Computational guidance method based on gated progressive neural network
By using a gated progressive neural network approach, the output of the old column task unit's output layer is introduced into the new column task unit, and weight parameters are set. This solves the parameter explosion problem of progressive neural networks when the task environment changes, enabling guidance tasks to quickly adapt to new environments and reducing training time and computational resource requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2023-07-27
- Publication Date
- 2026-07-24
AI Technical Summary
Existing data-based computational guidance methods require a large amount of data and time when the mission environment changes, and the lateral connection method of progressive neural networks leads to an explosive growth in network parameters, increasing training time and computational resource requirements.
A gated progressive neural network is adopted. By introducing the output of the old column task unit output layer into the input layer of the new column task unit and setting the weight parameters, the complexity of lateral connections is reduced. During retraining, the parameters of the old task unit are frozen, and only the parameters of the new task unit and the weight parameters are trained. The mean squared error is used as the loss function.
The network structure was simplified, the explosive growth of parameters was avoided, the amount of training computation was reduced, the guidance method was improved to adapt quickly to new environments, and time costs were saved.
Smart Images
Figure CN117195685B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a computational guidance method based on a gated progressive neural network, belonging to the field of guidance technology. Background Technology
[0002] Computational guidance can be mainly divided into two categories: model-based computational guidance and data-based computational guidance.
[0003] Model-based computational guidance methods rely on mathematical models of missile flight dynamics and guidance laws to find optimal solutions through optimization algorithms. However, this method requires accurate models, which are often difficult to obtain in practice due to factors such as flight environment and target dynamics. Data-driven computational guidance methods, on the other hand, learn from large amounts of historical data to extract useful information and make decisions. The advantage of this type of method is that it does not require accurate models and can handle complex or uncertain scenarios. For example, machine learning methods such as deep learning are representative of data-driven computational guidance. This method, by training on large amounts of data, can learn complex patterns in the guidance process and make effective decisions.
[0004] However, when the mission environment changes, the pre-trained guidance model does not perform ideally. Obtaining a guidance model suitable for a new guidance mission environment requires a large amount of data and time, which is impractical in real-world applications.
[0005] Existing technologies include incremental learning using progressive neural networks to replace traditional neural networks. By adding column networks, the old guidance tasks are not forgotten during the learning process of new guidance tasks, thereby reducing the amount of training data for guidance tasks in new environments and achieving accurate guidance in new environments. However, the special lateral connection method in progressive neural networks can cause the network parameters to grow explosively with the number of tasks, resulting in a sharp increase in the time cost of adapting to subsequent tasks.
[0006] Therefore, it is necessary to further study existing data-based computational guidance methods to address the aforementioned issues. Summary of the Invention
[0007] To overcome the above problems, the inventors conducted in-depth research and provided a computational guidance method based on a gated progressive neural network, comprising the following steps:
[0008] Using the training data from the original mission, a gated progressive neural network is pre-trained to obtain an initial network. The initial network outputs the bias acceleration of the aircraft, and the guidance law of the aircraft is obtained based on the bias acceleration.
[0009] When the guidance mission changes, the initial network is retrained using the training data of the new mission to obtain an updated network. The updated network is then used to output the aircraft bias acceleration, and the guidance law of the aircraft is obtained based on the bias acceleration.
[0010] In a preferred embodiment, the gated progressive neural network includes one or more columns of task units, each column of task units having an input layer, one or more hidden layers, and an output layer; there are lateral connections between different columns of task units.
[0011] In a preferred embodiment, the retraining includes the following steps:
[0012] Add new column task units to the initial network and establish lateral connections between the new column task units and the old column task units to form a new progressive neural network;
[0013] A new progressive neural network is trained using training data from the new task to obtain an updated network.
[0014] In a preferred embodiment, the input to the new column task unit input layer includes neural network input data and the outputs of all old column task unit output layers.
[0015] In a preferred embodiment, a weighting parameter is set in the input of the new column task unit input layer for the output of the old column task unit output layer.
[0016] In a preferred embodiment, the weighting parameter is an adaptive parameter.
[0017] In a preferred embodiment, during retraining, the parameters in the old task units are frozen.
[0018] In a preferred embodiment, the training data includes the flight state parameters and guidance law of the aircraft, wherein the flight state parameters are the input parameters of the neural network and the bias acceleration of the aircraft is the output parameter of the neural network.
[0019] In a preferred embodiment, during retraining, the mean squared error is used as the loss function for training new columns.
