A method for monitoring UAV communication terminals based on multi-source information fusion

Abstract

This invention discloses a method for monitoring UAV communication terminals based on multi-source information fusion, comprising the following steps: Step S1, Multi-source information acquisition: Parallel acquisition of multi-source heterogeneous data from the UAV communication terminal, wherein the multi-source heterogeneous data includes at least terminal transmit power, received signal strength (RSSI), signal-to-noise ratio (SNR), neighboring cell interference intensity, wireless link transmission bit error rate, terminal positioning data, flight status data, and environmental meteorological data; Step S2, Information fusion and feature extraction: Spatiotemporal alignment and normalization processing of the multi-source heterogeneous data acquired in Step S1, and inputting it into an information fusion model for feature layer fusion, outputting a set of fused feature vectors that comprehensively characterize the current operating state of the communication terminal. This invention enables more reliable monitoring of UAV communication terminals.

Description

Technical Field

[0001] This invention relates to the field of drone supervision, specifically to a method for supervising drone communication terminals based on multi-source information fusion. Background Technology

[0002] The UAV communication terminal is the core component for data interaction between the UAV and the ground control station, and the stability of its communication link directly determines whether the UAV mission can be carried out normally. Monitoring the UAV communication terminal essentially involves real-time monitoring of terminal operation data, assessing communication link quality, and promptly generating and implementing intervention measures to ultimately ensure the reliable transmission of critical data between the terminal and the ground control station, avoiding problems such as data loss, transmission interruption, or mission delays caused by link quality deterioration.

[0003] Existing regulatory methods for UAV communication terminals have significant shortcomings: they offer only one type of protection and fail to consider the unique characteristics of UAV communication scenarios. During flight, UAVs face complex situations such as dynamic environmental changes, flight status adjustments, and multi-source interference. Traditional solutions rely solely on single-dimensional data monitoring or fixed rules for protection, failing to comprehensively assess the impact of multi-source heterogeneous data on communication link quality. Therefore, this paper proposes a regulatory method for UAV communication terminals based on multi-source information fusion. Summary of the Invention

[0004] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0005] Step S1, Multi-source information acquisition: Parallel acquisition of multi-source heterogeneous data from the UAV communication terminal. The multi-source heterogeneous data includes at least terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, neighboring cell interference intensity, wireless link transmission bit error rate, terminal positioning data, flight status data, and environmental meteorological data.

[0006] Step S2, Information Fusion and Feature Extraction: The multi-source heterogeneous data collected in Step S1 is spatiotemporally aligned and normalized, and then input into the information fusion model for feature layer fusion, outputting a set of fused feature vectors that can comprehensively characterize the current operating state of the communication terminal.

[0007] Step S3, Communication Link Health Assessment: Based on the fused feature vector obtained in Step S2, a comprehensive communication link health index value HI is calculated using a preset terminal status assessment function. The process of obtaining HI is as follows: ;

[0008] in, Represents the fused feature vector. Let i be the i-th normalized eigenvalue. This represents the dynamic weight vector corresponding to each feature. The weight of the i-th feature and ;

[0009] Step S4, Generation and Execution of Monitoring Strategies: The communication link health index value HI calculated in step S3 is compared with multiple preset threshold intervals. Based on the interval it is in, the preset monitoring system automatically generates corresponding monitoring strategies. The monitoring strategies include parameter adjustment strategies, link switching strategies, or early warning strategies. The generated strategies are then sent to the UAV communication terminal or ground control station for execution.

[0010] Furthermore, in step S2, the information fusion model adopts the DS evidence theory model;

[0011] The feature layer fusion process includes: assigning the support probability of evidence from different data sources to propositions on different states of communication link quality based on the basic probability allocation function of DS evidence theory, synthesizing the support probabilities of all evidence, and outputting the synthesized probability distribution to form the fused feature vector.

[0012] The different data sources include at least terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, neighboring cell interference intensity, and bit error rate data.

[0013] The different states of the communication link quality include "good link quality", "average link quality" and "poor link quality".

[0014] Furthermore, in step S3, the dynamic weight vector It is not a fixed value, but is dynamically adjusted based on environmental meteorological data and the flight status data.

[0015] When environmental meteorological data indicates heavy rainfall or flight status data indicates that the UAV is in a high-speed maneuver, the weights of the features corresponding to the bit error rate and the neighboring cell interference intensity are automatically increased. Specifically:

[0016] Judgment criteria: Environmental meteorological data indicates rainfall ≥ 5 mm / h (heavy rainfall), or flight status data indicates UAV flight speed ≥ 15 m / s (high-speed maneuvering state);

[0017] Weight adjustment: When any of the above conditions are met, the weights of the features corresponding to the bit error rate are adjusted. Increase the weight of the features corresponding to the neighboring cell interference intensity by 20%-50%. Increase by 15%-30%, while proportionally reducing the weight of non-critical features such as terminal positioning data and terminal transmission power.

[0018] Furthermore, the terminal status evaluation function in step S3 further includes a time prediction term based on historical data. This prediction term is used to predict the communication link health index for a future period of time, thereby obtaining the future communication link health index, specifically:

[0019] ;

[0020] in, The current health index value (HI) is... These are the predicted future health values ​​obtained using a time series prediction algorithm based on historical HI value sequences. and The weighting coefficients and The pre-set regulatory system is based on The value generates the final regulatory strategy.

