes the inherent advantage of probabilistic models in handling uncertainty and noise, making the judgment result more robust; second, it combines historical state information, considering the continuity of intention evolution and avoiding the irrationality of state jumps; finally, and most importantly, its state transition probability is dynamically updated, which allows the model reasoning to closely integrate with the current real-time context for intention recognition, realizing context-aware intelligent reasoning, thereby significantly improving the accuracy and context adaptability of intention recognition in complex and dynamic power grid environments.

[0019] Furthermore, the step of dynamically updating the current state transition probability matrix based on the relative distance between the flight target and the nearest power facility includes: Based on the current position coordinates of the flight target and the facility coordinates of each power facility in the target area, determine the relative distance between the flight target and the nearest power facility; Based on the preset ground state transition probability, the facility sensitivity coefficient corresponding to the nearest power facility, and the relative distance, the first prior probability corresponding to the preset type hidden state in the previous state transition probability matrix is ​​updated to obtain the current state transition probability matrix.

[0020] This application further defines the specific method for dynamically updating the state transition probability matrix. Traditional Hidden Markov Models (HMMs) typically have static state transition probabilities, failing to reflect the direct impact of external environmental changes on intent evolution. This embodiment creatively embeds a spatial risk model of power grid security into the probability transition process. Specifically, it dynamically adjusts the probability of transitioning to a high-risk state (such as an "attack" state) in the model based on the real-time distance between the flight target and the nearest power facility, and the facility's preset sensitivity coefficient. The closer the distance and the more sensitive the facility, the higher the transition probability. This design has significant beneficial effects: it transforms the intent recognition model from a closed, static entity into a dynamic perception-reasoning system closely linked to the physical world. The model understands that the spatial event of "approaching a substation" itself implies an increased probability of threatening intent, thus enabling it to raise the alert level before the target exhibits strong aggressive behavior. This prior probability adjustment based on geospatial risk greatly enhances the system's predictability of potential attacks, achieving a leap from "passive response based on behavior" to "proactive prediction combined with environmental risk," further improving the accuracy of flight intent recognition for flight targets.

[0021] Secondly, embodiments of this application provide a flight target flight intention recognition system based on synesthetic data, including an acquisition module, a first evaluation module, a second evaluation module, a third evaluation module, an observation module, and a recognition module; The acquisition module is used to acquire the historical trajectory point set of the flight target and the real-time sensing data within a preset time period. The real-time sensing data includes the real-time speed sequence, wireless signal strength, wireless signal frequency band, and current position coordinates of the flight target. The first evaluation module is used to calculate and obtain the behavioral threat score of the flight target based on the real-time speed sequence and the historical trajectory point set; The second evaluation module is used to calculate the signal threat score of the flying target based on the wireless signal strength and the wireless signal frequency band. The third assessment module is used to calculate the geofence risk score of the flight target based on the current location coordinates; The observation module is used to construct an observation status set of the flight target based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensor data. The identification module is used to input the set of observed states into a preset hidden Markov model, so that the hidden Markov model generates the second hidden state of the flight target at the current moment through probabilistic deduction based on the current set of observed states and the first hidden state of the flight target at the previous moment, and then uses the second hidden state as the flight intention of the flight target at the current moment.

[0022] Furthermore, the identification module includes an observation probability determination unit, a priori probability determination unit, and an identification unit, comprising: The observation probability determination unit is used to input the set of observation states into a preset hidden Markov model, so that the hidden Markov model determines the observation probability of various hidden states of the flight target at the current moment based on the current set of observation states and the preset observation probability matrix. The observation probability matrix is ​​obtained by modeling based on a Gaussian mixture model. The prior probability determination unit is used to calculate the prior probability of each hidden state of the flight target at the current moment based on the first hidden state and the current state transition probability matrix. The current state transition probability matrix is ​​dynamically updated based on the relative distance between the flight target and the nearest power facility. The identification unit is used to determine the second hidden state of the flight target at the current moment based on each of the observation probabilities and each of the prior probabilities. Attached Figure Description

[0023] Figure 1 A flowchart illustrating a method for recognizing the flight intent of a flight target based on synesthetic data, provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a flight target flight intention recognition system based on synesthesia data provided in an embodiment of this application. Detailed Implementation

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

[0025] It should be noted that the step numbers in this document are only for the convenience of explaining the specific embodiments and are not intended to limit the order in which the steps are performed. In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0026] Example 1: like Figure 1 As shown, Embodiment 1 provides a method for recognizing the flight intention of a flight target based on synesthetic data, including steps S1-S6: Step S1: Obtain the historical trajectory point set of the flight target and the real-time sensing data within a preset time period. The real-time sensing data includes the real-time speed sequence, wireless signal strength, wireless signal frequency band, and current position coordinates of the flight target. Step S2: Calculate the behavioral threat score of the flight target based on the real-time velocity sequence and the historical trajectory point set; Step S3: Calculate the signal threat score of the flying target based on the wireless signal strength and the wireless signal frequency band; Step S4: Calculate the geofencing risk score of the flight target based on the current location coordinates; Step S5: Construct the observation state set of the flight target based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensor data; Step S6: Input the set of observed states into a preset hidden Markov model, so that the hidden Markov model generates the second hidden state of the flight target at the current moment through probabilistic deduction based on the current set of observed states and the first hidden state of the flight target at the previous moment, and then uses the second hidden state as the flight intention of the flight target at the current moment.

[0027] This application provides a method for identifying the flight intent of a flight target based on synesthetic data. First, it acquires multi-dimensional real-time data and historical trajectories of the flight target, and calculates threat scores from three independent dimensions: behavior, signal, and geofencing. This achieves a multi-faceted, quantitative assessment of the abnormal state of the flight target. Subsequently, this embodiment innovatively fuses these multi-dimensional quantitative scores with the original synesthetic data to construct an observation state set that comprehensively reflects the instantaneous state of the target. Finally, the observation state is input into a preset Hidden Markov Model (HMM). Utilizing the probabilistic inference capability of the HMM, combined with the target's previous intent state, the most probable flight intent at the current moment is inferred. Compared to traditional methods relying solely on single radar data or simple threshold alarms, the technical solution provided in this application, through deep fusion of multi-source heterogeneous data and dynamic inference based on probabilistic models, can more comprehensively and accurately understand the behavioral patterns and potential threats of flight targets. This effectively reduces false alarms and missed alarms caused by incomplete data or rigid rules, significantly improving the accuracy and intelligence level of identifying low-altitude security threats to the power grid, and providing a reliable decision-making basis for proactive defense and precise countermeasures of the power grid.

[0028] In a preferred embodiment, in step S1, a 5G-A sensing network and control platform is constructed. Based on this platform, real-time sensing data and historical trajectory point sets of low-altitude targets within the power grid are acquired. Specifically, the 5G-A sensing network and control platform includes a 5G-A integrated sensing terminal module and a low-altitude control platform. The 5G-A integrated sensing terminal module acquires the real-time coordinates, speed, and altitude information of low-altitude targets within the power grid and feeds it back to the low-altitude control platform. These low-altitude targets include, but are not limited to, drones. The core functions of the 5G-A sensing network and control platform include real-time monitoring of the drone's position, speed, altitude, trajectory, and other data, and based on data analysis, performing anomaly alarms and countermeasures.

