Guided shell high-altitude wind speed and direction estimation method based on GA-BP neural network
Patent Information
- Application Number
- CN202310847494.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-11
- Publication Date
- 2026-08-21
- Estimated Expiration
- 2043-07-11
AI Technical Summary
但是目前，尚未有能够同时满足战场快速性要求、抑制控制中的风力系数扰动、提高系统的鲁棒性的高空气象测量方法
[0082](1)将制导炮弹的动力学模型(解析方法)与人工神经网络(黑箱建模方法)相结合，形成灰箱模型；使用遗传算法对弹道数据进行初始处理，使他可以避开峰值陷阱更好的追踪实际风场，对于使用神经网络对实际风速采集达到更加准确效果。
Smart Images
Figure CN116956717B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of rapid high-altitude wind speed estimation technology for guided artillery shells, specifically involving a method for estimating high-altitude wind speed and direction of guided artillery shells based on a GA-BP neural network. Background Technology
[0002] Guided missiles possess the capability to strike enemy targets with high precision, making them a current hot research area in missile technology. However, these weapon systems are also highly complex, expensive, and suffer from low rate of fire and high operational costs. With the increasing variety and quantity of guided missiles, how to efficiently utilize them on the battlefield has become an urgent research topic. With the development of computer simulation technology, during the flight of guided projectiles, it is necessary to collect, logically analyze, calculate, and control their state and environmental characteristics (such as flight trajectory parameters and targets). The flight trajectory can be adjusted by modifying the projectile's aerodynamic characteristics. Simulation modeling of the control modules (satellites, inertial navigation, servos, seekers, etc.) of guided missiles, considering factors such as weather distribution characteristics and control device error characteristics, will provide a more accurate theoretical basis for the target impact point distribution in the operational use of guided missiles. As high-altitude flight weapons, wind speed and direction at high altitudes have a significant impact on the flight position and attitude information of guided missiles and rockets, which in turn significantly affects the final guidance accuracy. How to quickly measure meteorological data and make corrections and controls on the guidance process of guided missiles and rockets is currently the most pressing issue.
[0003] However, in actual battlefield combat, it is difficult to quickly collect external meteorological data that significantly affects the impact point of guided missiles and rockets. Conventional upper-air meteorological detection methods involve using weather balloons carrying electronic equipment to measure upper-air wind, temperature, and air pressure, ascending at a rate of approximately 4–6 meters per second to detect upper-air conditions from the bottom up. [1] When encountering convective weather, balloons are affected by downdrafts and gradually descend from high altitudes, thus increasing detection time. Guided missiles, on the other hand, reach maximum altitudes ranging from 15km to 30km during their flight phase, with some even reaching higher. While this prolonged detection method is accurate, it does not meet the speed requirements of actual combat.
[0004] Another method involves using meteorological rockets to detect upper-level winds. In this method, the radiosonde separates from the rocket during the ascent phase and continues its inertial motion to the apex. During descent, the parachute gradually opens, and after reaching a stable state, it falls under the traction of the radiosonde and drifts with the wind. The parachute becomes a tracer of the upper-level wind field. Based on the positional changes recorded by the radiosonde, the atmospheric wind field in the near-space region (20–60 km) can be retrieved. [2]This method is faster but less accurate than wind field calculations using weather balloons. However, it relies on parachute descent to infer data and requires waiting for the parachute to fall and recovering the recorder, which is not fast enough for real-time battlefield operations.
[0005] Real-time performance and rapid responsiveness are key dynamic performance characteristics of guided missiles and rockets. Improving the real-time performance and rapid response of guided missiles and rockets is an effective countermeasure against incoming, highly maneuverable targets. For aircraft guidance and control systems, rapidly acquiring external environmental information to correct control parameters and improve accuracy is paramount. However, currently, there is no high-altitude meteorological measurement method that can simultaneously meet the requirements of battlefield speed, suppress wind coefficient disturbances in control, and improve system robustness. Summary of the Invention
[0006] The purpose of this invention is to provide a rapid high-altitude meteorological detection method that can meet the requirements of battlefield speed, suppress high-altitude wind speed and direction disturbances in control, and improve the robustness of guidance systems.