[0021] Furthermore, the time-based prediction item based on historical data is predicted using a time series prediction algorithm. This time series prediction algorithm employs a dedicated prediction model, and the construction and deployment process of this prediction model includes the following steps:

[0022] Step 1: Data Preparation: Extract data from historical task logs and construct a training sample set. , where input Let the fused feature vector sequence of the k-th sample within the time window T be the label. For the future The actual HI value corresponding to a given moment;

[0023] The second step, model training: Construct a Long Short-Term Memory (LSTM) neural network model, where the number of neurons in the input layer corresponds to the dimension n of the fused feature vector, and the output layer consists of a single neuron; the mean squared error (MSE) is used as the loss function, specifically: ,in The model predicts the values, and N is the number of samples. The LSTM model is trained using the gradient descent algorithm to obtain the optimal model parameters.

[0024] The third step, model deployment: Integrate the trained LSTM model into the monitoring system to receive the latest fused feature vector sequence output in step S2 in real time and output the predicted value. .

[0025] Furthermore, the generation of the parameter adjustment strategy in step S4 specifically includes the following steps:

[0026] The first step is to construct a knowledge base for parameter adjustment strategies: the knowledge base stores tuples. ,in In flight mode, For the environment model, and These represent the historically optimal transmit power adjustment amount and modulation and coding scheme adjustment level, respectively.

[0027] The second step is to match the initial strategy: using the current HI value, flight status data, and environmental meteorological data as the joint query key, the K-nearest neighbor algorithm is used to find the K most similar tuples in the knowledge base, and the adjustment values ​​in these tuples are weighted and averaged to obtain the initial adjustment strategy. Specifically , This is the initial transmit power adjustment amount. Adjust the initial MCS level;

[0028] The third step, online adaptive optimization: fine-tuning the initial policy using a reinforcement learning framework; defining the state. ,action The reward function is a fine-tuning factor for the initial policy. The policy network is iteratively updated using the Actor-Critic algorithm to ultimately generate the execution policy. .

[0029] Furthermore, the following data preprocessing steps are included after step S1 and before step S2:

[0030] First step, credibility assessment: For the credibility score of the i-th data source Calculate its credibility score. The calculation formula is:

[0031] ;

[0032] in A score for whether the data value is within a reasonable physical range. The score is given to determine whether the instantaneous rate of change of the data is reasonable. This data is compared with other relevant data sources. Consistency score, , and These are the weighting coefficients;

[0033] Step Two, Data Repair: For Below the threshold Data The graph neural network (GNN) is used for repair.

[0034] A spatiotemporal graph is constructed using each data source as a node and the physical or statistical correlation between data sources as edges.

[0035] Each node's characteristics are defined as a combination of its historical time-series data and current data;

[0036] The spatiotemporal graph is input into a pre-trained GNN model, and the output feature of the target node i is taken as the repaired value. This is used to replace the original abnormal data.

[0037] Furthermore, it also includes step S5: visualization and feedback of regulatory effects; the communication link health index HI, the generated regulatory strategy, and the execution results are displayed graphically in real time on the human-computer interaction interface, and the link quality change data after the strategy execution is used as a feedback signal to adjust the dynamic weight vector in step S3. Alternatively, supervised learning and optimization updates can be performed using the terminal state evaluation function.

[0038] Compared with existing technologies, this invention has the following advantages: This UAV communication terminal monitoring method based on multi-source information fusion collects multi-source heterogeneous data from UAV communication terminals in parallel, including terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, neighboring cell interference intensity, wireless link transmission bit error rate, terminal positioning data, flight status data, and environmental meteorological data. This avoids the limitations of a single data dimension and can comprehensively capture the terminal's operating status and external influencing factors. Furthermore, it combines the DS evidence theory model for feature layer fusion, assigning support probabilities for different communication link quality propositions to different data sources and synthesizing them. This reduces misjudgments caused by single data errors or interference, and the output fused feature vector more accurately reflects the actual operating status of the communication terminal, improving the comprehensiveness and accuracy of monitoring. In the communication link health assessment stage, the dynamic weight vector is dynamically adjusted based on environmental meteorological data and flight status data. For example, when there is heavy rainfall or high-speed drone maneuvering, the weights of features corresponding to bit error rate and neighboring cell interference intensity are automatically increased, making the health assessment more in line with the actual scenario. At the same time, the terminal status assessment function introduces a time series prediction term based on LSTM neural network, combining the current health index value and the future health prediction value to calculate the final assessment value. This allows the monitoring system to predict the health trend of the communication link in advance, avoiding post-event remediation and achieving forward-looking and adaptive monitoring. In terms of monitoring strategy generation and execution, a parameter adjustment strategy knowledge base storing tuples is constructed. Using the current HI value, flight status data, and environmental meteorological data as the joint query key, the K-nearest neighbor algorithm is used to find similar tuples to obtain the initial adjustment strategy. Then, a reinforcement learning framework and the Actor-Critic algorithm are used to perform online adaptive fine-tuning of the initial strategy, generating differentiated strategies such as parameter adjustment strategies, link switching strategies, or early warning strategies. This can cope with all scenarios from minor link fluctuations to serious quality risks without manual intervention, improving the speed and accuracy of monitoring response. Furthermore, after multi-source information collection and before information fusion, a credibility score is calculated for each data source based on three dimensions: physical scope, instantaneous data change rate, and consistency with other data sources. Data below a threshold is repaired using a graph neural network (GNN). A spatiotemporal graph is constructed with each data source as a node and the correlation between data sources as edges. The repaired data is output using a pre-trained GNN model, avoiding interference from abnormal data in subsequent fusion and evaluation, and providing a reliable data foundation for the regulatory process. Simultaneously, this method includes a visualization and feedback mechanism for regulatory effects. The communication link health index (HI), generated regulatory strategies, and execution results are displayed graphically in real time. The link quality change data after strategy execution is used as a feedback signal to optimize the dynamic weight vector and terminal state evaluation function, forming a virtuous cycle of evaluation, execution, feedback, and optimization, enabling the regulatory system to continuously improve its accuracy with the accumulation of usage scenarios.Overall, compared to traditional single protection systems, this method achieves an upgrade from passive protection to active supervision, from fixed rules to dynamic adaptation, and from local assessment to global integration. It can effectively address the pain points of complex environments, variable states, and sensitive links in UAV communication scenarios, significantly improve the stability of communication links and the reliability of supervision, and is applicable to UAV application scenarios in multiple fields such as aerial photography, inspection, and emergency communication, making the system more worthy of promotion and use. Attached Figure Description