[0020] In a preferred embodiment, the training data is obtained in the following manner:
[0021] Establish a dynamic model of the aircraft;
[0022] Set the optimal control objective and obtain multiple sets of aircraft flight state variables and corresponding aircraft bias accelerations based on the dynamic model.
[0023] The beneficial effects of this invention include:
[0024] (1) It simplifies the PNN network, avoids the explosive growth of network parameters caused by the increase of tasks, and reduces the computational requirements for training;
[0025] (2) A gating mechanism is proposed to facilitate knowledge transfer and alleviate the weakness of knowledge transfer between different tasks due to network simplification.
[0026] (3) The guidance method can quickly adapt to new aerodynamic environments with limited new data and can rapidly adapt to mission requirements in new environments. Compared with traditional deep learning guidance, it saves a lot of time and greatly improves the deployment efficiency of guidance laws. Attached Figure Description
[0027] Figure 1 A schematic flowchart of a computational guidance method based on a gated progressive neural network according to a preferred embodiment of the present invention is shown.
[0028] Figure 2 A schematic diagram of a traditional progressive neural network architecture is shown.
[0029] Figure 3 A schematic diagram of a gated progressive neural network architecture according to a preferred embodiment of the present invention is shown;
[0030] Figure 4 The experimental structures in Example 1 and Comparative Example 1 are shown. Detailed Implementation
[0031] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present invention will become clearer and more apparent.
[0032] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.
[0033] According to the present invention, a computational guidance method based on a gated progressive neural network is provided, such as... Figure 1 As shown, it includes the following steps:
[0034] Using the training data from the original mission, a gated progressive neural network is pre-trained to obtain an initial network. The initial network outputs the bias acceleration of the aircraft, and the guidance law of the aircraft is obtained based on the bias acceleration.
[0035] When the guidance mission changes, the initial network is retrained using the training data of the new mission to obtain an updated network. The updated network is then used to output the aircraft bias acceleration, and the guidance law of the aircraft is obtained based on the bias acceleration.
[0036] According to the present invention, the gated progressive neural network includes one or more columns of task units, each task unit having an input layer, one or more hidden layers, and an output layer; there are lateral connections between different columns of task units.
[0037] In a preferred embodiment, when the training data of the original task contains multiple tasks, each task corresponds to a column of task units, and the training data of multiple tasks are trained sequentially.
[0038] Preferably, the input of the next column of task unit input layer includes the input data and the output of all the previous column of task unit output layers.
[0039] More preferably, the input of the input layer of the next column of task units has a weighting parameter set for the output of the output layers of the previous column of task units.
[0040] According to the present invention, the retraining includes the following steps:
[0041] Add new column task units to the initial network and establish lateral connections between the new column task units and the old column task units to form a new progressive neural network;
[0042] A new progressive neural network is trained using training data from the new task to obtain an updated network.
[0043] According to the present invention, the new column task unit has an input layer, one or more hidden layers, and an output layer;
[0044] Preferably, the new task unit has one input layer, two hidden layers, and one output layer. More preferably, the input layer has a dimension of 6, the hidden layers have a dimension of 32, and the output layer has a dimension of 1, with different layers connected by ReLU activation layers.
[0045] Furthermore, the input to the new column task unit input layer includes neural network input data and the outputs of all old column task unit output layers.
[0046] Traditional progressive neural networks (PNNs) add a new column when a new task is introduced, and laterally connect the results from different columns and layers to achieve compatibility between new and old tasks. Their structure is as follows: Figure 2As shown, however, due to the unique lateral connections, this structure will result in a larger network size, a lack of control over the degree of transition, an explosive increase in the number of parameters that need to be trained, a greater demand on computing resources and computing power, and an increase in training time.
[0047] Unlike traditional progressive neural networks (PNNs), in this invention, the output of the old column task unit output layer is used as one of the inputs of the new column task unit input layer, thereby reducing the complexity of lateral connections, reducing network size, and reducing computing power requirements.
[0048] Furthermore, to ensure that the lateral connections, even with reduced complexity, still provide sufficient knowledge transfer from older tasks, in this invention, preferably, the input of the new column task unit input layer includes a weight parameter for the output of the old column task unit output layer. This weight parameter adaptively controls the degree of connection, as shown in the following structure: Figure 3 As shown, the weight parameter is also called the gating weight, and the product of the output of the old column task unit output layer and the weight parameter is called the gating data.
[0049] Preferably, the weight parameters are adaptive parameters, thereby implementing a gating mechanism through the weight parameters, which allows for the selective transfer of knowledge from previous tasks. The network can autonomously determine the useful knowledge for new tasks and control the degree of knowledge transfer. This enables new tasks to learn more from the potential knowledge in previous tasks and reduces the amount of training sample data required to adapt to new tasks.