[0029] The real-time speed sequence is a collection of speed data formed by the 5G-A sensing network and control platform after continuous time-series sampling and preprocessing of low-altitude targets in the power grid. Its core function is to calculate the maximum speed change and speed mutation terms. It can generally be divided into two categories: core time-series data and auxiliary calculation data. All data is adapted to the 5G-A sensing module's 5-100 km / h drone speed perception range, as detailed below: 1. Core time series data: Instantaneous speed values ​​of UAVs under continuous timestamps, that is, the instantaneous speed of UAVs flying at low altitude at different sampling times, which is the basis of the speed sequence; it also includes the collection timestamp corresponding to each speed value, forming a time-speed time series correspondence, supporting the calculation of the maximum speed change value within the sliding window.

[0030] 2. Auxiliary calculation data: Normalized reference velocity value, used to standardize instantaneous velocity values, eliminate the differences in velocity dimensions between different flight scenarios and UAV types, and ensure the accuracy of velocity change value calculation.

[0031] The purpose of historical trajectory point sets is to calculate the trajectory entropy term. Trajectory entropy quantifies the randomness of flight trajectories by the probability of a UAV staying in a specific area / path segment. The calculation of its probability distribution requires sufficient historical trajectory data. Therefore, the time period of the trajectory point set must cover the effective flight trajectories of the UAV within the monitoring area to ensure the accuracy of the stay probability calculation. For example, in the scenario of 10kV distribution lines / ordinary power grid corridors (low-sensitivity areas in the power grid), the time period of the historical trajectory point set can be set to a sliding time window of approximately 30 minutes. In the scenario of temporary construction and inspection areas of the power grid, the time period of the historical trajectory point set can be set to a sliding time window of approximately 5 minutes.

[0032] It should be noted that the time period of the historical trajectory point set is a configurable parameter of the power grid low-altitude control platform, which can be further adjusted according to the actual scenario: for example, in complex electromagnetic environments / densely built areas (traditional radar failure areas), the time period of the corresponding area can be appropriately shortened (e.g., 5-10 minutes) to improve monitoring sensitivity; for open suburban power grid areas, it can be appropriately extended (e.g., 40-60 minutes) to further reduce the false judgment rate.

[0033] Furthermore, in step S2, calculating the behavioral threat score of the flight target based on the real-time velocity sequence and the historical trajectory point set includes: Based on the real-time velocity sequence, calculate the velocity change value of the flight target at each moment within the preset time period, and determine the maximum velocity change value from each velocity change value; The velocity mutation value of the flight target is calculated based on the maximum velocity change value and the preset normalized velocity. The probability of the flight target staying in each preset area is determined by analyzing the historical trajectory point set. Then, the trajectory entropy of the flight target is calculated based on each of the staying probabilities and the weighting factors corresponding to each preset area. The behavioral threat score of the flight target is calculated based on the velocity mutation value and the trajectory entropy.

[0034] This application further defines the calculation method for behavioral threat scoring. By analyzing real-time speed sequences, the "maximum speed change value" is extracted to quantify abnormal speed changes, effectively capturing suspicious behaviors such as sudden acceleration, sharp turns, or evasive maneuvers by the flight target. Simultaneously, by analyzing historical trajectory point sets to calculate "trajectory entropy," the randomness and unpredictability of flight paths can be quantified, thereby identifying unplanned flight path behaviors such as loitering and detours. Combining speed change values ​​reflecting instantaneous maneuvering anomalies with trajectory entropy reflecting long-term behavioral patterns constitutes the behavioral threat score, making the assessment of flight behavior no longer a single, instantaneous judgment, but a comprehensive evaluation that considers both short-term dynamic anomalies and long-term pattern anomalies. This assessment method can effectively distinguish between normal inspections (fixed flight path, stable speed) and potential threat behaviors (chaotic trajectory, speed changes), enhancing the depth and reliability of threat perception at the behavioral level and providing more robust and granular behavioral feature input for subsequent intent determination.

[0035] In a preferred embodiment, the maximum velocity change value of a low-altitude target of the power grid within a preset time period is obtained based on a real-time velocity sequence, and a velocity mutation term reflecting abnormal velocity fluctuations of the low-altitude target of the power grid is calculated based on the maximum velocity change value. .in, This represents the normalized velocity, or reference velocity, used to standardize velocity variation values. The maximum speed change value refers to the largest change in the velocity vector of a UAV within a very short time (usually on the order of seconds), including instantaneous rate changes caused by peak acceleration or sudden changes in direction. It reflects the limits of the UAV's maneuverability and is a key dynamic parameter for assessing its ability to evade detection or perform emergency tasks (such as obstacle avoidance and penetration). The maximum speed change value reflects abnormal fluctuations in the UAV's speed and can be obtained in real time through 5G-A sensing base stations. 5G-A sensing base stations provide high-precision speed measurement, and the data is accessed by a low-altitude control platform. The platform calculates speed changes in real time, for example, by sampling a UAV speed sequence and calculating the maximum difference value within a sliding window. For example, the result with the largest absolute value of all speed differences calculated within a sliding window of a preset length can be extracted as the maximum speed change value of a low-altitude target within that preset time period.

[0036] Then, based on the historical trajectory point set, the trajectory entropy term used to quantify the randomness of the flight trajectory of low-altitude targets in the power grid is obtained. A behavioral threat score is calculated based on the velocity mutation term and the trajectory entropy term to quantify the threat level of abnormal behavior of low-altitude targets in the power grid. For example, the formula for calculating the behavioral threat score is as follows: Among them, trajectory entropy Based on probability distribution The calculation quantifies the randomness or uncertainty of the drone's flight trajectory; the higher the entropy value, the more unpredictable the trajectory. The probability of a drone staying in a specific area or path segment can be calculated based on historical trajectory data. This represents a weighting factor used to balance the relative importance of velocity mutations and trajectory entropy, typically between 0.1 and 0.5, with a default of 0.25.

[0037] Generally, drones are characterized by low flight altitude, slow movement speed, and small radar cross-section, making them difficult to detect with traditional radar and posing a safety risk to power facilities, such as collisions or intentional damage. Manual inspections and video surveillance are inefficient, and insufficient communication network throughput leads to data delays, affecting the accuracy of analysis. 5G-A sensing technology, through the fusion of communication and sensing, achieves high-precision target detection (such as distance, speed, and angle) and real-time data transmission, providing a data foundation for calculating behavioral threat scores.

[0038] The aforementioned behavioral threat scoring function is used to assess the threat level of drone behavior, supporting alarm management and countermeasure linkage systems. It can utilize 5G-A sensing technology to solve the problem of low-altitude control of power grids, especially the monitoring, identification, and countermeasures against drones.