[0039] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0040] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0041] like Figure 1 As shown, this embodiment provides a technical solution: a method for monitoring unmanned aerial vehicle (UAV) communication terminals based on multi-source information fusion, comprising the following steps:

[0042] Step S1, Multi-source information acquisition: Parallel acquisition of multi-source heterogeneous data from the UAV communication terminal. The multi-source heterogeneous data includes at least terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, neighboring cell interference intensity, wireless link transmission bit error rate, terminal positioning data, flight status data, and environmental meteorological data.

[0043] Step S2, Information Fusion and Feature Extraction: The multi-source heterogeneous data collected in Step S1 is spatiotemporally aligned and normalized, and then input into the information fusion model for feature layer fusion, outputting a set of fused feature vectors that can comprehensively characterize the current operating state of the communication terminal.

[0044] Step S3, Communication Link Health Assessment: Based on the fused feature vector obtained in Step S2, a comprehensive communication link health index value HI is calculated using a preset terminal status assessment function. The process of obtaining HI is as follows: ;

[0045] in, Represents the fused feature vector. Let i be the i-th normalized eigenvalue. This represents the dynamic weight vector corresponding to each feature. The weight of the i-th feature and ;

[0046] Step S4, Generation and Execution of Monitoring Strategies: The communication link health index value HI calculated in step S3 is compared with multiple preset threshold intervals. Based on the interval it is in, the preset monitoring system automatically generates corresponding monitoring strategies. The monitoring strategies include parameter adjustment strategies, link switching strategies, or early warning strategies. The generated strategies are then sent to the UAV communication terminal or ground control station for execution.

[0047] In step S2, the information fusion model adopts the DS evidence theory model;

[0048] The feature layer fusion process includes: assigning the support probability of evidence from different data sources to propositions on different states of communication link quality based on the basic probability allocation function of DS evidence theory, synthesizing the support probabilities of all evidence, and outputting the synthesized probability distribution to form the fused feature vector.

[0049] The different data sources include at least terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, neighboring cell interference intensity, and bit error rate data.

[0050] The different states of the communication link quality include "good link quality", "average link quality" and "poor link quality".

[0051] First, its DS evidence theory model excels at handling uncertainties arising from multi-source information. For key data sources directly impacting communication link quality, such as terminal transmit power, received signal strength (RSSI), signal-to-noise ratio (SNR), neighboring cell interference intensity, and bit error rate, the model uses a basic probability allocation function to assign support probabilities for three distinct state propositions: "good link quality," "average link quality," and "poor link quality." This effectively mitigates potential errors, ambiguities, or judgment biases from different data sources, avoiding misjudgments caused by the limitations of a single data source. Second, by synthesizing evidence from the support probabilities of all data sources—rather than simply adding or filtering data—it fully integrates the effective information from each source. This allows the output probability distribution to more comprehensively and objectively reflect the true quality status of the communication link. The resulting fused feature vector accurately and comprehensively characterizes the current operating status of the communication terminal, providing reliable and accurate basic data support for subsequent communication link health assessments. This reduces health assessment biases caused by unscientific information integration, ensuring the rationality and relevance of subsequent regulatory strategies.

[0052] In step S3, the dynamic weight vector It is not a fixed value, but is dynamically adjusted based on environmental meteorological data and the flight status data.

