Unmanned aerial vehicle communication terminal supervision method based on multi-source information fusion

Abstract

The invention discloses an unmanned aerial vehicle communication terminal supervision method based on multi-source information fusion, and the method comprises the following steps: S1, multi-source information collection: collecting multi-source heterogeneous data of an unmanned aerial vehicle communication terminal in parallel, the multi-source heterogeneous data at least comprises terminal transmitting power, received signal strength (RSSI), signal-to-noise ratio (SNR), adjacent region interference strength, wireless link transmission error rate, terminal positioning data, flight state data and environmental meteorological data; and S2, information fusion and feature extraction: carrying out space-time alignment and normalization processing on the multi-source heterogeneous data collected in the step S1, inputting the data into an information fusion model for feature layer fusion, and outputting a group of fusion feature vectors capable of comprehensively representing the current operation state of the communication terminal. According to the invention, more reliable unmanned aerial vehicle communication terminal supervision can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle supervision, in particular to a method for supervising unmanned aerial vehicle communication terminals based on multi-source information fusion. BACKGROUND

[0002] The unmanned aerial vehicle communication terminal is the core component for realizing data interaction between the unmanned aerial vehicle and the ground control station, and the stability of the communication link thereof directly determines whether the unmanned aerial vehicle task can be normally carried out. The essence of supervising the unmanned aerial vehicle communication terminal is to generate and execute intervention measures in a timely manner by monitoring the terminal operation related data in real time, evaluating the communication link quality, so as to ultimately guarantee the reliable transmission of key data between the terminal and the ground control station, and avoid problems such as data loss, transmission interruption or task delay caused by deteriorated link quality. The existing supervision method for the unmanned aerial vehicle communication terminal has significant defects: the protection type is single, and the special unmanned aerial vehicle communication scene cannot be combined. During the flight of the unmanned aerial vehicle, complex situations such as dynamic changes of the environment, adjustment of the flight state, and multi-source interference need to be faced. The traditional scheme only relies on single-dimensional data supervision or fixed rules for protection, and cannot comprehensively evaluate the influence of multi-source heterogeneous data on the communication link quality. Therefore, the present application proposes a method for supervising unmanned aerial vehicle communication terminals based on multi-source information fusion. SUMMARY

[0003] The present application solves the above technical problems by the following technical solutions. The present application comprises the following steps: Step S1, multi-source information collection: parallel collection of multi-source heterogeneous data of the unmanned aerial vehicle communication terminal, wherein the multi-source heterogeneous data at least includes terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, adjacent cell interference intensity, wireless link transmission error rate, terminal positioning data, flight state data and environmental meteorological data; Step S2, information fusion and feature extraction: the multi-source heterogeneous data collected in step S1 is subjected to spatio-temporal alignment and normalization processing, and is input into an information fusion model for feature layer fusion, and a set of fusion feature vectors capable of comprehensively representing the current running state of the communication terminal is output; Step S3, communication link health degree evaluation: based on the fusion feature vector obtained in step S2, a comprehensive communication link health degree index value HI is calculated by a preset terminal state evaluation function, and the acquisition process of HI is: ; wherein, represents the fusion feature vector, is the i-th normalized feature value, represents a dynamic weight vector corresponding to each feature, is the weight of the i-th feature, and ; Step S4, regulatory policy generation and execution: comparing the communication link health index value HI calculated in step S3 with the preset multiple threshold intervals, generating the corresponding regulatory policy from the preset regulatory system according to the interval it is in, the regulatory policy including parameter adjustment strategy, link switching strategy or early warning strategy, and issuing the generated policy to the unmanned aerial vehicle communication terminal or ground control station for execution.

[0004] Further, in step S2, the information fusion model adopts a D-S evidence theory model; The process of feature layer fusion includes: based on the basic probability assignment function of D-S evidence theory, assigning the support probability of different state propositions of communication link quality to the evidence from different data sources, and synthesizing the support probability of all evidence to output the synthesized probability distribution result to form the fusion feature vector; The different data sources at least include terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, adjacent cell interference intensity, and bit error rate data; The different state propositions of the communication link quality include "good link quality", "general link quality", and "poor link quality".

[0005] Further, in step S3, the dynamic weight vector is not a fixed value, but is dynamically adjusted according to environmental meteorological data and the flight state data; When the environmental meteorological data indicates that there is heavy rain or the flight state data indicates that the unmanned aerial vehicle is in a high-speed maneuvering state, the weights of the features corresponding to the bit error rate and the adjacent cell interference intensity are automatically increased , specifically: Judgment condition: environmental meteorological data indicates rainfall≥5mm / h (heavy rain), or flight state data indicates that the flight speed of the unmanned aerial vehicle≥15m / s (high-speed maneuvering state); Weight adjustment: when any of the above conditions is met, the weight of the feature corresponding to the bit error rate is increased by 20%-50%, the weight of the feature corresponding to the adjacent cell interference intensity is increased by 15%-30%, and the weights of non-critical features such as terminal positioning data and terminal transmit power are reduced by the same proportion.

