Aircraft multi-objective performance constraint collaborative optimization method based on multi-dimensional virtual-real parameter mapping

Abstract

The application discloses a kind of multi-dimensional virtual real parameter mapping-based aircraft multi-target performance constraint collaborative optimization method, by constructing real parameter space and virtual parameter space, mission critical elements such as target arrival, obstacle risk, formation keeping error and minimum safety margin are uniformly organized as multidimensional real parameter input, and speed scheduling quantity is uniformly organized as virtual parameter output, so as to realize consistent parameter expression and unified decision interface across scene, across risk level, improve engineering applicability and reusability under task constraint scene;By obtaining interactive response data in a plurality of scene sets and establishing a trajectory sample set, the mapping network of real parameter to speed virtual parameter is trained using PPO reinforcement learning, so that the system does not need to explicitly construct complex artificial rules or segmented threshold control logic, and adaptive trade-off can be formed among arrival efficiency, formation keeping and speed smoothness and other multiple targets, thereby reducing the cost of artificial parameter debugging and improving the strategy generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of aerospace technology, specifically relating to a collaborative optimization method for multi-objective performance constraints of aircraft based on multi-dimensional virtual-real parameter mapping. Background Technology

[0002] In existing technologies, to achieve target arrival and obstacle avoidance, trajectory planning and tracking control based on waypoints, local obstacle avoidance methods based on artificial potential fields or velocity barriers, and constraint control methods based on model predictive control or online optimization are commonly used. For formation missions, common solutions include lead-follow, consistency control, and virtual structures, and formation maintenance is achieved through relative position error or distance maintenance. The above methods can achieve certain results in ideal or weakly disturbed environments, but in practical applications, multiple objectives and constraints are often strongly coupled: for example, to improve arrival efficiency, it is usually necessary to increase speed or shorten detour distance, but this may lead to reduced obstacle avoidance safety, increased formation error, insufficient minimum spacing, or frequent control saturation; while excessive conservatism can lead to extended mission time, increased energy consumption, or even target arrival failure. Especially in situations with dense obstacles, narrow passages, or wind disturbances, it is difficult to simultaneously guarantee the performance and safety constraints of traditional methods.

[0003] In recent years, deep learning and reinforcement learning have been used for aircraft decision-making and parameter tuning, aiming to improve environmental adaptability through data-driven approaches. However, in multi-objective and multi-constraint scenarios, the weight design of reward and cost functions often relies on experience, and is prone to unstable training, slow convergence, or performance sacrifices for safety. Furthermore, learning strategies trained in simulations still face insufficient generalization due to differences between virtual and real-world parameters. Existing technologies rarely employ a multi-dimensional real-parameter-virtual-parameter mapping as their core, structuring key real parameters such as targets, obstacles, formation errors, and safety margins as input, and outputting virtual parameters that can be directly used for control and planning module scheduling. Simultaneously, they achieve collaborative optimization and feasibility assurance of multi-objective performance and constraints within the same framework. Summary of the Invention

[0004] The purpose of this invention is to provide a collaborative optimization method for multi-objective performance constraints of aircraft based on multi-dimensional virtual-real parameter mapping, so as to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for collaborative optimization of multi-objective performance constraints of aircraft based on multi-dimensional virtual-real parameter mapping, comprising the following steps:

[0006] S1. Establish a mission space for aircraft formations to reach the target area and avoid obstacles in a two-dimensional plane; unify the target arrival criteria, obstacle safety criteria, formation maintenance and aircraft spacing safety criteria, and speed capability boundary criteria.

[0007] S2. Quantify and characterize the unified target arrival criteria, obstacle safety criteria, formation maintenance criteria, inter-aircraft spacing safety criteria, and speed capability boundary criteria in S1; organize the target-related real parameters, obstacle-related real parameters, formation maintenance error, and minimum safety margin of inter-aircraft spacing into a multi-dimensional real parameter vector, and output the speed scheduling virtual parameters constructed based on the speed capability boundary criteria as a single virtual parameter.

[0008] S3. Generate a set of scenarios, which must include the following scenarios: typical scenarios of target distance, typical scenarios of sparse and dense obstacles, and typical scenarios of static and dynamic obstacles; perform interactive simulation on the set of scenarios to collect trajectory samples, form a dataset for PPO training, and clean and unify it to ensure training stability.

[0009] S4. The PPO algorithm is used to train a mapping network with the real parameter vector constructed in S2 as input and the velocity virtual parameter as output. This enables the strategy to achieve coordinated optimization of arrival efficiency and formation error while satisfying safety constraints, and to suppress the execution non-smoothness problem caused by drastic speed changes.

[0010] S5. The trained mapping network is used for online rolling execution. Based on the real parameter vector, the virtual speed parameter is output in each control cycle. The network output is mapped to the final speed command that meets the speed boundary through boundary constraints based on the Sigmoid function.

[0011] Preferably, S1 includes:

[0012] S11. Determine target elements and arrival criteria. The target is given in the form of a target point and the radius of the target area. The position of the target point is denoted as... The radius of the target area is denoted as ;

[0013] S12. Obstacle safety criteria include determining obstacle elements and safe expansion distances; the set of obstacles is denoted as... ,in Indexing obstacles The number of obstacles The center of the obstacle. The equivalent radius is used; to standardize safety distances, an obstacle safety expansion distance is introduced. ,in This is a preset constant, determined jointly by the aircraft's geometric dimensions and positioning error, and satisfies... .in The equivalent radius of the aircraft. As the upper bound of the positioning error, obstacle discrimination is based on the equivalent radius. implement;

[0014] S13. Formation maintenance and aircraft spacing safety criteria specifically involve determining the formation shape and safety spacing constraint thresholds. Formation maintenance does not explicitly restrict the specific geometric formation, but rather uses an adjacency set constraint method: a pre-defined set of adjacency relationships is provided. and for each edge Set reference distance It is used for calculating formation holding error; at the same time, it sets the minimum safe distance threshold. Used for machine spacing safety judgment, among which This is a preset constant, determined jointly by the aircraft's geometric dimensions and positioning error, and satisfies... .in The equivalent radius of the aircraft. This is the upper bound of the positioning error;

[0015] S14. The speed capability boundary criterion specifically determines the speed capability boundary and control cycle, with the speed virtual parameter value range set as follows: And set the maximum acceleration. With maximum deceleration It is used for candidate velocity generation, feasible region filtering, and online backoff; the control cycle is... This is used for unified timing of sample collection and policy execution.

