Knowledge enhancement improvement-based BP neural network end guidance method

Abstract

The invention relates to a BP neural network end guidance method based on knowledge enhancement improvement, and belongs to the field of end guidance. A proportion plus navigation (PPN) guidance law is combined, and the guidance precision is improved. According to the method, through a knowledge-enhanced BP neural network, the future position and speed of a target are predicted by using the historical motion track of the target, and the flight path of an aircraft is adjusted. Specifically, the nonlinear feature learning ability of the BP network is improved by using knowledge enhancement, and the tracking and intercepting ability of the rapid maneuvering target is effectively improved by combining the proportion of the PPN guidance law and the navigation mechanism. By innovatively combining prior information of knowledge enhancement and the self-learning ability of the BP neural network, the method shows higher adaptability at a guidance terminal, and the hit precision of a guidance system can be remarkably improved.

Description

Technical Field

[0001] The invention belongs to the field of terminal guidance, and in particular relates to a BP neural network terminal guidance method based on knowledge enhancement improvement. Background Art

[0002] Terminal guidance is a technology that ensures the accuracy of the strike by correcting the course or speed of the aircraft in the final stage of flight. The terminal guidance stage is extremely short, and the traditional optimal guidance law requires a high amount of calculation and is difficult to meet the real-time requirements of the terminal guidance stage; and in complex environments, traditional guidance methods (such as proportional navigation and optimal guidance law) are difficult to adapt to dynamic and highly uncertain target maneuvers.

[0003] In response to the development trend of high speed and high maneuverability of space interception targets, the industry has proposed a terminal guidance method based on neural networks. On the basis of the traditional guidance law, a nonlinear relationship is established through the neural network to adaptively adjust the parameters of the guidance law. The BP network can generate guidance instructions in a very short time through the fast mapping relationship obtained, so as to improve the performance of the traditional guidance law under the conditions of overload control, line of sight angle dependence, and landing angle constraint.

[0004] Although the terminal guidance method based on neural network has many advantages, it also has the following defects:

[0005] 1) Reliance on a large amount of training data, difficulty in obtaining training data, prone to errors in data labeling, insufficient data diversity, insufficient accuracy, and long data training time.

[0006] 2) The model structure is complex, the computing resource requirements are high, and real-time performance is difficult to guarantee.

[0007] 3) Poor anti-interference ability, susceptibility to adversarial attacks, limited environmental adaptability, and poor robustness. Summary of the invention

[0008] The purpose of the present invention is to provide a terminal guidance method based on knowledge enhancement and improvement of BP neural network, which introduces knowledge enhancement and improvement of BP neural network, which uses the prior information provided by knowledge enhancement, provides more meaningful initial features through knowledge embedding, reduces the model's dependence on a large amount of training data, makes the guidance path optimization process more efficient, reduces the calculation time, and dynamically adjusts the guidance path by introducing historical error distribution and external environment knowledge. The BP network generates the optimal acceleration and attitude adjustment instructions in real time, improves the hit accuracy, and the new method can effectively enhance the guidance accuracy and robustness. It solves the technical problems in the prior art that the terminal guidance relies on a large amount of training data, the data training time is long, the accuracy is insufficient, and the robustness is poor.

[0009] In order to solve the above technical problems, the present invention provides a BP neural network terminal guidance method based on knowledge enhancement improvement.

[0010] The following steps are involved:

[0011] Step S1: construct a three-dimensional PPN guidance law aircraft and target interception model, in which the aircraft trajectory and the target trajectory finally intersect at a point in the three-dimensional space;

[0012] Step S2: construct BP neural network;

[0013] Step S3: Determine the error function and loss function of the BP neural network;

[0014] Step S4: using knowledge enhancement to improve the BP neural network, and obtaining a knowledge enhanced BP neural network model;

[0015] Step S5: using RAdam optimizer to adjust the learning rate of the knowledge-enhanced BP neural network and stabilize the training process of the knowledge-enhanced BP neural network;

[0016] Step S6: training the knowledge-enhanced BP neural network model, and using the trained knowledge-enhanced BP neural network model to verify the terminal guidance.

[0017] Furthermore, in step S1, a three-dimensional PPN guidance law aircraft and target interception model is constructed, the aircraft trajectory and the target trajectory are in the same plane, and the velocity direction of the aircraft and the velocity direction of the target have an intersection in this plane; according to the guidance relationship of the PNN guidance law, an iterative difference equation group of the aircraft and target positions is derived, and a data set is obtained using the equation group.