[0006] Further, the terminal state evaluation function in step S3 further includes a time prediction term based on historical data, which is used to predict the communication link health index in the future for a period of time, and obtain the future communication link health index, specifically: ; Wherein, HI is the health index value at the current time. is the future health prediction value obtained by the time series prediction algorithm based on the historical HI value series, and is the weighting coefficient and ; The default supervision system is based on The final supervision strategy is generated based on the value.

[0007] Furthermore, the time prediction item based on historical data is predicted using a time series prediction algorithm, and the time series prediction algorithm adopts a dedicated prediction model. The construction and deployment process of the prediction model includes the following steps: Step 1: Data preparation: Extract data from historical task logs and build a training sample set , where the input is the fusion feature vector sequence of the kth sample in the time window T, label For the future The actual HI value corresponding to the moment; The second step is model training: a long short-term memory (LSTM) neural network model is constructed, where the number of neurons in the input layer corresponds to the dimension n of the fused feature vector, and the output layer is a single neuron; the mean square error (MSE) is used as the loss function, specifically: ,in is the model prediction value, N is the number of samples, and the LSTM model is trained using the gradient descent algorithm to obtain the optimal model parameters; Step 3: Model deployment: Integrate the trained LSTM model into the supervision system to receive the latest fused feature vector sequence output from step S2 in real time and output the predicted value. .

[0008] Furthermore, the generation of the parameter adjustment strategy in step S4 specifically includes the following steps: The first step is to build a parameter adjustment strategy knowledge base: the knowledge base stores multiple groups ,in In flight mode, For ambient mode, and They are respectively the historically optimal transmit power adjustment amount and modulation and coding scheme adjustment level; The second step is to match the initial strategy: using the current HI value, flight status data and environmental weather data as joint query keys, the K nearest neighbor algorithm is used to find the most similar K multi-tuples in the knowledge base, and the weighted average of the adjustment amounts is performed to obtain the initial adjustment strategy. , specifically , is the initial transmit power adjustment amount, Adjust the level for the initial MCS; Step 3: Online adaptive optimization: Use the reinforcement learning framework to fine-tune the initial strategy; define the state ,action is the amount of fine-tuning of the initial strategy, and the reward function is ; Iteratively update the policy network through the Actor-Critic algorithm and finally generate the execution strategy .

[0009] Furthermore, after step S1 and before step S2, the following data preprocessing steps are included: Step 1: Credibility Assessment: Credibility score of the i-th data source Calculate the credibility score The calculation formula is: ; in A score for whether the data value is within a reasonable physical range, Score whether the instantaneous rate of change of the data is reasonable. For this data and other related data sources The consistency score, 、 and is the weight coefficient; The second step, data repair: Below threshold Data , using graph neural network GNN for repair; A spatiotemporal graph is constructed with each data source as a node and the physical or statistical correlations between data sources as edges; The characteristics of each node are defined as the combination of its historical time series data and current data; Input the spatiotemporal graph into the pre-trained GNN model and take the output feature of the target node i as the repaired value , used to replace the original abnormal data.

[0010] Furthermore, it also includes step S5: visualization and feedback of supervision effect; the communication link health index HI, the generated supervision strategy and the execution result are displayed in real time on the human-computer interaction interface in a graphical manner, and the link quality change data after the strategy is executed is used as a feedback signal to adjust the dynamic weight vector in step S3. Or terminal state evaluation function for supervised learning and optimization update.

