Dynamic unmanned aerial vehicle track optimization method based on brain-like prediction strategy

Abstract

The invention discloses a dynamic unmanned aerial vehicle track optimization method based on a brain-like prediction strategy, and the method achieves the high-precision modeling and dynamic perception of a complex flight environment through the fusion of a laser radar, a visual sensor and a geographic information database. A multi-target path optimization model oriented to cluster cooperation is constructed, and an MP-DMOEA / D algorithm based on environmental modulation synaptic plasticity is introduced. When environment change is detected, the system calls a multi-prediction mechanism of fusion transfer learning, adaptive regression and reinforcement learning to generate a high-quality initial solution, and rapid replanning and stable convergence are realized. The prediction strategy weight is adjusted through synaptic plasticity, and the algorithm can adaptively adjust the optimization direction according to the environmental fluctuation intensity. The method has the advantages of being high in real-time performance, high in adaptability, good in engineering implementability and the like, and is suitable for unmanned aerial vehicle cluster collaborative flight path planning tasks in a dynamic complex environment.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control and computational intelligence technology, specifically to a dynamic UAV trajectory optimization method based on a brain-like prediction strategy. Background Technology

[0002] In recent years, the rapid development of unmanned aerial vehicle (UAV) technology has driven its widespread application in fields such as military reconnaissance, emergency rescue, environmental monitoring, logistics distribution, and agricultural inspection. During missions, UAVs often need to plan their flight paths in complex and ever-changing environments to ensure flight safety, stability, and efficient mission execution. However, with the increasing complexity of mission environments and the diversification of application scenarios, traditional static flight path planning methods are no longer sufficient to meet the demands of real-time performance and multi-objective optimization. In practical missions, UAV swarms often need to consider multiple mutually constraining objectives simultaneously, such as obstacle avoidance safety, flight time, energy consumption, collaborative constraints, mission coverage, and communication stability. Furthermore, the external environment often changes dynamically over time, such as wind speed fluctuations, obstacle movement, target position changes, and communication link interference, which can cause flight path planning results to become invalid during mission execution, affecting mission continuity and efficiency.

[0003] In existing technologies, UAV route modeling relies on heuristic or evolutionary algorithm-based trajectory planning methods. Most of these methods assume a static or slowly changing environment, lacking real-time responsiveness to multi-objective coupling relationships in dynamic scenarios. This makes it difficult to achieve global optimization and rapid replanning of UAV swarms in complex environments. Therefore, there is an urgent need for a UAV trajectory optimization method that can adapt to various environmental conditions while balancing accuracy and efficiency in trajectory tracking. Summary of the Invention

[0004] Objective: To address the problems mentioned in the background art, this invention discloses a dynamic UAV trajectory optimization method based on a brain-like prediction strategy. By introducing an environment-modulated synaptic plasticity mechanism, the selection weights of the prediction strategy can adaptively adjust according to the intensity of environmental disturbances, achieving neural self-regulation of the prediction strategy in dynamic environments, rather than traditional static weighted fusion. This enables UAV swarms to rapidly generate and track a set of Pareto trajectories that comprehensively optimize multiple objectives such as range, safety, energy consumption, and collaborative performance in dynamic environments, thus achieving UAV flight path optimization tasks in complex environments.

[0005] Technical solution:

[0006] This invention discloses a dynamic UAV trajectory optimization method based on a brain-like prediction strategy, the method comprising the following steps:

[0007] S1: The drone collects and preprocesses environmental data to create a dynamic environmental map;

[0008] S2: Use a vector objective function for modeling, set decision variables, optimize objectives and constraints, and build a drone swarm trajectory planning model;

[0009] S3: Set the population size and initialize the MOEA / D algorithm;

[0010] S4: During swarm flight, for each subpopulation, calculate the comprehensive characteristic change representing the environmental state to obtain the synaptic change intensity Θ(t). If Θ(t) exceeds the threshold, call the MP-DMOEA / D algorithm to perform dynamic replanning of the multi-prediction strategy; if the environment has not changed, use the conventional MOEA / D algorithm for path optimization.

[0011] S5: Adaptive optimization of the prediction strategy is achieved by adjusting the weights. The MOEA / D algorithm generates a set of Pareto optimal solutions and selects the optimal path through non-dominated sorting and KneePoint optimization mechanism.

[0012] S6: Determine if the current drone position has reached the target point. If not, re-execute S4; if the drone has reached the target point, the path planning task is complete.

[0013] Furthermore, the decision variable in S2 is the sequence of waypoints of all UAVs in the cluster; the optimization objective includes the total flight cost of the cluster and the overall threat cost of the cluster; the constraints include cluster collaboration constraints.

[0014] Furthermore, the total flight cost of the cluster described in S2 is expressed as:

[0015]

[0016] The overall threat cost of the cluster is expressed as:

[0017]

[0018] Wherein, the difference between the two distances, d1-d2, represents the actual distance the UAV travels towards the target point, v is the constant speed of the UAV, and n is the number of track segments contained in the total track; Threat k (x) represents the probability that the drone is shot down by the kth threat source.