[0018] Furthermore, in step S2, the BP neural network is constructed by connecting an input layer, a first hidden layer, a second hidden layer and an output layer in series. The input data is input into the BP neural network from the input layer, passes through the first hidden layer and the second hidden layer, and is output from the output layer.

[0019] Furthermore, in step S3, the BP neural network error function for:

[0020]

[0021] in, Indicates the set speed of the aircraft. Indicates the speed of BP neural network output; , , Indicates the velocity components of the aircraft's set velocity along the X-axis, Y-axis, and Z-axis; , , Represents the velocity components of the BP neural network output along the X-axis, Y-axis, and Z-axis;

[0022] BP neural network loss function for:

[0023] .

[0024] Furthermore, in step S4, the knowledge enhancement constraint aircraft terminal guidance loss function is constrained from four aspects: aircraft acceleration and speed constraints, aircraft and target relative position constraints, flight area constraints and target maneuverability constraints; the BP neural network loss function is combined with the aircraft acceleration and speed constraints. , relative position constraints between the aircraft and the target , Flight area constraints and target mobility constraint combination , forming a knowledge-enhanced constraint loss function :

[0025]

[0026] in, is a hyperparameter that controls the weight of each constraint term in the total loss function.

[0027] Furthermore, the update rule of the RAdam optimizer is as follows:

[0028]

[0029] in: represents the parameter value to be optimized of the knowledge-enhanced BP neural network at time t+1, represents the parameter value to be optimized of the knowledge-enhanced BP neural network at time t, is the learning rate, is the modified first-order moment estimate, is the modified second-order moment estimate, is a correction factor to smooth gradient updates.

[0030] Compared with the prior art, the present invention has the following beneficial technical effects:

[0031] (1) The present invention uses knowledge enhancement to improve the nonlinear feature learning ability of the BP network, and combines the ratio of the PPN guidance law and the navigation mechanism to effectively improve the tracking and interception capabilities of fast maneuvering targets. By innovatively combining the prior information of knowledge enhancement with the self-learning ability of the BP neural network, the present invention exhibits stronger adaptability at the guidance end, and can significantly improve the hit accuracy of the guidance system.

[0032] (2) The fast mapping relationship obtained by training the BP network of the present invention can generate guidance instructions in a very short time. The BP network uses the prior information provided by knowledge enhancement to make the guidance path optimization process more efficient and reduce the calculation time.

[0033] (3) The present invention uses the RAdam optimizer to adjust the learning rate and correct the instability in small batch training, making the gradient update more stable and improving the hit accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0035] Figure 1 It is a schematic diagram of a three-dimensional flight trajectory model established based on the PNN guidance law of the present invention.

[0036] Figure 2 This is a schematic diagram of the knowledge-enhanced improved BP neural network structure of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] The present invention proposes a knowledge-enhanced improved BP neural network terminal guidance method, based on a general three-dimensional PPN (Proportional Navigation Plus Bias) guidance law, comprising the following steps:

[0039] Step S1: construct a three-dimensional PPN guidance law aircraft and target interception model, in which the aircraft trajectory and the target trajectory finally intersect at a point in the three-dimensional space.

[0040] The three-dimensional PPN guidance law aircraft and target interception model constructed by the present invention is as follows: Figure 1 As shown in the figure, the aircraft trajectory and the target trajectory are in the same plane, and the velocity direction of the aircraft and the velocity direction of the target have an intersection in this plane. A in the figure represents the intersection point of the aircraft trajectory and the target track. The coordinates of the intersection point A are expressed as express; , Indicates that the aircraft is Moment and Coordinates of time, aircraft Time coordinates Indicates aircraft Time coordinates express; , Indicates the target Moment and Coordinates of time, target Time coordinates Representation, goal Time coordinates express.

[0041] The flight distance of the aircraft in a time interval is :

[0042]

[0043] in, Indicates the aircraft speed, Indicates a time interval.

[0044] The flight distance of the target in a time interval is :

[0045]

[0046] in, Indicates the target speed.

