An unmanned aerial vehicle air-ground link quality prediction method based on a distributed improved whale optimization algorithm

Abstract

This invention discloses a method for predicting UAV air-to-ground link quality based on a distributed improved whale optimization algorithm, belonging to the field of UAV communication and intelligent prediction technology. Addressing the problem that link status in UAV image transmission, low-altitude inspection, and emergency communication is easily affected by flight status, air-to-ground distance, terrain obstruction, multipath propagation, and environmental disturbances, leading to large link quality fluctuations and insufficient prediction stability, this invention constructs a link quality prediction model based on the Spark distributed computing framework. This method collects UAV flight status, link monitoring, and environmental auxiliary data, and constructs link change characteristics after preprocessing; it uses an improved whale optimization algorithm to optimize random forest parameters, and combines Lévy flight initialization and Brownian motion perturbation to improve global search capabilities; finally, it utilizes Spark to compute the optimal parameter combination in parallel to establish the link quality prediction model. This invention can improve the accuracy, stability, and parameter optimization efficiency of UAV dynamic air-to-ground link quality prediction.

Description

Technical Field

[0001] This invention belongs to the field of UAV communication and wireless link state prediction technology, and relates to a method for predicting the quality of UAV air-to-ground links in dynamic air-to-ground link environments. More particularly, it relates to a method for predicting the quality of UAV air-to-ground links that combines distributed computing, Lévy flight initialization, an improved whale optimization algorithm, and a random forest model. Background Technology

[0002] With the development of applications such as low-altitude economy, intelligent inspection, emergency rescue, remote sensing mapping, and border patrol, drones have been widely used for image acquisition, video transmission, and real-time monitoring in complex environments. Unlike fixed ground communication scenarios, drone air-to-ground communication links are highly dynamic. During flight, the link quality of a drone is affected by factors such as flight altitude, flight speed, attitude changes, air-to-ground distance, obstruction, multipath propagation, terrain environment, and weather disturbances. This results in nonlinear, non-stationary, and time-varying characteristics in indicators such as signal-to-noise ratio, received signal strength, packet loss rate, bit error rate, transmission delay, and throughput.

[0003] Currently, existing methods for predicting UAV link quality mainly include empirical threshold-based methods, traditional statistical model methods, and single machine learning model methods. Empirical threshold-based methods typically set fixed thresholds based on received signal strength, signal-to-noise ratio, or packet loss rate to determine link reliability; however, these methods struggle to adapt to dynamic changes in link status under complex environments. Traditional statistical models, such as autoregressive models and autoregressive moving average models, are usually based on linear assumptions and cannot accurately characterize the nonlinear and multi-factor coupling characteristics of UAV air-to-ground links. While machine learning models such as random forests, support vector machines, and neural networks can improve prediction capabilities, their performance often depends on parameter selection; inappropriate parameter settings can easily lead to insufficient prediction stability or decreased generalization ability.

[0004] Whale optimization is a swarm intelligence optimization algorithm characterized by its simple structure, few parameters, and strong global search capability, making it suitable for optimizing prediction model parameters. However, traditional whale optimization algorithms still have certain shortcomings in UAV link quality prediction tasks: First, the initial population exhibits strong randomness, and the distribution of parameter combinations in the search space is insufficient, easily affecting the global search capability; second, under complex multi-source link data, the algorithm may get trapped in local optima, leading to unstable optimization results for prediction model parameters; third, when faced with large-scale UAV flight status data and link monitoring data, the single-machine iterative computation efficiency is low, making it difficult to meet the requirements of rapid modeling and real-time prediction.