[0019] Furthermore, the cluster coordination constraints include coordination time constraints and UAV physical constraints:

[0020] The coordination time constraint is expressed as:

[0021]

[0022]

[0023] The physical constraints of the drone are represented as follows:

[0024]

[0025] in, Let be the time interval for the optimal flight path of UAV j. Let be the time interval for the optimal trajectory of UAV i. Based on this constraint, it can be ensured that the UAV swarm has a non-empty cooperative time, during which any This can be used as the coordination time, and the specific flight parameters of the drone can be determined accordingly; These represent the drone's speed, acceleration, climb angle, horizontal turn angle, and flight altitude, respectively.

[0026] Furthermore, the drone swarm trajectory planning model is specifically as follows:

[0027] Furthermore, the comprehensive characteristic changes representing the environmental state described in S4 include: the cluster average centroid change rate ΔC t Cluster target spatial distribution drift D t The sparse reconstruction error E of the environmental map rec,t The synaptic change intensity Θ(t) is calculated by weighting the cluster centroid change rate, target distribution drift, and environmental sparse reconstruction error.

[0028] Furthermore, as described in S4, the MP-DMOEA / D algorithm is invoked for dynamic replanning of the multi-prediction strategy:

[0029] The MP-DMOEA / D algorithm is used to dynamically predict the subpopulation using multiple prediction strategies, generating a candidate solution set Q. A two-layer screening strategy is then used to optimize population construction and capture potential environmental change trends.

[0030] The transfer learning prediction strategy retrieves three similar environments with mobile threat characteristics from the historical task archives. Using the PCA-GFK and Sinkhorn optimal transfer methods, the optimal cluster trajectory scheme is mapped to the current environment, generating a set of candidate scheme QTLs that focus on historical successful avoidance experience.

[0031] The adaptive fuzzy linear regression prediction strategy extracts the centroid sequence of the cluster track over the last 5 time steps to fit the movement trend of the centroid, fuzzily selects the excitation step length rs, and predicts the large-scale maneuver of the cluster as a whole according to the formula Post+1=Post+rs*Dirt, generating candidate schemes QLS.

[0032] The reinforcement learning prediction strategy inputs the population sequence of the past 20 generations into a pre-trained Transformer. The Transformer uses a dispersed-to-convergence strategy to generate a set of proposal clusters and then reassembles the candidate solutions QRL near the target point. In the archive maintenance phase, the archived solutions are discretized into grids through the ε-dominance mechanism, retaining the optimal or representative solution in each grid, eliminating redundant paths, and injecting Gaussian perturbation path solutions.

[0033] Furthermore, the specific steps of S5 are as follows:

[0034] The effectiveness of each prediction strategy in generating solutions under the current environmental changes is evaluated. Policies with good performance are incentivized (i.e., their selection weights are increased), while those with poor performance are suppressed (i.e., their weights are reduced but their opportunities are preserved), achieving adaptive strategy optimization. The MOEA / D algorithm generates a set of Pareto optimal solutions, and the optimal path is selected through non-dominated sorting and KneePoint optimization mechanisms. The predicted strategies update synaptic weights and normalize them into selection probabilities.

[0035]

[0036] Where K=3 is the total number of strategies, namely TL-P, LS-P, and RL-P; θmin=0.1 is the minimum selection probability to prevent strategies from being permanently eliminated due to temporary performance fluctuations.

[0037] Beneficial effects:

[0038] 1. This invention establishes a UAV swarm trajectory planning model suitable for dynamic environments. It directly models swarm collaboration requirements such as collision avoidance, formation maintenance, and energy consumption balance as optimization objectives. Through a vector objective function, it achieves a comprehensive balance of multi-dimensional performance indicators such as path length, flight time, safety, and energy consumption, ensuring the feasibility and superiority of the swarm-level path planning results.

[0039] 2. This invention uses the MOEA / D algorithm to dynamically replan the route of the UAV, captures potential environmental change trends to achieve rapid updates of the path planning scheme, and further improves the real-time performance and adaptability of the model, making this invention more suitable for path planning scenarios with complex environments and high efficiency requirements.

[0040] 3. This invention is based on the MP-DMOEA / D algorithm, introduces a brain-like prediction mechanism, simulates the synaptic plasticity and incremental learning characteristics of the brain, integrates multiple prediction models to form a hybrid prediction system, evaluates the effectiveness of each prediction strategy in generating solutions in this round of environmental changes, and adaptively selects the optimal prediction strategy when the environment undergoes linear changes, periodic fluctuations or sudden disturbances, so as to quickly generate a high-quality initial solution set, avoid random restarts, and further improve the efficiency of replanning and prediction accuracy.

[0041] 4. This invention can rapidly output a set of Pareto optimal paths after dynamic environmental changes, providing command and control systems with multiple path schemes with different emphases and supporting multi-level flexible decision-making. It further improves the method's real-time performance, adaptability, and engineering feasibility, making it suitable for intelligent UAV swarm path planning tasks in complex dynamic environments and possessing broad application prospects. Attached Figure Description

[0042] Figure 1 This is a diagram illustrating the overall control framework of the present invention.