[0007] The technical solution to achieve the purpose of this invention is: a method for estimating high-altitude wind speed and direction of guided projectiles based on GA-BP neural networks, comprising the following steps:
[0008] Step (1): Conduct actual flight marking experiments, collect the flight attitude angle, velocity, acceleration and position of the gliding extended range guided projectile at each sampling moment in the experimental environment, and record the mass of the guided projectile. Obtain wind speed and wind direction information at different altitudes by releasing weather balloons.
[0009] Step (2): Establish ballistic motion model and aerodynamic model;
[0010] Step (3): Obtain motion parameters at each sampling moment using the onboard sensors and calculate theoretical ballistic data;
[0011] Step (4): Build a backpropagation neural network, take the wind speed and direction at each sampling height in the ballistic coordinate system as the output, take the flight attitude angle, flight speed, acceleration and theoretical ballistic data collected at each sampling point of the guided projectile as the input, then use the genetic algorithm to preprocess and optimize the data, train the backpropagation neural network model, and save the training model after the neural network model is trained.
[0012] Step (5): At the location where measurements are required, launch an identical glide-extended-range guided projectile and record the flight attitude angle, velocity, acceleration, position, and thrust information at each sampling moment;
[0013] Step (6): Use the launch velocity of the guided projectile, the projectile's flight torque, and the flight attitude angle, flight speed, and acceleration information collected at each sampling point as input to the trained artificial neural network model to calculate the output wind speed vector.
[0014] Furthermore, step (2) establishes the motion model in the ballistic ground coordinate system as follows:
[0015]
[0016]
[0017]
[0018]
[0019]
[0020]
[0021]
[0022]
[0023]
[0024] The direct formula for calculating the dynamic equilibrium angle is:
[0025]
[0026]
[0027]
[0028]
[0029] The parameters in the direct calculation formula are:
[0030]
[0031] x2=(a a -a b )b y V r 2
[0032]
[0033]
[0034]
[0035]
[0036] V rx =V x -W x
[0037] V ry =V y
[0038] V rz =V z -W z
[0039] In the formula, V x V is the velocity along the horizontal axis of the gliding missile. y V is the velocity along the longitudinal axis of the glider. z Let m be the vertical velocity of the glider, t be the mass of the glider, ρ be the air density, W be the wind speed, and V be the velocity along the vertical axis. r The x, y, and z values represent the actual velocity, respectively, and the C value represents the displacement in each direction. x0 This indicates the drag coefficient and angle of attack is 0, S represents the characteristic area, and C... x C represents the drag coefficient. y C represents the lift coefficient. z Magnus force coefficient, α e The dynamic equilibrium angle is represented by γ, the rotation angle of the projectile axis coordinate system is represented by d, the projectile diameter is represented by l, and the projectile length is represented by m. y The yaw moment is represented by m. z This represents the static moment coefficient.
[0040] Furthermore, step (4) of building the backpropagation neural network specifically involves: determining the number of input layer nodes, the number of hidden layer neurons, the activation function, and the number of output layer nodes of the backpropagation BP neural network model, and optimizing the algorithm; including the following steps:
[0041] Step (41): Input layer node settings
[0042] Based on the motion model of the controlled glide vehicle and live-fire data, the influencing factors were determined to be range x, flight altitude y, sideslip z, glide vehicle velocity V, and thrust P, where the glide vehicle velocity V is decomposed into V0... x V y and V z Three components, a total of 7 variables; determine the number of nodes in the input layer of the network, n=7;
[0043] Step (42): Setting up hidden layer nodes
[0044] In a three-layer network, with one hidden layer, there is an empirical formula relating the number of neurons in the neural network l, the number of neurons in the input layer n, and the number of neurons in the output layer m:
[0045]
[0046] a is an integer between 0 and 10, and the hidden layer node l is 15;
[0047] Step (43): Setting the number of output layer nodes
[0048] The backpropagation BP neural network is used to identify the wind speed W at different heights, and the number of nodes in the output layer is selected as m=1.