[0005] To address the aforementioned problems, this invention proposes a method for predicting the quality of UAV air-to-ground links. This method constructs a multi-source state feature vector based on UAV flight status, air-to-ground link monitoring status, and environmental auxiliary information. It then utilizes the Spark distributed computing framework to compute fitness values ​​corresponding to different random forest parameter combinations in parallel. A Lévy flight mechanism is employed to enhance the search diversity of the initial whale population. Parameter optimization is achieved by combining the shrinking encirclement, foraging, and spiral update strategies of the whale optimization algorithm. Furthermore, a Brownian motion perturbation mechanism is introduced to update superior parameter combinations and replace inferior ones, thereby reducing the probability of getting trapped in local optima. Finally, a UAV air-to-ground link quality prediction model is constructed to predict future link quality. Summary of the Invention

[0006] In view of this, the technical problem to be solved by the present invention is to provide a method for predicting the air-to-ground link quality of unmanned aerial vehicles (UAVs), aiming to overcome the problems existing in the link quality prediction method, such as insufficient characterization of multi-source link states, reliance on manual experience in setting model parameters, low efficiency in parameter optimization, and susceptibility to local optima in the optimization process. The technical solution adopted by the present invention is as follows:

[0007] A method for predicting UAV air-to-ground link quality based on a distributed improved whale optimization algorithm includes the following steps:

[0008] Step 1) Collect multi-source status data during the flight of the UAV. The multi-source status data includes UAV flight status data, air-to-ground link monitoring data and environmental auxiliary information. Synchronize the collected multi-source status data according to timestamps and store the synchronized data in a distributed file system to form a UAV air-to-ground link quality prediction dataset.

[0009] Step 2) Preprocess and feature construct the UAV air-to-ground link quality prediction dataset. The preprocessing includes outlier removal, missing value completion, normalization, and categorical variable encoding. The feature constructing includes the construction of current state features, state change features, and historical time window statistical features, thereby obtaining UAV air-to-ground link state features, which are used to characterize the dynamic change law of air-to-ground link quality during UAV flight.

[0010] Step 3) Construct the parameter search space of the random forest prediction model. Use the random forest model as the UAV air-to-ground link quality prediction model. Encode the parameters to be optimized in the random forest model into the position vectors of individual whales. Construct a whale population so that each individual whale corresponds to a set of candidate random forest parameter combinations.

[0011] Step 4) Construct a whale population initialization module based on Levy flight. First, randomly generate the first whale individual, and then use the Levy flight mechanism to generate subsequent whale individuals, so that the whale individuals produce positional changes at different scales in the parameter search space, thereby expanding the initial parameter search range and enhancing population diversity.

[0012] Step 5) Construct a fitness calculation module based on distributed computing. Use the Spark distributed computing framework to distribute the random forest parameter combinations corresponding to different whale individuals to different computing nodes for parallel training. Each computing node trains the random forest model based on the corresponding random forest parameter combination, calculates the link quality prediction error on the validation sample set, determines the fitness value of each whale individual based on the link quality prediction error, and obtains the current optimal parameter combination through distributed aggregation.

[0013] Step 6) Construct an improved whale optimization parameter update module. Based on the current optimal parameter combination, random probability, and coefficient vector, iteratively update the random forest parameter combination represented by the individual whale. The iterative update includes shrinking enclosing update, searching and foraging update, and spiral update, which are used to realize the global search and local refinement of the random forest parameter combination.

[0014] Step 7) Construct a Brownian motion perturbation module, sort the parameter combinations in the current whale population according to their fitness values, determine the current optimal and second-best parameter combinations, apply Brownian motion random perturbation to the current optimal and second-best parameter combinations, generate new candidate parameter combinations, and replace the parameter combinations with lower fitness in the current whale population with the new candidate parameter combinations to enhance search diversity and reduce the probability of getting trapped in local optima.

[0015] Step 8) Construct a UAV air-to-ground link quality prediction model. When the number of iterations reaches the preset maximum number of iterations, or the fitness value meets the preset convergence condition, output the optimal random forest parameter combination. Based on the optimal random forest parameter combination, construct the UAV air-to-ground link quality prediction model, and input the UAV air-to-ground link state characteristics before the time to be predicted into the UAV air-to-ground link quality prediction model, and output the link quality prediction results for future times.