[0053] When environmental meteorological data indicates heavy rainfall or flight status data indicates that the UAV is in a high-speed maneuver, the weights of the features corresponding to the bit error rate and the neighboring cell interference intensity are automatically increased. Specifically:

[0054] Judgment criteria: Environmental meteorological data indicates rainfall ≥ 5 mm / h (heavy rainfall), or flight status data indicates UAV flight speed ≥ 15 m / s (high-speed maneuvering state);

[0055] Weight adjustment: When any of the above conditions are met, the weights of the features corresponding to the bit error rate are adjusted. Increase the weight of the features corresponding to the neighboring cell interference intensity by 20%-50%. Increase by 15%-30%, while proportionally reducing the weight of non-critical features such as terminal positioning data and terminal transmission power;

[0056] By freeing the dynamic weight vector W in step S3 from fixed values ​​and adjusting it in real time based on environmental meteorological data and flight status data, this design precisely matches the actual characteristics of "dynamic changes in environment and flight status" in UAV communication scenarios. For example, when environmental meteorological data shows heavy rainfall, rainwater will attenuate and interfere with wireless signal transmission, leading to an increase in the bit error rate of wireless link transmission; when flight status data shows that the UAV is in a high-speed maneuvering state, the signal connection between the terminal and the base station is prone to instability, and the impact of neighboring cell interference intensity on link quality will increase significantly. At this time, automatically increasing the weights of the features corresponding to bit error rate and neighboring cell interference intensity allows these two factors that have a key impact on the current link quality to occupy a more reasonable proportion in the calculation of the health index HI, avoiding the drawbacks of a "one-size-fits-all" approach due to fixed weights that would cause the evaluation results to deviate from the actual link status. This dynamic adjustment mechanism ensures that communication link health assessments are no longer divorced from real-world scenarios. Instead, they closely follow changes in the environment and flight status, accurately focusing on key influencing factors. This results in outputting HI values ​​that are more in line with reality, providing a more reliable assessment basis for the monitoring system to generate targeted monitoring strategies (such as parameter adjustments and early warnings) in subsequent steps. This effectively reduces the problem of inappropriate monitoring strategies caused by assessment biases, further ensuring the effectiveness and flexibility of UAV communication link monitoring.

[0057] Step S3, the terminal status evaluation function further includes a time prediction term based on historical data. This prediction term is used to predict the communication link health index for a future period of time, thereby obtaining the future communication link health index, specifically:

[0058] ;

[0059] in, The current health index value (HI) is... These are the predicted future health values ​​obtained using a time series prediction algorithm based on historical HI value sequences. and The weighting coefficients and The pre-set regulatory system is based on The final regulatory strategy for value generation;

[0060] The formula combines the current health index with the predicted future health value based on the historical HI value series (obtained through a time series prediction algorithm) to form the final HI' value used to generate regulatory strategies. The core advantage of this design is:

[0061] Anticipating link risks in advance: Moving away from the traditional passive assessment model that only looks at the current state, it can perceive the health trend of communication links in advance over a period of time, avoiding the problem of lagging regulatory strategies when link quality suddenly deteriorates;

[0062] Optimize the predictability of the strategy: The monitoring system is based on the HI' value generation strategy that integrates the current state and future trends, rather than relying solely on the current state. This makes the strategy more forward-looking and allows for early intervention measures (such as adjusting parameters in advance and preparing for link switching) to reduce the probability of link interruption or quality degradation.

[0063] Adapting to dynamic drone scenarios: Drones are often in high-speed movement and changing environments (such as moving from sunny weather to rainy weather), and link quality is prone to rapid fluctuations. The time prediction item can accurately capture this dynamic trend, making the evaluation results more in line with the actual communication needs of drones.

[0064] Taking the use of drones for field power line inspection as an example, the specific parameters are set as follows:

[0065] The time series forecasting algorithm uses an LSTM model to predict the future time series Δt = 60 seconds. value;

[0066] Weighting coefficients α = 0.5 (current state weight), β = 0.5 (future prediction weight);

[0067] Current communication link status of the drone at a certain moment: =0.7 (The preset HI value of 0.6-1.0 is the range of good link quality, and the current state is good).

[0068] The LSTM model predicts the next 60 seconds based on a fusion of feature vector sequences from historical data (such as "environmental transition from sunny to light rain" and "drone gradually approaching high-voltage lines (potential for increased interference from neighboring cells)" over the past 5 minutes). =0.45 (0.4-0.6 is the preset value for "average link quality", and below 0.4 is "poor". The predicted value is close to the threshold for "poor").

[0069] If the design of this case is not adopted (only used) Evaluate)

[0070] The regulatory system is based solely on At t=0.7 (good), the current link is deemed to require no intervention, and no adjustment strategy is generated. However, after 60 seconds, as the light rain intensifies and interference from neighboring cells increases, the actual HI value drops to 0.38 (poor). Only then does the system urgently generate an early warning or parameter adjustment strategy, which may have already caused problems such as interruption of inspection data transmission and screen lag, affecting the efficiency of the inspection task.

[0071] First calculate =0.5×0.7+0.5×0.45=0.575, which falls within the "moderate link quality" range. The monitoring system is based on... Assessment: While the current link is good, it is highly likely to deteriorate to the "poor" level within the next 60 seconds, requiring advance intervention. Therefore, an automatic strategy was generated, such as "slightly increase the terminal's transmit power (ΔPtx = 2dBm) and lower the MCS level from 10 to 8 (to improve anti-interference capability)," and sent to the UAV communication terminal. After 60 seconds, the actual HI value remained at 0.55 (still within the "normal" range) due to the advance adjustment, and the link quality did not deteriorate to "poor," ensuring stable transmission of inspection data and avoiding the risk of mission interruption.