[0039] The following is a specific implementation method for calculating behavioral threat scores: The first step is to collect data in real time. Since the 5G-A sensing base station senses the drone's status such as distance, speed, and angle, the data is transmitted to the management platform via TCP / UDP protocol. The data preprocessing, such as noise removal (e.g., GPS drift) and format conversion (binary to text), is performed to ensure the quality of the data input into the formula. Structured data (speed, position) is stored in a relational database (such as MySQL), and trajectory data is stored in a non-relational database (such as MongoDB) to support fast querying.

[0040] The second step is to call the behavioral threat scoring function. Based on real-time velocity sequences and historical trajectory point sets, a behavioral threat score is calculated to quantify the threat level of abnormal behavior of low-altitude targets in the power grid.

[0041] Furthermore, when the behavioral threat score exceeds a threshold, the platform generates an alert and marks it as abnormal behavior, indicating its location on the map. The alert information includes time, coordinates, and threat score details for easy backtracking.

[0042] In summary, the proposed method for dynamic identification of threat intent of low-altitude targets in power grids constructs a 5G-A sensing network and control platform. Based on this platform, it acquires real-time velocity sequences and historical trajectory point sets of low-altitude targets in power grids. Then, it obtains velocity mutation terms reflecting abnormal velocity fluctuations of low-altitude targets and trajectory entropy terms quantifying the randomness of their flight trajectories. Based on these terms, it obtains a behavioral threat score to quantify the threat level of abnormal behavior of low-altitude targets in power grids. This method leverages the high-precision velocity measurement advantages of 5G-A integrated sensing technology, improving monitoring accuracy and efficiently identifying abnormal UAV behavior, thus overcoming the shortcomings of traditional radar in monitoring low-altitude targets in power grids.

[0043] In one possible implementation, step S3, calculating the signal threat score of the flight target based on the wireless signal strength and the wireless signal frequency band, includes: The first signal threat index of the flight target is determined by comparing the wireless signal strength with a preset signal strength threshold. Determine whether the wireless signal frequency band is within a preset list of illegal frequency bands, and determine the second signal threat index of the flight target based on the determination result; The signal threat score of the flight target is calculated based on the first signal threat index and the second signal threat index.

[0044] This application specifies the calculation method for signal threat scoring, integrating two key indicators—Radio Signal Strength Indicator (RSSI) and signal frequency band—for threat determination. By comparing real-time signal strength with a preset threshold, abnormal signal enhancement (potentially indicating relaying or close-range manipulation) or abnormal signal weakening (potentially an attempt to evade detection) can be identified. Simultaneously, by checking whether the signal frequency band belongs to a preset list of illegal frequency bands, "black flight" drones using unauthorized frequency bands for communication can be directly identified. These two indicators are fused through logical judgment to generate multi-level signal threat scores, achieving refined hierarchical assessment of abnormal communication signals. Compared to traditional methods that rely solely on signal strength or only perform frequency band blacklist / whitelist filtering, this embodiment's dual-judgment mechanism significantly improves the accuracy and robustness of identifying illegal or malicious drone signals, effectively combating simple signal camouflage or power modulation, and providing reliable signal-dimensional evidence for subsequent intent identification.

[0045] In a preferred embodiment, a signal strength threshold and a list of illegal frequency bands are first predefined. Then, raw RSSI and wireless signal frequency band data are collected using a 5G-A integrated sensing terminal module, and the data is analyzed based on the wireless signal strength and the signal strength threshold. The magnitude of the signal and whether the wireless signal frequency band belongs to the list of illegal frequency bands are used to calculate a signal threat score to quantify the threat level of abnormal signals from low-altitude targets in the power grid based on sensor data.

[0046] In low-altitude airspace management of power grids, signal strength thresholds need to be dynamically adjusted according to the environment. A single base station's sensing coverage must meet the requirements of a 300-meter low-altitude airspace with sub-meter sensing accuracy. Signal strength thresholds are typically set based on baseline signal strength, such as the RSSI range for normal drone communication. In complex electromagnetic environments, δ can be set to -90dBm. The illegal frequency band list, or blacklist, usually includes common unauthorized drone frequency bands, such as 2.4GHz or 5.8GHz, and may also extend to DIY drone frequency bands.

[0047] For example, signal threat scoring .in, This indicates an indicator function that returns 1 if the condition is met, and 0 otherwise. This function indicates whether the signal frequency band belongs to a predefined list of illegal frequency bands. It is used to assess the risk of signal anomalies from drones or other low-altitude flying objects and is a key component of 5G-A sensing technology in the power grid's low-altitude control system. Specifically, a signal threat score of 0 indicates no threat, normal signal strength, and a legal frequency band; a score of 01 indicates a low-level threat, with either abnormal signal strength or an illegal frequency band, but not both occurring simultaneously; and a score of 02 indicates a high-level threat, with both abnormal signal strength and an illegal frequency band, requiring immediate triggering of alarms or countermeasures.

[0048] Furthermore, when a drone enters a controlled area, such as near a substation or transmission line, the 5G-A base station acquires signal data such as location, speed, and frequency band through its sensing function. After calculation using a signal threat scoring function, if the signal threat score is ≥1, the system marks it as a suspicious target and displays real-time warning information such as the alarm device identifier and alarm type, while also reviewing historical data. For example, in an electronic fence area, if a drone flying illegally triggers a signal threat score of 2, the platform issues a real-time alarm. In summary, in low-altitude power grid control, this embodiment provides an efficient and quantitative threat assessment method based on signal threat scoring, offering the following technical advantages: 1. Quickly identify unauthorized drones to prevent them from damaging power facilities, such as through deliberate interference or collisions, thereby achieving disaster prevention, mitigation, and safety technologies.

[0049] 2. Compared with traditional radar, 5G-A sensing technology, combined with signal threat scoring, enables low-cost and accurate monitoring.

[0050] 3. As a typical application of synesthetic data fusion, it supports dynamic decision-making and embodies the integration of "communication-sensing-computing".

[0051] Furthermore, in step S4, calculating the geofencing risk score of the flight target based on the current location coordinates includes: Calculate the Euclidean distances between the flight target and each boundary coordinate point in the preset electronic fence boundary based on the current position coordinates, and determine the shortest Euclidean distance. The shortest Euclidean distance is input into a preset inverse proportional function to calculate the geofencing risk score of the flight target.

[0052] This application defines a method for calculating geofence risk scores. The risk score is calculated by substituting the shortest Euclidean distance from the current location of a flying target to the boundary of the geofence into an inverse proportional function. The advantage of this method lies in its intuitive and dynamically continuous evaluation results: when the target is infinitely far from the no-fly zone, the score approaches zero, indicating negligible risk; as the target approaches or even touches the geofence boundary, the score increases sharply and even approaches infinity (in practice, a very small constant is usually introduced to avoid division by zero), thus clearly reflecting the security principle of "the closer the distance, the higher the risk" in a quantitative form. This risk scoring model based on continuous distance provides better early warning of potential intrusion risks than a simple "enter / not enter" Boolean judgment and quantifies and classifies the risks at different distances. This enables the security system to achieve risk-sensitive adaptive responses, improving the accuracy of subsequent identification of the flying target's flight intent.