[0043] Figure 2 This is a flowchart of the control method of the present invention;

[0044] Figure 3 This is a schematic diagram of the MP-DMOEA / D optimization of the present invention;

[0045] Figure 4 This is a comparison chart of the trajectory planning of this invention and DNSGA-II;

[0046] Figure 5 This is a graph showing the change in heading angle of the planned flight path under different preferences for static threat conditions in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] like Figure 1-2 As shown, this invention discloses a dynamic UAV trajectory optimization method based on a brain-like prediction strategy. The method includes the following steps: Step 1: Using the lidar, visual sensors and external geographic information system carried by the UAV swarm, collect surrounding environmental data and preprocess it, fuse and construct a dynamic environmental map including obstacles, threats and no-fly zones, and perform real-time environmental modeling.

[0049] Five quadcopter UAVs, each equipped with a GPS module, inertial measurement unit (IMU), lidar sensor, and wireless communication module. The ground control station is equipped with the swarm trajectory planning system described in this invention. The system includes an environmental perception module, a mission planning module, a neuromorphic prediction and optimization module, and an execution control module. These modules interact with each other via a communication bus. The optimizer uses an Intel i7 processor and runs the algorithm of this invention.

[0050] The communication link uses a 5G-based data transmission link for data interaction between the ground station and the drone cluster.

[0051] Step 2: Establish a UAV swarm trajectory planning model using a vector objective function. The decision variables are the sequence of track points of all UAVs in the swarm. The optimization objectives include at least: total swarm range cost, overall swarm threat cost, and constraints including swarm coordination constraints, which include coordination time constraints and UAV physical constraints. Perform UAV swarm trajectory planning modeling and set the starting point and target point.

[0052] Decision variables are encoded using node sequence encoding. Each track is represented as: S={P0,P1,...,P...} T}, where P0(X0,Y0,Z0) is the starting point, P L (X) T ,Y T Z T Let P1 be the target point, and P be the distance from P1 to P2. T-1 These serve as intermediate waypoints. For a swarm of 5 drones, a complete solution is a set of 5 such waypoints.

[0053] The objective function includes total range cost, threat cost, mission time cost, and cluster coordination cost.

[0054] Optimizing the path planning of a drone from its starting point to its target point primarily involves the total flight cost, which requires the drone to travel the shortest distance along its current flight path before changing its flight attitude.

[0055] The distance from the starting point A of the flight path to the target point T is:

[0056]

[0057] The distance from the end of the track B to the target point T is:

[0058]

[0059] When a drone swarm performs a mission, the difference between two distances, d1-d2, represents the actual distance the drone travels to approach the target point. The maximum distance the drone can approach the target is d = v·∆t·n. To more intuitively evaluate the trajectory performance, a normalized and quantitative analysis is performed on the total range cost objective function:

[0060]

[0061] Minimize the probability of the entire cluster being detected or hit by a threat source. Based on a real-time environmental model fusion of LiDAR and map data, calculate and sum the threat probabilities of each UAV's flight path. These threats include, but are not limited to, terrain obstacles, enemy defense systems, adverse weather conditions, and other aircraft.

[0062] The probability that a drone is shot down by the k-th (k = 1, 2, ..., K) threat source can be expressed as:

[0063]

[0064] Where x represents the coordinate vector of the UAV's current position, u k H indicates the central location of the threat. k This indicates the scope of the threat source's influence.

[0065] The threat probability of the k-th threat source to the i-th (i=1, 2, ..., n) track segment can be expressed as:

[0066]

[0067] because Then the first The threat probability of a drone can be expressed as: Since the threats are independent of each other, the probability of a drone being threatened can be expressed as:

[0068]

[0069] There exists a timeframe that satisfies the time coordination constraint of multiple drones, allowing all drones to arrive at the target point simultaneously. This coordination time constraint can be expressed as:

[0070]

[0071]

[0072] in, Let be the time interval for the optimal flight path of UAV j. Let be the time interval for the optimal flight path of UAV i. Based on this constraint, it can be ensured that the UAV swarm has a non-empty cooperative time, during which any This can be used as the coordination time, and the specific flight parameters of the drone can be determined accordingly.

[0073] Consider the pitch angle of the UAV, denoted as P(x,y,z,θ,γ), where x,y,z are the position coordinates of the UAV, and θ is the pitch angle of the UAV.

[0074] γ is the horizontal turning angle, and γ is the vertical climb angle or descent angle.

[0075] When considering turning angle constraints, the maximum turning angle is denoted as θ. max The i-th waypoint S i The horizontal rotation angle is θ i Therefore, θ i ≤θ max Consider three consecutive waypoints S. i-1 ,S i ,S i+1 track segment S i-1 ,S i and S i ,S i+1 The projections of the resulting vectors onto the horizontal plane are respectively (x i-1 -x i ,y i-1 -y i ) and (x i+1 -x i ,y i+1 -y i At this time, θ i It can be represented as:

[0076]

[0077] The formula describes the waypoint P i The horizontal turning angle at a given point is relative to the positions of its two preceding and following track points, ensuring that it does not exceed the maximum turning angle θ. max This means that the drone can only turn within a specified maximum turning angle in the horizontal plane.

