An Online Recognition Method for Target Maneuvers in an Air Combat Simulation Environment

Abstract

The present invention relates to a method for online recognition of target maneuvers in an air combat simulation environment, including calculating target trajectory feature information; equally dividing the target continuous maneuver trajectory feature information sequence into several subsequences; sequentially inputting the several subsequences into a trained maneuver conversion subsequence recognition network to obtain a set of maneuver conversion subsequences; inputting each maneuver conversion subsequence into a trained maneuver conversion point positioning network to recognize the data bits of the maneuver conversion points; calculating each maneuver conversion point by using the sequential number and the data bits of the maneuver conversion points to obtain an ordered set of maneuver conversion points; segmenting the target continuous maneuver trajectory by using the recognition result of the target trajectory maneuver conversion points to obtain a set of multi-segment single maneuver trajectories; and inputting each single maneuver trajectory into a trained single maneuver recognition network for maneuver category recognition to obtain a maneuver category prediction label sequence. This method solves the problems of difficult recognition of continuous multi-segment unknown maneuver trajectories of targets and the convergence and timeliness of large-scale training samples.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer simulation and artificial intelligence, and particularly relates to a method for online recognition of target maneuvers in an air combat simulation environment. Background Art

[0002] Both air combat confrontation games and air combat simulation systems simulate the entire combat process of fighter jets in a detailed and realistic manner by means of computer simulation. In order to effectively improve the authenticity and ease of operation of the user experience, it is necessary to conduct simulation design from the perspective of actual air combat. In actual air combat with high dynamics and strong confrontation, the combat intentions of both sides are often achieved through a series of maneuvers. Therefore, online recognition of target maneuvers is beneficial to predicting the tactical intentions of the target and is the basis for achieving in-depth situation awareness and intelligent decision-making. For air combat confrontation games and air combat simulation systems, online recognition of target maneuvers is the key to improving their human-computer interaction and intelligent level, and is of great significance for the fidelity and credibility of simulation design.

[0003] Since the air combat sides change their postures through continuous maneuvering behaviors, the target flight trajectory data obtained by our airborne sensors may contain one or more maneuver segments. Online recognition of air combat target maneuvers is a data mining process that extracts key feature information from the target trajectory data obtained by airborne sensors and identifies one or more maneuver actions contained in the currently obtained target trajectory data as quickly and accurately as possible. Online and accurately recognizing air combat target maneuver actions can create tactical advantages for our aircraft, improve the efficiency of air combat decision-making, and thus gain air combat superiority.

[0004] The existing research on this problem generally applies to the recognition of single maneuver trajectories that have been segmented according to maneuver segments. Or some trajectory segmentation rules are artificially introduced to perform recognition after segmenting continuous maneuver trajectories. Such methods have obvious limitations and do not meet the requirements for recognizing unknown continuous maneuver trajectories of targets in air combat confrontation. Currently, the recognition methods mainly focus on the application of theories such as expert systems, support vector machines, and Bayesian networks. The recognition of maneuver actions based on expert systems relies on the prior knowledge of domain experts. A certain maneuver recognition rule base only corresponds to a certain type of specific air combat mission scenario, so its application is difficult to promote. The support vector machine algorithm needs to combine decision trees or construct multiple classifiers when solving the maneuver recognition problem, and it is difficult to implement for large-scale training samples. The Bayesian network can learn causal relationships and is an ideal model for integrating prior knowledge and data, but this method is computationally complex, resulting in a long recognition time and difficult to meet the timeliness requirements of online recognition.

[0005] In summary, the current recognition method has problems such as difficulty in recognizing continuous multi-segment unknown maneuver trajectories of targets, poor convergence of large-scale training samples, and long recognition time. Summary of the Invention

[0006] To solve the above problems existing in the prior art, the present invention provides a method for online recognition of target maneuvers in an air combat simulation environment. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0007] An embodiment of the present invention provides a method for online recognition of target maneuvers in an air combat simulation environment, including the steps of:

[0008] S1. Calculate target trajectory feature information for the airborne sensor target trajectory data containing multiple continuous maneuvers to obtain a target continuous maneuver trajectory feature information sequence;

[0009] S2. Average the target continuous maneuver trajectory feature information sequence into a fixed length and divide it into several subsequences;

