A Spacecraft Target Allocation Method Based on Improved BAS Algorithm and Optimized Neural Network
By improving the BAS algorithm and optimizing the BP neural network, setting the update method for step size and whisker length, obtaining the optimal initial weights and thresholds, and constructing a spacecraft target allocation neural network model, the problem of poor stability of the BP neural network in spacecraft target allocation is solved, and real-time and accurate target allocation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF CONTROL ENG
- Filing Date
- 2022-10-31
- Publication Date
- 2026-05-26
AI Technical Summary
In existing technologies, BP neural networks have poor stability in spacecraft target allocation problems, are prone to getting trapped in local optima, and are difficult to achieve real-time, optimal target allocation.
By improving the BAS algorithm and optimizing the BP neural network, setting the update method for step size and whisker length, obtaining the optimal initial weights and thresholds, constructing a spacecraft target allocation neural network model, and optimizing it using the training dataset.
It improves the generalization and learning capabilities of neural networks, solves the problems of slow convergence speed and easy getting trapped in local extrema, and realizes real-time, accurate and applicable target allocation for spacecraft.
Smart Images
Figure CN115755942B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of spacecraft countermeasures technology, and specifically relates to a spacecraft target allocation method based on an improved BAS algorithm optimized neural network. Background Technology
[0002] For spacecraft swarm combat missions in a space environment, it is necessary to establish a spacecraft adversarial system model. Traditional "OODA" theory primarily divides space adversarial theory into a perception system, a judgment system, a decision-making system, and a control system. The main task of the decision-making system is to allocate targets based on the situational assessment results from the perception system and the adversarial rules in the knowledge base. In a red-blue force adversarial environment, target allocation involves assigning red-team spacecraft to blue-team spacecraft to perform adversarial tasks based on the situational awareness results.
[0003] Traditional real-time solutions to the objective assignment problem are mostly limited to conventional algorithms such as implicit enumeration, branch and bound, cutting plane method, and dynamic programming. While these algorithms are relatively simple in concept, their programming implementation is cumbersome, and their convergence speed is slow as the number of objectives increases, making them unsuitable for handling objective assignment problems with high dimensionality. With the development of computer technology, some novel heuristic optimization algorithms have provided new ideas and methods for solving the objective assignment problem.
[0004] For the target allocation problem involving multiple spacecraft, the Beetle Antennae Search (BAS) algorithm, proposed by Jiang et al. in 2017, is a novel bio-inspired intelligent algorithm inspired by the foraging behavior of longhorn beetles. When foraging, beetles use their two antennae to detect the strength of odors and then move towards the direction of the stronger odor. Through repeated detection and movement, they find the food. This algorithm has a simple optimization mechanism, gradually approaching the optimal solution through continuous iteration. It also boasts advantages such as low computational cost and the ability to perform optimization on a single beetle. Traditional BAS algorithms use a fixed decay coefficient for step size updates, resulting in excessively rapid step size decay and low search accuracy in the early stages, and excessively slow step size decay and slow local search efficiency in the middle and later stages. It also suffers from low solution accuracy, susceptibility to local optima, and sensitivity to parameter settings. When using a backpropagation (BP) neural network to solve the target allocation problem, improper selection of initial weights and thresholds can easily lead to the network getting trapped in local optima, sometimes failing to obtain a stable solution, or even causing the network to fail to converge. Summary of the Invention
[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a spacecraft target allocation method based on an improved BAS algorithm optimized neural network. This method solves the problems of poor stability and easy getting trapped in local optima in the prior art, effectively improves network performance, and achieves real-time and optimal target allocation for spacecraft.
[0006] The technical solution of this invention is:
[0007] A spacecraft target allocation method based on an improved BAS algorithm and optimized neural network includes the following steps:
[0008] 1) Construct a spacecraft target allocation sample dataset, in which each sample data includes the state parameters of each spacecraft in the red and blue spacecraft cluster and the corresponding target allocation scheme;
[0009] 2) Construct the topology of the spacecraft target assignment neural network;
[0010] 3) Set the step size and whisker length update method of the BAS algorithm to obtain the optimal initial weights and thresholds of the spacecraft target allocation neural network;
[0011] 4) Use the spacecraft target allocation sample dataset to train the spacecraft target allocation neural network to obtain the optimal initial weights and thresholds, and use the trained model to predict spacecraft target allocation.