[0016] Preferably, S1 further includes:

[0017] S15. Establish calculation standards for status acquisition quantities and intermediate quantities, and acquire the position and velocity of each aircraft in each control cycle. And calculate the position of the formation's center of mass:

[0018] ;

[0019] Among them, it means Current location of the aircraft This indicates the number of aircraft in the formation, and the position of the center of mass is used for the calculation of real parameters of target distance, obstacle net safe distance, and directional components.

[0020] Preferably, S2 includes:

[0021] S21. Construct the target real parameters based on the centroid position. With the target point Calculate target distance And using the unit vector pointing from the centroid to the target as the direction information, its components on the two axes of the two-dimensional coordinate system are taken as... and It is used to simultaneously characterize the distance of a target and its relative direction;

[0022] S22. Construct real parameters for obstacles, for the obstacle set. Calculate the net safe distance from the center of mass to each obstacle, and denote the net safe distance corresponding to the smallest one as . And take the unit vector component pointing to the most dangerous obstacle as , , used to characterize the degree of proximity to an obstacle and its relative direction;

[0023] S23. Construct a formation that preserves error, for a pre-given edge set. Calculate the pairwise distances on the edges And construct a dimensionless formation to maintain error:

[0024] ;

[0025] in, The number of aircraft in the formation; This is a set of adjacency relationships for formation, used to specify pairs of aircraft that need to maintain relative distance constraints; Let be an edge in the edge set, representing the first edge. frame and the first The two aircraft form a pair of adjacent constraint relationships; The number of elements in the edge set; The current moment; and The first frame and the first The aircraft at all times A two-dimensional position vector; For the first frame and the first The aircraft at all times The actual distance; For the corresponding edge The expected distance; To standardize the scale parameters when configuring reference distances; To standardize the scale parameters when configuring reference distances; To maintain accuracy in formation;

[0026] S24. Construct a minimum safety margin: Calculate the distance between any two machines in the formation and take the minimum value as the minimum machine spacing. Then, compare this minimum machine spacing with a threshold. The minimum safety margin is obtained by performing normalized comparison. .when Determine if the minimum safety clearance constraint is satisfied, when It was determined that there was a risk of touching or crossing the boundary.

[0027] Preferably, S2 further includes:

[0028] S25. Form the real parameter vector and define the virtual parameters, and organize the features in S2 into a real parameter vector:

[0029] ;

[0030] in: This indicates the distance between the formation's center of mass and the target point; and This represents the components of the unit vector pointing from the formation's center of mass to the target point on the two axes of the planar coordinate system. This indicates the net safe distance between the formation's center of mass and the nearest obstacle; and This represents the components of the unit vector pointing from the formation's center of mass to the nearest obstacle on the two axes of the planar coordinate system. This indicates that the formation maintains its error. Indicates the minimum safety margin;

[0031] The virtual parameter space defines only the velocity virtual parameter. It is used for velocity scheduling and multi-objective constraint collaborative optimization; pose information is used for calculation. However, it does not directly enter the output definition of the virtual-real mapping, thereby avoiding the dimensional expansion and weakening of interpretability caused by the direct mapping of pose and control variables.

[0032] Preferably, S3 includes:

[0033] S31. Scene set generation, target point Randomly sample within the mission area; number of obstacles Sampling within a preset range, obstacle radius With position Generate in a uniform or clustered distribution; in dynamic obstacle scenarios, assign uniform motion or preset trajectory to some obstacles, so that they may intersect or cross near the formation track, thereby covering typical dynamic obstacle avoidance scenarios.

[0034] S32. Initial formation conditions sampling: the initial position is randomly perturbed around an initial center, and the initial velocity is... Internal sampling;

[0035] S33. Trajectory Acquisition and Termination Criteria, based on control cycle. The propulsion system calculates the real parameter vector at each step. Output the virtual speed parameter and propulsion status, and record it. In addition to reward and constraint states, termination conditions include reaching the target area, touching the obstacle boundary, or crossing the distance between the aircraft;

[0036] S34. Sample cleaning and scaling to eliminate numerical divergence and obviously unachievable trajectories; normalize or truncate each component of the real parameters according to a unified range.

[0037] Preferably, S4 includes:

[0038] S41. Policy network input and output definition: The policy network uses a real parameter vector. Input and output speed recommendations And by using boundary constraints, the output is guaranteed to be located at... The value network estimates the state value using the same input, which is then used to calculate the advantage function.

[0039] S42, PPO update method, uses PPO cut target to limit the change of the ratio of new and old policies, avoids the instability of training caused by excessive policy update; adopts small batch multi-round update to improve sample utilization efficiency, and monitors success rate and average reward in the validation scenario. After meeting the performance requirements, the model parameters are exported and saved for deployment.

[0040] S43. The reward function is designed in blocks, which decompose the reward into arrival progress blocks, time cost blocks, formation maintenance blocks, safety constraint blocks, velocity smoothing blocks, and terminal blocks. These blocks are defined and combined separately, and the total reward satisfies the following:

[0041] ;

[0042] in, Indicates progress reward. Indicates time-cost reward. This indicates that the formation will maintain its reward. Indicates speed smoothing reward, This indicates a safety penalty item. This indicates a terminal reward.