[0016] Furthermore, step 1) specifically includes the following steps:

[0017] Step 1.1) Multi-source status data acquisition. Acquire UAV flight status data, air-to-ground link monitoring data, and environmental auxiliary information during UAV flight. The UAV flight status data includes at least one of the following: flight altitude, flight speed, flight acceleration, attitude angle, heading angle, air-to-ground distance, and trajectory position. The air-to-ground link monitoring data includes at least one of the following: signal-to-noise ratio, received signal strength, reference signal received power, packet loss rate, bit error rate, transmission delay, throughput, and retransmission count. The environmental auxiliary information includes at least one of the following: weather condition, terrain type, occlusion status, line-of-sight / non-line-of-sight status, and mission scenario type.

[0018] Step 1.2) Data Time Synchronization and Storage. The collected multi-source state data is aligned according to timestamps. For data with different sampling frequencies, resampling or interpolation is used for synchronization to form a unified time series sample. The synchronized data is then stored in a distributed file system to support subsequent distributed reading and parallel computing.

[0019] Furthermore, step 2) specifically includes the following steps:

[0020] Step 2.1) Data Preprocessing. The UAV air-to-ground link quality prediction dataset undergoes outlier removal, missing value completion, normalization, and categorical variable encoding. For continuous variables such as flight altitude, flight speed, air-to-ground distance, signal-to-noise ratio, received signal strength, packet loss rate, transmission delay, and throughput, normalization is used to eliminate the influence of different units on model training. For categorical variables such as weather state, terrain type, occlusion status, line-of-sight / non-line-of-sight status, and mission scenario type, encoding is used to convert them into numerical features that the model can process.

[0021] Step 2.2) Construction of Current State Features. Based on the standardized multi-source state data, current state features are constructed. These current state features are used to describe the UAV's flight state, air-to-ground link monitoring state, and environmental state at the current moment.

[0022] Step 2.3) Construction of State Change Features. Based on the changes in the UAV's flight state and air-to-ground link state between adjacent time points, state change features are constructed. These state change features are used to describe at least one of the following: changes in flight altitude, flight speed, air-to-ground distance, signal-to-noise ratio, received signal strength, packet loss rate, and transmission delay.

[0023] Step 2.4) Construction of historical time window statistical features. Based on a historical time window of preset length, at least one of the following air-to-ground link monitoring features within the window is statistically analyzed to describe the fluctuation pattern of link quality over a past period.

[0024] Step 2.5) Formation of UAV air-to-ground link state features. The current state features, state change features, and historical time window statistical features are combined to obtain the UAV air-to-ground link state features, which are used for training and prediction of the subsequent UAV air-to-ground link quality prediction model.

[0025] Furthermore, step 3) specifically includes the following steps:

[0026] Step 3.1) Construction of the Random Forest Prediction Model. A random forest model is used as the UAV air-to-ground link quality prediction model to establish the mapping relationship between the UAV air-to-ground link state characteristics and the link quality prediction results at future times.

[0027] Step 3.2) Random Forest Parameter Encoding. The parameters to be optimized in the random forest model are encoded as position vectors for individual whales, so that each whale corresponds to a set of candidate random forest parameter combinations. The parameters to be optimized include at least one of the following: number of decision trees, maximum tree depth, maximum number of candidate features, and minimum number of leaf node samples.

[0028] Step 3.3) Setting the Parameter Search Space. Based on the range of values ​​for the random forest model parameters, set the positional boundaries of individual whales so that they can search within the preset parameter search space.

[0029] Furthermore, step 4) specifically includes the following steps:

[0030] Step 4.1) Generation of the first whale individual. Randomly generate the first whale individual and use it as a set of initial parameters for the random forest model.

[0031] Step 4.2) Levy Flight Initialization. Based on the position of the previous whale individual, subsequent whale individuals are generated through the Levy flight mechanism, causing the whale individuals to undergo positional changes at different scales in the parameter search space.

[0032] Step 4.3) Boundary Correction. Perform boundary detection on the generated whale individual positions. If the whale individual position exceeds the preset parameter search space, correct it to the corresponding parameter range.

[0033] Furthermore, step 5) specifically includes the following steps:

[0034] Step 5.1) Construction of training and validation sample sets. The UAV air-to-ground link quality prediction dataset is divided into a training sample set and a validation sample set. The training sample set is used for training the random forest model, and the validation sample set is used to calculate the link quality prediction error.