[0072] The time-based forecasting item based on historical data is predicted using a time series forecasting algorithm. This time series forecasting algorithm employs a dedicated forecasting model, and the construction and deployment process of this forecasting model includes the following steps:

[0073] Step 1: Data Preparation: Extract data from historical task logs and construct a training sample set. , where input Let the fused feature vector sequence of the k-th sample within the time window T be the label. For the future The actual HI value corresponding to a given moment;

[0074] The second step, model training: Construct a Long Short-Term Memory (LSTM) neural network model, where the number of neurons in the input layer corresponds to the dimension n of the fused feature vector, and the output layer consists of a single neuron; the mean squared error (MSE) is used as the loss function, specifically: ,in The model predicts the values, and N is the number of samples. The LSTM model is trained using the gradient descent algorithm to obtain the optimal model parameters.

[0075] The third step, model deployment: Integrate the trained LSTM model into the monitoring system to receive the latest fused feature vector sequence output in step S2 in real time and output the predicted value. ;

[0076] Taking a power line inspection scenario using drones as an example, suppose a power line inspection drone needs to operate in a suburban high-voltage line area. This area often experiences situations such as "increased humidity in the afternoon → signal attenuation" and "drone flying around the tower → interference fluctuations in neighboring areas." An LSTM model is needed to predict future signal interference. seconds The value, specific process, and effect are as follows:

[0077] Based on the constructed LSTM prediction model;

[0078] Data preparation: Data was extracted from the power line inspection historical task logs of this type of drone over the past 6 months, with a time window of T=5 minutes (feature vectors were collected every 10 seconds for each sample). It contains 30 consecutive fused feature vectors with a feature dimension of n=8, corresponding to 8 core features: terminal transmit power, RSSI, SNR, neighboring cell interference intensity, bit error rate, terminal positioning accuracy, flight speed, and ambient humidity.

[0079] Label Set to end after the sample time window The true HI value at a given second (e.g., for a sample) The feature sequence corresponding to 13:00:00-13:04:50, The actual HI value is 13:05:30, and a training set containing 120,000 valid samples is finally constructed.

[0080] Model Training: An LSTM neural network is built with 8 neurons in the input layer (matching feature dimension n), 2 hidden layers (64 neurons per layer, using the ReLU activation function), and 1 neuron in the output layer (output...). Using MSE as the loss function, the Adam gradient descent algorithm (learning rate = 0.001) was used to train the model. After 100 iterations, the MSE of the training set decreased from the initial 0.15 to 0.009, and the MSE of the validation set stabilized within 0.012. The model converged and had good generalization ability.

[0081] Model Deployment: The trained LSTM model is integrated into the UAV ground monitoring system. Every 10 seconds, the system automatically receives the latest fused feature vector output from step S2 and concatenates it in real time into a feature sequence of the most recent 5 minutes (i.e., The input is then fed into an LSTM model, which outputs the forecast for the next 40 seconds within 20 milliseconds. value.

[0082] Application performance comparison (i.e., with and without LSTM model), LSTM model (using traditional ARIMA algorithm): If only ARIMA algorithm is used for prediction This algorithm can only predict based on a single historical HI value and cannot incorporate multiple feature changes such as "increased ambient humidity (from 60% to 85%), increased flight speed (from 4 m / s to 7 m / s), and increased neighboring cell interference intensity (from -88 dBm to -75 dBm)". For example, at 13:00:00, the current... (Preset) (For "link quality is good"), ARIMA predicts 40 seconds later. (Still rated as good), but due to increased humidity leading to signal attenuation and increased interference from neighboring cells, the actual HI value dropped to 0.41 after 40 seconds (preset). The link quality was "poor" (already close to the "poor" threshold). The monitoring system did not intervene in advance, resulting in a 15-second delay in the transmission of inspection images from the drone, which affected the efficiency of line fault identification.

[0083] With an LSTM model: At 13:00:00, the monitoring system inputs the feature sequence of the LSTM model. The data already includes the characteristic trends of "continuously increasing ambient humidity, faster flight speed, and enhanced interference from neighboring regions." Based on these correlations, the LSTM model outputs the forecast for the next 40 seconds. Substitute (Pre- The system detected a "moderate link quality" error and immediately generated a strategy to increase terminal transmit power by 2.5 dBm and lower the modulation and coding scheme (MCS) level from 11 to 9 (to enhance anti-interference capabilities), which was then sent to the drone. After 40 seconds, the actual HI value remained at 0.57 (still within the "moderate" range), with no image transmission interruptions, ensuring the continuous and stable operation of the power line inspection mission.

[0084] The generation of the parameter adjustment strategy in step S4 specifically includes the following steps:

[0085] The first step is to construct a knowledge base for parameter adjustment strategies: the knowledge base stores tuples. ,in In flight mode, For the environment model, and These represent the historically optimal transmit power adjustment amount and modulation and coding scheme adjustment level, respectively.