[0043] Preferably, S5 includes:

[0044] S51. In each control cycle, acquire the real parameters of target-related parameters, obstacle-related parameters, formation holding error, and minimum safety margin, and assemble them into a real parameter vector according to a uniform dimension and fixed order as input to the strategy network.

[0045] S52. Input the real parameter vector into the converged policy network to obtain the network speed output, and map the network output to a final speed command that satisfies the speed boundary through boundary constraints based on the Sigmoid function.

[0046] ;

[0047] in, For discrete control step index, For a moment Target speed, The lower bound of the velocity parameter is... The upper bound of the speed parameter is dummy. For the Sigmoid function, Forward mapping function of policy network, For policy network parameters, For a moment The real parameter vector, To assist in velocity observation, it is used to characterize the velocity information at the current or previous moment.

[0048] The technical effects and advantages of this invention are as follows: 1. By constructing a real parameter space and a virtual parameter space, key task elements such as target arrival, obstacle risk, formation maintenance error and minimum safety margin are uniformly organized into multi-dimensional real parameter inputs, and speed scheduling is uniformly organized into virtual parameter outputs, thereby achieving consistent parameter expression and unified decision interface across scenarios and risk levels, and improving the engineering applicability and reusability under task constraints.

[0049] 2. By acquiring interactive response data and establishing a trajectory sample set in a multi-scenario set, and using PPO reinforcement learning to train the mapping network from actual parameters to velocity virtual parameters, the system can form an adaptive trade-off between multiple objectives such as arrival efficiency, formation maintenance and velocity stability without explicitly constructing complex manual rules or segmented threshold control logic. This reduces the cost of manual parameter debugging and improves the generalization ability of the strategy.

[0050] 3. By adopting boundary constraints based on the Sigmoid function for the speed output of the policy network, and constructing an online constraint discrimination and penalty mechanism based on the net safety distance of the obstacle and the minimum safety margin, key safety constraints are incorporated into the reward-driven collaborative optimization framework. This allows the speed output to be directly evaluated during online execution and converge to the output region that meets the safety requirements during the training phase, thereby effectively reducing the risk of boundary crossing and improving the safety and stability of closed-loop operation.

[0051] 4. By decomposing the reward function into sub-items such as arrival progress, time cost, formation maintenance, speed smoothing, safety penalty, and terminal reward and punishment, and conducting collaborative optimization training, the aircraft can actively reduce its speed to restore safety when approaching obstacles, and increase its speed to improve arrival efficiency when it is safe and the target is far away. This achieves dynamic coordination of efficiency and safety within the same framework, enhancing the continuity, interpretability, and reliability of strategy deployment. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the overall process of the present invention.

[0053] Figure 2 This is a schematic diagram illustrating the composition of the real parameter vector and the mapping relationship between the real and virtual parameters of the present invention.

[0054] Figure 3 The figure shows the performance comparison results of the method of this invention and the MPC speed planning method in typical scenarios. Detailed Implementation

[0055] 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, and 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.

[0056] This invention provides, for example Figures 1-3 The method for collaborative optimization of multi-objective performance constraints of an aircraft based on multi-dimensional virtual-real parameter mapping, as shown, includes the following steps:

[0057] S1. Define the mission space for the aircraft formation to reach the target area and avoid obstacles in a two-dimensional plane, establish target elements, obstacle elements, and capability boundary constraint elements, and determine the set of state variables collected online. The target elements include at least the target point position. With the radius of the target area The obstacle elements include at least a set of obstacles. ,in For the center of the obstacle, The equivalent radius is used, and a safe expansion distance for the obstacle is set. The capability boundary constraint elements include at least the velocity range. Maximum acceleration With maximum deceleration The safety constraint elements include at least the minimum inter-machine distance value. .

[0058] Attitude and velocity information of each aircraft is collected during each control cycle. And thus calculate the position of the formation's center of mass. With the scalar velocity of the center of mass This serves as the foundational input for subsequent real parameter construction, constraint determination, and velocity scheduling.

[0059] S2. Establish a multi-dimensional real parameter set to characterize target arrival progress, obstacle risk, formation maintenance and minimum safety margin between aircraft, and establish a velocity virtual parameter for aircraft motion; construct a real parameter space from the real parameter set and form a real parameter vector, construct a virtual parameter space from the velocity virtual parameter and form a virtual parameter output, thereby forming a virtual-real parameter mapping object of multi-dimensional real parameters and single velocity virtual parameter.

[0060] The real parameters include at least: target distance and target direction components, net safe distance to the nearest obstacle and obstacle direction components, formation holding error, and minimum safety margin. The real parameters are organized into a real parameter vector. ,satisfy:

[0061] ;

[0062] in, This represents the distance between the formation's center of mass and the target point, used to characterize the progress towards the target. and This represents the components of the unit vector pointing from the formation's center of mass to the target point on the two axes of the planar coordinate system, used to characterize the target's relative direction; This represents the net safe distance between the formation's center of mass and the nearest obstacle, which is the difference between the distance from the formation's center of mass to the boundary of the nearest obstacle and the safe distance threshold. and This represents the components of the unit vector pointing from the formation's center of mass to the nearest obstacle on the two axes of the planar coordinate system, used to characterize the relative direction of the obstacle; This indicates the formation retention error, used to represent the degree of consistency between the current formation and the desired formation. This represents the minimum safety margin between aircraft, used to characterize the degree of leeway between any two aircraft in the formation relative to the safety threshold. A non-negative margin indicates that the minimum safety distance constraint is met.

[0063] The virtual parameter space defines only the velocity virtual parameter. It is used to output speed scheduling quantities to achieve collaborative optimization of multiple objectives and multiple constraints.