[0011] Compared with the prior art, the unmanned aerial vehicle communication terminal supervision method based on multi-source information fusion has the following advantages: the method collects terminal transmission power, received signal strength RSSI, signal-to-noise ratio SNR, adjacent cell interference intensity, wireless link transmission error rate, terminal positioning data, flight state data and environmental meteorological data and other multi-source heterogeneous data in parallel, avoids the limitation of single data dimension, can comprehensively capture the terminal running state and external influencing factors, and combines the D-S evidence theory model for feature layer fusion, assigns support probabilities of different data sources to different state propositions of the communication link quality, and synthesizes them, reduces the misjudgment caused by single data error or interference, and the output fusion feature vector can more truly reflect the actual running state of the communication terminal, improving the comprehensiveness and accuracy of supervision. In the communication link health degree evaluation link, the dynamic weight vector is dynamically adjusted according to the environmental meteorological data and the flight state data, for example, when there is heavy rain or high-speed maneuvering of the unmanned aerial vehicle, the weights of the error rate and the adjacent cell interference intensity corresponding features are automatically increased, so that the health degree evaluation is more suitable for the actual scene, and the terminal state evaluation function introduces a time series prediction term based on an LSTM neural network, combines the current health degree index value and the future health degree prediction value to calculate the final evaluation value, so that the supervision system can predict the health trend of the communication link in advance, avoid post-repair, realize forward-looking and adaptive supervision. In the supervision strategy generation and execution aspect, a parameter adjustment strategy knowledge base storing multiple groups is constructed, the current HI value, flight state data and environmental meteorological data are used as joint query keys, the K nearest neighbor algorithm is used to find similar multiple groups to obtain an initial adjustment strategy, and the reinforcement learning framework and Actor-Critic algorithm are used to perform online adaptive fine-tuning on the initial strategy to generate parameter adjustment strategies, link switching strategies or warning strategies and other differentiated strategies, which can cope with full-scene problems from slight link fluctuations to serious quality risks without human intervention, improving the response speed and accuracy of supervision. In addition, after multi-source information collection and before information fusion, the data source credibility score is calculated from three dimensions of physical range, data instantaneous change rate and consistency with other data sources through credibility evaluation, and the data below the threshold is repaired by a graph neural network GNN, a space-time graph is constructed with each data source as a node and the correlation between data sources as an edge, and the pre-trained GNN model is used to output the repaired data, avoiding the interference of abnormal data on subsequent fusion and evaluation, and providing a reliable data basis for the supervision link. At the same time, the method also includes a supervision effect visualization and feedback link, the communication link health index HI, the generated supervision strategy and the execution result are displayed in a graphical way in real time, and the link quality change data after strategy execution is used as a feedback signal to optimize the dynamic weight vector and the terminal state evaluation function, forming a virtuous cycle of evaluation, execution, feedback and optimization, so that the supervision system continuously improves the accuracy with the use scene.Overall, compared with the traditional single protection system, the method realizes the upgrading from passive protection to active supervision, from fixed rules to dynamic adaptation, from local evaluation to global fusion, can effectively cope with the pain points of complex environment, variable state and sensitive link in the unmanned aerial vehicle communication scene, significantly improves the communication link stability and supervision reliability, and is suitable for unmanned aerial vehicle application scenes such as aerial photography, inspection, emergency communication and the like, so that the system is more worthy of popularization and use. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is the overall flowchart of the present application. DETAILED DESCRIPTION

[0013] The embodiments of the present application will be described in detail below, and the embodiments are implemented on the premise of the technical scheme of the present application, and detailed implementation modes and specific operation processes are given, but the protection scope of the present application is not limited to the following embodiments.

[0014] As shown in Figure 1 , the present embodiment provides a technical scheme: a UAV communication terminal supervision method based on multi-source information fusion, comprising the following steps: Step S1, multi-source information acquisition: parallel acquisition of multi-source heterogeneous data of a UAV communication terminal, wherein the multi-source heterogeneous data at least includes terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, adjacent cell interference intensity, wireless link transmission error rate, terminal positioning data, flight state data and environmental meteorological data; Step S2, information fusion and feature extraction: the multi-source heterogeneous data collected in step S1 is subjected to time-space alignment and normalization processing, and is input into an information fusion model for feature layer fusion, and a set of fusion feature vectors capable of comprehensively representing the current running state of the communication terminal is output; Step S3, communication link health degree evaluation: based on the fusion feature vector obtained in step S2, a comprehensive communication link health degree index value HI is calculated by a preset terminal state evaluation function, and the acquisition process of HI is: ; wherein, the fusion feature vector is represented by , the i-th normalized feature value is represented by , the dynamic weight vector corresponding to each feature is represented by , the weight of the i-th feature is represented by ; and Step S4, regulatory policy generation and execution: comparing the communication link health index value HI calculated in step S3 with the preset multiple threshold intervals, generating the corresponding regulatory policy according to the interval it is in by the preset regulatory system, the regulatory policy including parameter adjustment strategy, link switching strategy or early warning strategy, and issuing the generated policy to the unmanned aerial vehicle communication terminal or ground control station for execution.

[0015] In step S2, the information fusion model adopts a D-S evidence theory model; The process of the feature layer fusion includes: based on the basic probability assignment function of the D-S evidence theory, assigning the support probability of different state propositions of the communication link quality to the evidence from different data sources, and synthesizing the support probability of all evidence to output the synthesized probability distribution result to form the fusion feature vector; The different data sources at least include terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, adjacent cell interference intensity and bit error rate data; The different state propositions of the communication link quality include "good link quality", "general link quality" and "poor link quality"; Firstly, the D-S evidence theory model adopted is good at dealing with the uncertainty problem of multi-source information. For the key data sources of terminal transmit power, received signal strength RSSI, signal-to-noise ratio SNR, adjacent cell interference intensity and bit error rate that directly affect the communication link quality, the model can assign the support probability of three explicit state propositions of "good link quality", "general link quality" and "poor link quality" to each data source through the basic probability assignment function, effectively resolving the possible errors, ambiguities or judgment biases of different data sources, and avoiding misjudgment caused by the limitations of a single data source. Secondly, by synthesizing the support probability of all data sources instead of simply adding or filtering data, the effective information of each data source can be fully integrated, the probability distribution result output is more comprehensive and objective to reflect the real quality state of the communication link, and the fusion feature vector formed can accurately and comprehensively represent the current running state of the communication terminal, providing reliable and accurate basic data support for the communication link health assessment in the subsequent steps, reducing the health assessment deviation caused by unscientific information integration, and ensuring the rationality and pertinence of the subsequent regulatory policy generation.