[0078] The maximum climb angle or glide angle constraint can be expressed as denoted by γ. max , No. track point P i The ascent or descent angle is γ i Therefore, γ i ≤γ max .

[0079] Consider two consecutive waypoints S i-1 ,S i track segment S i-1 ,S i The resulting vector is (x i-1 -x i ,y i-1 -y i ,z i-1 -z i At this time, γ i It can be represented as:

[0080]

[0081] The above formula expresses the waypoint P i The relative position of the climb or glide angle at a given point to its previous waypoint should be considered to ensure that it does not exceed the maximum climb or glide angle γ. max .

[0082] The regulations stipulate that drones must fly within a designated altitude range, ensuring vertical altitude restrictions. The minimum and maximum flight altitude constraints can be expressed as follows, with the minimum flight altitude denoted as w. min The maximum flight altitude is w max Therefore, we have w min ≤w≤w max W represents the drone's flight altitude. This constraint ensures that the drone flies within the specified altitude range.

[0083] In summary, a dynamic multi-objective optimization model for the online flight path planning problem of unmanned aerial vehicles (UAVs) can be established:

[0084]

[0085] The physical constraints are:

[0086] in, These represent the drone's speed, acceleration, climb angle, horizontal turn angle, and flight altitude, respectively.

[0087] Step 3: Set the population size, initialize the weight vector and neighborhood structure of the MOEA / D algorithm, and divide the population into multiple subpopulations, each subpopulation representing a cluster track allocation scheme;

[0088] (3) Initialize the MOEA / D algorithm population to provide a search space for multi-objective optimization;

[0089] After completing the UAV swarm environment modeling and multi-objective task construction, it is necessary to construct the initial population for the MOEA / D algorithm to establish an effective search space and provide basic support for subsequent swarm cooperative trajectory optimization. This step, through systematic population initialization operations, ensures the feasibility of the trajectory scheme while taking into account the diversity and structural distribution of the population.

[0090] (31) The encoding method of the solution is determined to be cluster node sequence encoding;

[0091] Considering the high-dimensionality, collaborative, and strongly constrained characteristics of UAV swarm trajectory planning, this invention employs an extended node sequence encoding method as the individual representation method. Each individual (i.e., a complete swarm trajectory scheme) is represented as:

[0092] X={S1,S2,...,S M}, where M is the number of drones in the cluster, Si ={P0,P1,...,P T} represents the flight path of the i-th UAV, which consists of a sequence of starting point, several intermediate waypoints, and ending point. This encoding method intuitively reflects the collaborative structure of the swarm, facilitates the execution of mutation operations for individual UAV paths and cross operations for swarm collaboration, and can flexibly adapt to flight areas with different mission complexities.

[0093] (32) Define the population size and decompose the subproblems;

[0094] The population size is set to N, which is determined based on the cluster size, task complexity, and available computing resources, typically ranging from 100 to 200. A larger population size results in greater solution diversity, but also increases computational overhead. Each individual corresponds to a MOEA / D subproblem, and the population covers the entire multi-objective solution space.

[0095] (33) Weight vector generation and subproblem decomposition;

[0096] To decompose the cluster multi-objective trajectory optimization problem into multiple single-objective subproblems that can be processed in parallel, this invention employs a systematic weight vector generation strategy. This involves uniformly generating N weight vectors λ¹, λ², ..., λ₀ on the unit hyperplane of the m-dimensional target space. ^N , where λ 1 =(λ1 1 ,λ2 1 ,...,λ m 1 And satisfy ∑λj 1 =1. Each weight vector defines the bias relationship between different optimization objectives (such as range, threat, time, and cooperation cost) in a subproblem. Each weight vector corresponds to a subproblem, and the objective function is constructed by weighted Tchebycheff, so that MOEA / D can optimize path solutions with different objective preferences in parallel.

[0097] (34) Initial individual path generation strategy and objective function evaluation;

[0098] In the individual path generation phase, this invention comprehensively utilizes traditional heuristic path planning algorithms to generate some high-quality initial paths, which are then randomly combined into cluster schemes and injected into the population, accounting for approximately 25%. Successful trajectory schemes from similar scenarios are retrieved from the historical task database, and after Gaussian perturbation, they are added to the population, accounting for approximately 15%, achieving brain-like preloading of prior experience. The remaining individuals are randomly generated within the feasible solution space to ensure population diversity.

[0099] (35) Neighborhood structure construction promotes local collaborative search.

[0100] Based on the similarity of weight vectors, a neighborhood structure is constructed for each subproblem. By calculating the Euclidean distance between weight vectors, the T nearest neighbor subproblems are determined for each subproblem i, forming a neighborhood B(i). The neighborhood size T is typically set to 10%-20% of the population size. This neighborhood structure promotes information sharing and collaborative optimization among individuals during subsequent evolution, and improves the algorithm's convergence efficiency through a local update strategy.

[0101] This step, through the systematic initialization of the population, ensures path feasibility while taking into account the diversity and structural distribution of solutions, providing a solid solution space foundation for subsequent multi-objective evolutionary optimization.