[0064] S3. Abstract the formation preservation requirement into adjacency constraints and construct the formation preservation error to avoid explicit restrictions on specific geometric formations. A pre-defined set of formation adjacency relationships is provided. and for each edge Set the desired distance ;in It can be uniformly configured as a single scale parameter Alternatively, the configuration can be grouped by edge type to meet formation flexibility requirements. The actual distance on the edge set is calculated at any given time. Constructing dimensionless formations to maintain error ,satisfy:

[0065] ;

[0066] in, The number of aircraft in the formation; This is a set of adjacency relationships for formation, used to specify pairs of aircraft that need to maintain relative distance constraints; Let be an edge in the edge set, representing the first edge. frame and the first The two aircraft form a pair of adjacent constraint relationships; The number of elements in the edge set; The current moment; and The first frame and the first The aircraft at all times A two-dimensional position vector; For the first frame and the first The aircraft at all times The actual distance; For the corresponding edge The expected distance; To standardize the scale parameters when configuring reference distances; To standardize the scale parameters when configuring reference distances; To maintain formation error, it is used to characterize the overall deviation of actual distances on the edge set from the reference distance. The smaller the value, the more consistent it is with the desired formation.

[0067] Simultaneously construct the minimum safety margin in the machine room. At that moment Calculate the distance between any two aircraft in the formation, take the minimum distance between them as the minimum aircraft spacing, and then compare this minimum aircraft spacing with the minimum safe spacing threshold. The safety margin is obtained by performing normalization comparison. .when When the minimum safety distance constraint is satisfied; The system determines when there is a risk of touching or crossing the boundary. The resulting formation maintains its error. With minimum safety margin All operations rely solely on pairwise distances, exhibit translation and rotation invariance, and their computational complexity increases with the size of the edge set. Linear growth, facilitating online deployment.

[0068] S4. Generate a scene set within the task space and establish an interactive acquisition process to obtain a trajectory sample set for reinforcement learning training. Run the process in each scene according to a control cycle, recording the actual parameter vectors at each time step. Speed ​​virtual parameter output Reward signals and next-moment parameters Trajectory samples are generated. Abnormal trajectory samples are removed. Abnormal trajectories include at least those with numerical divergence, significantly excessive velocity or acceleration, obstacle crossing, or severely excessive inter-machine spacing, which are considered unusable, in order to ensure the data consistency and training stability of the trajectory sample set.

[0069] S5. Constructing the policy network and setting boundary constraints at the output end. The policy network takes the real parameter vector at the current moment and auxiliary velocity observations as input, directly outputs the velocity virtual parameters at the corresponding moment, and limits the velocity within a preset velocity boundary through the output sigmoid function, obtaining an online deployable real parameter to velocity virtual parameter mapping policy, used for online periodic output of velocity virtual parameters to achieve multi-objective performance constraint collaborative optimization. The velocity virtual parameters are given by the following formula:

[0070] ;

[0071] in, For discrete control step index, For a moment Speed ​​virtual parameter output value, As the lower bound of velocity, For the upper limit of speed, For the Sigmoid function, Forward mapping function of policy network, For policy network parameters, For a moment The real parameter vector, To assist in velocity observation, it is used to characterize the velocity information at the current or previous moment.

[0072] S6. Train the mapping network parameters using the PPO algorithm. This allows for the synergistic optimization of arrival efficiency and speed stability under constraints of obstacle safety, minimum spacing, and formation maintenance. The PPO training process includes trajectory acquisition, reward estimation, advantage calculation, and trimming updates, limiting the policy update magnitude and maintaining training stability. The reward function is decomposed into several sub-terms, defined separately, and the total reward is obtained by weighted summation.

[0073] The reward function is defined in blocks as follows:

[0074] (a) Progress item :

[0075] ;

[0076] (b) Time cost item :

[0077] ;

[0078] (c) Formation maintenance :

[0079] ;

[0080] (d) Velocity smoothing term :

[0081] ;

[0082] (e) Safety penalties :

[0083] Define the boundary quantity: , The safety penalty is:

[0084] ;

[0085] (f) Terminal item :

[0086] Define arrival indication Failure Indication The terminal item is:

[0087] ;

[0088] (g) Total Reward :

[0089] ;

[0090] in: For the current moment, To control the cycle; The distance from the formation's center of mass to the target point; Weights for progress items; Weighted by time cost; To maintain formation accuracy, To maintain the weight of the formation; The target speed at this moment, This is the upper limit of speed. Speed ​​smoothing weights; The net safe distance from the formation's center of gravity to the nearest obstacle; This represents the minimum safety margin between machines; and These are the obstacle contact limit and the machine-to-machine distance contact limit, respectively. and These are the obstacle safety penalty weight and the distance-to-machine safety penalty weight, respectively. Arrival instructions; To reach the terminal reward weight, The penalty weight for a failed terminal.

[0091] After training convergence, the online deployment phase is executed on a rolling basis according to the control cycle: calculation It constructs the constraint evaluation quantities required for online discrimination, which are used to generate safety penalty terms and terminal trigger signals; the speed virtual parameters are directly output by the policy network. The final target velocity is obtained by processing it with the Sigmoid function. ;Will The data is sent to the motion execution module and executed in a loop until the arrival criterion is met or the failure termination condition is triggered.

[0092] As a preferred embodiment of the present invention, the sample construction process includes the following steps:

[0093] S41. Determine the basic scene within the task space, and form a scene set for interactive data collection from the basic scene.

[0094] S42. In each scenario, the system runs in a rolling manner according to the control cycle, records the actual parameter vector, velocity virtual parameter output, reward signal and the actual parameter of the next moment at each moment, forms a trajectory sample, and accumulates the trajectory sample set through multiple rounds of interaction.

[0095] S43. Perform multiple rounds of interactive data collection in each scenario to enable the trajectory to cover two typical state transitions: rapid advancement state and risk recovery state, thereby improving the coverage of trajectory samples for different task stages and constraint states.

[0096] S44. Perform a consistency check on the collected trajectory data to detect whether there are numerical anomalies, control variable anomalies, or constraint state anomalies in the trajectory, and mark the unusable trajectory.