[0053] In a preferred embodiment, the shortest Euclidean distance from the current location of the low-altitude target of the power grid to the boundary of the electronic fence is obtained, and a geofence risk score is calculated based on the shortest Euclidean distance to quantify the threat level of the low-altitude target of the power grid approaching the no-fly zone.

[0054] For example, geofencing risk scoring .in, Indicates the current location of the drone The shortest Euclidean distance to the electronic fence boundary F reflects the spatial relationship between the drone and the no-fly zone and is the core basis for risk assessment. This is a very small positive number, typically between 0.001 and 0.1, primarily used to avoid a zero denominator while ensuring function continuity. Observing the geofence risk scoring function reveals an inverse relationship between distance and geofence risk. When a drone approaches the fence boundary, the score increases to 1 / ε (maximum value), indicating the highest risk; conversely, when the drone moves away from the fence, the score decreases to 0, and the risk becomes negligible. Furthermore, the smaller the distance, the faster the score increases, aligning with the security requirement of high risk near the boundary. Simultaneously, the platform allows for the creation of custom polygonal no-fly zones, with parameters such as height and location set to form the geofence boundary F.

[0055] Furthermore, by acquiring the drone's real-time pose (such as longitude, latitude, altitude, speed, etc.) through a 5G-A sensing base station, calculations are performed. When the geofence risk score exceeds the threshold (e.g., 0.5), a real-time alarm is triggered, such as a platform pop-up and a red map icon.

[0056] In summary, this geofencing risk scoring function transforms spatial distance into an actionable risk indicator, intuitively reflecting spatial threats. It is computationally efficient, enables a distance-driven scoring mechanism, and shows that risks increase exponentially at close range, meeting the needs of security scenarios. It is also beneficial for prioritizing the handling of high-risk targets and optimizing resource allocation.

[0057] In one possible implementation, step S5, constructing the observation state set of the flight target based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensor data, includes: Based on a preset first weighting coefficient, the behavioral threat score, the signal threat score, and the geofence risk score are weighted and summed to obtain a multimodal feature fusion threat score; Based on the real-time velocity sequence in the real-time sensing data, the velocity change vector of the flying target at the current moment is calculated; The flight altitude of the flight target at the current moment is determined based on the current position coordinates in the real-time sensing data; Based on the wireless signal strength and frequency band in the real-time sensing data, the signal characteristics of the flying target at the current moment are determined; By combining the multimodal feature fusion threat score, the velocity mutation vector, the flight altitude, and the signal features, an observation state set of the flight target is constructed.

[0058] This application's embodiments specifically define the construction method of the observation state set. Instead of simply inputting threat scores from various dimensions directly into the model, it employs an innovative two-stage fusion strategy. In the first stage, threat scores from the behavioral, signal, and geographical dimensions are fused into a comprehensive "multimodal feature fusion threat score" through weighted summation. This score provides a macroscopic quantification of the overall anomaly level of the target. In the second stage, this comprehensive score is combined with more physically meaningful micro-features (such as velocity mutation vectors, flight altitude, and signal features) directly extracted from the original synesthetic data to jointly constitute the observation state set. This construction method combines high-level semantic information (fusion score) and low-level physical features (original data features), providing the Hidden Markov Model with extremely rich and comprehensive observation input. The high-level score provides a rapid judgment of threat level, while the low-level features retain detailed information distinguishing different intent categories, enabling the model to perform probability matching and intent inference more accurately, significantly improving the accuracy of intent classification.

[0059] In a preferred embodiment, a hidden state set and an observation state set are predefined, and the Hidden Markov Model (HMM) is constructed based on these sets. The observation state set is defined according to four core categories of indicators: multimodal feature fusion threat score, velocity mutation vector, altitude, and signal characteristics. It integrates observable data from all dimensions of behavior, velocity, space, and signal in low-altitude target monitoring of the power grid. The core function of the observation state set is to transform the multi-dimensional, real-time observable data collected by 5G-A sensing into standardized, scenario-based observation data that supports the HMM inferring the true threat intent of drones. Together with the hidden state set, it constructs the HMM.

[0060] The specific method for defining the observation state set includes the following steps: obtaining a multimodal feature fusion threat score based on behavioral threat score, signal threat score, and geofence risk score; defining the observation state set based on the multimodal feature fusion threat score, velocity mutation vector, altitude, and signal features. Specifically, the velocity mutation vector is mainly used to detect typical characteristics of attack intent, such as sudden acceleration / turning of the target. Altitude refers to the target's current flight altitude, with a focus on detecting abnormally low-altitude close-flying. Signal features include communication frequency bands (whether they are on blacklisted bands) and scattering features (millimeter-wave sensing to determine payload type). The multimodal feature fusion threat score is a weighted value of the threat score, signal threat score, and geofence risk score. The velocity mutation vector is the rate of change of velocity per unit time. It can be obtained by first measuring the three-dimensional velocity components of the UAV in real time through radar waveform design and multi-base station collaboration using a 5G-A integrated sensing network, and then calculating the three-dimensional instantaneous acceleration through the three-dimensional velocity components, thus obtaining the velocity mutation vector. That is, the velocity mutation vector is equal to the square root of the sum of the squares of the instantaneous accelerations along the X, Y, and Z axes.

[0061] The specific method for defining the hidden state set includes the following steps: obtaining the true intentions of low-altitude targets in the power grid that cannot be directly observed, including four categories: inspection, observation, attack, and escape; and defining the hidden state set based on the four categories of true intentions. Specifically, the hidden state represents the true intentions that the UAV cannot directly observe, and there are four categories: 1. Patrol: The target is performing a pre-set power inspection task (such as flying along a planned route).

[0062] 2. Observation (Recon): The target hovers or flies at low speed over sensitive areas (such as over substations) and may collect data.

[0063] 3. Attack: The target approaches core power facilities (such as high-voltage towers and transformers) at high speed, or carries dangerous loads.

[0064] 4. Escape: The target attempts to leave the scene quickly after the alarm is triggered (e.g., by making a sharp turn or diving).