[0102] Step 4: During swarm flight, continuously sense changes in the location of threat sources and the appearance of new obstacles. For each subpopulation, calculate the comprehensive characteristic changes representing the environmental state, including: the swarm's average centroid change rate ΔC. t Cluster target spatial distribution drift D t The sparse reconstruction error E of the environmental map rec,t

[0103] In traditional neuromorphic synaptic plasticity, the formula is usually:

[0104]

[0105] Among them, w ij Synaptic strength (connection weight), x i y j It represents the activity of neurons before and after learning, η is the learning rate, and λ is the forgetting factor to prevent over-enhancement.

[0106] The formula for calculating the intensity of synaptic changes is defined as follows:

[0107]

[0108] Where ΔCt is the rate of change of the center of mass, and D t It is the target distribution drift, E rec t is the sparse reconstruction error, and α1+α2+α3=1 is the weighting coefficient, which controls the contribution of the three signals to the environmental intensity.

[0109] The intensity of synaptic change is denoted as Θ(t). The larger Θ(t) is, the more drastic the environmental change. This invention adapts to the intensity of environmental change by using the principle of synaptic intensity change (i.e., adaptive adjustment of the prediction strategy amplitude).

[0110] The linear modulation form of f(Θ) is defined as:

[0111]

[0112] Here, α is the modulation coefficient. The stronger the environmental change, the larger f(Θ) becomes, and the more obvious the synaptic adjustment.

[0113] Effectiveness index δ k t

[0114]

[0115]

[0116] Among them, Epred k t It is the average prediction error of the k-th predictor, Eavg t It is the average error of all predictors. The difference between them reflects the relative performance of the predictors. A positive performance that is better than average will increase the probability of selection, while a negative performance that is worse than average will decrease the probability of selection.

[0117] Therefore, we changed the synaptic strength update to an environment-modulated synaptic plasticity (EMSP) model:

[0118]

[0119] Where, Δw ij Let f(t) represent the change in synaptic strength, and f(Θ(t)) be the modulation function for the intensity of environmental change.

[0120] The synaptic change intensity Θ(t) is obtained by weighting the above indicators. If Θ(t) exceeds the threshold, it is determined that the environment has changed significantly, triggering dynamic reprogramming.

[0121] Step 5: As Figure 3 As shown, when an environmental change is detected, the MP-DMOEA / D algorithm performs dynamic replanning of multiple prediction strategies to generate a new initial cluster trajectory scheme in parallel:

[0122] (51) Transfer learning prediction strategy

[0123] Three similar environments exhibiting characteristics of encountering mobile threats were retrieved from historical mission archives. Using the described PCA-GFK and Sinkhorn optimal transmission methods, the optimal swarm path schemes from these historical environments were mapped to the current environment, generating a set of candidate schemes Q that emphasize historical successful avoidance experience. TL .

[0124] And based on historical evolution information, calculate the dynamic feature weight vector w=(w1,w2,…,w D The new environment representation can be obtained through weighted operations as follows:

[0125]

[0126] Here, Diag(w) is a diagonal matrix composed of weight vectors. This weighting mechanism highlights the importance of key dimensions, thereby reducing the interference of irrelevant features in subsequent mapping processes.

[0127] Select the k environments {X1, X2, ..., Xk} that are most similar to the current environment from the historical archives. k} and perform principal component analysis (PCA) on them respectively. The formula for solving the low-dimensional basic matrix is:

[0128]

[0129] Where, P∈R D×d Let be the projection matrix, and d be the target dimension reduction dimension. Then, the geodesic flow kernel (GFK) is used to align the manifolds of the source and target domains. Its definition is as follows:

[0130]

[0131] Here, Φ(λ) represents the interpolation mapping from the source subspace to the target subspace. This kernel function describes the smooth transition from the source environment to the target environment in the manifold space.

[0132] To further enhance the alignment effect under nonlinear distributions, t-SNE and the Sinkhorn optimal transport mechanism are introduced. t-SNE in the weighted space X... t w Nonlinear low-dimensional manifold structures are extracted from the data, while Sinkhorn optimal transmission achieves efficient distribution alignment by introducing an entropy regularization term into the transmission matrix. Its objective function is as follows:

[0133]

[0134] Where, γ∈R n×m Let C be the transfer matrix, and H(γ) be the cost matrix between the source and target samples. i,j γ ij logγ ij Let ε be the entropy regularization term and ε be the regularization coefficient. The constraint set U(a,b) is defined as follows:

[0135]

[0136] Where a and b are the distribution vectors of the source and target domains, respectively. Finally, the aligned historical context is represented as:

[0137]

[0138] The migrated individuals are mapped to the target environment using a trend correction strategy. To ensure feasibility, boundary pruning and repair operations are performed on the predicted individuals, as shown below:

[0139]

[0140] Ultimately, the repaired individuals were incorporated into the updated subpopulation P. t+1 This forms a predictive solution set for the new environment.