[0012] Preferably, in step 3), the step size and beard length update method of the BAS algorithm are set, specifically, the step size is updated using the following expression:
[0013]
[0014] In the formula, δ t δ represents the step size in the t-th iteration; 0 Indicates the initial step size; t max Indicates the maximum number of iterations; δ min Indicates the minimum step size;
[0015] The length of the beard is updated using the following expression:
[0016] d t =c t δ t
[0017] In the formula, d t c represents the length of the whisker in the t-th iteration; t It is a constant, taken as 2.5.
[0018] Preferably, in step 3), obtaining the optimal initial weights and thresholds of the spacecraft target allocation neural network specifically includes:
[0019] (31) Randomly select the initial weights and thresholds of the spacecraft target neural network as the initial position of the beetle's centroid, and initialize the beetle's forward step size to δ. 0 The direction vector of the longhorn beetle's movement is initialized as follows:
[0020]
[0021] In the formula, rand(k,1) is a k-dimensional random vector formed by the left and right whiskers of the longhorn beetle, and ||rand(k,1)|| is its modulus;
[0022] (32) Define the fitness function of the BAS algorithm;
[0023] (33) Save the initial position of the centroid of the longhorn beetle in bestX, calculate the fitness function value corresponding to the position, and save it in bestY;
[0024] (34) Update the positions of the longhorn beetle's left and right whiskers according to the following formula:
[0025]
[0026] In the formula, x t Let the centroid of the longhorn beetle be the location of the beetle's center of mass after t iterations. Let t be the position of the longhorn beetle's left whisker. Let t be the position of the right whisker of the longhorn beetle after t iterations;
[0027] (35) Update the position of the longhorn beetle's centroid according to the following formula:
[0028]
[0029] In the formula, f l t and Let be the fitness function values at the positions of the left and right whiskers of the longhorn beetle after t iterations, respectively. sgn(C) is the sign function. When C > 0, sgn(C) = 1; when C = 0, sgn(C) = 0; when C < 0, sgn(C) < 0.
[0030] Calculate the fitness function value at the current position. If the fitness function value is less than bestY, then update bestX and bestY; otherwise, do not update.
[0031] (36) Determine whether the iteration termination condition is met. If it is met, output the current centroid position of the beetle as the initial weights and threshold of the spacecraft target allocation neural network. Otherwise, return to step (34).
[0032] Preferably, the expression for the fitness function is:
[0033]
[0034] Where T is the sample size, s i This represents the number of unequal elements between the output value and the actual value of the i-th sample.
[0035] Preferably, in step 1), each sample data in the spacecraft target allocation sample dataset includes: the three-dimensional coordinates, velocity, deflection angle and combat range of each spacecraft in the red spacecraft cluster, the three-dimensional coordinates, velocity, deflection angle and threat level of each spacecraft in the blue spacecraft cluster, and the blue spacecraft assigned to each red spacecraft.
[0036] Preferably, in step 2), the topology of the spacecraft target allocation neural network is constructed as follows: it includes an input layer, a hidden layer and an output layer. The input layer parameters are the three-dimensional coordinates, velocity, deflection angle and adversarial range of each spacecraft in the red spacecraft cluster, and the three-dimensional coordinates, velocity, deflection angle and threat level of each spacecraft in the blue spacecraft cluster. The output layer parameters are the blue spacecraft assigned to each red spacecraft.
[0037] Preferably, the number of input layer neurons in the spacecraft target allocation neural network is 6*(M+N), where M is the number of red spacecraft and N is the number of blue spacecraft, and the number of output layer neurons is M.
[0038] Preferably, the activation function of the hidden layer of the spacecraft target allocation neural network is the Sigmoid function.
[0039] Preferably, the activation function of the output layer of the spacecraft target allocation neural network is the Softmax function.
[0040] Preferably, in step (4), the spacecraft target allocation neural network for obtaining the optimal initial weights and thresholds is trained using the spacecraft target allocation sample dataset, and the training error loss function used is the trianlm function.
[0041] The advantages of this invention compared to the prior art are:
[0042] (1) This invention utilizes the BAS algorithm to optimize the initial connection weights and thresholds of the BP neural network, thereby improving the generalization ability, learning ability and overall search efficiency of the neural network, and solving the problems of slow convergence speed and easy trapping in local extrema of the BP neural network.