[0086] The second step is to match the initial strategy: using the current HI value, flight status data, and environmental meteorological data as the joint query key, the K-nearest neighbor algorithm is used to find the K most similar tuples in the knowledge base, and the adjustment values ​​in these tuples are weighted and averaged to obtain the initial adjustment strategy. Specifically , This is the initial transmit power adjustment amount. Adjust the initial MCS level;

[0087] The third step, online adaptive optimization: fine-tuning the initial policy using a reinforcement learning framework; defining the state. ,action The reward function is a fine-tuning factor for the initial policy. The policy network is iteratively updated using the Actor-Critic algorithm to ultimately generate the execution policy. ;

[0088] Leveraging historical experience to reduce the initial strategic uncertainty: by constructing storage tuples The parameter adjustment strategy knowledge base solidifies the "scenario-optimal adjustment amount" correspondence verified in historical tasks, avoiding the problem of "detaching from actual experience and starting from scratch" when generating traditional strategies. (Flight status mode) The introduction of (environmental model) allows historical experience to be accurately linked to the current scenario, ensuring that the initial strategy has basic rationality.

[0089] Rapidly match scenarios and improve policy response speed: Using the current HI value, flight status data, and environmental meteorological data as joint query keys, the K-nearest neighbor algorithm is used to quickly find the K most similar tuples in the knowledge base, and then the initial policy is obtained by weighted averaging of the adjustment amounts. .

[0090] This process does not require complex real-time modeling and calculation, and can quickly output a near-optimal strategy that adapts to the current scenario, meeting the requirement of "rapid response to link fluctuations" in UAV communication links (such as generating an initial strategy in seconds).

[0091] Real-time adaptive fine-tuning ensures policy accuracy: The initial policy is fine-tuned using a reinforcement learning framework (Actor-Critic algorithm) by defining the "current state". "Fine-tuning actions" "Reward function" This allows the strategy to dynamically optimize based on real-time link quality changes. If the Hierarchical Index (HI) improves after fine-tuning, a positive reward is given, and the algorithm reinforces the adjustment direction; if the HI decreases, the strategy is adjusted, resulting in the final execution strategy. It can adapt to subtle changes in the current scenario (such as sudden minor interference or fine-tuning of flight attitude) and avoid adjustment deviations in the initial strategy due to minor differences in the scenario.

[0092] Suppose a power line inspection drone is inspecting suburban power lines, and its state at time t is as follows: Communication link health (Preset) (Link quality is "average," needs to be adjusted to a better range) Flight status "Low-speed cruise" (speed 4m / s, altitude 80m), environmental meteorological data "Partly cloudy, light breeze (wind speed 2m / s)" requires parameter adjustment strategies generated through the workflow in this case. The specific process and effects are as follows:

[0093] The parameter adjustment strategy based on this case:

[0094] Step 1: Call the parameter adjustment strategy knowledge base: The knowledge base stores historically verified tuples, such as the following 3 tuples most similar to the current scenario (filtered by the K-nearest neighbor algorithm, K=3):

[0095] Multivariate group 1: , , (MCS level downgraded by 1 level);

[0096] Group 2: , , , , ;

[0097] Multivariate group 3: , , , , ;

[0098] Step 2: Matching the initial strategy Based on the weighted average rule of the K-nearest neighbors algorithm (the higher the similarity, the greater the weight; here, all three tuples have high similarity to the current scene, and each has a weight of 1 / 3), the initial adjustment is calculated:

[0099] Initial transmit power adjustment ;

[0100] Initial MCS Adjustment Level ;

[0101] Obtain the initial strategy ;

[0102] Step 3: Online Adaptive Optimization Generation : Launch the reinforcement learning framework and define the current state Initial fine-tuning actions (Attempt to slightly increase the transmission power); after fine-tuning, Moment reward function (Positive reward indicates that the fine-tuning was effective).

[0103] The Actor-Critic algorithm iteratively optimizes based on this reward, ultimately determining the optimal fine-tuning action. Generate execution strategy .

[0104] If the strategy in this case is not generated, then the traditional "fixed adjustment rule" (such as "when HI is between 0.4 and 0.6, the fixed adjustment rule") is used. ), after execution While there has been an improvement, it has not reached the optimal performance for the current scenario because the fixed rules have not been incorporated into historical scenarios of "cloudy skies and light winds + low-speed cruising". The experience of significantly improving HI in real time was not adjusted according to real-time HI changes, resulting in the link quality only maintaining the lower limit of the "normal" range, and there is still a risk of subsequent fluctuations.

[0105] If a strategy for this case is generated: Execute back, ,and At time (10 seconds after fine-tuning), HI stabilized at 0.58, which is close to the "good" range. This strategy not only relies on historical experience to avoid insufficient adjustments, but also adapts to subtle interferences in the current scenario (such as slight signal fluctuations caused by a light breeze) through real-time fine-tuning. This ensures that parameter adjustments can accurately improve link quality, guarantee the stable transmission of inspection images and data, and avoid task interruptions due to insufficient link quality.

[0106] The following data preprocessing steps are included after step S1 and before step S2:

[0107] First step, credibility assessment: For the credibility score of the i-th data source Calculate its credibility score. The calculation formula is:

[0108] ;

[0109] in A score for whether the data value is within a reasonable physical range. The score is given to determine whether the instantaneous rate of change of the data is reasonable. This data is compared with other relevant data sources. Consistency score, , and These are the weighting coefficients;

[0110] Step Two, Data Repair: For Below the threshold Data The graph neural network (GNN) is used for repair.

[0111] A spatiotemporal graph is constructed using each data source as a node and the physical or statistical correlation between data sources as edges.