[0065] Defining a set of hidden states is the core prerequisite and key logical foundation for achieving accurate and dynamic identification of the threat intent of low-altitude targets in power grids. It standardizes, contextualizes, and hierarchically classifies the true threat intent that cannot be directly observed by UAVs, providing clear state dimensions and judgment targets for Hidden Markov Model (HMM) modeling, multimodal feature fusion implementation, and dynamic evolution judgment of threat intent. It directly improves the accuracy of dynamic identification of threat intent of low-altitude targets in power grids from five dimensions: resolving the ambiguity of traditional identification, building the core framework for algorithm modeling, achieving accurate mapping between features and intents, adapting to differentiated judgments in power grid scenarios, and supporting dynamic tracking of intents. Generally, traditional low-altitude target monitoring in power grids can only determine whether a target is abnormal through a single feature (such as position or speed), but it cannot accurately distinguish the true threat intent behind the anomaly, easily leading to misjudgments (such as misjudging an observation as an attack) and missed judgments (such as classifying an observation before an attack as normal), resulting in low identification accuracy. This embodiment explicitly defines the hidden state set into four categories: inspection, observation, attack, and escape. It transforms the vague and abstract behavioral intentions of drones in power grid scenarios into clear, mutually exclusive, and determinate state categories. Each state category corresponds to typical behaviors and threat levels in low-altitude security scenarios within the power grid. This classification upgrades threat intent recognition from vague anomaly judgments to precise intent classification, directly reducing the false positive and false negative rates, and is a fundamental prerequisite for improving recognition accuracy.

[0066] Furthermore, the flight target flight intention recognition method also includes: before calculating the multimodal feature fusion threat score, calculating the cumulative trajectory offset distance of the flight target based on the historical trajectory point set and the preset reference inspection route trajectory sequence, and weighting and summing the behavioral threat score, the signal threat score, the geofence risk score and the cumulative trajectory offset distance according to the preset second weight coefficient to obtain the multimodal feature fusion threat score.

[0067] This embodiment additionally incorporates the key feature of "cumulative trajectory deviation distance" into the calculation of the multimodal feature fusion threat score. Whether the trajectory deviates from the preset legitimate inspection route is one of the most direct and powerful pieces of evidence for judging the legitimacy of the flight target's intent. This embodiment calculates the trajectory deviation distance through a weighted dynamic time warping algorithm, which can effectively align and compare the real-time trajectory with the reference trajectory in the time dimension. Even if there are differences in speed and pauses, the overall deviation can be accurately calculated. This quantified trajectory anomaly index is then weighted and fused with the threat scores of the other three dimensions, so that the final "multimodal feature fusion threat score" includes information from the four core dimensions of behavior, signal, geography, and trajectory. This improvement makes feature fusion more comprehensive, especially enhancing the detection capability of covert threats such as "disguised normal flight but actually deviating from the flight path." An aircraft exhibiting normal speed and normal signal but with a severely deviated trajectory will also have its threat score increased due to the contribution of the trajectory dimension, thus being identified as an anomaly by the system earlier and more accurately. This compensates for the deficiencies of single or a few dimensional features, further improving the reliability and comprehensiveness of the entire intent recognition system.

[0068] In a preferred embodiment, real-time coordinates, speed, and altitude information of a low-altitude target within the power grid are acquired, and a real-time flight trajectory sequence T of the target is obtained based on this information. Specifically, the position, speed, and altitude of the drone are acquired in real-time using a 5G-A terminal module. The module supports lightweight design and high-precision sensing, sensing drone speeds within the range of 5-100 km / h to ensure the accuracy of the real-time flight trajectory sequence T. Then, a reference inspection route trajectory sequence R is acquired. The real-time flight trajectory sequence and the reference inspection route trajectory sequence are processed using a weighted dynamic time warping algorithm to obtain the cumulative trajectory offset distance. Specifically, the reference inspection route trajectory sequence is acquired based on the route management functions of the low-altitude control platform, such as route import and preset routes. The cumulative trajectory offset distance is then calculated through the following steps: 1. Obtain each trajectory point in the real-time flight trajectory sequence. Reference points in the corresponding reference inspection route trajectory sequence The Euclidean distance between the points.

[0069] 2. Obtain the cumulative distance of trajectory offset by calculating and minimizing the weighted sum of Euclidean distances between points on the path. For example, according to the function... Among them, the weighting coefficient By point-based Distance to the nearest power facility E Calculations show that the sensitivity to anomaly detection increases with proximity to power facilities, enhancing the sensitivity of critical areas, such as those near substations or transmission lines where the risk of deviation is higher. This is an adjustable parameter used to control the sensitivity of the weighting coefficient to distance changes. The DDTW function is an efficient method for identifying abnormal trajectories. It strengthens the monitoring of key areas of the power grid through a weighting mechanism and can be seamlessly integrated into the low-altitude management platform of the 5G-A sensing system. By utilizing terminal module data, platform analysis functions, and countermeasure linkage, it can achieve end-to-end anomaly detection and promote the intelligentization of low-altitude safety of the power grid.

[0070] Furthermore, after obtaining the cumulative trajectory offset distance, it can be determined whether the cumulative trajectory offset distance is greater than the preset safe distance threshold. If so, it is determined that the low-altitude target of the power grid deviates from the reference inspection route trajectory and is marked as an abnormal trajectory.

[0071] In this embodiment, abnormal trajectory identification is not an independent detection step, but rather a process that deeply integrates abnormal trajectory features into the entire dynamic identification of threat intent of low-altitude targets in the power grid. It becomes a crucial dimension of multimodal feature fusion, upgrading threat intent judgment from a three-dimensional approach of "speed, signal, and geography" to a four-dimensional approach of "speed, signal, geography, and trajectory." This results in more comprehensive features and more accurate intent determination. The fusion method is mainly reflected in: 1. As a basic marker of abnormal behavior: The deviation of the trajectory from the regular inspection route is an intuitive feature of the drone's "illegal inspection". The marked abnormal trajectory will directly serve as a reference for the behavior threat score, improve the drone's behavior threat score, and make the judgment of abnormal behavior more convincing. 2. Integrating Multimodal Feature Fusion Threat Scoring: The trajectory offset features corresponding to abnormal trajectories, together with behavioral threat scores, signal threat scores, and geofence risk scores, participate in the calculation of multimodal feature fusion threat scores, supplementing the definition of the observation state set with core features of the trajectory dimension; 3. Supporting the Hidden Markov Model's Intent Determination: The multimodal feature fusion threat score is the key basis for defining the set of observed states, and the set of observed states and the set of hidden states (inspection, observation, attack, escape) together construct the Hidden Markov Model. When a drone is marked as having an abnormal trajectory, it will directly improve the multimodal feature fusion threat score, allowing the model to more accurately distinguish it from the normal inspection state, and thus more accurately determine whether it is in the observation (hovering around), attack (high-speed approach), or escape (leaving after triggering an alarm) state, fundamentally improving the accuracy and timeliness of threat intent recognition.

[0072] In one possible implementation, in step S6, inputting the set of observed states into a preset Hidden Markov Model (HMM) so that the HMM generates the second hidden state of the flight target at the current moment through probabilistic deduction based on the current set of observed states and the first hidden state of the flight target at the previous moment, includes: The set of observed states is input into a preset hidden Markov model so that the hidden Markov model determines the observation probability of various hidden states of the flight target at the current moment based on the current set of observed states and the preset observation probability matrix. The observation probability matrix is ​​obtained by modeling based on a Gaussian mixture model. Based on the first hidden state, the prior probabilities corresponding to various hidden states of the flight target at the current moment are calculated through the current state transition probability matrix, wherein the current state transition probability matrix is ​​dynamically updated based on the relative distance between the flight target and the nearest power facility; The second hidden state of the flight target at the current moment is determined based on each of the observed probabilities and each of the prior probabilities.