[0043] (2) This invention improves and optimizes the update formulas for step size and beard length of the BAS algorithm. The improved BAS-BP model has improved in terms of target allocation accuracy and convergence speed. This invention is highly practical and can obtain more real-time and accurate prediction results. It has good applicability to spacecraft target allocation problems and provides a more effective intelligent allocation method for spacecraft swarm confrontation missions. Attached Figure Description
[0044] Figure 1 This is a schematic diagram of the spacecraft target allocation method based on the improved BAS algorithm and optimized neural network of the present invention. Detailed Implementation
[0045] The features and advantages of the present invention will become clearer and more explicit through the following detailed description.
[0046] This invention addresses the problem of target allocation for multiple spacecraft. It fully considers the selection of initial weights and thresholds for the BP neural network and combines it with an improved BAS algorithm to overcome the shortcomings of the existing BP neural network, such as poor stability and susceptibility to local optima. The result is a new method for spacecraft target allocation, namely a spacecraft target allocation method based on the beetle whisker algorithm, which aims to achieve real-time and optimal target allocation for spacecraft in a rapidly changing space combat environment.
[0047] Based on the above considerations, this invention discloses a spacecraft target allocation method based on the beetle whisker algorithm. For example... Figure 1 As shown, this target allocation method will proceed according to the following three steps: a, b, and c.
[0048] a. In step a, based on the spacecraft confrontation simulation data, a dataset of 200 sample data for spacecraft target allocation tasks is obtained. Considering that the confrontation environment is M red spacecraft versus N blue spacecraft, each sample data includes the position, speed and threat rating of the red and blue spacecraft, as well as the existing target allocation scheme, a target allocation dataset is established, and the dataset is randomly divided into a training set and a test set.
[0049] b. In step b, initialize the parameters of the BP neural network model and the BAS algorithm, construct the BAS-BP neural network model for target allocation in the spacecraft swarm adversarial problem, randomly generate the initial connection weights and thresholds of the neural network as the initial spatial position of the beetle, and calculate the fitness value at this time. The BAS algorithm finds the optimal initial connection weights and thresholds through iterative updates and obtains the optimal fitness value, thus constructing the optimal BAS-BP neural network model.
[0050] In this embodiment, the BP neural network structure and parameters are first determined, that is, the BP neural network model parameters are initialized based on the number of red and blue spacecraft, as follows: the number of input layer neurons is 6*(M+N), where 6 represents that each spacecraft has 6 parameters. The 6 parameters for the red spacecraft are 3D coordinates, velocity, deflection angle, and adversarial range, and the 6 parameters for the blue spacecraft are 3D coordinates, velocity, deflection angle, and threat level; based on experience, the number of hidden layer neurons is set to 100; the number of output layer neurons is M, representing the allocation result for each red spacecraft; the hidden layer neuron type is Sigmoid; the output layer neuron type is Softmax; the training error loss function is the trainlm function; and the performance evaluation metric is mean squared error.
[0051] In this embodiment, the step size decay and beard length decay of the BAS algorithm were optimized and improved. The improved step size formula is as follows:
[0052]
[0053] In the formula, δ t δ represents the step size in the t-th iteration; 0 Indicates the initial step size; t max Indicates the maximum number of iterations; δ min This represents the minimum step size. Based on the aforementioned step size update strategy, the whisker length formula used in this embodiment is:
[0054] d t =c t δ t
[0055] In the formula, c t Let be a constant, taken as 2.5;
[0056] The improved BAS algorithm described above is used to find the optimal initial weights and thresholds of the BP neural network, and then applied to the configured network to train the final spacecraft target allocation model.
[0057] The BAS algorithm simulates the foraging behavior of longhorn beetles to find the optimal solution. The biological principle is as follows: longhorn beetles do not know the exact location of food during foraging; they search for food based on the intensity of the odor sense detected by their left and right antennae. If the right antennae detect a stronger odor than the left, the beetle moves to the right; otherwise, it moves to the left, until it finds food. In the longhorn beetle whisker search algorithm, the beetle is simplified to a centroid, with its left and right antennae located on either side of this centroid. The beetle's orientation is randomly initialized, and the optimal solution is found through iteration by adjusting the search radius and step size. Figure 1 The improved BAS algorithm for optimizing the initial connection weights and thresholds of a BP neural network is presented in detail. The specific optimization process is as follows:
[0058] Step 1: Initialize the relevant parameters of the longhorn beetle. For the optimal search problem in k-dimensional space, let the centroid of the longhorn beetle be x, and the left whisker vector be x. l The right-hand vector is x r All are k-dimensional vectors. The BP network structure is 6*(M+N)-100-M, where 6*(M+N) is the number of neurons in the input layer, 100 is the number of neurons in the hidden layer, M is the number of neurons in the output layer, and the dimension of the search space is k = 6*(M+N)*100+100*M+100+M;
[0059] Create a random vector in the direction of the longhorn beetle:
[0060]
[0061] In the formula, rand(k, 1) is a k-dimensional random vector formed by the left and right whiskers of the longhorn beetle, and ||rand(k, 1)|| is its modulus. Therefore, the coordinates of the left and right whiskers are expressed as:
[0062]
[0063] Where, x t To determine the centroid position of the longhorn beetle after t iterations, d t Let be the distance between the left and right sides after t iterations.