[0112] Each node's characteristics are defined as a combination of its historical time-series data and current data;

[0113] The spatiotemporal graph is input into a pre-trained GNN model, and the output feature of the target node i is taken as the repaired value. This is used to replace the original abnormal data;

[0114] Suppose a power line inspection drone collects multi-source data in parallel at 14:30:00, among which the "wireless link transmission bit error rate" data source (denoted as...) Abnormal values ​​were collected due to temporary electromagnetic interference from the sensor. (In this scenario, the distance between the drone and the ground station is 1.2km, the environment is sunny with a light breeze, and the normal range of the bit error rate should be 0.001-0.005. Simultaneously collected RSSI=-62dBm and SNR=30dB, the bit error rate based on physical correlation should not exceed 0.004.) This requires processing through the preprocessing procedure described in this case. The specific process and results are as follows:

[0115] Perform a credibility assessment and set the weighting coefficients for the credibility assessment. (Physical range scoring weight) (Weight of the rate of change score) (Consistency score weight), unreliable data threshold :

[0116] calculate The hardware performance of the drone's communication terminal determines that the physical upper limit of the bit error rate is 0.01%, and outliers... Far exceeding the upper limit, therefore ;

[0117] calculate The bit error rate at the previous moment (14:29:50) was 0.003, and the instantaneous change rate was... This far exceeds the reasonable range of variation (≤0.005), therefore... ;

[0118] calculate : Assuming a normal data source for the same period (RSSI=-62dBm, SNR=30dB), according to the historical data correlation model, when RSSI=-62dBm and SNR=30dB, the average bit error rate is 0.0028. d_5=0.07 severely contradicts this pattern, therefore... ;

[0119] Substitute into the formula to calculate the credibility score: well below the threshold ,determination This data is unreliable and requires repair.

[0120] Perform GNN data repair and construct a spatiotemporal graph:

[0121] Using eight data sources—terminal transmit power, RSSI, SNR, neighboring cell interference intensity, bit error rate, terminal positioning data, flight speed, and ambient humidity—as nodes, edges are constructed between nodes based on physical relationships (e.g., SNR↑→Bit error rate↓, RSSI↑→SNR↑). The feature of each node is defined as "historical time-series data within the past minute (one value every 10 seconds) + current data" (the current data for the bit error rate node is 0.07, while the data for the other nodes are normal).

[0122] Model Repair: The spatiotemporal graph is input into a pre-trained GNN model (this model has been trained with 100,000 normal inspection data points from drones and can accurately learn the correlation patterns between various data sources). The model uses normal data from other nodes (such as RSSI=-62dBm, SNR=30dB) and the historical trend of the bit error rate (the bit error rate has been stable at 0.002-0.003 in the past minute) to output the bit error rate repaired value. ;

[0123] Application effect comparison: Preprocessing without rights in this case: Abnormal bit error rate The process proceeds directly to step S2, the information fusion stage. When the DS evidence theory model calculates the support probability of "link quality status," an abnormally high bit error rate causes the support probability of "poor link quality" to artificially inflate from the normal 0.1 to 0.8, resulting in severe distortion of the output fused feature vector. Consequently, in step S3, when calculating HI, the bit error rate feature (weight 0.2) is erroneously substituted, ultimately resulting in HI = 0.31 (below the "poor link quality" threshold of 0.4). The monitoring system then mistakenly generates an "emergency switch to backup communication link" strategy. However, the actual link quality is good; this erroneous strategy will cause a 2-second communication interruption, resulting in the loss of insulator defect images during inspection, requiring the drone to return and retake the images, thus delaying the mission.

[0124] Preprocessing for this case: Corrected bit error rate In step S2, the DS evidence theory model calculates the support probability based on normal data. The support probability for "good link quality" is 0.75, and the output fused feature vector accurately reflects the actual link status. In step S3, HI is calculated to be 0.73 (which is in the "good link quality" range of 0.6-1.0). The monitoring system does not need to generate adjustment strategies, but only maintains the current communication parameters. Inspection data (such as images of line towers and insulator detection data) are continuously and stably transmitted without data loss or interruption, ensuring that the power inspection task is completed as planned.

[0125] It also includes step S5: visualization and feedback of regulatory effects; the communication link health index HI, the generated regulatory strategy and execution results are displayed graphically in real time on the human-computer interaction interface, and the link quality change data after the strategy execution is used as a feedback signal to adjust the dynamic weight vector in step S3. Alternatively, supervised learning and optimization updates can be performed using the terminal state evaluation function.

[0126] Furthermore, 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. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0127] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0128] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for monitoring unmanned aerial vehicle (UAV) communication terminals based on multi-source information fusion, characterized in that, Includes the following steps: Step S1, Multi-source information acquisition: Parallel acquisition of multi-source heterogeneous data from the UAV communication terminal. The multi-source heterogeneous data includes at least terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, neighboring cell interference intensity, wireless link transmission bit error rate, terminal positioning data, flight status data, and environmental meteorological data. Step S2, Information Fusion and Feature Extraction: The multi-source heterogeneous data collected in Step S1 is spatiotemporally aligned and normalized, and then input into the information fusion model for feature layer fusion, outputting a set of fused feature vectors that can comprehensively characterize the current operating state of the communication terminal. Step S3, Communication Link Health Assessment: Based on the fused feature vector obtained in Step S2, a comprehensive communication link health index value HI is calculated using a preset terminal status assessment function. The process of obtaining HI is as follows: ; in, Represents the fused feature vector. Let i be the i-th normalized eigenvalue. This represents the dynamic weight vector corresponding to each feature. The weight of the i-th feature and ; Step S4, Generation and Execution of Monitoring Strategies: The communication link health index value HI calculated in step S3 is compared with multiple preset threshold intervals. Based on the interval it is in, the preset monitoring system automatically generates corresponding monitoring strategies. The monitoring strategies include parameter adjustment strategies, link switching strategies, or early warning strategies. The generated strategies are then sent to the UAV communication terminal or ground control station for execution.