[0049] Step (44): Normalize the data.
[0050] Using min-max normalization, all sample data are mapped to the interval [0,1]; Step (45): Network parameter settings
[0051] The logarithmic sigmoid function was chosen as the hidden layer node transfer function, the linear purelin function was chosen as the output node transfer function, and the trainlm function of Levenbrg_Marquardt was chosen as the training function of the BP neural network algorithm. The number of hidden nodes was set to 15, the maximum number of iterations was 1000, the error threshold was 0.000001, and the learning rate was 0.01.
[0052] Step (46): Population optimization settings: The number of generations of the population is set to 50, the population size is set to 5, the number of optimization parameters is set to 241, and the boundary of the optimization variables is set to [-1, 1];
[0053] Step (47): Optimize the BP neural network using a genetic algorithm function.
[0054] Furthermore, the normalization formula in step (44) is:
[0055]
[0056] In the formula, x j For the original input data, x min For the minimum input data, x max For the maximum input data, The input data is normalized.
[0057] Furthermore, step (47) of optimizing the BP neural network using the genetic algorithm function specifically includes the following steps:
[0058] Step (471): Population initialization
[0059] The individual encoding method is real number encoding. Each individual is transformed into a real number string, which consists of four parts: the connection weights between the input layer and the hidden layer, the hidden layer threshold, the connection weights between the hidden layer and the output layer, and the output layer threshold. Each individual contains all the weights and thresholds of the neural network. Given that the network structure is known, it constitutes a BP neural network with a determined structure, weights, and thresholds.
[0060] Step (472) Fitness Function
[0061] Based on the initial weights and thresholds of the BP neural network obtained for each individual, the BP neural network is trained using training data to predict the system output. The absolute value of the error between the predicted output and the expected output, E, is used as the individual fitness value F, calculated using the following formula:
[0062]
[0063] In the formula, n is the number of network output nodes; y i Let o be the expected output of the i-th node in the BP neural network; i Let k be the predicted output for the i-th node, and k be the coefficient.
[0064] Step (473) Select Operation
[0065] The genetic algorithm uses a roulette wheel selection method, which is a selection strategy based on fitness ratios. The selection probability p for each individual i is as follows:
[0066] f i =k / F i
[0067]
[0068] In the formula, F i Let k be the fitness value of individual i, calculated as the reciprocal of the fitness value before individual selection; k is a coefficient; N is the number of individuals in the population.
[0069] Step (474): Cross operation
[0070] Using the real number crossover method, the k-th chromosome a k and the l-th chromosome a l The crossover operation at position j is as follows:
[0071] a kj =a kj (1-b)+a lj b
[0072] a lj =a lj (1-b)+a kj b
[0073] In the formula, b is a random number between [0, 1];
[0074] Step (475): Mutation operation
[0075] Select the j-th gene a of the i-th individual ij To perform mutation, the mutation operation method is as follows:
[0076]
[0077] In the formula, a max For gene a ij The upper bound; a max For gene a ij The lower bound; f(g) = r2(1-g / G) max ) 2 r2 is a random number; g is the current iteration number; G max is the maximum number of evolutions; r is a random number between [0, 1].
[0078] Furthermore, the wind speed vector obtained in step (6) is specifically as follows:
[0079] After the optimized backpropagation (BP) neural network has been trained, the connection weights w3 from the intermediate layers to the input layer and the hidden layer thresholds b2 are obtained. Assuming that M prediction samples x(k) are given at this point and the activation function is g(x), the network output y(k) is:
[0080] y(k)=g(w3x(k)+b2),k=1,2,...,M.