[0073] This application's embodiments specifically disclose the internal mechanism of Hidden Markov Model (HMM) intention reasoning, clarifying the intention reasoning process as a combination of two probabilistic pieces of information: first, the "observation probability" calculated based on the current set of observed states and the observation probability matrix (modeled by a Gaussian mixture model), which reflects the likelihood of the target being in each hidden intention state under the various observed features; second, the "prior probability" calculated based on the intention state at the previous moment and the dynamically updated state transition probability matrix, which reflects the trend of the target transitioning to each intention state after considering historical states and the current environmental context (reflected in the dynamic transition probability). Finally, the most likely intention state is determined by combining these two probabilities. The advantages of this mechanism are: first, it utilizes the inherent advantage of probabilistic models in handling uncertainty and noise, making the judgment result more robust; second, it combines historical state information, considering the continuity of intention evolution and avoiding the irrationality of state jumps; finally, and most importantly, its state transition probability is dynamically updated, which allows the model reasoning to closely integrate with the current real-time context for intention recognition, realizing context-aware intelligent reasoning, thereby significantly improving the accuracy and context adaptability of intention recognition in complex and dynamic power grid environments.

[0074] Furthermore, the step of dynamically updating the current state transition probability matrix based on the relative distance between the flight target and the nearest power facility includes: Based on the current position coordinates of the flight target and the facility coordinates of each power facility in the target area, determine the relative distance between the flight target and the nearest power facility; Based on the preset ground state transition probability, the facility sensitivity coefficient corresponding to the nearest power facility, and the relative distance, the first prior probability corresponding to the preset type hidden state in the previous state transition probability matrix is ​​updated to obtain the current state transition probability matrix.

[0075] This application further defines the specific method for dynamically updating the state transition probability matrix. Traditional Hidden Markov Models (HMMs) typically have static state transition probabilities, failing to reflect the direct impact of external environmental changes on intent evolution. This embodiment creatively embeds a spatial risk model of power grid security into the probability transition process. Specifically, it dynamically adjusts the probability of transitioning to a high-risk state (such as an "attack" state) in the model based on the real-time distance between the flight target and the nearest power facility, and the facility's preset sensitivity coefficient. The closer the distance and the more sensitive the facility, the higher the transition probability. This design has significant beneficial effects: it transforms the intent recognition model from a closed, static entity into a dynamic perception-reasoning system closely linked to the physical world. The model understands that the spatial event of "approaching a substation" itself implies an increased probability of threatening intent, thus enabling it to raise the alert level before the target exhibits strong aggressive behavior. This prior probability adjustment based on geospatial risk greatly enhances the system's predictability of potential attacks, achieving a leap from "passive response based on behavior" to "proactive prediction combined with environmental risk," further improving the accuracy of flight intent recognition for flight targets.

[0076] In a preferred embodiment, the ground state transition probability is first obtained during the probabilistic deduction process. Facility sensitivity coefficient The coordinates of the nearest power facility to the flight target are used to calculate the current dynamic transition probability matrix of the Hidden Markov Model based on the ground state transition probability, facility sensitivity coefficient, and the coordinates of the nearest power facility.

[0077] For facility sensitivity coefficients, substations can be set to 0.8, and transmission towers can be set to 0.3. The ground state transition probability is used to control the baseline probability of an "attack" state occurring. Preferably, the ground state transition probability is learned from real data, such as pilot application records, and can be initially set to 0.3, then dynamically adjusted based on machine learning.

[0078] Based on the dynamic transition probability matrix and the first hidden state of the previous time step, the threat intent of low-altitude targets in the power grid is identified, and the prior probabilities of various hidden states of low-altitude targets in the power grid at the current time step are determined.

[0079] For example, through the state transition matrix Calculate the dynamic transition probabilities (i.e., the prior probabilities) of various hidden states in the Hidden Markov Model. Here, E represents the coordinates of the nearest power facility. Indicates the type of the second hidden state. Let P represent the type of the first hidden state, and let P represent the dynamic transition probability from the first hidden state to the second hidden state.

[0080] Specifically, when = An attack can be understood as an unauthorized drone intruding into a no-fly zone, such as illegally entering the vicinity of power grid facilities, with a transfer probability of . Drone location The closer to the coordinates of the nearest power facility, the smaller the `dist` value, the larger the exponent, approaching 1, and the higher the attack probability. This reflects the intuitive logic that the intrusion risk increases as distance decreases. The system can issue an early warning when the attack probability increases. When not under attack, such as when cooperative drones are operating within a suitable flight zone, the transition probability is a constant probability, for example... .

[0081] In the Hidden Markov Model of this embodiment, state i in the state transition matrix refers to the hidden state of the low-altitude target (drone) in the power grid. Specifically, it is any one of the four types of real threat intent in the predefined set of hidden states. It is the current / target state after the transition during the state transition process, corresponding to the "previous state" before the transition, and serves as the core target for dynamic transition probability calculation. When state i is inspection, observation, or escape, the transition probability of non-attack state i is a constant probability (e.g., the initial value of the ground state transition probability is 0.3). Only the attack state i will dynamically change with the power grid scenario parameters. In general, the dynamic transition probability is dynamically bound. This system enables 500kV substations to raise their alert level 1km away, and 10kV distribution lines to activate their defense only within 200m. The state transition matrix can be used to model the probabilistic dynamics of drone intrusion behavior. In particular, when combined with electronic fences and countermeasure linkage systems, it provides a quantitative model of drone intrusion risk, using distance-dependent probabilities for attack state judgment and simplifying non-attack states to constant probabilities. This improves the intelligence level of intrusion early warning and provides reliable protection for power grid security.

[0082] The observation probability matrix can be modeled using a Gaussian mixture model. A low threat score from multimodal feature fusion, low velocity mutation vector, stable altitude, and legitimate signal features define an inspection state; a high threat score from multimodal feature fusion, a sudden increase in velocity mutation vector, a sharp decrease in altitude, and abnormal signal features define an attack state. Specifically, the observation probability matrix, as a core component parameter of the Hidden Markov Model (HMM), supports the model in probabilistic intent deduction. It is the core quantitative parameter for accurately determining the threat intent of low-altitude targets in the power grid using the HMM in this invention. Modeled by a Gaussian mixture model, it represents the probability value of a certain unobservable hidden state (the true threat intent of the UAV) corresponding to a certain observable state. Its core function is to probabilistically match the multimodal observable features of the UAV with four types of true intent: inspection, observation, attack, and escape. This, combined with the dynamic transition probability, forms a dual constraint, upgrading the intent deduction of the HMM from "qualitative judgment" to "quantitative calculation," ultimately improving the accuracy, objectivity, and real-time performance of dynamic threat intent identification.