[0141] (52) Adaptive fuzzy linear regression prediction strategy

[0142] The centroid sequence of the cluster tracks over the most recent five time steps was extracted. The movement trend of the centroids was fitted using the least squares method. Based on the drastic nature of the environmental changes and the stability of the trend, a fuzzy selection of the excitation step length rs was made. The formula Pos was then used. t+1 =Pos t +rs*Dir t It predicts that the cluster will undergo significant maneuvers and generates candidate solutions Q. LS .

[0143] The population center of the most recent k environments is {Pc} t-k Pc t−k+1 ,…,Pc t}, and construct a historical center point sequence based on these points. Each center point P ct P is a vector of dimension D. ct The calculation formula is as follows:

[0144]

[0145] Among them, P ct Pop represents the population centroid. t Let |Pop be the non-dominated solution set at time t. t | represents the population size. For each decision variable dimension d (d=1,2,…,D), the least squares fitted linear model is as follows:

[0146]

[0147] Where, α d β is the slope, representing the rate of change of dimension d. d ϵ is the intercept. i It represents the residual.

[0148] Using this linear model, the value of the center point at the next time step t+1 is predicted, as shown in the formula:

[0149]

[0150] The vector form is represented as:

[0151]

[0152] In this formula, |Pop t | is the population size, x t =(x 1 t ,x 2 t ,…,x D t ) represents the solution at time t, and the direction of motion of the center point at time t (Dir) t It can be calculated using the following formula:

[0153]

[0154] Based on the above formula, the non-dominated solution set Pos can be obtained. t Position at time t+1:

[0155]

[0156] Where rs is the step size along the direction of movement. To avoid violating the continuity assumption due to extreme step sizes, this invention provides three different step size values ​​rs = 0.5, 1.0, and 1.5, which correspond to the degree of drastic change in the environment, respectively.

[0157] (53) Reinforcement learning prediction strategy

[0158] The sequence of the past 20 generations of the population was input into a pre-trained Transformer. The model captured that, under crosswind conditions, a disperse-to-converge strategy effectively balances threat avoidance and formation maintenance. Therefore, it generated a set of candidate schemes Q suggesting that the cluster first disperse to reduce the risk of being collectively detected, and then regroup near the target point. RL .

[0159] In the algorithm, the historical population sequence within the sliding window is:

[0160]

[0161] As TransformerT θ The input is transformed by the encoding function ϕ(⋅), and the output is the predicted population at the next time step:

[0162]

[0163] To measure the relationship between environmental change and population diversity, a variance vector v is defined for each dimension of the decision variable. t With scalar diversity index S t :

[0164]

[0165] Att t This represents differential attention, used to adaptively determine the proportion of predicted solutions introduced into the new population:

[0166]

[0167] After the predicted solution is trimmed and repaired according to the formula boundary, it is merged with the elite solution and a new generation of initial population is obtained through non-dominated sorting and crowding selection:

[0168]

[0169] Furthermore, the Transformer model is fine-tuned using low-rank adaptive (LoRA) training, incrementally training only with data from the most recent k generations:

[0170]

[0171] Sequence modeling is used to capture population evolution trends. Adaptive weights and differential attention control the proportion of predicted solutions, and a lightweight fine-tuning mechanism is combined to achieve rapid response to dynamic environments while ensuring population diversity and convergence performance.

[0172] Step 6: Subsequently, a brain-like synaptic stimulation mechanism is simulated. By evaluating the effectiveness of each prediction strategy in generating a plan in this round of environmental changes, incentives are added to strategies with good performance, i.e., their selection weight is increased, while inhibition is applied to strategies with poor performance, i.e., their weight is reduced but their opportunities are retained, thus achieving adaptive optimization of strategies.

[0173] Update the synaptic weights and normalize them to selection probabilities:

[0174]

[0175]

[0176] The key to MP-MOEA / D lies in how to dynamically adjust the application weights of each prediction strategy in the subpopulation. To adaptively select better prediction strategies, a roulette wheel betting mechanism is used to dynamically update the weights ω. t =[ω t TL ,ω t NS ,ω t RL ], and normalize it to the selection probability θ t The formula is as follows:

[0177]

[0178] Where K=3 represents the total number of strategies (TL-P, LS-P, RL-P);

[0179] θ min =0.1 is the minimum selection probability to prevent the strategy from being permanently eliminated due to temporary performance fluctuations.

[0180] Step 7: Inject the high-quality initial solutions generated by the preferred prediction strategy into the MOEA / D algorithm framework for further optimization. Through neighborhood crossover and mutation among subpopulations, and considering the constraints of all individual UAVs and the overall swarm, merge the candidate solutions generated by the three prediction strategies, and use non-dominated sorting to select elite individuals to replace the lower-quality individuals in the original population. This evolves a new generation of swarm trajectory solution set (Pareto solution set).

[0181] Step 8: Maintain a diverse set of non-dominated solutions using the archive management mechanism, and automatically select a set of cluster paths that are optimal among multiple targets from the Pareto solution set through the knee point optimization mechanism, and distribute them to each UAV for execution.