[0050] According to the present invention, during the retraining process, the parameters in the old task unit are frozen, and only the parameters and weight parameters in the new task unit are trained.
[0051] For example, the initial network trained on the original task training data has K-1 tasks, i.e., K-1 columns of task units. When retraining the initial network using training data for a new task, the Kth column of task units is added, and the input I of the Kth column of task units... K Represented as:
[0052]
[0053] Where I represents the training data for the new task, and O i This represents the output of the i-th task unit. This represents the weight parameters from the i-th task unit to the K-th task unit.
[0054] Furthermore, the weighting parameter is expressed as:
[0055] in, Let be the weight matrix of the i-th old column task unit. Let be the bias matrix of the i-th old column task unit. Both the weight matrix and the bias matrix are trainable parameters, which allows the network to determine how to transfer knowledge from the previous columns.
[0056] In a preferred embodiment, during retraining, the mean squared error is used as the loss function for training new columns. Represented as:
[0057]
[0058] Where m is the number of samples in the training data. This is the output of the m-th sample.
[0059] In a preferred embodiment, the lateral connection consists of a fully connected layer and an activation layer. The dimension of the fully connected layer is the same as the dimension of the current column input layer, and the activation layer uses the Sigmoid function to achieve adaptive learning of the weight parameters.
[0060] According to the present invention, new task units can be continuously added for new tasks, thereby continuously enhancing the adaptability of the network.
[0061] According to the present invention, the training data includes the flight state quantities and guidance laws of the aircraft, wherein the flight state quantities are the input quantities of the neural network and the bias acceleration of the aircraft is the output quantity of the neural network.
[0062] Since the flight state variables and the guidance law of the aircraft are highly nonlinear, directly mapping the flight state variables to the optimal guidance law will lead to a significant decrease in prediction performance when the state deviation is small. In order to improve the robustness of guidance, this invention adopts the bias proportional guidance law and uses the aircraft bias acceleration as the output of the neural network.
[0063] Furthermore, based on the aircraft offset, the final navigation command is obtained, expressed as:
[0064]
[0065] Where N is the proportionality coefficient and v is the missile velocity. Let a be the rate of change of the line-of-sight angle between the projectile and the target. b This refers to the aircraft's bias acceleration.
[0066] Furthermore, the flight state quantities include the initial position of the aircraft, the speed of the aircraft, the trajectory tilt angle of the aircraft, the line-of-sight angle of the aircraft, the initial trajectory tilt angle, and the target position.
[0067] In a preferred embodiment, the training data for the original task includes the following tasks:
[0068] Overload-constrained mission: Under limited angle of attack, the aircraft reaches the target position at maximum terminal velocity;
[0069] This task is the most common and typical.
[0070] More preferably, the training data for the original task also includes one or more of the following tasks:
[0071] Overload and landing angle constraint mission: Under the constraints of angle of attack and landing, the aircraft reaches the target position at the maximum terminal velocity.
[0072] Overload and time-constrained missions require the aircraft to reach the target position at maximum terminal velocity under conditions of limited angle of attack and flight time.
[0073] Overload, angle of attack, and flight time constraints: Under conditions of limited angle of attack, angle of attack, and flight time, the aircraft reaches the target position at its maximum terminal velocity.
[0074] In a preferred embodiment, the training data is obtained in the following manner:
[0075] Establish a dynamic model of the aircraft;
[0076] By setting the optimal control objective, multiple sets of aircraft flight state variables and corresponding aircraft bias accelerations are obtained through optimal problem-solving methods.
[0077] In a preferred embodiment, the dynamic model is represented as follows:
[0078]
[0079] Where (x,y) represents the position of the aircraft in the plane, and V is the velocity of the aircraft. Let L be the trajectory inclination angle, P be the aircraft lift, α be the aircraft thrust, m be the aircraft mass, g be the gravitational acceleration, and D be the aircraft drag. sp For propellant specific impulse.
[0080] Wherein, the gravitational acceleration g is:
[0081]
[0082] G is the Earth's gravitational constant, M is the Earth's mass, and R is the Earth's gravitational constant. g Where is the Earth's radius, and h is the aircraft's flight altitude.
[0083] The lift L of the aircraft is:
[0084] The aircraft drag D is:
[0085] S represents the characteristic area of the aircraft, and q represents the dynamic pressure of the aircraft. ρ is the air density. C is the derivative of the lift coefficient with respect to the angle of attack. D0 Zero-lift drag coefficient, This is the induced drag coefficient. All values are functions of altitude or Mach number; specific values can be obtained by interpolation after consulting tables.