[0097] S45. Remove abnormal trajectory samples. The abnormal trajectories include at least those with numerical divergence, significantly excessive velocity or acceleration, obstacle crossing, or serious cross-limits between machines, which are unusable, in order to ensure the data consistency and training stability of the trajectory sample set.

[0098] As a preferred embodiment of the present invention, the process of constructing the policy network and the output boundary constraints includes the following steps:

[0099] S51 takes the current real parameter vector and auxiliary velocity observation as input to the policy network, directly outputs the velocity virtual parameter, and uses it as the basis for generating the target velocity.

[0100] S52 transforms the velocity and acceleration capability constraints into calculable velocity boundaries and velocity change rate boundaries, which serve as output constraints at the policy network output end. That is, the output of the policy network is mapped to the target velocity through the Sigmoid function, and the target velocity is used for online closed-loop execution.

[0101] As a preferred embodiment of the present invention, the PPO iterative training process includes the following steps:

[0102] S61. Set initial values ​​for the strategy network and value network and set iteration termination conditions, wherein the termination conditions include at least an average return increment threshold or a maximum iteration round threshold.

[0103] S62. Collect multiple interaction trajectories and calculate the reward and advantage functions in each iteration;

[0104] S63. Update network parameters based on PPO cutout target and fit value network parameters;

[0105] S64. When training is unstable, perform parameter rollback, adjust exploration intensity, adjust shearing coefficient, or adjust reward weight to ensure that the training process converges in the direction that satisfies the set of consistent constraints.

[0106] S65. Output the policy network parameters after training convergence and form a deployable real parameter and velocity virtual parameter mapping strategy for online periodic output of velocity virtual parameters to achieve multi-objective performance constraint collaborative optimization.

[0107] Working principle: Step 1: Establish a mission space in a two-dimensional plane for aircraft formations to reach the target area and avoid obstacles, and unify the target arrival criteria, obstacle safety criteria, formation maintenance and aircraft spacing safety criteria, and speed capability boundary criteria, so that subsequent real parameter extraction, virtual parameter output, sample collection, PPO training and online deployment can be completed under a consistent standard. Figure 1 This is a flowchart illustrating the overall process of the method described in this invention.

[0108] Step 1.1: Determine target elements and arrival criteria. The target is given in the form of a target point and the radius of the target area, and the position of the target point is denoted as... The radius of the target area is denoted as The mission is considered complete when the formation's center of gravity enters the target area, and this information is used for terminal rewards and statistical indicator extraction.

[0109] Step 1.2: Determine obstacle elements and safe expansion distances. The set of obstacles is denoted as... ,in The center of the obstacle. This is the equivalent radius. To standardize safety distances, an obstacle safety expansion distance is introduced. Obstacles are determined based on equivalent radius. Execution. For dynamic obstacle scenarios, the obstacle center... Despite the changes over time, the same safety expansion standard is still used for calculating net safety distance and determining boundary conditions.

[0110] Step 1.3: Determine the formation shape and safety distance constraint thresholds. The formation does not explicitly restrict the specific geometric shape, but is described using adjacency set constraints: a pre-defined set of adjacency relationships is provided. and for each edge Set reference distance It is used for calculating formation holding error; at the same time, it sets the minimum safe distance threshold. It is used for safety judgment of machine spacing.

[0111] Step 1.4: Determine the speed capability boundary and control cycle. The range of values ​​for the speed virtual parameter is set to... And set the maximum acceleration. With maximum deceleration It is used for candidate velocity generation, feasible region filtering, and online backoff; the control cycle is... This is used for unified timing of sample collection and policy execution.

[0112] Step 1.5: Establish the calculation specifications for state acquisition quantities and intermediate quantities. Acquire the position and velocity of each aircraft in each control cycle. And calculate the position of the formation's center of mass:

[0113] ;

[0114] Among them, it means Current location of the aircraft This indicates the number of aircraft in the formation. The center of mass position is used for calculating real parameters such as target distance, obstacle safety distance, and directional components.

[0115] To ensure the reproducibility of the embodiments, this embodiment adopts a unified threshold and parameter configuration, as shown in Table 1.

[0116] Table 1 Parameter Configuration Values

[0117]

[0118] Step 2: Organize target arrival, obstacle risk, formation maintenance and minimum safety margin into a multi-dimensional real parameter input vector, and organize the speed scheduling quantity into a single virtual parameter output, thereby forming a virtual-real parameter mapping object of real parameters and speed virtual parameters.

[0119] Step 2.1: Construct target-related real parameters. Based on the centroid position. With the target point Calculate target distance And using the unit vector pointing from the centroid to the target as the direction information, its components on the two axes of the two-dimensional coordinate system are taken as... and It is used to simultaneously characterize the distance of a target and its relative direction.

[0120] Step 2.2: Construct obstacle-related real parameters. For the obstacle set... Calculate the net safe distance from the center of mass to each obstacle, and denote the net safe distance corresponding to the smallest one as . And take the unit vector component pointing to the most dangerous obstacle as , It is used to characterize the degree of proximity to an obstacle and its relative direction.

[0121] Step 2.3: Construct formation preservation error. For a pre-given edge set... Calculate the pairwise distances on the edges And construct a dimensionless formation to maintain error:

[0122] ;

[0123] in, The number of aircraft in the formation; This is a set of adjacency relationships for formation, used to specify pairs of aircraft that need to maintain relative distance constraints; Let be an edge in the edge set, representing the first edge. frame and the first The two aircraft form a pair of adjacent constraint relationships; The number of elements in the edge set; The current moment; and The first frame and the first The aircraft at all times A two-dimensional position vector; Represents the L2 norm; For the first frame and the first The aircraft at all times The actual distance; For the corresponding edge The expected distance; To standardize the scale parameters when configuring reference distances; To maintain accuracy in formation.