[0016] In step S3, the dynamic weight vector is not a fixed value, but is dynamically adjusted according to the environmental meteorological data and the flight state data; When the environmental meteorological data indicates that there is heavy rain or the flight state data indicates that the unmanned aerial vehicle is in a high-speed maneuvering state, the weights of the features corresponding to the bit error rate and the adjacent cell interference intensity are automatically increased , specifically: Judgment condition: environmental meteorological data indicates rainfall ≥ 5mm / h (heavy rain), or flight state data indicates that the flight speed of the UAV ≥ 15m / s (high-speed maneuvering state); Weight adjustment: when any of the above conditions is met, the weight of the error rate corresponding feature is increased by 20%-50% to increase the weight of the adjacent cell interference intensity corresponding feature by 20%-50% by 15%-30%, while proportionally reducing the weights of non-key features such as terminal positioning data and terminal transmission power; The dynamic weight vector W in step S3 is no longer limited to a fixed value, but is adjusted in real time based on environmental meteorological data and flight state data, which accurately matches the actual characteristics of the dynamic changes in the environment and flight state in the UAV communication scenario. For example, when the environmental meteorological data shows heavy rain, the rain will cause attenuation interference to the wireless signal transmission, resulting in an increase in the wireless link transmission error rate. When the flight state 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 influence of adjacent cell interference intensity on link quality will increase significantly. At this time, increasing the weights of the error rate and adjacent cell interference intensity corresponding features can make these two key factors affecting the current link quality occupy a more reasonable proportion in the calculation of the health index HI, avoiding the drawbacks of fixed weights that can lead to evaluation results deviating from the actual link state. This dynamic adjustment mechanism allows the communication link health evaluation to no longer deviate from the actual scenario, but to accurately focus on key factors that change with the environment and flight state, thereby outputting a HI value that is more in line with the actual situation, providing more reliable evaluation basis for the subsequent step of generating targeted regulatory strategies (such as parameter adjustment, early warning, etc.) for the regulatory system, effectively reducing the problem of inappropriate regulatory strategies caused by evaluation bias, and further ensuring the effectiveness and flexibility of UAV communication link regulation.

[0017] The terminal state evaluation function in step S3 further includes a time prediction term based on historical data, which is used to predict the communication link health index in the future for a period of time. The future communication link health index is obtained, specifically: ; wherein, HI is the health index value at the current time, is the future health prediction value obtained based on the historical HI value sequence through a time series prediction algorithm, and are weighting coefficients and ; the preset regulatory system generates the final regulatory strategy based on the value; By the formula, the health index at the current time is combined with the future health prediction value based on the historical HI value sequence (obtained by the time series prediction algorithm) to form the final HI' value used to generate the supervision strategy. The core benefits of this design are: Early prediction of link risk: Unlike the traditional passive evaluation mode that only looks at the current state, it can predict the health trend of the communication link in the future period of time, avoiding the problem of delayed response of the supervision strategy when the link quality suddenly deteriorates; Optimized strategy foresight: The supervision system generates strategies based on the HI' value that combines the current state and future trend, rather than only relying on the current state, making the strategy more forward-looking and enabling early intervention measures (such as adjusting parameters in advance, preparing for link switching), reducing the probability of link interruption or quality degradation; Adapt to dynamic scenarios of unmanned aerial vehicles: Unmanned aerial vehicles often operate in high-speed moving and changing environments (such as transitioning from sunny to rainy areas), and the link quality is prone to rapid fluctuations. The time prediction term can accurately capture this dynamic trend, making the evaluation results more suitable for the needs of actual communication scenarios of unmanned aerial vehicles; Taking the unmanned aerial vehicle performing a field power inspection task as an example, the specific parameters are set as follows: The time series prediction algorithm uses an LSTM model to predict the value in the future Δt = 60 seconds; The weighting coefficients are α = 0.5 (current state weight) and β = 0.5 (future prediction weight); The current communication link state of the unmanned aerial vehicle at a certain time: = 0.7 (preset HI value 0.6-1.0 for good link quality, currently in good condition); The LSTM model predicts the = 0.45 (preset 0.4-0.6 for "average link quality", below 0.4 for "poor", and the predicted value is close to the threshold of "poor").

[0018] If the design of the present case is not adopted (only use evaluation) The supervision system only relies on t = 0.7 (good), judging that the current link does not need intervention and no adjustment strategy is generated. However, after 60 seconds, as the rain intensifies and the adjacent area interference increases, the actual HI value drops to 0.38 (poor), at which point the system urgently generates a warning or parameter adjustment strategy, which may have caused problems such as interruption of inspection data transmission, screen freezing, and impact on the efficiency of the inspection task.

[0019] First, calculate =0.5×0.7+0.5×0.45=0.575, which is in the range of “normal link quality”. Judgment: While the current link is good, there's a high probability it will degrade to "poor" within the next 60 seconds, necessitating proactive intervention. Therefore, a policy was automatically generated, such as "slightly increase terminal transmit power (ΔPtx = 2dBm) and reduce the modulation and coding scheme (MCS) level from 10 to 8 (to improve anti-interference capabilities)," and distributed to the drone's communication terminal. After 60 seconds, the actual HI value remained at 0.55 (still in the "fair" range) due to this proactive adjustment. Link quality did not deteriorate to "poor," ensuring stable transmission of inspection data and avoiding the risk of mission interruption.