[0064] Step 2: Set the step size and beard length. The step size controls the longhorn beetle's area search capability. The initial step size should be as large as possible to cover a sufficiently large search area, and subsequent step sizes should be gradually reduced to ensure finer search.
[0065] Step 3: Determine the fitness function. The root mean square error of the test data is used as the fitness evaluation function. The fitness function is:
[0066]
[0067] Where T is the number of samples in the training set, s i Let be the number of unequal elements (ranging from 0 to T, and being an integer) between the output value and the actual value (both are M-dimensional vectors) of the i-th sample. Therefore, the position where the fitness function value is minimum when the algorithm stops iterating is the optimal solution to the problem.
[0068] Step 4: Initialize the longhorn beetle's position. Randomly select the initial position of the longhorn beetle and save it in bestX.
[0069] Step 5: Calculate the fitness function. Calculate the fitness function for the initial position and store it in bestY.
[0070] Step 6: Update the positions of the longhorn beetle's left and right whiskers.
[0071] Step 7: Update the beetle position. Calculate the fitness function value f(x) for the left and right whiskers respectively. r ) and f(x l The position of the longhorn beetle is updated based on this value, that is, the weights and thresholds of the BP neural network are adjusted, and the fitness function value at the current position is calculated. If the fitness function value is less than bestY, then bestX and bestY are updated.
[0072] Step 8: Iteration Termination Judgment. Determine whether the fitness function value has reached the set precision or the maximum number of iterations has been reached. If the conditions are met, the optimal solution has been obtained; otherwise, return to S6 to continue the iteration.
[0073] Step 9: Optimal Solution Generation. When the algorithm stops iterating, the optimal solution is stored in bestX, which represents the optimal initial weights and thresholds of the BP neural network. The obtained optimal initial weights and thresholds are then substituted into the BP network for training, ultimately yielding the spacecraft target allocation model.
[0074] c. In step c, the training error loss function is the trianlm function. The improved BAS-BP neural network model is trained using the input parameter matrix (input) and output parameter matrix (output) of the target allocation offline training dataset. The BP neural network continuously corrects the weights through gradient descent until the error reaches the set accuracy or the number of iterations exceeds the maximum number of iterations, thereby obtaining the optimal initial connection weights and thresholds, constructing the optimal improved BAS-BP neural network model. Then, the input parameter matrix (input) of the training set and test set is used as the input parameters of the optimal BAS-BP neural network model, and then the target allocation is performed for the spacecraft swarm adversarial task.
[0075] In summary, this invention addresses the shortcomings of BP neural networks, such as slow convergence and susceptibility to local optima, by improving the initial connection weights and thresholds of the BAS algorithm to optimize the BAS-BP model. The improved BAS algorithm is a single-particle search algorithm, significantly increasing the convergence speed of the BAS-BP model. Furthermore, improvements and optimizations to the update formulas for the step size and whisker length of the BAS algorithm enhance the target allocation and convergence speed of the improved BAS-BP model. This invention is highly practical, providing more real-time and accurate vector allocation results, and is well-suited for spacecraft target allocation problems, offering a more effective intelligent allocation method for spacecraft swarm warfare missions.
[0076] The examples provided above are merely for illustrating the purpose of this invention and are not intended to limit the scope of the invention. Any person skilled in the art can make possible variations and modifications to the technical solutions of this invention using the methods and techniques disclosed above, without departing from the spirit and scope of the invention. The scope of this invention is defined by the appended claims. All equivalent substitutions and modifications made without departing from the spirit and principles of this invention should be covered within the scope of this invention.
[0077] The contents not described in detail in this specification are common knowledge to those skilled in the art.