[0182] Under the MOEA / D framework, the system underwent approximately 50 generations of rapid evolution, further optimizing trajectory details through neighborhood crossover mutation. After evolution, the system obtained a Pareto solution set containing 15 non-dominated swarm trajectory schemes. Using a knee-point optimization mechanism, the optimal scheme across the four objectives was automatically selected. This scheme was then rapidly distributed to all UAVs in the swarm via the communication link.

[0183] Step 9: Determine if the current drone position has reached the target point. If not, re-enter Step 3 for environmental perception and path update, and perform dynamic optimization. When the drone reaches the target point, the path planning task is completed. This forms a closed-loop autonomous control system encompassing environmental perception, predictive decision-making, self-motivation, and self-monitoring.

[0184] During path planning and optimization, the system needs to monitor the drone's current position in real time and determine whether it has reached the target point to decide whether to continue environmental perception and path update operations. When the spatial distance between the drone's current position and the target point is less than a certain threshold, it can be determined that the drone has successfully reached the target area.

[0185]

[0186] Among them, the threshold The specific value is determined by the mission accuracy requirements, sensor positioning error, and target area size. For precise operations (such as indoor positioning and formation landing), a value of 0.1–1 meter is recommended; for wide-area path planning or more forgiving scenarios (such as field reconnaissance), a value of 1–3 meters can be used.

[0187] If the target has not yet been reached, the system re-enters step 3 of the environmental change detection process to begin a new round of environmental perception, judgment, and path optimization. This mechanism forms a dynamic closed loop, ensuring that the path always matches the latest environmental conditions, thereby ensuring flight safety and mission completion.

[0188] If the target has been reached, the current path planning task is considered successfully completed. The system will then terminate the path planning process and disable the optimization module and path update logic. At this point, the system can proceed with subsequent task flows (such as landing, data upload, or task switching) or remain in a hovering, ready state.

[0189] This invention not only ensures the continuity and real-time nature of the path planning process, but also achieves automatic closed-loop control of the entire flight mission through the control logic of updating if the destination is not reached and terminating upon arrival, which significantly improves the intelligence and execution efficiency of the system.

[0190] Finally, to verify the effectiveness of this invention, experimental data will be compared with the planning results of DNSGA-II under static threat and sudden threat scenarios. The environmental area is assumed to be 100km × 100km, with the origin at (0km, 0km). In the static threat scenario experiment, the static threat sources are K1 (-30km, -10km) and K2 (-15km, -10km), the UAV's starting point coordinates are (-10km, -18km), the target point coordinates are (-40km, -4km), and the initial heading angle is the direction of the line connecting the starting point and the target point.

[0191] Figure 4 The images show the flight paths of this invention and DNSGA-II under different static threat conditions and preferences. It can be seen that the MP-DMOEA / D algorithm, due to its multi-prediction optimization mechanism based on neuromorphic synaptic plasticity, plans for threats in advance, making it safer at the threat source and producing a smoother planned flight path, thus improving the operational stability of the UAV.

[0192] Figure 5 The curves show the heading angle variation of the planned trajectory under different static threat conditions. The heading angle variation of the planned trajectory is smaller, resulting in better UAV operation performance.

[0193] Table 1

[0194]

[0195] Table 1 shows the flight time, survival probability, and average change in heading angle of the planned paths under static threat conditions for the two algorithms. Flight time reflects the path length, survival probability reflects the safety level of the UAV, and average change in heading angle reflects the overall stability of the UAV during flight. It can be seen that the MP-DMOEA / D algorithm outperforms the DNSGA-II algorithm in all aspects of the planned path. The MP-DMOEA / D algorithm has an average optimization time of 275ms, satisfying t... s =1s planning time requirement.

[0196] Since sudden threats are a type of threat whose timing, location, and magnitude cannot be accurately predicted by the UAV system, the trajectory planning algorithm is required to quickly update threat source information and plan a corresponding reasonable trajectory when dealing with such threat sources. In the experiment, sudden threat 3 will appear at the location (-35km, -8km) at 150s, and its threat range parameter σ x and σ y Both are 2.1km.

[0197] Table 2

[0198]

[0199] Table 2 shows the flight time, survival probability, and average change in heading angle for the two algorithms under emergency conditions. It can be seen that in favor of (ω1, ω2)... Τ =(0.7, 0.3) Τ In this scenario, the survival probabilities of the MP-DMOEA / D algorithm-planned trajectory and the DNSGA-II algorithm-planned trajectory are similar, but the MP-DMOEA / D algorithm shortens the flight time and reduces the change in heading angle; in the preference (ω1, ω2) Τ =(0.3, 0.7) Τ In this scenario, the MP-DMOEA / D algorithm demonstrates superior performance across various trajectory planning metrics. The MP-DMOEA / D algorithm has an average optimization time of 306ms, satisfying the condition t... s =1s planning time requirement.