[0020] The time prediction item based on historical data is predicted using a time series prediction algorithm. The time series prediction algorithm adopts a dedicated prediction model. The construction and deployment process of the prediction model includes the following steps: Step 1: Data preparation: Extract data from historical task logs and build a training sample set , where the input is the fusion feature vector sequence of the kth sample in the time window T, label For the future The actual HI value corresponding to the moment; The second step is model training: a long short-term memory (LSTM) neural network model is constructed, where the number of neurons in the input layer corresponds to the dimension n of the fused feature vector, and the output layer is a single neuron; the mean square error (MSE) is used as the loss function, specifically: ,in is the model prediction value, N is the number of samples, and the LSTM model is trained using the gradient descent algorithm to obtain the optimal model parameters; Step 3: Model deployment: Integrate the trained LSTM model into the supervision system to receive the latest fused feature vector sequence output from step S2 in real time and output the predicted value. ; Taking the UAV power inspection scenario as an example, suppose a power inspection UAV needs to operate in a suburban high-voltage line area. In this area, there are often situations such as "increased humidity in the afternoon → signal attenuation" and "UAV flying around the tower → interference fluctuations in neighboring areas". The LSTM model needs to be used to predict the future seconds The specific process and effects are as follows: Based on the constructed LSTM prediction model; Data preparation: Extract data from the power inspection history logs of this type of drone in the past 6 months, set the time window T = 5 minutes (collect a fusion feature vector every 10 seconds, and each sample The fusion feature vector includes 30 continuous fusion feature vectors, the feature dimension n=8, and corresponds to eight core features of "terminal transmission power, RSSI, SNR, adjacent cell interference intensity, bit error rate, terminal positioning accuracy, flight speed, and environmental humidity". Label The true HI value at the moment of 30 seconds after the end of the sample time window (for example, the true HI value at 13:00:00-13:04:50 is 0.95, and the true HI value at 13:05:30 is 0.41) is set as the true HI value at the moment of 30 seconds after the end of the sample time window (for example, the true HI value at 13:00:00-13:04:50 is 0.95, and the true HI value at 13:05:30 is 0.41). The true HI value at the moment of 30 seconds after the end of the sample time window (for example, the true HI value at 13:00:00-13:04:50 is 0.95, and the true HI value at 13:05:30 is 0.41) is set as the true HI value at the moment of 30 seconds after the end of the sample time window (for example, the true HI value at 13:00:00-13:04:50 is 0.95, and the true HI value at 13:05:30 is 0.41). The feature sequence corresponding to 13:00:00-13:04:50, The true HI value at the moment of 30 seconds after the end of the sample time window (for example, the true HI value at 13:00:00-13:04:50 is 0.95, and the true HI value at 13:05:30 is 0.41) is set as the true HI value at the moment of 30 seconds after the end of the sample time window (for example, the true HI value at 13:00:00-13:04:50 is 0.95, and the true HI value at 13:05:30 is 0.41).

[0021] Model training: build an LSTM neural network, the number of input layer neurons=8 (match the feature dimension n), set 2 layers of hidden layers (64 neurons in each layer, use ReLU activation function), and the output layer is one neuron (output ); use MSE as the loss function, use Adam gradient descent algorithm (learning rate=0.001) to train the model, after 100 iterations, the training set MSE decreases from the initial 0.15 to 0.009, and the validation set MSE is stable within 0.012, the model converges and has good generalization ability.

[0022] Model deployment: integrate the trained LSTM model into the unmanned aerial vehicle ground supervision system, the system automatically receives the latest fusion feature vector output by step S2 every 10 seconds, real-time splicing into the feature sequence of the last 5 minutes (i.e. ) and input into the LSTM model, and the model outputs the value of the future 40 seconds within 20 milliseconds.

[0023] Application effect comparison (i.e. with or without LSTM model), LSTM model (using traditional ARIMA algorithm): if only ARIMA algorithm is used to predict , the algorithm can only predict based on the historical HI value single sequence, and cannot combine multiple feature changes such as "environmental humidity rising (from 60% to 85%), flight speed increasing (from 4m / s to 7m / s), adjacent cell interference intensity rising (from -88dBm to -75dBm)". For example, at 13:00:00, the current (pre-set "link quality is good"), ARIMA predicts that 40 seconds later (still judged as good), but actually due to the increase of humidity leading to signal attenuation and adjacent cell interference enhancement, the real HI value decreases to 0.41 (pre-set "link quality is poor", which is close to the threshold of poor quality), the supervision system does not intervene in advance, resulting in 15 seconds of inspection image transmission jam of the unmanned aerial vehicle, affecting the line fault identification efficiency.