[0124] Step 2.4: Construct the minimum safety margin. Calculate the distance between any two machines in the formation and take the minimum value as the minimum machine spacing. Then, compare this minimum machine spacing with a threshold value. The minimum safety margin is obtained by performing normalized comparison. .when Determine if the minimum safety clearance constraint is satisfied, when A risk of boundary breaching or exceeding the boundary has been identified. Safety margin will serve as one of the key criteria for safety penalties and feasibility screening.

[0125] Step 2.5: Form the real parameter vector and define the imaginary parameters. Organize the above elements into a real parameter vector:

[0126] ;

[0127] in: This indicates the distance between the formation's center of mass and the target point; and This represents the components of the unit vector pointing from the formation's center of mass to the target point on the two axes of the planar coordinate system. This indicates the net safe distance between the formation's center of mass and the nearest obstacle; and This represents the components of the unit vector pointing from the formation's center of mass to the nearest obstacle on the two axes of the planar coordinate system. This indicates that the formation maintains its error. This indicates the minimum safety margin.

[0128] The virtual parameter space defines only the velocity virtual parameter. It is used for velocity scheduling and multi-objective constraint collaborative optimization. Pose information is used for calculation. However, it does not directly enter the output definition of the virtual-real mapping, thereby avoiding the dimensional expansion and weakening of interpretability caused by the direct mapping of pose and control variables.

[0129] Step 3: Generate a scene set covering typical situations such as target distance, sparse or dense obstacles, and static or dynamic conditions. Conduct interactive simulation to collect trajectory samples, forming a dataset for PPO training. Then, clean and standardize the dataset to ensure training stability.

[0130] Step 3.1: Scene set generation. Target point Random sampling is performed within the task area, target radius. Use the configuration in Table 1. Number of obstacles Sampling within a preset range, obstacle radius With position Generates obstacles in a uniform or clustered distribution; in dynamic obstacle scenarios, some obstacles are given uniform motion or preset trajectories, so that they may intersect or cross near the formation track, thereby covering typical dynamic obstacle avoidance scenarios.

[0131] Step 3.2: Initial formation condition sampling. The initial position is randomly perturbed around an initial center, and the initial velocity is... Internal sampling; this sampling method is used to cover state distributions such as slight deviations from the expected formation, large formation error deviations, and local congestion, thereby improving the strategy's adaptability to error states.

[0132] Step 3.3: Trajectory Acquisition and Termination Criteria. Based on the control cycle. The propulsion system calculates the real parameter vector at each step. Output the virtual speed parameter and propulsion status, and record it. In addition, there are reward and constraint states. Termination conditions include reaching the target area, hitting an obstacle, or exceeding the distance between machines, which are used for terminal rewards and penalties and marking of failed samples during the training process.

[0133] Step 3.4: Sample Cleaning and Scale Consistency. Remove numerically divergent and obviously unachievable trajectories; normalize or truncate each component of the real parameters to a uniform range. For example, normalize the distance component to the maximum task scale, and maintain the orientation component within a uniform range. Error and safety margin are normalized to a threshold scale to ensure that the policy network maintains a stable input scale across the entire scenario; samples near the boundary are collected with appropriate encryption to improve the reliability of policy decisions under high-risk conditions.

[0134] Step 4: Use the PPO algorithm to train the mapping network from actual parameters to velocity virtual parameters, so that the strategy can achieve synergistic optimization of arrival efficiency and formation error while satisfying safety constraints, and suppress the execution non-smoothness problem caused by drastic speed changes.

[0135] Step 4.1: Define the input and output of the policy network. The policy network uses a vector of real parameters. Input and output speed recommendations And by using boundary constraints, the output is guaranteed to be located within the specified range. The value network estimates the state value with the same input, which is used for advantage function calculation.

[0136] Step 4.2: PPO Update Method. PPO (Plan-Off Pairing) is used to limit the change in the ratio of the old and new policies, preventing large policy updates from causing training instability. Mini-batch, multi-round updates are employed to improve sample utilization efficiency. Success rate and average reward are monitored in the validation scenario. Once performance requirements are met, the model parameters are exported and saved for deployment.

[0137] Step 4.3: Block Design of Reward Function. The reward is broken down into arrival progress block, time cost block, formation maintenance block, safety constraint block, velocity smoothing block, and terminal block, which are defined and combined separately. The total reward satisfies:

[0138] ;

[0139] in, Indicates progress reward. Indicates time-cost reward. This indicates that the formation will maintain its reward. Indicates speed smoothing reward, This indicates a safety penalty item. This indicates a terminal reward.

[0140] Step 5: Use the trained mapping network for online rolling execution. Based on the real parameter vector, output the virtual velocity parameter in each control cycle. After processing by the Sigmoid function, send it to the motion execution module.

[0141] Step 5.1: In each control cycle, acquire real parameters such as target-related parameters, obstacle-related parameters, formation holding error, and minimum safety margin, and arrange them into a real parameter vector according to a unified dimension and fixed order as input to the strategy network.

[0142] Step 5.2: Input the real parameter vector into the converged policy network to obtain the network speed output, and map the network output to a final speed command that satisfies the speed boundary through boundary constraints based on the Sigmoid function.

[0143] ;

[0144] in, For discrete control step index, For a moment Target speed, As the lower bound of velocity, For the upper limit of speed, For the Sigmoid function, Forward mapping function of policy network, For policy network parameters, For a moment The real parameter vector, To assist in velocity observation, it is used to characterize the velocity information at the current or previous moment.

[0145] Step 5.3: Set the target speed The data is sent to the motion execution module, and during online execution, safety-related quantities such as the nearest obstacle net safe distance and minimum safety margin are continuously calculated. At the same time, the arrival criteria and failure termination conditions are determined. When the arrival criteria are met, the task segment is terminated; when the failure termination condition is triggered, the task is terminated.