[0086] In a preferred embodiment, the optimal control objective in the overload constraint task is set as follows:
[0087]
[0088]
[0089] x(t f )=x T ,y(t f )=y T
[0090] |α|<α max
[0091] Among them, V f α represents the speed at which the aircraft reaches the target position. max x(t) is the upper bound of the absolute value of the angle of attack α. f ),y(t f x represents the desired position of the aircraft. T ,y T The location of the target.
[0092] In a preferred embodiment, the optimal control objective in the overload and landing constraint tasks is set as follows:
[0093]
[0094]
[0095] x(t f )=x T ,y(t f )=y T ,θ(t f )=θ f
[0096] |α|<α max
[0097] Among them, V f α represents the speed at which the aircraft reaches the target position. max x(t) is the upper bound of the absolute value of the angle of attack α. f ),y(tf x represents the desired position of the aircraft. T ,y T Let θ(t) be the position of the target. f θ is the desired landing angle of the aircraft. f This refers to the terminal trajectory inclination angle of the aircraft.
[0098] In a preferred embodiment, the optimal control objective in the overload and time-of-flight constrained task is set as follows:
[0099]
[0100]
[0101] x(t f )=x T ,y(t f )=y T ,t f =t d
[0102] |α|<α max
[0103] Among them, V f α represents the speed at which the aircraft reaches the target position. max x(t) is the upper bound of the absolute value of the angle of attack α. f ),y(t f x represents the desired position of the aircraft. T ,y T For the position of the target, t f t is the expected flight time of the aircraft. d This represents the expected attack time for the aircraft.
[0104] In a preferred embodiment, the optimal control objective in the overload, landing, and time-of-flight constraints is set as follows:
[0105]
[0106]
[0107] x(t f )=x T ,y(t f )=y T ,t f =t d
[0108] |α|<α max
[0109] Among them, V f α represents the speed at which the aircraft reaches the target position.max x(t) is the upper bound of the absolute value of the angle of attack α. f ),y(t f x represents the desired position of the aircraft. T ,y T For the position of the target, t f t is the expected flight time of the aircraft. d This represents the expected attack time for the aircraft.
[0110] In this invention, no specific method is limited for solving the optimal problem. Those skilled in the art can freely choose based on experience, such as using the General-Purpose Pseudospectral Optimal Control Software (GPOPS-Ⅱ) method to solve the bias acceleration.
[0111] Example
[0112] Example 1
[0113] Experiments were conducted, with three original tasks: Task 1 "Overload Constraint Task", Task 2 "Overload and Landing Angle Constraint Task", and Task 3 "Overload and Flight Time Constraint Task"; and one new task, Task 4 "Overload, Landing Angle, and Flight Time Constraint Task", was set up. Computational guidance based on a gated progressive neural network was performed, including the following steps:
[0114] Using the training data from the original mission, a gated progressive neural network is pre-trained to obtain an initial network. The initial network outputs the bias acceleration of the aircraft, and the guidance law of the aircraft is obtained based on the bias acceleration.
[0115] When the guidance mission changes, the initial network is retrained using the training data of the new mission to obtain an updated network. The updated network is then used to output the aircraft bias acceleration, and the guidance law of the aircraft is obtained based on the bias acceleration.
[0116] The gated progressive neural network consists of three columns of task units, each of which has an input layer, two hidden layers, and an output layer.
[0117] The retraining includes the following steps: adding new column task units to the initial network, establishing lateral connections between the new column task units and the old column task units, passing gating data between the lateral connections to form a new progressive neural network; training the new progressive neural network using training data of the new task to obtain an updated network.
[0118] In the new progressive neural network, the input to the new column task unit is represented as:
[0119]
[0120] The weight parameters are expressed as
[0121] Training data was obtained in the following ways:
[0122] Establish a dynamic model of the aircraft;
[0123] By setting the optimal control objective, multiple sets of aircraft flight state variables and corresponding aircraft bias accelerations are obtained through optimal problem-solving methods.