[0024] With LSTM model: At 13:00:00, the supervision system inputs the feature sequence of the LSTM model The characteristic trends of "continuous increase in environmental humidity, increased flight speed, and increased interference in neighboring areas" are already included in the LSTM model. Based on these association rules, the LSTM model outputs the next 40 seconds. ; Substitution (Pre The monitoring system immediately generated a policy to increase terminal transmit power by 2.5dBm and lower the modulation and coding scheme (MCS) level from 11 to 9 (to enhance anti-interference capabilities) and sent it to the drone. Forty seconds later, the actual HI value remained at 0.57 (still in the "fair" range), with no image transmission interruptions, ensuring the continuous and stable progress of power inspections.

[0025] The generation of the parameter adjustment strategy in step S4 specifically includes the following steps: The first step is to build a parameter adjustment strategy knowledge base: the knowledge base stores multiple groups ,in In flight mode, For ambient mode, and They are respectively the historically optimal transmit power adjustment amount and modulation and coding scheme adjustment level; The second step is to match the initial strategy: using the current HI value, flight status data and environmental weather data as joint query keys, the K nearest neighbor algorithm is used to find the most similar K multi-tuples in the knowledge base, and the weighted average of the adjustment amounts is performed to obtain the initial adjustment strategy. , specifically , is the initial transmit power adjustment amount, Adjust the level for the initial MCS; Step 3: Online adaptive optimization: Use the reinforcement learning framework to fine-tune the initial strategy; define the state ,action is the amount of fine-tuning of the initial strategy, and the reward function is ; Iteratively update the policy network through the Actor-Critic algorithm and finally generate the execution strategy ; Relying on historical experience to reduce the blindness of initial strategies: by building storage multi-groups The parameter adjustment strategy knowledge base solidifies the "scenario-optimal adjustment amount" correspondence that has been verified to be effective in historical tasks, avoiding the problem of "departing from actual experience and starting from scratch" when generating traditional strategies. Flight status mode), The introduction of the environmental mode allows historical experience to be accurately associated with the current scenario, ensuring that the initial strategy is reasonably based.

[0026] Quickly match the scene and improve the response speed of the strategy: use the current HI value, flight state data, and environmental meteorological data as the joint query key, quickly find the K most similar multiple tuples in the knowledge base through the K nearest neighbor algorithm, and then take a weighted average of the adjustment amount to obtain the initial strategy .

[0027] This process does not require complex real-time modeling calculations and can quickly output an approximately optimal strategy that adapts to the current scenario, meeting the demand for rapid response to link fluctuations (such as generating an initial strategy within seconds).

[0028] Real-time adaptive fine-tuning to ensure strategy accuracy: use the reinforcement learning framework (Actor-Critic algorithm) to fine-tune the initial strategy, define the "current state ", "fine-tuning action ", and "reward function ", so that the strategy can be dynamically optimized according to real-time link quality changes. If the HI improves after fine-tuning, a positive reward is obtained, and the algorithm will strengthen the adjustment direction. If the HI decreases, the strategy is adjusted, and the final execution strategy can adapt to subtle changes in the current scenario (such as sudden small disturbances and flight attitude fine-tuning), avoiding adjustment deviations caused by minor differences in the initial strategy.

[0029] Suppose a power inspection UAV is conducting a line inspection in the suburbs, and the state at time t is as follows: communication link health (preset as "average link quality" and needs to be adjusted to a better range), flight state "low-speed cruising" (speed 4 m / s, height 80 m), and environmental meteorological data "cloudy, light wind (wind speed 2 m / s)", the parameter adjustment strategy needs to be generated through the process of this case, and the specific process and effects are as follows: Based on the parameter adjustment strategy of this case: First step: call the parameter adjustment strategy knowledge base: the knowledge base stores multiple tuples that have been historically verified, such as the following 3 multiple tuples that are most similar to the current scenario (selected through the K nearest neighbor algorithm, K=3): Multiple tuple 1: , , (MCS level down by 1 level); Multiple tuple 2: , , , , ; tuple 3: , , , , ; Step 2: Match the initial strategy : According to the weighted average rule of the K-nearest neighbor algorithm (the higher the similarity, the greater the weight. Here, the three multi-tuples are all highly similar to the current scene, and the weight is 1 / 3), calculate the initial adjustment amount: Initial transmit power adjustment ; Initial MCS adjustment level ; Get the initial strategy ; Step 3: Online adaptive optimization generation : Start the reinforcement learning framework and define the current state , initial fine-tuning action (Try to increase the transmit power slightly); after fine-tuning, Moment , the reward function (Positive rewards indicate that fine-tuning is effective).

[0030] The Actor-Critic algorithm iteratively optimizes based on this reward and ultimately determines the optimal fine-tuning action , generate execution strategy .

[0031] If there is no strategy generation for this case, the traditional "fixed adjustment rule" (such as "when HI is 0.4-0.6, fixed adjustment rule") is used. ”), after execution Although there is improvement, it does not reach the optimal effect in the current scenario because the fixed rule does not combine the historical "cloudy breeze + low-speed cruising" scenario. The experience of more obvious improvement of HI in real time was not achieved, and fine-tuning was not made according to the real-time HI changes, resulting in the link quality being maintained at the lower limit of the "general" range, and there is still a risk of subsequent fluctuations.