[0081] Compared with the prior art, the significant advantages of this invention are:
[0082] (1) The dynamic model (analytical method) of the guided projectile is combined with the artificial neural network (black box modeling method) to form a gray box model; the genetic algorithm is used to perform initial processing on the ballistic data so that it can avoid peak traps and better track the actual wind field, and achieve a more accurate effect for the actual wind speed acquisition using the neural network.
[0083] (2) The method is simple and accurate, and can avoid the determination of the thrust coefficient and drag coefficient of the guided projectile, as well as the modeling of the relationship between the flight speed and azimuth and its input thrust and the influence of external minor factors.
[0084] (3) Compared with the wind speed estimation method based on weather balloons (which requires slow ascent and hovering to measure wind speed), the wind speed estimation method proposed in this invention can quickly estimate three-dimensional wind speed / direction by collecting data after the flight process, and the tracking is fast, which meets the requirements of battlefield speed.
[0085] (4) Since the wind speed estimation method proposed in this invention is based on the dynamic model of the guided projectile and the gray box model established by using artificial neural network, it avoids the systematic error caused by insufficient consideration of actual flight state parameters due to oversimplification of the dynamic equation of the guided projectile. Therefore, compared with the method that only uses the dynamic equation model of the guided projectile, it has higher measurement accuracy and practical value.
[0086] (5) The wind speed estimation method proposed in this invention only requires one calibration experiment to complete the training of the artificial neural network, which can then be used to measure environmental wind speed. In contrast, existing methods using weather balloons and weather rockets require waiting for the data acquisition instrument to fall and be retrieved for analysis of wind speed and direction. Therefore, the wind speed acquisition method proposed in this invention is more convenient and easier to implement. Attached Figure Description
[0087] Figure 1 This is a schematic diagram of a backpropagation BP neural network.
[0088] Figure 2 This is a flowchart of a backpropagation BP neural network optimized based on a genetic algorithm.
[0089] Figure 3 This is a MATLAB network model of a backpropagation BP neural network optimized based on a genetic algorithm.
[0090] Figure 4 The fitness change curve of the trained neural network.
[0091] Figure 5 Compare the predictions for the test set results.
[0092] Figure 6 This represents the absolute error of the test set results.
[0093] Figure 7 This represents the percentage error in the test set results.
[0094] Figure 8 A comparison chart of actual wind speed and wind speed fitted by a neural network. Detailed Implementation
[0095] The present invention will now be described in further detail with reference to the accompanying drawings.
[0096] A rapid high-altitude meteorological detection method that meets battlefield speed requirements, suppresses high-altitude wind speed and direction disturbances in control, and improves the robustness of guidance systems. It includes the following steps:
[0097] Step 1: Conduct actual flight marking experiments, collect the flight attitude angle, velocity, acceleration, and position of the gliding extended-range guided projectile at each sampling moment in the experimental environment, and record the mass of the guided projectile. Obtain wind speed and direction information at different altitudes by releasing weather balloons.
[0098] Step 2: Establish ballistic kinematic and aerodynamic models;
[0099] Step 3: Obtain the motion parameters at each sampling moment using the onboard sensors and calculate the theoretical ballistic data;
[0100] Step 4: Build a backpropagation neural network. Take the wind speed and direction at each sampling height in the ballistic coordinate system as the output, and take the flight attitude angle, flight speed, acceleration and theoretical ballistic data collected at each sampling point of the guided projectile as the input. Then, use a genetic algorithm to preprocess and optimize the data, train the backpropagation neural network model, and save the training model after the neural network model is trained.
[0101] Step 5: At the location where measurements are required, launch an identical glide-extended-range guided projectile and record the flight attitude angle, velocity, acceleration, position, and thrust information at each sampling moment.
[0102] Step 6: Use the launch velocity of the guided projectile, the projectile's flight torque, and the flight attitude angle, flight speed, acceleration, and other information collected at each sampling point as input to the trained artificial neural network model to calculate the output wind speed vector.