[0146] Step 6: Construct an experimental scenario set, use the MPC speed planning method as a comparison baseline, and conduct comparative verification under the same scenario, the same threshold configuration and the same time sequence settings. Extract performance indicators such as arrival time and formation error to verify the performance of the algorithm of this invention.

[0147] Step 6.1: Typical Scenario Setup. Construct three types of typical scenarios:

[0148] (1) Sparse obstacle and distant target scenario: There are few obstacles and they are sparsely distributed. The target is far away. This is used to compare the rationality of arrival efficiency and speed output.

[0149] (2) Obstacle-dense static passage scenario: There are many obstacles and they form passage constraints, which are used to compare the safety maintenance capability and the formation maintenance error control capability;

[0150] (3) Dynamic obstacle intersection scenario: Some obstacles pass through the formation path at a constant speed or intersect near the target, which is used to compare the stability under dynamic risks.

[0151] Step 6.2: Index extraction and statistical standards, recording arrival time for each experiment. Average formation error Minimum obstacle net safety distance Minimum safety margin The results of multiple experiments were used to generate a curve showing the formation error changing over time. The results are as follows: Figure 3 As shown, the blue curve represents the result of the method of this invention, and the red curve represents the result of the MPC speed planning method.

[0152] Step 6.3: Results Record Table. Multiple experiments were conducted on the three typical scenarios, and the average values ​​were calculated. The results are recorded in Table 2.

[0153] Table 2. Results Records and Comparisons for Typical Scenarios

[0154]

[0155] As shown in Table 2, in scenarios with sparse obstacles and distant targets, the method of this invention can shorten the arrival time with a lower formation error level. In scenarios with dense obstacles and static passage, the method of this invention can maintain a positive net safety distance and minimum safety margin, while the comparative methods show that the safety margin approaches or touches the limit in some scenarios. In scenarios with dynamic obstacles, the method of this invention reduces the average formation error while maintaining a safe distance from obstacles, demonstrating the ability to synergistically optimize arrival efficiency, safety, and formation maintenance under dynamic risks.

[0156] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-objective performance constraint-based collaborative optimization of aircraft based on multi-dimensional virtual-real parameter mapping, characterized in that: Includes the following steps: S1. Establish a mission space for aircraft formations to reach the target area and avoid obstacles in a two-dimensional plane; Unify the criteria for target arrival, obstacle safety, formation maintenance and aircraft spacing safety, and speed capability boundary criteria; S2. Quantify and characterize the unified target arrival criteria, obstacle safety criteria, formation maintenance criteria, inter-aircraft spacing safety criteria, and speed capability boundary criteria in S1; organize the target-related real parameters, obstacle-related real parameters, formation maintenance error, and minimum safety margin of inter-aircraft spacing into a multi-dimensional real parameter vector, and output the speed scheduling virtual parameters constructed based on the speed capability boundary criteria as a single virtual parameter. S2 includes: S21. Construct the target real parameters based on the centroid position. With the target point Calculate target distance And using the unit vector pointing from the centroid to the target as the direction information, its components on the two axes of the two-dimensional coordinate system are taken as... and It is used to simultaneously characterize the distance of a target and its relative direction; S22. Construct real parameters for obstacles, for the obstacle set. Calculate the net safe distance from the center of mass to each obstacle, and denote the net safe distance corresponding to the smallest one as . And take the unit vector component pointing to the most dangerous obstacle as , , used to characterize the degree of proximity to an obstacle and its relative direction; S23. Construct a formation that preserves error, for a pre-given edge set. Calculate the pairwise distances on the edges And construct a dimensionless formation to maintain error: ； in, The number of aircraft in the formation; This is a set of adjacency relationships for formation, used to specify pairs of aircraft that need to maintain relative distance constraints; Let be an edge in the edge set, representing the first edge. frame and the first The two aircraft form a pair of adjacent constraint relationships; The number of elements in the edge set; The current moment; and The first frame and the first The aircraft at all times A two-dimensional position vector; Represents the L2 norm; For the first frame and the first The aircraft at all times The actual distance; For the corresponding edge The expected distance; To standardize the scale parameters when configuring reference distances; To maintain accuracy in formation; S24. Construct a minimum safety margin: Calculate the distance between any two machines in the formation and take the minimum value as the minimum machine spacing. Then, compare this minimum machine spacing with a threshold. The minimum safety margin is obtained by performing normalized comparison. ,when Determine if the minimum safety clearance constraint is satisfied, when It has been determined that there is a risk of touching or crossing the boundary; S25. Form the real parameter vector and define the virtual parameters, and organize the features in S2 into a real parameter vector: ； in: This indicates the distance between the formation's center of mass and the target point; and This represents the components of the unit vector pointing from the formation's center of mass to the target point on the two axes of the planar coordinate system. This indicates the net safe distance between the formation's center of mass and the nearest obstacle; and This represents the components of the unit vector pointing from the formation's center of mass to the nearest obstacle on the two axes of the planar coordinate system. This indicates that the formation maintains its error. Indicates the minimum safety margin; The virtual parameter space defines only the velocity virtual parameter. It is used for velocity scheduling and multi-objective constraint collaborative optimization; pose information is used for calculation. However, it does not directly enter the output definition of the virtual-real mapping, thereby avoiding the dimensional expansion and weakening of interpretability caused by the direct mapping of pose and control variables; S3. Generate a set of scenarios, which must include the following scenarios: typical scenarios of target distance, typical scenarios of sparse and dense obstacles, and typical scenarios of static and dynamic obstacles; perform interactive simulation on the set of scenarios to collect trajectory samples, form a dataset for PPO training, and clean and unify it to ensure training stability. S4. The PPO algorithm is used to train a mapping network with the real parameter vector constructed in S2 as input and the velocity virtual parameter as output. This enables the strategy to achieve coordinated optimization of arrival efficiency and formation error while satisfying safety constraints, and to suppress the execution non-smoothness problem caused by drastic speed changes. S5. The trained mapping network is used for online rolling execution. Based on the real parameter vector, the virtual speed parameter is output in each control cycle. The network output is mapped to the final speed command that meets the speed boundary through boundary constraints based on the Sigmoid function.