[0035] Step 5.2) Distributed Task Allocation. Using the Spark distributed computing framework, the random forest parameter combinations corresponding to different whale individuals are allocated to different computing nodes, enabling multiple random forest model training tasks to be executed in parallel.

[0036] Step 5.3) Fitness value calculation. Each computing node trains a random forest model based on the corresponding random forest parameter combination and calculates the link quality prediction error on the validation sample set. The fitness value of each individual whale is determined based on the link quality prediction error.

[0037] Step 5.4) Global Optimal Parameter Combination Update. The fitness values ​​of each computing node are aggregated through distributed aggregation operations to obtain the optimal parameter combination in the current iteration, which is then used as the basis for subsequent whale optimization parameter updates.

[0038] Furthermore, step 6) specifically includes the following steps:

[0039] Step 6.1) Parameter update condition judgment. Based on the current optimal parameter combination, random probability, and coefficient vector, determine whether the current whale individual adopts a shrinking encirclement, searching for food, or spiral update method.

[0040] Step 6.2) Shrinking the Enclosure Update. When the shrinking enclosure condition is met, guide the current random forest parameter combination to move closer to the current optimal parameter combination to enhance local exploitation capabilities.

[0041] Step 6.3) Search and foraging update. When the search and foraging conditions are met, a global search is performed based on a randomly selected combination of parameters to expand the parameter search range.

[0042] Step 6.4) Spiral Update. When the spiral update condition is met, a refined search is performed around the current optimal parameter combination to improve the accuracy of parameter optimization.

[0043] Furthermore, step 7) specifically includes the following steps:

[0044] Step 7.1) Whale population sorting. Sort the parameter combinations in the current whale population according to their fitness values ​​to determine the current optimal and second-best parameter combinations.

[0045] Step 7.2) Brownian motion perturbation. Apply Brownian motion random perturbation to the current optimal and second-best parameter combinations to generate new candidate parameter combinations.

[0046] Step 7.3) Replacement of low-fit individuals. Replace the low-fit parameter combinations in the current whale population with new candidate parameter combinations to enhance population search diversity.

[0047] Furthermore, step 8) specifically includes the following steps:

[0048] Step 8.1) Output the optimal random forest parameter combination. When the number of iterations reaches the preset maximum number of iterations, or the fitness value meets the preset convergence condition, the optimal random forest parameter combination is output.

[0049] Step 8.2) Construction of UAV air-to-ground link quality prediction model. A UAV air-to-ground link quality prediction model is constructed based on the optimal random forest parameter combination.

[0050] Step 8.3) Link Quality Prediction. Input the UAV air-to-ground link state characteristics before the time to be predicted into the UAV air-to-ground link quality prediction model, and output the link quality prediction results for the future time. Attached Figure Description

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0052] Figure 1 This is a flowchart of the method;

[0053] Figure 2 This is a diagram illustrating the overall model framework of this method;

[0054] Figure 3 A schematic diagram of the UAV multi-source state data acquisition and synchronization module constructed using this method;

[0055] Figure 4 A schematic diagram of the UAV air-to-ground link state feature construction module constructed by this method;

[0056] Figure 5 A schematic diagram of the Spark distributed fitness computation module constructed using this method;

[0057] Figure 6 A schematic diagram of the Levi flight initialization module constructed using this method;

[0058] Figure 7 A schematic diagram of the improved whale optimization parameter update module constructed using this method;

[0059] Figure 8 A schematic diagram of the Brownian motion perturbation module constructed using this method;

[0060] Figure 9 A schematic diagram of the UAV air-to-ground link quality prediction output module constructed using this method;

[0061] Figure 10 A schematic diagram illustrating how link quality prediction results assist in UAV image backhaul decision-making. Detailed Implementation

[0062] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0063] This invention provides a method for predicting the quality of air-to-ground links for unmanned aerial vehicles (UAVs), the flowchart of which is shown below. Figure 1 As shown, the overall model framework diagram of this method is as follows: Figure 2 As shown, the method includes the following steps:

[0064] Step 1) Collect multi-source status data during the drone's flight;

[0065] Step 2) Construct the UAV air-to-ground link state characteristics;

[0066] Step 3) Construct the parameter search space for the random forest prediction model;

[0067] Step 4) Construct a whale population initialization module based on Levi's flight;

[0068] Step 5) Construct a fitness calculation module based on distributed computing;

[0069] Step 6) Construct an improved whale optimization parameter update module;

[0070] Step 7) Construct the Brownian motion perturbation module;

[0071] Step 8) Construct a UAV air-to-ground link quality prediction model.