[0103] Step 2 describes the establishment of a 4D motion model of the controlled glide vehicle in the ground coordinate system:
[0104]
[0105]
[0106]
[0107]
[0108]
[0109]
[0110]
[0111]
[0112]
[0113] The direct formula for calculating the dynamic equilibrium angle is:
[0114]
[0115]
[0116]
[0117]
[0118] The parameters in the direct calculation formula are:
[0119]
[0120] x2=(a a -a b )b y V r 2 (15)
[0121]
[0122]
[0123]
[0124] V rx =V x -W x (19)
[0125] V ry =V y (20)
[0126] V rz =V z -W z (twenty one)
[0127] In the formula, V x V is the velocity along the horizontal axis of the gliding missile. y V is the velocity along the longitudinal axis of the glider. z Let m be the vertical velocity of the glider, t be the mass of the glider, ρ be the air density, W be the wind speed, and V be the velocity along the vertical axis. r The x, y, and z values represent the actual velocity, respectively, and the C value represents the displacement in each direction. x0 S represents the drag coefficient (angle of attack is 0), S represents the characteristic area, and C represents the drag coefficient. x C represents the drag coefficient. y C represents the lift coefficient. z Magnus force coefficient, α e The dynamic equilibrium angle is represented by γ, the rotation angle of the projectile axis coordinate system is represented by d, the projectile diameter is represented by l, and the projectile length is represented by m. y The yaw moment is represented by m. zThis represents the static moment coefficient.
[0128] Based on the model and sample data described in step 4, a suitable backpropagation (BP) neural network model is first constructed: the number of input layer nodes, the number of hidden layer neurons, the activation function, and the number of output layer nodes of the backpropagation (BP) neural network model are determined, specifically including the following steps:
[0129] Step 4.1: Input Layer Node Settings
[0130] Based on the 4D motion model of the controlled glide missile and live-fire data, the influencing factors were determined to be range x, flight altitude y, sideslip z, glide missile velocity V, and thrust P. The glide missile velocity V can be specifically decomposed into V0... x V y and V z Three components, a total of 7 variables. The number of nodes in the input layer of the network is determined to be n = 7;
[0131] Step 4.2: Setting up hidden layer nodes
[0132] For general pattern recognition problems, a three-layer network can solve the problem well. In a three-layer network, the hidden layer is located at one layer, and there is an empirical formula relating the number of neural network neurons l, the number of input layer neurons n, and the number of output neurons m:
[0133]
[0134] Since a is an integer between 0 and 10, we set the hidden layer node l to be 15.
[0135] Step 4.3: Set the number of output layer nodes;
[0136] The backpropagation (BP) neural network is used to identify the wind speed W at different altitudes, and the number of nodes in the output layer is selected as m=1.
[0137] Step 4.4: Normalize the data;
[0138] The min-max normalization method is used to map all sample data to the interval [0,1].
[0139] The normalization formula is as follows:
[0140]
[0141] In the formula, x j For the original input data, x min For the minimum input data, x max For the maximum input data, x j *The input data is normalized;
[0142] Step 4.5: Network parameter settings;
[0143] The logarithmic sigmoid function is chosen as the hidden layer node transfer function, the linear purelin function is set as the output node transfer function, and the trainlm function of Levenbrg_Marquardt is set as the training function of the BP neural network algorithm.