2. The method for collaborative optimization of multi-objective performance constraints of an aircraft based on multi-dimensional virtual-real parameter mapping according to claim 1, characterized in that: S1 includes: S11. Determine target elements and arrival criteria. The target is given in the form of a target point and the radius of the target area. The position of the target point is denoted as... The radius of the target area is denoted as ; S12. Obstacle safety criteria include determining obstacle elements and safe expansion distances; the set of obstacles is denoted as... ,in Indexing obstacles The number of obstacles The center of the obstacle. The equivalent radius is used; to standardize safety distances, an obstacle safety expansion distance is introduced. ,in This is a preset constant, determined jointly by the aircraft's geometric dimensions and positioning error, and satisfies... ,in The equivalent radius of the aircraft. As the upper bound of the positioning error, obstacle discrimination is based on the equivalent radius. implement; S13. Formation maintenance and aircraft spacing safety criteria specifically involve determining the formation shape and safety spacing constraint thresholds. Formation maintenance does not explicitly restrict the specific geometric formation, but rather uses an adjacency set constraint method: a pre-defined set of adjacency relationships is provided. ,in, Indicates the number of aircraft in the formation. and Number the aircraft and index it; and assign each edge Set reference distance It is used for calculating formation holding error; at the same time, it sets the minimum safe distance threshold. Used for machine spacing safety judgment, among which This is a preset constant, determined jointly by the aircraft's geometric dimensions and positioning error, and satisfies... ,in The equivalent radius of the aircraft. This is the upper bound of the positioning error; S14. The speed capability boundary criterion specifically determines the speed capability boundary and control cycle, with the speed virtual parameter value range set as follows: ,in The lower bound of the velocity parameter is... Set an upper bound for the velocity parameter and set the maximum acceleration. With maximum deceleration It is used for candidate velocity generation, feasible region filtering, and online backoff; the control cycle is... This is used for unified timing of sample collection and policy execution.

3. The method for collaborative optimization of multi-objective performance constraints of an aircraft based on multi-dimensional virtual-real parameter mapping according to claim 2, characterized in that: S1 further includes: S15. Establish calculation standards for status acquisition quantities and intermediate quantities, and acquire the position of each aircraft in each control cycle. With speed ,in For aircraft identification index, Given the current moment, calculate the position of the formation's center of mass: ； Among them, it means Current location of the aircraft This indicates the number of aircraft in the formation, and the position of the center of mass is used for the calculation of real parameters of target distance, obstacle net safe distance, and directional components.

4. The method for collaborative optimization of multi-objective performance constraints of an aircraft based on multi-dimensional virtual-real parameter mapping according to claim 1, characterized in that: S3 includes: S31. Scene set generation, target point Randomly sample within the mission area; number of obstacles Sampling within a preset range, obstacle radius With position Generate in a uniform or clustered distribution; in dynamic obstacle scenarios, assign uniform motion or preset trajectory to some obstacles, so that they may intersect or cross near the formation track, thereby covering typical dynamic obstacle avoidance scenarios. S32. Initial formation conditions sampling: the initial position is randomly perturbed around an initial center, and the initial velocity is... Internal sampling; S33. Trajectory Acquisition and Termination Criteria, based on control cycle. The propulsion system calculates the real parameter vector at each step, outputs the virtual velocity parameters and propulsion status, and records the real parameter vector at the current moment. With velocity virtual parameter The real parameter vector at the next time step In addition to reward and constraint states, termination conditions include reaching the target area, touching the obstacle boundary, or crossing the distance between the aircraft; S34. Sample cleaning and scaling to eliminate numerical divergence and obviously unachievable trajectories; normalize or truncate each component of the real parameters according to a unified range.

5. The method for collaborative optimization of multi-objective performance constraints of an aircraft based on multi-dimensional virtual-real parameter mapping according to claim 1, characterized in that: S4 includes: S41. Policy network input and output definition: The policy network uses a real parameter vector. Input and output speed recommendations And by using boundary constraints, the output is guaranteed to be located within the specified range. The value network estimates the state value using the same input, which is then used to calculate the advantage function. S42, PPO update method, uses PPO cut target to limit the change of the ratio of new and old policies, avoids the instability of training caused by excessive policy update; adopts small batch multi-round update to improve sample utilization efficiency, and monitors success rate and average reward in the validation scenario. After meeting the performance requirements, the model parameters are exported and saved for deployment. S43. The reward function is designed in blocks, which decompose the reward into arrival progress blocks, time cost blocks, formation maintenance blocks, safety constraint blocks, velocity smoothing blocks, and terminal blocks. These blocks are defined and combined separately, and the total reward satisfies the following: ； in, Indicates progress reward. Indicates time-cost reward. This indicates that the formation will maintain its reward. Indicates speed smoothing reward, This indicates a safety penalty item. This indicates a terminal reward.

6. The method for collaborative optimization of multi-objective performance constraints of an aircraft based on multi-dimensional virtual-real parameter mapping according to claim 1, characterized in that: S5 includes: S51. In each control cycle, acquire the real parameters of target-related parameters, obstacle-related parameters, formation holding error, and minimum safety margin, and assemble them into a real parameter vector according to a uniform dimension and fixed order as input to the strategy network. S52. Input the real parameter vector into the converged policy network to obtain the network speed output, and map the network output to a final speed command that satisfies the speed boundary through boundary constraints based on the Sigmoid function. ； in, For discrete control step index, For a moment Target speed, As the lower bound of velocity, For the upper limit of speed, For the Sigmoid function, Forward mapping function of policy network, For policy network parameters, For a moment The real parameter vector, To assist in velocity observation, it is used to characterize the velocity information at the current or previous moment.

Images