[0072] Furthermore, step 1) specifically includes the following steps:

[0073] Step 1.1) as Figure 3 As shown, the data collected during the UAV's flight includes flight status data, air-to-ground link monitoring data, and environmental auxiliary information. The UAV flight status data includes at least one of the following: flight altitude, flight speed, flight acceleration, attitude angle, heading angle, air-to-ground distance, and trajectory position. The air-to-ground link monitoring data includes at least one of the following: signal-to-noise ratio, received signal strength, reference signal received power, packet loss rate, bit error rate, transmission delay, throughput, and retransmission count. The environmental auxiliary information includes at least one of the following: weather condition, terrain type, occlusion status, line-of-sight / non-line-of-sight status, and mission scenario type.

[0074] Step 1.2) as Figure 3 As shown, the collected multi-source state data is synchronized according to timestamps, and the synchronized data is stored in a distributed file system. For data with different sampling frequencies, resampling or interpolation methods are used to align them to a uniform time scale.

[0075] Furthermore, step 2) specifically includes the following steps:

[0076] Step 2.1) Perform outlier removal, missing value completion, normalization, and categorical variable encoding on the UAV air-to-ground link quality prediction dataset to obtain standardized multi-source state data.

[0077] Step 2.2) Construct current state features, state change features, and historical time window statistical features based on standardized multi-source state data. The current state features describe the UAV flight state, air-to-ground link monitoring state, and environmental state at the current moment; the state change features describe the changing trends of the UAV flight state and air-to-ground link state between adjacent moments; and the historical time window statistical features describe the fluctuation patterns of link quality over a past period.

[0078] Step 2.3) as Figure 4 As shown, the current state features, state change features, and historical time window statistical features are combined into UAV air-to-ground link state features, which are used for training and prediction of the UAV air-to-ground link quality prediction model.

[0079] In this embodiment, the UAV air-to-ground link state feature vector is shown in formula (1):

[0080]

[0081] in, Indicates time The characteristics of the drone's flight status Indicates time air-to-ground link monitoring characteristics Indicates time Environmental auxiliary features, This indicates the characteristics of flight state changes at adjacent time points. This indicates the characteristics of link state changes at adjacent time points. This represents a statistical measure of link monitoring characteristics within a historical time window of length W.

[0082]

[0083]

[0084] By using formulas (1), (2) and (3), the prediction model can simultaneously utilize the current flight status of the UAV, the current link status, environmental disturbance information, short-term change trends and historical fluctuation patterns, thereby enhancing its ability to characterize dynamic air-to-ground link quality changes.

[0085] Furthermore, step 3) specifically includes the following steps:

[0086] Step 3.1) Use the random forest model as the UAV air-to-ground link quality prediction model.

[0087] Step 3.2) Encode the parameters to be optimized in the random forest model as position vectors of individual whales. The parameters to be optimized include at least one of the following: number of decision trees, maximum tree depth, maximum number of candidate features, and minimum number of leaf node samples.

[0088] Step 3.3) Based on the range of parameters of the random forest model, set the location boundaries of individual whales and construct a whale population.

[0089] In this embodiment, the first The combination of random forest parameters corresponding to each individual whale is expressed by formula (4):

[0090]

[0091] in, Indicates the first The combination of random forest parameters corresponding to each individual whale. Indicates the number of decision trees. Indicates the maximum depth of the subtree. Represents the maximum number of candidate features. This represents the minimum number of leaf nodes.

[0092] Furthermore, step 4) specifically includes the following steps:

[0093] Step 4.1) as Figure 6 As shown, the first individual whale is randomly generated and used as a set of initial parameters for the random forest model.