[0144]
[0145] Step 4.6: Population optimization settings;
[0146]
[0147] Step 4.7 Optimization of the Backpropagation (BP) Neural Network using Genetic Algorithm
[0148] The optimization of a backpropagation (BP) neural network using a genetic algorithm consists of three parts: BP neural network structure determination, genetic algorithm optimization, and BP neural network prediction. The BP neural network structure determination part determines the BP neural network structure based on the number of input and output parameters of the fitting function, thereby determining the length of each individual in the genetic algorithm. The genetic algorithm optimization part uses a genetic algorithm to optimize the weights and thresholds of the BP neural network. Each individual in the population contains all the network's weights and thresholds. The individual's fitness value is calculated using a fitness function, and the genetic algorithm finds the individual corresponding to the optimal fitness value through selection, crossover, and mutation operations. The BP neural network prediction part uses the optimal individuals obtained by the genetic algorithm to assign initial weights and thresholds to the network. After training, the network outputs a prediction function. The weights and thresholds of the neural network are generally randomly initialized to random numbers in the range [-0.5, 0.5]. The training results of the network are the same; the introduction of the genetic algorithm is to optimize for the best initial weights and thresholds.
[0149] Step 4.7.1 Population Initialization
[0150] The individual encoding method is real number encoding, where each individual is transformed into a real number string, consisting of four parts: the connection weights between the input layer and the hidden layer, the hidden layer threshold, the connection weights between the hidden layer and the output layer, and the output layer threshold. Each individual contains all the weights and thresholds of the neural network. Given a known network structure, a BP neural network with a defined structure, weights, and thresholds can be constructed.
[0151] Step 4.7.2 Fitness Function
[0152] Based on the initial weights and thresholds of the BP neural network obtained for each individual, the BP neural network is trained using training data to predict the system output. The absolute value of the error between the predicted output and the expected output, E, is used as the individual fitness value F, calculated using the following formula:
[0153]
[0154] In the formula, n is the number of network output nodes; y i Let o be the expected output of the i-th node in the BP neural network; i Let k be the predicted output of the i-th node, and k be the coefficient.
[0155] Step 4.7.3 Select Operation
[0156] Genetic algorithm selection operations include various methods such as roulette wheel selection and tournament selection. This scheme chooses the roulette wheel selection method, which is a selection strategy based on fitness ratios, where each individual i has a selection probability p. The method is as follows:
[0157] f i =k / F i (25)
[0158]
[0159] In the formula, F i Let be the fitness value of individual i. Since a smaller fitness value is better, the reciprocal of the fitness value is taken before individual selection; k is the coefficient; N is the number of individuals in the population.
[0160] Step 4.7.4 Cross Operation
[0161] Since individuals are encoded using real numbers, the crossover operation uses the real number crossover method. The k-th chromosome a... k and the l-th chromosome a l The crossover operation at position j is as follows:
[0162]
[0163] In the formula, b is a random number between [0, 1].
[0164] Step 4.7.5 Mutation Operation
[0165] Select the j-th gene a of the i-th individual ij To perform mutation, the mutation operation method is as follows:
[0166]
[0167] In the formula, a max For gene a ij The upper bound; a max For gene aij The lower bound; f(g) = r2(1-g / G) max ) 2 r2 is a random number; g is the current iteration number; G max is the maximum number of evolutions; r is a random number between [0, 1].
[0168] Step 6.1: Implement parameter identification.
[0169] Wind speed identification for controlled projectile flight altitude includes the following steps:
[0170] After the optimized backpropagation (BP) neural network has been trained, the connection weights w3 from the intermediate layers to the input layer and the hidden layer thresholds b2 are obtained. Assuming that M prediction samples x(k) are given at this point and the activation function is g(x), the network output y(k) is:
[0171] y(k)=g(w3x(k)+b2),k=1,2,...,M (29)
[0172] Verify the output results: According to the method described in step 6: Based on the model and sample data, compare the height wind speed samples with the wind speed height samples identified by the input parameters. Divide the 9175 sets of data samples along the entire trajectory, randomly allocate 9100 sets of data as the training subset, and use the last 75 sets of data as the test set. Judge the training effect of the entire network by the mean square error, absolute error, and error percentage of the test set error.
[0173] Example 1
[0174] Using the azimuth data of gliding guided projectiles, a genetic algorithm was used to optimize a backpropagation neural network to identify the wind speed w. A total of 9175 data samples along the entire trajectory were divided, with 9100 sets randomly assigned as the training set and the remaining 75 sets used as the test set. The training effect of the entire network was analyzed by evaluating the mean squared error, absolute error, and percentage error of the test set.