[0032] If a strategy is generated for this case: execute back, ,and At this moment (10 seconds after fine-tuning), HI is stable at 0.58, which is close to the "good" range ( This strategy not only relies on historical experience to avoid under-adjustments, but also adapts to subtle interference in the current scenario (such as slight signal fluctuations caused by 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 mission interruptions caused by insufficient link quality.

[0033] After step S1 and before step S2, the following data preprocessing steps are further included: First step, credibility assessment: for the i-th data source, the credibility score is The calculation is as follows: The calculation formula is: ; Wherein is the score of whether the data value is in a reasonable physical range, is the score of whether the instantaneous change rate of the data is reasonable, is the consistency score of the data with other related data sources , , and are weight coefficients; Second step, data repair: for data below the threshold , the graph neural network GNN is used for repair; A space-time graph is constructed with each data source as a node and the physical or statistical correlation between data sources as an edge; The feature of each node is defined as the combination of its historical time series data and current data; The space-time graph is input into the pre-trained GNN model, and the output feature of the target node i is taken as the repaired value , which is used to replace the original abnormal data; Suppose a power inspection unmanned aerial vehicle collects multi-source data at 14:30:00, and the "wireless link transmission error rate" data source (denoted as ) collects an abnormal value due to temporary electromagnetic interference of the sensor (in this scenario, the unmanned aerial vehicle is 1.2 km away from the ground station, the environment is sunny and windy, the normal range of error rate should be 0.001-0.005, and the RSSI=-62dBm and SNR=30dB collected at the same period, according to the physical association, the error rate should not exceed 0.004), which needs to be processed through the preprocessing process of the case, and the specific process and effect are as follows: Performing credibility assessment sets the weight coefficients of credibility assessment (the physical range score weight), (the change rate score weight), (the consistency score weight), and the unreliable data threshold : Calculate : the hardware performance of the unmanned aerial vehicle communication terminal determines that the upper limit of the error rate is 0.01, and the abnormal value is far beyond the upper limit, so ; Compute : the error rate at the previous moment (14:29:50) is 0.003, the instantaneous change rate , far beyond the reasonable change range (≤0.005), so ; Compute : is the normal data source at the same time (RSSI=-62dBm, SNR=30dB), according to the historical data association model, when RSSI=-62dBm, SNR=30dB, the average error rate is 0.0028, d_5=0.07 is in serious conflict with this rule, so ; Substitute the formula to calculate the credibility score: , which is much lower than the threshold , it is determined as unreliable data, and repair needs to be performed.

[0034] Perform GNN data repair and construct a spatio-temporal graph: Take "terminal transmit power, RSSI, SNR, adjacent interference intensity, error rate, terminal positioning data, flight speed, and environmental humidity" as nodes, and build edges between nodes according to physical association (such as SNR↑→error rate↓, RSSI↑→SNR↑); The characteristics of each node are defined as "historical time series data in the past 1 minute (1 value every 10 seconds) + current time data" (the current data of the error rate node is 0.07, and the data of the other nodes are normal); Model repair: input the spatio-temporal graph into the pre-trained GNN model (this model has been trained with 100,000 normal UAV inspection data and can accurately learn the association rules between data sources), and the model outputs the repaired value of the error rate based on the normal data of other nodes (such as RSSI=-62dBm, SNR=30dB) and the historical trend of the error rate (the error rate has been stable at 0.002-0.003 for the past 1 minute); Application effect comparison: without the preprocessing of this case, the abnormal error rate directly enters the information fusion link of step S2. In the D-S evidence theory model, the abnormally high value of the error rate will make the support probability of "poor link quality" from the normal 0.1 to 0.8, causing the output fusion feature vector to be severely distorted; Then in step S3, the error value is brought in by the error rate feature (weight 0.2) when calculating HI, and finally HI=0.31 (lower than the "poor link quality" threshold 0.4), the supervision system generates the "emergency switch to backup communication link" strategy. But the actual link quality is good, this error strategy will cause the communication to be interrupted for 2 seconds, causing the insulator defect image to be lost during the inspection process, and the UAV needs to return to take pictures again, delaying the task progress.

[0035] There is a pre-processing of the present case: the bit error rate after repair Enter step S2, D-S evidence theory model based on normal data to calculate support probability, the support probability of "link quality good" is 0.75, the output fusion feature vector accurately reflects the actual link state; Step S3 calculates HI=0.73 (in "link quality good" interval 0.6-1.0), the regulatory system does not need to generate adjustment strategy, only maintain the current communication parameters, the inspection data (such as line tower image, insulator detection data) continue to transmit stably, no data loss or interruption, guarantee the power inspection task to complete the plan.

[0036] Also includes step S5: regulatory effect visualization and feedback; the communication link health index HI, the generated regulatory strategy and the execution result are displayed in the man-machine interface in a graphical way in real time, and the link quality change data after the strategy execution is used as a feedback signal to supervise the learning and optimization update of the dynamic weight vector Or terminal state evaluation function in step S3.

[0037] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise specifically limited.