[0047] According to the guidance relationship of the PNN guidance law, the iterative difference equations of the aircraft and target positions are derived, and the data set is obtained using the equations:

[0048]

[0049] in, express The target sight rotation angle at any moment, express The target sight rotation angle at any moment, It represents the rotation angle increment per unit time. Indicates aircraft The rotation angle at the moment, Indicates aircraft The rotation angle at the moment, represents the proportionality coefficient, express The angle between the trajectory direction of the aircraft and the direction of the line connecting the aircraft and the target at the moment, , , express The target obtained by updating the iterative equations at any time is The coordinate values ​​along the X-axis, Y-axis and Z-axis in three-dimensional space at any moment. , , express The coordinate values ​​of the aircraft along the X-axis, Y-axis and Z-axis in three-dimensional space at this moment. , , express The coordinate values ​​of the aircraft along the X-axis, Y-axis and Z-axis in three-dimensional space at this moment. , , express The coordinate values ​​of the target along the X-axis, Y-axis and Z-axis in three-dimensional space at the moment. , , express The coordinate values ​​of the target along the X-axis, Y-axis and Z-axis in three-dimensional space at the moment. From intersection point A to aircraft Time location The distance From intersection A to target Time location distance.

[0050] When the aircraft and the target are Relative distance between moments Less than threshold When The angle between the trajectory direction of the aircraft and the direction of the line connecting the aircraft and the target at the moment Approximate rotation angle increment per unit time The error is relatively large, so the relative distance between the aircraft and the target needs to be corrected using the following formula:

[0051]

[0052] in, Indicates the calculated theoretical distance between the aircraft and the target. express The angle between the trajectory direction of the aircraft and the direction of the line connecting the aircraft and the target at the moment, express The angle between the target trajectory direction and the direction of the line connecting the target and the aircraft at the moment, the threshold It can be set reasonably according to needs.

[0053] After the above iterative difference equations, we get Time data ,in The coordinates of the target at time K are derived based on the actual motion trajectory of the target. Multiple sets of simulation results at different times are used express, The dataset that serves as input to the neural network.

[0054] Step S2: Construct BP neural network.

[0055] Specifically, Figure 2 As shown, the BP neural network is composed of an input layer, a first hidden layer, a second hidden layer and an output layer connected in series. The input data is input into the BP neural network from the input layer, passes through the first hidden layer and the second hidden layer, and is output from the output layer.

[0056] The input layer contains 6 neurons ,in, Indicates neurons, j ∈[ 1,6] The first neuron Corresponding to the data set , the first neuron Corresponding to the data set , the second neuron Corresponding to the data set , the third neuron Corresponding to the data set , the 4th neuron Corresponding to the data set , the 5th neuron Corresponding to the data set , the sixth neuron Corresponding to the data set The outputs of the input layer neurons are fed into the first hidden layer.

[0057] The first hidden layer contains 30 neurons, and the activation function of each neuron is a tan-sigmoid function. neurons, i ∈[ 1,30] , whose output The calculation process can be expressed as:

[0058]

[0059] in, is the first hidden layer The output of a neuron, It is the first of the six neurons in the input layer. neurons and the first hidden layer of 30 neurons The weight of a neuron. It is the first of the six neurons in the input layer. The output of the neuron (i.e., the first Quantity). It is the first hidden layer of 30 neurons. The bias term of a neuron.

[0060] It is the tan-sigmoid activation function, that is, the hyperbolic tangent sigmoid function, expressed as follows:

[0061]

[0062] or

[0063]

[0064] in, is the base of natural logarithms, is the input value, is the hyperbolic tangent function. The output range of this function is between [-1,1].

[0065] Calculate the output of each neuron in the first hidden layer, and finally get the output vector of the first hidden layer The output vector of the first hidden layer will be used as the input of the second hidden layer.

[0066] The second hidden layer contains 20 neurons, and the activation function of each neuron is a tan-sigmoid function. neurons, whose output The calculation process can be expressed as:

[0067]

[0068] in, is the second hidden layer The output of a neuron. It is the first hidden layer of 30 neurons connected neurons and the second hidden layer of 20 neurons The weight of a neuron. It is the first of the 30 neurons in the first hidden layer. The output of a neuron. It is the 20 neurons in the second hidden layer. The bias term of a neuron.

[0069] Calculate the output of each neuron in the second hidden layer, and finally get the output vector of the second hidden layer The output vector of the second hidden layer serves as the input of the output layer.

[0070] The output layer contains 3 neurons, and the activation function of each neuron is linear (for regression problems). neurons, whose output The calculation process can be expressed as:

[0071]

[0072] in, It is the first of the three neurons in the output layer. The output of a neuron. It is the connection of the 20 neurons in the second hidden layer. The neurons and the output layer's 3 neurons The weight of a neuron. It is the 20 neurons in the second hidden layer. The output of a neuron It is the first of the three neurons in the output layer. The bias term of a neuron.