[0094] Step 4.2) Use the Levy flight mechanism to generate subsequent whale individuals, so that some whale individuals will have large-scale positional changes in the parameter search space.

[0095] Step 4.3) Correct the boundaries of the generated whale individual positions to make them satisfy the range of values ​​in the random forest parameter search space.

[0096] In this embodiment, the Levi flight initialization process can be expressed by formula (5):

[0097]

[0098] in, Indicates the first The position vectors of individual whales This represents the position vector of the previous individual whale. Indicates the step size control factor. This indicates Levi's flight stride. This represents the Lévy distribution parameter.

[0099] In one implementation, the Levy flight step length is expressed by formula (6) using the Mantegna method:

[0100]

[0101] in, and Let be a random variable that follows a normal distribution. Through equations (5) and (6), some individual whales can generate large-scale positional jumps, thereby expanding the search range of initial parameters and enhancing population diversity.

[0102] Furthermore, step 5) specifically includes the following steps:

[0103] Step 5.1) as Figure 5 As shown, the UAV air-to-ground link quality prediction dataset is divided into a training sample set and a validation sample set.

[0104] Step 5.2) Using the Spark distributed computing framework, the random forest parameter combinations corresponding to different whale individuals are distributed to different computing nodes for parallel training.

[0105] Step 5.3) Each computing node trains a random forest model based on the corresponding combination of random forest parameters and calculates the link quality prediction error on the validation sample set.

[0106] Step 5.4) Determine the fitness value of each individual whale based on the link quality prediction error, and obtain the current optimal parameter combination through distributed aggregation.

[0107] In this embodiment, the first The fitness value of an individual whale is expressed by formula (7):

[0108]

[0109] in, Show the first The fitness value of an individual whale. Indicates the first The link quality prediction error of the random forest model corresponds to each individual whale. To prevent constants with a denominator of zero.

[0110] When the link quality prediction target is a single continuous variable, the link quality prediction error can be expressed by formula (8):

[0111]

[0112] in, Indicates the number of validation samples. Indicates the first The true link quality value of each sample Indicates the first Predicted link quality value for each sample.

[0113] When the link quality prediction objective is a comprehensive objective of multiple indicators, the link quality prediction error can be expressed by formula (9):

[0114]

[0115] In the formula, This indicates the signal-to-noise ratio prediction error. This indicates the error in packet loss rate prediction. Indicates the transmission delay prediction error. This indicates the throughput prediction error. These represent the weighting coefficients for the corresponding prediction errors.

[0116] Furthermore, step 6) specifically includes the following steps:

[0117] Step 6.1) as Figure 7 As shown, based on the current optimal parameter combination, random probability, and coefficient vector, it is determined whether the current whale individual adopts a shrinking encirclement, searching for food, or spiraling update method.

[0118] Step 6.2) When the shrinking encirclement condition is met, guide the current random forest parameter combination to move closer to the current optimal parameter combination in order to enhance local exploitation capabilities.

[0119] Step 6.3) When the search and foraging conditions are met, a global search is performed based on the randomly selected parameter combination to expand the parameter search range.

[0120] Step 6.4) When the spiral update condition is met, perform a refined search around the current optimal parameter combination to improve the accuracy of parameter optimization.

[0121] In this embodiment, the coefficient vector is shown in formulas (10), (11), and (12):

[0122]

[0123]

[0124]

[0125] in, and Represents the coefficient vector. Representing an interval random vectors within, This represents a control parameter that decreases linearly with the number of iterations. Indicates the current iteration number. Indicates the maximum number of iterations;

[0126] When random probability and When using shrink wrapping update, it can be expressed by formula (13) and formula (14):

[0127]

[0128]

[0129] in, This represents the current position vector of the individual whale. This represents the current optimal parameter combination. This represents the distance between the current parameter combination and the optimal parameter combination.

[0130] When random probability and At that time, the search-foraging update is adopted, as shown in formulas (15) and (16):

[0131]

[0132]

[0133] in, This represents the location vector of a randomly selected individual whale.

[0134] When random probability At that time, a spiral update is used, as shown in formulas (17) and (18):

[0135]

[0136]

[0137] in, This represents the distance between the current individual whale and the current optimal combination of parameters. Represents the helical shape constant. Representing an interval Random numbers within.