Claims
1. A spacecraft target assignment method based on an improved BAS algorithm optimized neural network, characterized in that, Includes the following steps: 1) Construct a spacecraft target allocation sample dataset, in which each sample data includes the state parameters of each spacecraft in the red and blue spacecraft cluster and the corresponding target allocation scheme; 2) Construct the topology of the spacecraft target assignment neural network; 3) Set the step size and whisker length update method of the BAS algorithm to obtain the optimal initial weights and thresholds of the spacecraft target allocation neural network; 4) Train the spacecraft target allocation neural network to obtain the optimal initial weights and thresholds using the spacecraft target allocation sample dataset, and use the trained model to predict spacecraft target allocation. In step 3), the step size and beard length update methods of the BAS algorithm are set. Specifically, the step size is updated using the following expression: wherein represents the step size at the i th iteration; represents the initial step size; represents the maximum number of iterations; represents the minimum step size; The length of the beard is updated using the following expression: wherein represents the length of the hair at the iteration; is a constant and takes the value 2.
5.
2. The spacecraft target assignment method of claim 1, wherein, In step 3), obtaining the optimal initial weights and thresholds of the spacecraft target allocation neural network specifically includes: (31) Randomly select the initial weights and thresholds of the spacecraft target neural network as the initial position of the bumblebee centroid, initialize the step length of the bumblebee as and initialize the direction vector of the bumblebee as: wherein is formed by the left antenna of the beetle is a random vector, is a module thereof; (32) Define the fitness function of the BAS algorithm; (33) save the initial position of the borer centroid in , calculate the fitness function value corresponding to the position, and save it in ; (34) Update the positions of the longhorn beetle's left and right whiskers according to the following formula: wherein is the iteration the centroid position of the cerambycid, is the iteration the left antenna position of the cerambycid, is the iteration the right antenna position of the cerambycid; (35) Update the position of the longhorn beetle's center of mass according to the following formula: wherein and are the fitness function values at the left and right antenna positions of the day-beetle after the th iteration, respectively, is a sign function, when ; is a sign function, when ; is a sign function, when ; And calculate the fitness function value at the current position. If the fitness function value is less than 1, then... Then update and Otherwise, it will not be updated; (36) Determine whether the iteration termination condition is met. If it is met, output the current centroid position of the beetle as the initial weights and threshold of the spacecraft target allocation neural network. Otherwise, return to step (34).
3. The spacecraft target allocation method based on an improved BAS algorithm optimized neural network according to claim 2, characterized in that, The expression for the fitness function is: in, For the sample size, For the first The number of unequal elements in the output value and the actual value of each sample.
4. The spacecraft target allocation method based on an improved BAS algorithm optimized neural network according to claim 3, characterized in that, In step 1), each sample data in the spacecraft target allocation sample dataset includes: the three-dimensional coordinates, velocity, deflection angle and combat range of each spacecraft in the red spacecraft cluster, the three-dimensional coordinates, velocity, deflection angle and threat level of each spacecraft in the blue spacecraft cluster, and the blue spacecraft assigned to each red spacecraft.
5. A spacecraft target allocation method based on an improved BAS algorithm optimized neural network according to claim 4, characterized in that, In step 2), the topology of the spacecraft target allocation neural network is constructed as follows: it includes an input layer, a hidden layer and an output layer. The input layer parameters are the three-dimensional coordinates, velocity, deflection angle and adversarial range of each spacecraft in the red spacecraft cluster, and the three-dimensional coordinates, velocity, deflection angle and threat level of each spacecraft in the blue spacecraft cluster. The output layer parameters are the blue spacecraft assigned to each red spacecraft.
6. The spacecraft target allocation method based on an improved BAS algorithm optimized neural network according to claim 5, characterized in that, The number of neurons in the input layer of the spacecraft target allocation neural network is: ,in The number of Red Team spacecraft. The number of blue team spacecraft and the number of output layer neurons are... .
7. A spacecraft target allocation method based on an improved BAS algorithm optimized neural network according to any one of claims 1 to 6, characterized in that, The hidden layer activation function of the spacecraft target allocation neural network is the Sigmoid function.
8. A spacecraft target allocation method based on an improved BAS algorithm optimized neural network according to any one of claims 1 to 6, characterized in that, The activation function of the output layer of the spacecraft target allocation neural network is the Softmax function.
9. A spacecraft target allocation method based on an improved BAS algorithm optimized neural network according to any one of claims 1 to 6, characterized in that, In step (4), the spacecraft target allocation neural network that obtains the optimal initial weights and thresholds is trained using the spacecraft target allocation sample dataset, and the training error loss function used is the trianlm function.