[0073] is the 𝑝𝑢𝑟𝑒l𝑖𝑛 activation function, expressed as follows:

[0074]

[0075] Calculate the output of each neuron in the output layer, and finally get the output vector of the output layer Among them, the output Corresponds to the velocity component of the output along the X-axis , output Corresponds to the velocity component of the output along the Y axis , output Corresponds to the velocity component of the output along the Z axis . This output vector is the final output of the BP neural network.

[0076] Step S3: Determine the error function and loss function of the BP neural network:

[0077] BP neural network error function for:

[0078]

[0079] in, Indicates the set speed of the aircraft. Indicates the speed of BP neural network output; , , Indicates the velocity components of the aircraft's set velocity along the X-axis, Y-axis, and Z-axis; , , Represents the velocity components of the BP neural network output along the X-axis, Y-axis, and Z-axis.

[0080] BP neural network loss function for:

[0081]

[0082] Step S4: Use knowledge enhancement to improve the BP neural network and obtain a knowledge enhanced BP neural network model.

[0083] Knowledge enhancement is used to constrain the loss function of the terminal guidance of the aircraft to ensure the trajectory accuracy of the aircraft and meet the physical limitations of the aircraft itself and the mission requirements. Knowledge enhancement constrains the loss function of the terminal guidance of the aircraft from four aspects: the acceleration and speed constraints of the aircraft, the relative position constraints between the aircraft and the target, the flight area constraints, and the target maneuverability constraints.

[0084] The acceleration and speed constraints of the aircraft can ensure that the maneuverability of the aircraft does not exceed the physical limit. The acceleration and speed constraints of the aircraft are calculated by the following formula:

[0085]

[0086] in, represents the acceleration and velocity constraints of the aircraft, represents the maximum value function, Indicates the minimum safe distance between the aircraft and the target. Indicates the absolute value operation. Indicates the distance between the current aircraft and the target.

[0087] The relative position constraint between the aircraft and the target can ensure that the aircraft can accurately approach the target within a reasonable time. The relative position constraint between the aircraft and the target is calculated by the following formula:

[0088]

[0089] in, Represents the relative position constraint between the aircraft and the target, represents the target position predicted based on the target trajectory and target position, Indicates the current position of the aircraft.

[0090] The flight area constraint can ensure that the aircraft trajectory is within the allowed flight area. The flight area constraint is calculated by the following formula:

[0091]

[0092] in, represents the flight area constraint, Indicates the distance between the current position of the aircraft and the center of the flight area. Indicates the maximum radius or boundary of the flight area.

[0093] The target maneuverability constraint takes into account the maneuverability of the guided target and optimizes the flight trajectory of the aircraft. The target maneuverability constraint is calculated by the following formula:

[0094]

[0095] in, represents the target mobility constraint, represents the speed of the target, Indicates the maximum radius or boundary of the flight area.

[0096] The BP neural network loss function is combined with the acceleration and speed constraints of the aircraft, the relative position constraints between the aircraft and the target, the flight area constraints and the target maneuverability constraints to form a knowledge-enhanced constraint loss function. :

[0097]

[0098] in, is a hyperparameter that controls the weight of each constraint term in the total loss function.

[0099] Step S5: Use RAdam optimizer to adjust the learning rate of the knowledge-enhanced BP neural network and stabilize the training process of the knowledge-enhanced BP neural network.

[0100] In terminal guidance, especially when the target is maneuvering rapidly, the training of BP neural network is often troubled by gradient explosion or gradient disappearance. The present invention uses RAdam optimizer to help adjust the learning rate of each parameter of knowledge-enhanced BP neural network, correct the instability in small batch training, make the gradient update more stable, and solve the problem of unstable gradient update. During the training process, RAdam optimizer will continuously adjust the learning rate, so as to maintain a stable training process with few samples.

[0101] The update rule of the RAdam optimizer is as follows:

[0102]

[0103] in: represents the parameter value to be optimized of the knowledge-enhanced BP neural network at time t+1, represents the parameter value to be optimized of the knowledge-enhanced BP neural network at time t, is the learning rate, is the modified first-order moment estimate, is the modified second-order moment estimate, is a correction factor to smooth gradient updates.

[0104] The final knowledge-enhanced neural network gradient update step is:

[0105] (1) Calculate the current gradient and use the gradient descent method to minimize the loss function by updating the parameters along the negative gradient direction.

[0106] (2) Use the update formula of the RAdam optimizer to adjust the gradient.

[0107] (3) Use the adjusted gradient to update the network parameters so that the loss function (including prediction error and constraint terms) is minimized to ensure that the trajectory of the aircraft meets the predetermined target.