2. The method for monitoring UAV communication terminals based on multi-source information fusion according to claim 1, characterized in that: In step S2, the information fusion model adopts the DS evidence theory model; The feature layer fusion process includes: assigning the support probability of evidence from different data sources to propositions on different states of communication link quality based on the basic probability allocation function of DS evidence theory, synthesizing the support probabilities of all evidence, and outputting the synthesized probability distribution to form the fused feature vector. The different data sources include at least terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, neighboring cell interference intensity, and bit error rate data. The different states of the communication link quality include "good link quality", "average link quality" and "poor link quality".

3. The method for monitoring UAV communication terminals based on multi-source information fusion according to claim 1, characterized in that: In step S3, the dynamic weight vector It is not a fixed value, but is dynamically adjusted based on environmental meteorological data and the flight status data. When environmental meteorological data indicates heavy rainfall or flight status data indicates that the UAV is in a high-speed maneuver, the weights of the features corresponding to the bit error rate and the neighboring cell interference intensity are automatically increased. Specifically: Judgment criteria: Environmental meteorological data indicates that the rainfall is greater than or equal to the preset rainfall, or flight status data indicates that the drone's flight speed is greater than the preset speed; Weight adjustment: When any of the above conditions are met, the weights of the features corresponding to the bit error rate are adjusted. Increase the preset ratio to adjust the weight of the features corresponding to the neighboring cell interference intensity. Increase the preset ratio, while proportionally reducing the weight of non-critical features such as terminal positioning data and terminal transmission power.

4. The method for monitoring UAV communication terminals based on multi-source information fusion according to claim 1, characterized in that: In step S3, the terminal status evaluation function further includes a time prediction term based on historical data. This prediction term is used to predict the communication link health index for a future period of time, thereby obtaining the future communication link health index. .

5. The method for monitoring UAV communication terminals based on multi-source information fusion according to claim 4, characterized in that: The time-based forecasting item based on historical data is predicted using a time series forecasting algorithm. This time series forecasting algorithm employs a dedicated forecasting model, and the construction and deployment process of this forecasting model includes the following steps: Step 1: Data Preparation: Extract data from historical task logs and construct a training sample set. ; The second step, model training: Construct a Long Short-Term Memory (LSTM) neural network model, with the number of neurons in its input layer corresponding to the dimension n of the fused feature vector, and a single neuron in its output layer; use the mean squared error (MSE) as the loss function and the gradient descent algorithm to train the LSTM model to obtain the optimal model parameters; The third step, model deployment: Integrate the trained LSTM model into the monitoring system to receive the latest fused feature vector sequence output from step S2 in real time and output the predicted value. .

6. The method for monitoring UAV communication terminals based on multi-source information fusion according to claim 1, characterized in that: The generation of the parameter adjustment strategy in step S4 specifically includes the following steps: The first step is to construct a knowledge base for parameter adjustment strategies: the knowledge base stores tuples. ,in In flight mode, For environment mode, and These represent the historically optimal transmit power adjustment amount and modulation and coding scheme adjustment level, respectively. The second step is to match the initial strategy: using the current HI value, flight status data, and environmental meteorological data as the joint query key, the K-nearest neighbor algorithm is used to find the K most similar tuples in the knowledge base, and the adjustment values ​​in these tuples are weighted and averaged to obtain the initial adjustment strategy. ; The third step, online adaptive optimization: fine-tuning the initial policy using a reinforcement learning framework; defining the state. ,action The reward function is a fine-tuning factor for the initial policy. The policy network is iteratively updated using the Actor-Critic algorithm to ultimately generate the execution policy. .

7. The method for monitoring UAV communication terminals based on multi-source information fusion according to claim 1, characterized in that: The following data preprocessing steps are included after step S1 and before step S2: First step, credibility assessment: For the credibility score of the i-th data source Perform calculations; Step Two, Data Repair: For Below the threshold Data The graph neural network (GNN) is used for repair. A spatiotemporal graph is constructed using each data source as a node and the physical or statistical correlation between data sources as edges. Each node's characteristics are defined as a combination of its historical time-series data and current data; The spatiotemporal graph is input into a pre-trained GNN model, and the output feature of the target node i is taken as the repaired value. This is used to replace the original abnormal data.

8. The method for monitoring unmanned aerial vehicle (UAV) communication terminals based on multi-source information fusion according to any one of claims 1-7, characterized in that: It also includes step S5: visualization and feedback of regulatory effects; the communication link health index HI, the generated regulatory strategy and execution results are displayed graphically in real time on the human-computer interaction interface, and the link quality change data after the strategy execution is used as a feedback signal to adjust the dynamic weight vector in step S3. Alternatively, supervised learning and optimization updates can be performed using the terminal state evaluation function.

Images