[0038] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms is not necessarily for the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, the skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of different embodiments or examples without contradiction.

[0039] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A method for monitoring UAV communication terminals based on multi-source information fusion, characterized in that: The following steps are involved: Step S1, multi-source information collection: parallel collection of multi-source heterogeneous data of 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 strength, 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 are subjected to spatiotemporal alignment and normalization, and input into the information fusion model for feature layer fusion, outputting a set of fused feature vectors that can comprehensively characterize the current operating status of the communication terminal; Step S3, communication link health evaluation: 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 evaluation function. The process of obtaining HI is as follows: ; in, represents the fused feature vector, is the normalized eigenvalue of the i-th eigenvalue, Represents the dynamic weight vector corresponding to each feature, is the weight of the i-th feature and ; Step S4, supervision strategy generation and execution: The communication link health index value HI calculated in step S3 is compared with multiple preset threshold intervals. The preset supervision system automatically generates a corresponding supervision strategy based on the interval in which it is located. The supervision strategy includes a parameter adjustment strategy, a link switching strategy or an early warning strategy, and the generated strategy is 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 is characterized by: In step S2, the information fusion model adopts the DS evidence theory model; The feature layer fusion process includes: assigning support probabilities for different communication link quality status propositions to evidence from different data sources based on the basic probability distribution function of DS evidence theory, synthesizing the support probabilities of all evidence, and outputting the synthesized probability distribution result 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 strength and bit error rate data; The different status propositions of the communication link quality include "good link quality", "fair link quality" and "poor link quality".

3. The method for monitoring UAV communication terminals based on multi-source information fusion according to claim 1 is characterized by: In step S3, the dynamic weight vector It is not a fixed value, but is dynamically adjusted according to the environmental meteorological data and the flight status data; When the environmental meteorological data indicates that there is heavy rainfall or the flight status data indicates that the UAV is in a high-speed maneuvering state, the weights of the characteristics corresponding to the bit error rate and the neighboring area interference intensity are automatically increased. , specifically: Judgment conditions: 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 weight of the bit error rate corresponding feature is adjusted. Increase the preset ratio and adjust the weight of the feature corresponding to the neighboring area interference intensity Increase the preset ratio and proportionally reduce 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 is characterized by: The terminal status evaluation function in step S3 further includes a time prediction item based on historical data, which is used to predict the communication link health index in the future period of time and obtain the future communication link health index. .

5. The method for monitoring UAV communication terminals based on multi-source information fusion according to claim 4 is characterized in that: The time prediction item based on historical data is predicted using a time series prediction algorithm. The time series prediction algorithm adopts a dedicated prediction model. The construction and deployment process of the prediction model includes the following steps: Step 1: Data preparation: Extract data from historical task logs and build a training sample set ; The second step is model training: a long short-term memory (LSTM) neural network model is constructed, in which the number of neurons in the input layer corresponds to the dimension n of the fused feature vector, and the output layer is a single neuron; the mean square error (MSE) is used as the loss function, and the gradient descent algorithm is used to train the LSTM model to obtain the optimal model parameters; Step 3: Model deployment: Integrate the trained LSTM model into the supervision 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 is characterized by: The generation of the parameter adjustment strategy in step S4 specifically includes the following steps: The first step is to build a parameter adjustment strategy knowledge base: the knowledge base stores multiple groups ,in In flight mode, For ambient mode, and They are respectively the historically optimal transmit power adjustment amount and modulation and coding scheme adjustment level; The second step is to match the initial strategy: using the current HI value, flight status data and environmental weather data as joint query keys, the K nearest neighbor algorithm is used to find the most similar K multi-tuples in the knowledge base, and the weighted average of the adjustment amounts is performed to obtain the initial adjustment strategy. ; Step 3: Online adaptive optimization: Use the reinforcement learning framework to fine-tune the initial strategy; define the state ,action is the amount of fine-tuning of the initial strategy, and the reward function is ; Iteratively update the policy network through the Actor-Critic algorithm and finally generate the execution strategy .

7. The method for monitoring UAV communication terminals based on multi-source information fusion according to claim 1 is characterized by: After step S1 and before step S2, the following data preprocessing steps are also included: Step 1: Credibility Assessment: Credibility score of the i-th data source Perform calculations; The second step, data repair: Below threshold Data , using graph neural network GNN for repair; A spatiotemporal graph is constructed with each data source as a node and the physical or statistical correlations between data sources as edges; The characteristics of each node are defined as the combination of its historical time series data and current data; Input the spatiotemporal graph into the pre-trained GNN model and take the output feature of the target node i as the repaired value , used to replace the original abnormal data.

8. The method for monitoring UAV communication terminals based on multi-source information fusion according to any one of claims 1 to 7, characterized in that: The step S5 also includes: visualization and feedback of supervision effect; the communication link health index HI, the generated supervision strategy and the execution result are displayed in real time on the human-computer interaction interface in a graphical manner, and the link quality change data after the strategy is executed is used as a feedback signal to adjust the dynamic weight vector in step S3. Or terminal state evaluation function for supervised learning and optimization update.

Images