Claims

1. A dynamic UAV trajectory optimization method based on a brain-like prediction strategy, characterized in that, The method includes the following steps: S1: The drone collects and preprocesses environmental data to create a dynamic environmental map; S2: Use a vector objective function for modeling, set decision variables, optimize objectives and constraints, and build a drone swarm trajectory planning model; S3: Set the population size and initialize the MOEA / D algorithm; S4: During swarm flight, for each subpopulation, calculate the comprehensive characteristic change representing the environmental state to obtain the synaptic change intensity Θ(t). If Θ(t) exceeds the threshold, call the MP-DMOEA / D algorithm to perform dynamic replanning of the multi-prediction strategy; if the environment has not changed, use the conventional MOEA / D algorithm for path optimization. S5: Adaptive optimization of the prediction strategy is achieved by adjusting the weights. The MOEA / D algorithm generates a set of Pareto optimal solutions and selects the optimal path through non-dominated sorting and KneePoint optimization mechanism. S6: Determine if the current drone position has reached the target point. If not, re-execute S4; if the drone has reached the target point, the path planning task is complete.

2. The dynamic UAV trajectory optimization method based on a brain-like prediction strategy according to claim 1, characterized in that, The decision variable in S2 is the sequence of waypoints of all UAVs in the cluster; the optimization objective includes the total flight cost of the cluster and the overall threat cost of the cluster; the constraints include cluster collaboration constraints.

3. The dynamic UAV trajectory optimization method based on a brain-like prediction strategy according to claim 2, characterized in that, The total flight cost of the cluster described in S2 is expressed as follows: ； The overall threat cost of the cluster is expressed as: ； Wherein, the difference between the two distances, d1-d2, represents the actual distance the UAV travels towards the target point, v is the constant speed of the UAV, and n is the number of track segments contained in the total track; Threat k (x) represents the probability that the drone is shot down by the kth threat source.

4. The dynamic UAV trajectory optimization method based on a brain-like prediction strategy according to claim 2, characterized in that, The cluster coordination constraints include coordination time constraints and UAV physical constraints: The coordination time constraint is expressed as: ； ； The physical constraints of the drone are represented as follows: ； in, Let be the time interval for the optimal flight path of UAV j. Let be the time interval for the optimal trajectory of UAV i. Based on this constraint, it can be ensured that the UAV swarm has a non-empty cooperative time, during which any This can be used as the coordination time, and the specific flight parameters of the drone can be determined accordingly; These represent the drone's speed, acceleration, climb angle, horizontal turn angle, and flight altitude, respectively.

5. The dynamic UAV trajectory optimization method based on a brain-like prediction strategy according to claim 4, characterized in that, The drone swarm trajectory planning model is specifically as follows: .

6. The dynamic UAV trajectory optimization method based on a brain-like prediction strategy according to claim 1, characterized in that, The comprehensive characteristic changes in environmental state described in S4 include: the cluster average centroid change rate ΔC t Cluster target spatial distribution drift D t The sparse reconstruction error E of the environmental map rec,t The synaptic change intensity Θ(t) is calculated by weighting the cluster centroid change rate, target distribution drift, and environmental sparse reconstruction error.

7. The dynamic UAV trajectory optimization method based on a brain-like prediction strategy according to claim 6, characterized in that, S4 describes the use of the MP-DMOEA / D algorithm for dynamic replanning of multiple prediction strategies: The MP-DMOEA / D algorithm is used to dynamically predict the subpopulation using multiple prediction strategies, generating a candidate solution set Q. A two-layer screening strategy is then used to optimize population construction and capture potential environmental change trends. The transfer learning prediction strategy retrieves three similar environments with mobile threat characteristics from the historical task archives. Using the PCA-GFK and Sinkhorn optimal transfer methods, the optimal cluster trajectory scheme is mapped to the current environment, generating a set of candidate scheme QTLs that focus on historical successful avoidance experience. The adaptive fuzzy linear regression prediction strategy extracts the centroid sequence of the cluster track over the last 5 time steps to fit the movement trend of the centroid, fuzzily selects the excitation step length rs, and predicts the large-scale maneuver of the cluster as a whole according to the formula Post+1=Post+rs*Dirt, generating candidate schemes QLS. The reinforcement learning prediction strategy inputs the population sequence of the past 20 generations into a pre-trained Transformer. The Transformer uses a dispersed-to-convergence strategy to generate a set of proposal clusters and then reassembles the candidate solutions QRL near the target point. In the archive maintenance phase, the archived solutions are discretized into grids through the ε-dominance mechanism, retaining the optimal or representative solution in each grid, eliminating redundant paths, and injecting Gaussian perturbation path solutions.

8. The dynamic UAV trajectory optimization method based on a brain-like prediction strategy according to claim 1, characterized in that, The specific steps for S5 are as follows: The effectiveness of each prediction strategy in generating solutions under the current environmental changes is evaluated. Policies with good performance are incentivized (i.e., their selection weights are increased), while those with poor performance are suppressed (i.e., their weights are reduced but their opportunities are preserved), achieving adaptive strategy optimization. The MOEA / D algorithm generates a set of Pareto optimal solutions, and the optimal path is selected through non-dominated sorting and KneePoint optimization mechanisms. The predicted strategies update synaptic weights and normalize them into selection probabilities. ； Where K=3 is the total number of strategies, namely TL-P, LS-P, and RL-P; θmin=0.1 is the minimum selection probability to prevent strategies from being permanently eliminated due to temporary performance fluctuations.

Images