[0108] Step S6: training the knowledge-enhanced BP neural network model, and using the trained knowledge-enhanced BP neural network model to verify the terminal guidance.

[0109] Initialize the weights of the input layer, the first hidden layer, and the second hidden layer through uniform distribution.

[0110] Knowledge-enhanced BP neural network model based on data set is the input and the output is , , , the value calculated in step S2 is As a data set, 80% of the data is used as a training set to train the knowledge-enhanced BP neural network model, and 20% of the data is used as a test set to verify the trained knowledge-enhanced BP neural network model terminal guidance. If the model effect needs to be verified (such as hyperparameter tuning), 10% to 15% can be divided from the training set for the verification set.

[0111] By validating the dataset metrics, it is possible to improve the accuracy of flight trajectory prediction using a smaller sample size and achieve precise control signal prediction.

[0112] The knowledge-enhanced BP neural network of the present invention can generate guidance instructions in a very short time through the fast mapping relationship obtained through training. The BP network uses the prior information provided by knowledge enhancement to make the guidance path optimization process more efficient, reduce the calculation time and realize accurate control signal prediction.

[0113] The present invention uses RAdam optimizer to adjust the learning rate, correct the instability in small batch training, make the gradient update more stable, and improve the hit accuracy.

[0114] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A BP neural network terminal guidance method based on knowledge enhancement improvement, characterized in that: The method comprises the following steps: Step S1: construct a three-dimensional PPN guidance law aircraft and target interception model, in which the aircraft trajectory and the target trajectory finally intersect at a point in the three-dimensional space; Step S2: construct BP neural network; Step S3: Determine the error function and loss function of the BP neural network; Step S4: using knowledge enhancement to improve the BP neural network, and obtaining a knowledge enhanced BP neural network model; Step S5: using RAdam optimizer to adjust the learning rate of the knowledge-enhanced BP neural network and stabilize the training process of the knowledge-enhanced BP neural network; Step S6: training the knowledge-enhanced BP neural network model, and using the trained knowledge-enhanced BP neural network model to verify the terminal guidance.

2. The BP neural network terminal guidance method based on knowledge enhancement improvement as claimed in claim 1 is characterized in that: In step S1, a three-dimensional PPN guidance law aircraft and target interception model is constructed, the aircraft trajectory and the target trajectory are in the same plane, and the speed direction of the aircraft and the speed direction of the target have an intersection in this plane; According to the guidance relationship of the PNN guidance law, the iterative difference equations of the aircraft and target positions are derived, and the data set is obtained using the equations.

3. The BP neural network terminal guidance method based on knowledge enhancement improvement as claimed in claim 1 is characterized in that: In step S2, the BP neural network is constructed by connecting an input layer, a first hidden layer, a second hidden layer and an output layer in series. The input data is input into the BP neural network from the input layer, and is output from the output layer after passing through the first hidden layer and the second hidden layer.

4. The BP neural network terminal guidance method based on knowledge enhancement improvement as claimed in claim 1 is characterized in that: In step S3, the BP neural network error function for: in, Indicates the set speed of the aircraft. Indicates the speed of BP neural network output; , , Indicates the velocity components of the aircraft's set velocity along the X-axis, Y-axis, and Z-axis; , , Represents the velocity components of the BP neural network output along the X-axis, Y-axis, and Z-axis; BP neural network loss function for: 。 5. The BP neural network terminal guidance method based on knowledge enhancement improvement as claimed in claim 4 is characterized in that: In step S4, the knowledge enhancement constrains the loss function of the terminal guidance of the aircraft, and constrains it from four aspects: the acceleration and speed constraints of the aircraft, the relative position constraints between the aircraft and the target, the flight area constraints, and the target maneuverability constraints; the BP neural network loss function is combined with the acceleration and speed constraints of the aircraft , relative position constraints between the aircraft and the target , Flight area constraints and target mobility constraint combination , forming a knowledge-enhanced constraint loss function : in, is a hyperparameter that controls the weight of each constraint term in the total loss function.

6. The BP neural network terminal guidance method based on knowledge enhancement improvement as claimed in claim 1 is characterized in that: The update rule of the RAdam optimizer is as follows: in: express The parameter values ​​to be optimized of the BP neural network enhanced by knowledge at all times, express The parameter values ​​to be optimized of the BP neural network enhanced by knowledge at all times, is the learning rate, is the modified first-order moment estimate, is the modified second-order moment estimate, is a correction factor to smooth the gradient update.

Images