[0175] Figure 1 This is a schematic diagram of a backpropagation (BP) neural network, which shows the change process of the backpropagation neural network and the numerical representation of each node. Figure 2 This is a flowchart of a backpropagation (BP) neural network optimized based on a genetic algorithm. Figure 3 This is a MATLAB network model of a backpropagation (BP) neural network optimized based on a genetic algorithm. Figure 4 The fitness curve of the trained neural network shows that the fitness decreases rapidly in the fourth and twenty-third iterations to reach a suitable fitness value, and the error is controlled within a reasonable range. Figure 5By comparing the prediction results on the test set, it can be seen that the network output can follow the changing test set very well, and the prediction trend is the same. Figure 6 The absolute error of the test set results is between -0.3 and 0.25. For wind speeds around 0 to 25, this represents a very small absolute error, indicating good tracking performance. Figure 7 The graph shows the percentage error of the test set results. As can be seen from the graph, for non-linearly changing wind speeds, the percentage error is controlled between -2.5% and 4%. In actual engineering, the possible error is between -5% and 5%, which fully meets the engineering standards and satisfies the accuracy requirements for wind speed identification. Figure 8 The graph shows a comparison between actual wind speed and wind speed fitted by the neural network. As can be seen from the graph, the estimation method of this invention is basically consistent with the actual wind speed.
Claims
1. A method for estimating high-altitude wind speed and direction of guided projectiles based on GA-BP neural networks, characterized in that, Includes the following steps: Step (1): Conduct actual flight marking experiments, collect the flight attitude angle, velocity, acceleration and position of the gliding extended range guided projectile at each sampling moment in the experimental environment, and record the mass of the guided projectile. Obtain wind speed and direction information at different altitudes by releasing weather balloons. Step (2): Establish ballistic motion model and aerodynamic model; Step (3): Obtain motion parameters at each sampling moment using the onboard sensors and calculate theoretical ballistic data; Step (4): Build a backpropagation neural network, take the wind speed and direction at each sampling height in the ballistic coordinate system as the output, take the flight attitude angle, flight speed, acceleration and theoretical ballistic data collected at each sampling point of the guided projectile as the input, then use the genetic algorithm to preprocess and optimize the data, train the backpropagation neural network model, and save the training model after the neural network model is trained. Step (5): At the location where measurements are required, launch an identical glide extended-range guided projectile and record the flight attitude angle, velocity, acceleration, position, and thrust information at each sampling moment; Step (6): Use the launch velocity of the guided projectile, the projectile's flight torque, and the flight attitude angle, flight speed, and acceleration information collected at each sampling point as input to the trained artificial neural network model to calculate the output wind speed vector; Step (2) establishes the motion model in the ballistic ground coordinate system as follows: ， ， ， ， ， ， ， ， ， The direct formula for calculating the dynamic equilibrium angle is: ， ， ， The parameters in the direct calculation formula are: ， ， ， ， ， ， ， ， In the formula, The velocity of the glider in the horizontal direction is... The velocity along the longitudinal axis of the glider missile. Let m be the vertical velocity of the gliding missile, m be the mass of the missile body, and t be the time. Indicates air density, Indicates wind speed. Indicates actual speed. Indicates the magnitude of displacement in each direction. This indicates that the drag coefficient and angle of attack are 0. Represents the characteristic area. Indicates the drag coefficient. Indicates the lift coefficient. Magnus force coefficient Indicates the dynamic equilibrium angle. This indicates the angle of rotation of the spring axis coordinate system. Indicates the diameter of the bullet. Indicates bullet length, The yaw moment is represented by m. z This represents the static moment coefficient.