[0138] Furthermore, step 7) specifically includes the following steps:

[0139] Step 7.1) as Figure 8 As shown, the parameter combinations in the current whale population are sorted according to their fitness values ​​to determine the current optimal and second-best parameter combinations.

[0140] Step 7.2) Apply Brownian motion random perturbation to the current optimal and suboptimal parameter combinations to generate new candidate parameter combinations.

[0141] Step 7.3) Replace the low-fitness parameter combinations in the current whale population with new candidate parameter combinations.

[0142] In this embodiment, the Brownian motion random perturbation can be expressed by formulas (19) and (20):

[0143]

[0144]

[0145] in, This represents the current optimal parameter combination. This represents the current suboptimal parameter combination. and

[0146] This represents the candidate parameter combinations generated by Brownian motion perturbation. This represents the disturbance intensity coefficient. and Let represent a random vector that follows a standard normal distribution. This represents element-wise multiplication. and These represent the upper and lower bounds of the parameter search space, respectively.

[0147] Furthermore, step 8) specifically includes the following steps:

[0148] Step 8.1) as Figure 9 As shown, when the number of iterations reaches the preset maximum number of iterations, or the fitness value meets the preset convergence condition, the optimal combination of random forest parameters is output.

[0149] Step 8.2) Construct a UAV air-to-ground link quality prediction model based on the optimal random forest parameter combination.

[0150] Step 8.3) Input the UAV air-to-ground link state characteristics before the time to be predicted into the UAV air-to-ground link quality prediction model, and output the link quality prediction results for the future time.

[0151] In this embodiment, as Figure 10 As shown, the UAV air-to-ground link quality prediction model can be expressed by formula (21):

[0152]

[0153] in express The state feature vector of the UAV air-to-ground link constructed at any time. Showing future moments The link quality prediction results Indicates the optimal combination of parameters A deterministic random forest prediction model.

[0154] The link quality prediction results include at least one of the following: signal-to-noise ratio, received signal strength, packet loss rate, bit error rate, transmission delay, throughput, comprehensive link quality value, or link quality level at future time.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art can still make various adjustments to its form and details, as long as they do not depart from the scope defined by the claims of the present invention.

Claims

1. A method for predicting the quality of UAV air-to-ground links based on a distributed improved whale optimization algorithm, characterized in that, Includes the following steps: Step 1: Collect multi-source status data during UAV flight: Collect multi-source status data during UAV flight, including UAV flight status data, air-to-ground link monitoring data, and environmental auxiliary information; synchronize the collected multi-source status data according to timestamps, and store the synchronized data in a distributed file system to form a UAV air-to-ground link quality prediction dataset. Step 2, construct UAV air-to-ground link state features: perform outlier removal, missing value completion, normalization, and categorical variable encoding on the UAV air-to-ground link quality prediction dataset obtained in Step 1 to obtain standardized multi-source state data; based on the standardized multi-source state data, construct current state features, state change features, and historical time window statistical features to obtain UAV air-to-ground link state features, which are used to characterize the dynamic change pattern of the link during UAV flight. Step 3, construct the parameter search space of the random forest prediction model: use the random forest model as the UAV air-to-ground link quality prediction model, encode the parameters to be optimized in the random forest model into the position vectors of individual whales, and construct a whale population; where each individual whale corresponds to a set of candidate random forest parameter combinations. Step 4: Construct a whale population initialization module based on Levy flight: Randomly generate the first whale individual, and use the Levy flight mechanism to generate subsequent whale individuals, so that the whale individuals produce positional changes at different scales in the parameter search space, thereby expanding the initial parameter search range and enhancing population diversity. Step 5: Construct a fitness calculation module based on distributed computing: Using the Spark distributed computing framework, the random forest parameter combinations corresponding to different whale individuals are distributed to different computing nodes for parallel training; each computing node trains a random forest model based on the corresponding random forest parameter combination and calculates the link quality prediction error on the validation sample set; the fitness value of each whale individual is determined based on the link quality prediction error, and the current optimal parameter combination is obtained through distributed aggregation. Step 6: Construct an improved whale optimization parameter update module: Based on the current optimal parameter combination, random probability, and coefficient vector, iteratively update the random forest parameter combination represented by the individual whale; the iterative update includes shrinking enclosing update, searching and foraging update, and spiral update, so as to realize global search and local refinement of random forest parameter combination. Step 7, construct the Brownian motion perturbation module: sort the parameter combinations in the current whale population according to their fitness values ​​to determine the current optimal and second-best parameter combinations; apply Brownian motion random perturbation to the current optimal and second-best parameter combinations to generate new candidate parameter combinations, and use the new candidate parameter combinations to replace the parameter combinations with lower fitness in the current whale population to enhance search diversity and reduce the probability of getting trapped in local optima. Step 8: Construct a UAV air-to-ground link quality prediction model: When the number of iterations reaches the preset maximum number of iterations, or the fitness value meets the preset convergence condition, output the optimal random forest parameter combination; construct a UAV air-to-ground link quality prediction model based on the optimal random forest parameter combination, and input the UAV air-to-ground link state characteristics before the time to be predicted into the UAV air-to-ground link quality prediction model, and output the link quality prediction results for future times.