[0083] Example 2: like Figure 2 As shown, Embodiment 2 provides a flight target flight intention recognition system based on synesthesia data, including an acquisition module 10, a first evaluation module 20, a second evaluation module 30, a third evaluation module 40, an observation module 50, and a recognition module 60; The acquisition module 10 is used to acquire the historical trajectory point set of the flight target and the real-time sensing data within a preset time period. The real-time sensing data includes the real-time speed sequence, wireless signal strength, wireless signal frequency band, and current position coordinates of the flight target. The first evaluation module 20 is used to calculate and obtain the behavioral threat score of the flight target based on the real-time speed sequence and the historical trajectory point set; The second evaluation module 30 is used to calculate and obtain the signal threat score of the flight target based on the wireless signal strength and the wireless signal frequency band; The third assessment module 40 is used to calculate the geofence risk score of the flight target based on the current location coordinates; The observation module 50 is used to construct an observation state set of the flight target based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensor data. The identification module 60 is used to input the set of observed states into a preset hidden Markov model, so that the hidden Markov model generates the second hidden state of the flight target at the current moment through probabilistic deduction based on the current set of observed states and the first hidden state of the flight target at the previous moment, and then uses the second hidden state as the flight intention of the flight target at the current moment.

[0084] Furthermore, the first evaluation module 20 calculates a behavioral threat score for the flight target based on the real-time velocity sequence and the historical trajectory point set, including: Based on the real-time velocity sequence, calculate the velocity change value of the flight target at each moment within the preset time period, and determine the maximum velocity change value from each velocity change value; The velocity mutation value of the flight target is calculated based on the maximum velocity change value and the preset normalized velocity. The probability of the flight target staying in each preset area is determined by analyzing the historical trajectory point set. Then, the trajectory entropy of the flight target is calculated based on each of the staying probabilities and the weighting factors corresponding to each preset area. The behavioral threat score of the flight target is calculated based on the velocity mutation value and the trajectory entropy.

[0085] Furthermore, the second evaluation module 30 calculates a signal threat score for the flying target based on the wireless signal strength and the wireless signal frequency band, including: The first signal threat index of the flight target is determined by comparing the wireless signal strength with a preset signal strength threshold. Determine whether the wireless signal frequency band is within a preset list of illegal frequency bands, and determine the second signal threat index of the flight target based on the determination result; The signal threat score of the flight target is calculated based on the first signal threat index and the second signal threat index.

[0086] Furthermore, the third assessment module 40 calculates a geofencing risk score for the flight target based on the current location coordinates, including: Calculate the Euclidean distances between the flight target and each boundary coordinate point in the preset electronic fence boundary based on the current position coordinates, and determine the shortest Euclidean distance. The shortest Euclidean distance is input into a preset inverse proportional function to calculate the geofencing risk score of the flight target.

[0087] In one possible implementation, the observation module 50 constructs an observation state set for the flight target based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensor data, including: Based on a preset first weighting coefficient, the behavioral threat score, the signal threat score, and the geofence risk score are weighted and summed to obtain a multimodal feature fusion threat score; Based on the real-time velocity sequence in the real-time sensing data, the velocity change vector of the flying target at the current moment is calculated; The flight altitude of the flight target at the current moment is determined based on the current position coordinates in the real-time sensing data; Based on the wireless signal strength and frequency band in the real-time sensing data, the signal characteristics of the flying target at the current moment are determined; By combining the multimodal feature fusion threat score, the velocity mutation vector, the flight altitude, and the signal features, an observation state set of the flight target is constructed.

[0088] Furthermore, the flight target flight intent recognition system also includes a fourth evaluation module, which is used to calculate the cumulative trajectory offset distance of the flight target based on the historical trajectory point set and the preset reference inspection route trajectory sequence before calculating the multimodal feature fusion threat score; the observation module 50 is used to perform a weighted summation of the behavioral threat score, the signal threat score, the geofence risk score and the cumulative trajectory offset distance based on a preset second weighting coefficient to obtain the multimodal feature fusion threat score.

[0089] In one possible implementation, the identification module 60 includes an observation probability determination unit, a prior probability determination unit, and an identification unit, comprising: The observation probability determination unit is used to input the set of observation states into a preset hidden Markov model, so that the hidden Markov model determines the observation probability of various hidden states of the flight target at the current moment based on the current set of observation states and the preset observation probability matrix. The observation probability matrix is ​​obtained by modeling based on a Gaussian mixture model. The prior probability determination unit is used to calculate the prior probability of each hidden state of the flight target at the current moment based on the first hidden state and the current state transition probability matrix. The current state transition probability matrix is ​​dynamically updated based on the relative distance between the flight target and the nearest power facility. The identification unit is used to determine the second hidden state of the flight target at the current moment based on each of the observation probabilities and each of the prior probabilities.

[0090] Furthermore, the step of dynamically updating the current state transition probability matrix based on the relative distance between the flight target and the nearest power facility includes: Based on the current position coordinates of the flight target and the facility coordinates of each power facility in the target area, determine the relative distance between the flight target and the nearest power facility; Based on the preset ground state transition probability, the facility sensitivity coefficient corresponding to the nearest power facility, and the relative distance, the first prior probability corresponding to the preset type hidden state in the previous state transition probability matrix is ​​updated to obtain the current state transition probability matrix.

[0091] This application provides a flight target intention recognition system based on synesthetic data. First, it acquires multi-dimensional real-time data and historical trajectories of the flight target, and calculates threat scores from three independent dimensions: behavior, signal, and geofencing, achieving a multi-faceted and quantitative assessment of the flight target's abnormal state. Then, this embodiment innovatively fuses these multi-dimensional quantitative scores with the original synesthetic data to construct an observation state set that comprehensively reflects the target's instantaneous state. Finally, the observation state is input into a preset Hidden Markov Model (HMM). Utilizing the probabilistic inference capability of the HMM, combined with the target's intention state at the previous moment, the most probable flight intention at the current moment is inferred. Compared to traditional methods relying solely on single radar data or simple threshold alarms, the technical solution provided in this application, through deep fusion of multi-source heterogeneous data and dynamic inference based on probabilistic models, can more comprehensively and accurately understand the behavioral patterns and potential threats of flight targets. This effectively reduces false alarms and missed alarms caused by incomplete data or rigid rules, significantly improving the accuracy and intelligence level of identifying low-altitude security threats to the power grid, and providing a reliable decision-making basis for the power grid's proactive defense and precise countermeasures.

[0092] For a more detailed explanation of the working principle and procedures of this embodiment, please refer to the relevant description in Embodiment 1.

[0093] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application for those skilled in the art.

Claims

1. A method for recognizing the flight intention of a flight target based on synesthetic data, characterized in that, include: Acquire the historical trajectory point set of the flight target and real-time sensing data within a preset time period. The real-time sensing data includes the real-time speed sequence, wireless signal strength, wireless signal frequency band, and current position coordinates of the flight target. The behavioral threat score of the flight target is calculated based on the real-time velocity sequence and the historical trajectory point set. The signal threat score of the flying target is calculated based on the wireless signal strength and the wireless signal frequency band. The geofencing risk score of the flight target is calculated based on the current location coordinates; Based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensor data, a set of observation states for the flight target is constructed. The set of observed states is input into a preset hidden Markov model, so that the hidden Markov model generates the second hidden state of the flight target at the current moment through probabilistic deduction based on the current set of observed states and the first hidden state of the flight target at the previous moment, and then uses the second hidden state as the flight intention of the flight target at the current moment.