2. The method according to claim 1, characterized in that, Step (4) Building the backpropagation neural network specifically involves: determining the number of input layer nodes, the number of hidden layer neurons, the activation function, and the number of output layer nodes of the backpropagation BP neural network model, and optimizing the algorithm; including the following steps: Step (41): Setting up input layer nodes Based on the motion model of the controlled glide vehicle and live-fire data, the influencing factors were determined to be range x, flight altitude y, sideslip z, glide vehicle velocity V, and thrust P, where the glide vehicle velocity V was decomposed into... , as well as Three components, a total of 7 variables; determine the number of nodes in the input layer of the network, n=7; Step (42): Setting up hidden layer nodes In a three-layer network, the hidden layer is positioned as one layer, and the number of neural networks is... and the number of neurons in the input layer Number of output neurons There is an empirical formula between them: ， Let the hidden layer nodes be integers between 0 and 10. It is 15; Step (43): Setting the number of output layer nodes Using a backpropagation BP neural network to measure wind speed at different altitudes The identification process determines the number of nodes in the output layer, m=1. Step (44): Normalize the data. The min-max normalization method is used to map all sample data to the interval [0,1]. Step (45): Network parameter settings The logarithmic sigmoid function was chosen as the hidden layer node transfer function, the linear purelin function was chosen as the output node transfer function, and the trainlm function of Levenbrg_Marquardt was chosen as the training function of the BP neural network algorithm. The number of hidden nodes was set to 15, the maximum number of iterations was 1000, the error threshold was 0.000001, and the learning rate was 0.
01. Step (46): Population optimization settings: The number of generations of the population is set to 50, the population size is set to 5, the number of optimization parameters is set to 241, and the boundary of the optimization variables is set to [-1, 1]; Step (47): Optimize the BP neural network using a genetic algorithm function.
3. The method according to claim 2, characterized in that, The normalization formula in step (44) is: ， In the formula, The original input data, For the minimum input data, For the maximum input data, The input data is normalized.
4. The method according to claim 3, characterized in that, Step (47) Optimizing the BP neural network using the genetic algorithm function specifically includes the following steps: Step (471): Population initialization The individual encoding method is real number encoding. Each individual is transformed into a real number string, which consists of four parts: the connection weights between the input layer and the hidden layer, the hidden layer threshold, the connection weights between the hidden layer and the output layer, and the output layer threshold. Each individual contains all the weights and thresholds of the neural network. Given that the network structure is known, it constitutes a BP neural network with a determined structure, weights, and thresholds. Step (472) Fitness Function Based on the initial weights and thresholds of the BP neural network obtained for each individual, the BP neural network is trained using training data to predict the system output. The absolute value of the error between the predicted output and the expected output, E, is used as the individual fitness value F, calculated using the following formula: ， In the formula, Output the number of nodes in the network; For the BP neural network, the first The expected output of each node; For the first The predicted output of each node, For coefficients; Step (473) Select Operation The genetic algorithm uses a roulette wheel selection method, which is a selection strategy based on fitness ratios, where each individual... Selection probability The method is as follows: ， ， In the formula, For individuals The fitness value is calculated by taking the reciprocal of the fitness value before individual selection; For coefficients; The number of individuals in the population; Step (474): Cross operation Using the real number crossover method, the first Chromosomes and the Chromosomes ,exist The bit crossover operation method is as follows: ， In the formula, It is a random number between [0, 1]; Step (475): Mutation operation Select the first The individual's first One gene To perform mutation, the mutation operation method is as follows: ， In the formula, For genes The upper bound; For genes The lower bound; ; `g` is a random number; `g` is the current iteration number. This represents the maximum number of evolutions. It is a random number between [0, 1].
5. The method according to claim 4, characterized in that, The wind speed vector obtained in step (6) is as follows: After training the optimized backpropagation (BP) neural network, the connection weights from the intermediate layers to the input layer of the model are obtained. and hidden layer threshold Assume that M prediction samples are given at this time. The activation function is Then the network output for: 。