2. The method for predicting UAV air-to-ground link quality based on the distributed improved whale optimization algorithm according to claim 1, characterized in that, The multi-source state data mentioned in steps 1 and 2 includes UAV flight state data, air-to-ground link monitoring data, and environmental auxiliary information. The UAV flight state data includes at least one of flight altitude, flight speed, flight acceleration, attitude angle, heading angle, air-to-ground distance, and trajectory position information. The air-to-ground link monitoring data includes at least one of signal-to-noise ratio, received signal strength, reference signal received power, packet loss rate, bit error rate, transmission delay, throughput, and retransmission count. The environmental auxiliary information includes at least one of weather state, terrain type, occlusion state, line-of-sight / non-line-of-sight state, and mission scenario type. The current state feature is used to describe the UAV flight state, air-to-ground link monitoring state, and environmental state at the current moment. The state change feature is used to describe the changing trend of UAV flight state and air-to-ground link state between adjacent moments. The historical time window statistical feature is used to describe the fluctuation pattern of link quality over a past period.

3. The method for predicting UAV air-to-ground link quality based on the distributed improved whale optimization algorithm according to claim 1, characterized in that, The parameters to be optimized in the random forest model described in steps 3 and 4 include at least one of the following: number of decision trees, maximum tree depth, maximum number of candidate features, and minimum number of leaf node samples. The position vector of each whale individual corresponds to a set of random forest parameter combinations. The Levy flight mechanism is used to generate subsequent whale individuals based on the position of the previous whale individual, so that some whale individuals produce large-scale position changes in the parameter search space, thereby expanding the initial search range of the random forest parameter combinations.

4. The method for predicting UAV air-to-ground link quality based on the distributed improved whale optimization algorithm according to claim 1, characterized in that, The fitness value mentioned in steps 5 and 6 is determined by the link quality prediction error, which includes at least one of the following: signal-to-noise ratio prediction error, received signal strength prediction error, packet loss rate prediction error, bit error rate prediction error, transmission delay prediction error, throughput prediction error, or multi-index comprehensive prediction error; the shrinking encirclement update is used to guide the current random forest parameter combination to move closer to the current optimal parameter combination, the search foraging update is used to perform a global search based on the randomly selected parameter combination, and the spiral update is used to perform a refined search around the current optimal parameter combination.

5. The method for predicting UAV air-to-ground link quality based on the distributed improved whale optimization algorithm according to claim 1, characterized in that, The Brownian motion perturbation module described in steps 7 and 8 is used to randomly perturb the current optimal and suboptimal parameter combinations to generate new candidate parameter combinations, and to replace the parameter combinations with lower fitness in the current whale population with the new candidate parameter combinations; the link quality prediction results include at least one of the following at future time: signal-to-noise ratio, received signal strength, packet loss rate, bit error rate, transmission delay, throughput, comprehensive link quality value, or link quality level.

Images