[0124] The loss function for the new column is expressed as:
[0125]
[0126] The dynamic model is expressed as follows:
[0127]
[0128]
[0129]
[0130]
[0131] In overload constraint tasks, the optimal control objective is set as follows:
[0132]
[0133] x(t f )=x T ,y(t f )=y T
[0134] |α|<α max
[0135] In overload and drop angle constraint tasks, the optimal control objective is set as follows:
[0136]
[0137]
[0138] x(t f )=x T ,y(t f )=y T ,θ(t f )=θ f
[0139] |α|<α max
[0140] In overload and time-constrained missions, the optimal control objective is set as follows:
[0141]
[0142]
[0143] x(t f )=x T ,y(t f )=y T ,t f =t d
[0144] |α|<α max
[0145] The optimal control objective in overload, angle of fall, and time of flight constraints is set as follows:
[0146]
[0147]
[0148] x(t f )=x T ,y(t f )=y T ,t f =t d
[0149] |α|<α max
[0150] The size of the obtained training data is shown in Table 1.
[0151] Table 1
[0152] Number of data entries Task 1 500 Task 2 500 Task 3 500 Task 4 12
[0153] Run 500 Monte Carlo simulations. Specifically, within the initial conditions range in Table 2, randomly select the flight state variables of the aircraft, use the update network to obtain the bias acceleration of the aircraft, and obtain the guidance law of the aircraft based on the bias acceleration.
[0154] Table 2
[0155]
[0156] Comparative Example 1
[0157] Using the training data obtained in Example 1, the same simulation experiment as in Example 1 was conducted using a traditional progressive neural network (PNN).
[0158] Experimental Example
[0159] Comparing the terminal errors in Example 1 and Comparative Example 1, the results are as follows: Figure 4 As shown in the figure, ΔX is the horizontal miss distance; ΔY is the plumb bob miss distance; ΔV is the velocity error; Δθ is the attack angle error; and Δt is the time-of-flight error. It can be seen from the figure that Example 1 and Comparative Example 1 show almost identical performance in terms of terminal error, both achieving effective training even with only a small amount of data in Task 4, demonstrating the effectiveness of the method in Example 1 in knowledge transfer. However, as the number of tasks increases, the structure of Comparative Example 1 becomes overly complex, leading to significant time consumption for training new columns. The training time of Example 1 and Comparative Example 1 is compared, and the results are shown in Table 3.
[0160] Table 3
[0161] Task 1 Task 2 Task 3 Task 4 Example 1 922.74s 948.47s 1016.12s 11.94s Comparative Example 1 665.81s 1146.29s 1738.85s 33.58s
[0162] As can be seen from Table 3, the method in Example 1 effectively avoids the phenomenon of network parameter inflation, thereby enabling rapid adaptation to new tasks.
[0163] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front," and "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship in the working state of this invention, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0164] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0165] The present invention has been described above with reference to preferred embodiments; however, these embodiments are merely exemplary and illustrative. Various substitutions and modifications can be made to the present invention based on these embodiments, all of which fall within the scope of protection of the present invention.
Claims
1. A computational guidance method based on a gated progressive neural network, characterized in that, Includes the following steps: Using the training data from the original mission, a gated progressive neural network is pre-trained to obtain an initial network. The initial network outputs the bias acceleration of the aircraft, and the guidance law of the aircraft is obtained based on the bias acceleration. When the guidance mission changes, the initial network is retrained using the training data of the new mission to obtain an updated network. The updated network is used to output the bias acceleration of the aircraft, and the guidance law of the aircraft is obtained based on the bias acceleration. The gated progressive neural network includes one or more columns of task units, each column of task units having an input layer, one or more hidden layers, and an output layer; there are lateral connections between different columns of task units; The retraining includes the following steps: Add new column task units to the initial network and establish lateral connections between the new column task units and the old column task units to form a new progressive neural network; A new progressive neural network is trained using training data from the new task to obtain an updated network; The input to the new column task unit input layer includes neural network input data and the output of all old column task unit output layers. In the input of the new column task unit input layer, a weight parameter is set for the output of the old column task unit output layer; The weight parameters are adaptive parameters.
2. The computational guidance method based on a gated progressive neural network according to claim 1, characterized in that, During retraining, the parameters in the old task units are frozen.
3. The computational guidance method based on a gated progressive neural network according to claim 1, characterized in that, The training data includes the aircraft's flight state variables and guidance laws, where the flight state variables are the input to the neural network and the aircraft's bias acceleration is the output of the neural network.
4. The computational guidance method based on a gated progressive neural network according to claim 1, characterized in that, During retraining, mean squared error is used as the loss function for training new columns.
5. The computational guidance method based on a gated progressive neural network according to claim 1, characterized in that, The training data was obtained through the following methods: Establish a dynamic model of the aircraft; Set the optimal control objective and obtain multiple sets of aircraft flight state variables and corresponding aircraft bias accelerations based on the dynamic model.