2. The method for recognizing the flight intention of a flight target based on synesthetic data as described in claim 1, characterized in that, The step of calculating the behavioral threat score of the flight target based on the real-time velocity sequence and the historical trajectory point set includes: Based on the real-time velocity sequence, calculate the velocity change value of the flight target at each moment within the preset time period, and determine the maximum velocity change value from each velocity change value; The velocity mutation value of the flight target is calculated based on the maximum velocity change value and the preset normalized velocity. The probability of the flight target staying in each preset area is determined by analyzing the historical trajectory point set. Then, the trajectory entropy of the flight target is calculated based on each of the staying probabilities and the weighting factors corresponding to each preset area. The behavioral threat score of the flight target is calculated based on the velocity mutation value and the trajectory entropy.

3. The method for recognizing the flight intention of a flight target based on synesthetic data as described in claim 1, characterized in that, The step of calculating the signal threat score of the flying target based on the wireless signal strength and the wireless signal frequency band includes: The first signal threat index of the flight target is determined by comparing the wireless signal strength with a preset signal strength threshold. Determine whether the wireless signal frequency band is within a preset list of illegal frequency bands, and determine the second signal threat index of the flight target based on the determination result; The signal threat score of the flight target is calculated based on the first signal threat index and the second signal threat index.

4. The method for recognizing the flight intention of a flight target based on synesthetic data as described in claim 1, characterized in that, The process of calculating the geofencing risk score of the flight target based on the current location coordinates includes: Calculate the Euclidean distances between the flight target and each boundary coordinate point in the preset electronic fence boundary based on the current position coordinates, and determine the shortest Euclidean distance. The shortest Euclidean distance is input into a preset inverse proportional function to calculate the geofencing risk score of the flight target.

5. The method for recognizing the flight intention of a flight target based on synesthetic data as described in claim 1, characterized in that, The step of constructing the observation state set of the flight target based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensing data includes: Based on a preset first weighting coefficient, the behavioral threat score, the signal threat score, and the geofence risk score are weighted and summed to obtain a multimodal feature fusion threat score; Based on the real-time velocity sequence in the real-time sensing data, the velocity change vector of the flying target at the current moment is calculated; The flight altitude of the flight target at the current moment is determined based on the current position coordinates in the real-time sensing data; Based on the wireless signal strength and frequency band in the real-time sensing data, the signal characteristics of the flying target at the current moment are determined; By combining the multimodal feature fusion threat score, the velocity mutation vector, the flight altitude, and the signal features, an observation state set of the flight target is constructed.

6. The method for recognizing the flight intention of a flight target based on synesthetic data as described in claim 5, characterized in that, The flight target flight intent recognition method further includes: before calculating the multimodal feature fusion threat score, calculating the cumulative trajectory offset distance of the flight target based on the historical trajectory point set and the preset reference inspection route trajectory sequence, and weighting and summing the behavioral threat score, the signal threat score, the geofence risk score and the cumulative trajectory offset distance according to the preset second weight coefficient to obtain the multimodal feature fusion threat score.

7. The method for recognizing the flight intention of a flight target based on synesthetic data as described in claim 1, characterized in that, The step of inputting the set of observed states into a preset hidden Markov model, so that the hidden Markov model generates the second hidden state of the flight target at the current moment through probabilistic deduction based on the current set of observed states and the first hidden state of the flight target at the previous moment, includes: The set of observed states is input into a preset hidden Markov model so that the hidden Markov model determines the observation probability of various hidden states of the flight target at the current moment based on the current set of observed states and the preset observation probability matrix. The observation probability matrix is ​​obtained by modeling based on a Gaussian mixture model. Based on the first hidden state, the prior probabilities corresponding to various hidden states of the flight target at the current moment are calculated through the current state transition probability matrix, wherein the current state transition probability matrix is ​​dynamically updated based on the relative distance between the flight target and the nearest power facility; The second hidden state of the flight target at the current moment is determined based on each of the observed probabilities and each of the prior probabilities.

8. The method for recognizing the flight intention of a flight target based on synesthetic data as described in claim 7, characterized in that, The step of dynamically updating the current state transition probability matrix based on the relative distance between the flight target and the nearest power facility includes: Based on the current position coordinates of the flight target and the facility coordinates of each power facility in the target area, determine the relative distance between the flight target and the nearest power facility; Based on the preset ground state transition probability, the facility sensitivity coefficient corresponding to the nearest power facility, and the relative distance, the first prior probability corresponding to the preset type hidden state in the previous state transition probability matrix is ​​updated to obtain the current state transition probability matrix.

9. A flight target flight intention recognition system based on synesthetic data, characterized in that, It includes an acquisition module, a first evaluation module, a second evaluation module, a third evaluation module, an observation module, and an identification module; The acquisition module is used to acquire the historical trajectory point set of the flight target and the real-time sensing data within a preset time period. The real-time sensing data includes the real-time speed sequence, wireless signal strength, wireless signal frequency band, and current position coordinates of the flight target. The first evaluation module is used to calculate and obtain the behavioral threat score of the flight target based on the real-time speed sequence and the historical trajectory point set; The second evaluation module is used to calculate the signal threat score of the flying target based on the wireless signal strength and the wireless signal frequency band. The third assessment module is used to calculate the geofence risk score of the flight target based on the current location coordinates; The observation module is used to construct an observation status set of the flight target based on the behavioral threat score, the signal threat score, the geofence risk score, and the real-time sensor data. The identification module is used to input the set of observed states into a preset hidden Markov model, so that the hidden Markov model generates the second hidden state of the flight target at the current moment through probabilistic deduction based on the current set of observed states and the first hidden state of the flight target at the previous moment, and then uses the second hidden state as the flight intention of the flight target at the current moment.

10. A flight target flight intention recognition system based on synesthetic data as described in claim 9, characterized in that, The identification module includes an observation probability determination unit, a prior probability determination unit, and an identification unit. include: The observation probability determination unit is used to input the set of observation states into a preset hidden Markov model, so that the hidden Markov model determines the observation probability of various hidden states of the flight target at the current moment based on the current set of observation states and the preset observation probability matrix. The observation probability matrix is ​​obtained by modeling based on a Gaussian mixture model. The prior probability determination unit is used to calculate the prior probability of each hidden state of the flight target at the current moment based on the first hidden state and the current state transition probability matrix. The current state transition probability matrix is ​​dynamically updated based on the relative distance between the flight target and the nearest power facility. The identification unit is used to determine the second hidden state of the flight target at the current moment based on each of the observation probabilities and each of the prior probabilities.

Images