[0010] S3. Input the several subsequences into a trained maneuver conversion subsequence recognition network in sequence to identify whether there is a maneuver conversion point in each subsequence, obtain a set of maneuver conversion subsequences, and record the sequence number of each maneuver conversion subsequence in the target continuous maneuver trajectory;

[0011] S4. Input each maneuver conversion subsequence into a trained maneuver conversion point positioning network to identify the data position of the maneuver conversion point in the maneuver conversion subsequence;

[0012] S5. Calculate each maneuver conversion point using the sequence number and the data position of the maneuver conversion point to obtain an ordered set of maneuver conversion points;

[0013] S6. Remove the characteristic parameters of the first moment point in the target continuous maneuver trajectory to form a new target continuous maneuver trajectory, and repeat steps S2 to S5 to obtain several new maneuver conversion points;

[0014] S7. Move each new maneuver conversion point one bit backward to return it to the real data position of the target continuous maneuver trajectory, obtain a new ordered set of maneuver conversion points, and take the union of the new ordered set of maneuver conversion points and the ordered set of maneuver conversion points to obtain the recognition result of the target trajectory maneuver conversion point;

[0015] S8. Segment the target continuous maneuver trajectory using the recognition result of the target trajectory maneuver conversion point to obtain a set of multi-segment single maneuver trajectories;

[0016] S9. Input each set of the single maneuver trajectories into the trained single maneuver recognition network for maneuver category recognition to obtain a maneuver category prediction label sequence.

[0017] In an embodiment of the present invention, step S1 includes:

[0018] Establish a geographic coordinate system, an inertial coordinate system, and a target track coordinate system, where the inertial coordinate system moves with the target, and the direction from the origin of the target track coordinate system to the x-axis is the target velocity vector direction and is consistent with the longitudinal axis of the aircraft body;

[0019] Calculate the track inclination angle, the track deviation angle change rate, the track inclination angle change rate, the altitude change rate, and the horizontal displacement change rate of the airborne sensor target track data in the geographic coordinate system, the inertial coordinate system, and the target track coordinate system to obtain characteristic information;

[0020] Perform normalization processing on the characteristic information to obtain the target continuous maneuver trajectory characteristic information sequence.

[0021] In an embodiment of the present invention, the track deviation angle of the target at time t is:

[0022]

[0023] where Δx(t), Δy(t), and Δz(t) are the position changes of the target in the geographic coordinate system at two consecutive times;

[0024] The track inclination angle of the target at time t is:

[0025]

[0026] where Δx(t), Δy(t), and Δz(t) are the position changes of the target in the geographic coordinate system at two consecutive times;

[0027] Δx(t) = x(t) - x(t - 1), Δy(t) = y(t) - y(t - 1), Δz(t) = z(t) - z(t - 1),

[0028] where (x(t), y(t), z(t)) is the position coordinate of the target in the geographic coordinate system at time t, and (x(t - 1), y(t - 1), z(t - 1)) is the position coordinate of the target in the geographic coordinate system at time t - 1;

[0029] The horizontal displacement of the target at time t is:

[0030]

[0031] where (x 0 ,y0 , z 0 ) is the initial coordinate of the target, and (x(t), y(t), z(t)) is the position of the target in the geographic coordinate system at time t;

[0032] The rate of change of the track deviation angle of the target at time t is:

[0033] The rate of change of the track inclination angle of the target at time t is:

[0034] The rate of change of the height of the target at time t is:

[0035] The rate of change of the horizontal displacement of the target at time t is:

[0036] where h is the simulation step size.

[0037] In an embodiment of the present invention, the trained maneuver conversion subsequence recognition network, the trained maneuver conversion point positioning network, and the trained single maneuver recognition network all include a mask layer, a long short-term memory network time series feature extraction layer, and a Softmax layer connected in sequence.

[0038] In an embodiment of the present invention, the network model of the long short-term memory network time series feature extraction layer is:

[0039] i t = σ(W i x t + U i h t-1 + b i )

[0040] f t = σ(W f x t + U f h t-1 + b f )

[0041] o t = σ(W o x t + U o h t-1 + b o )

[0042] e t = tanh(W e x t + U e h t-1 + b e )

[0043] c t = f t ⊙ c t-1 + i t ⊙ e t

[0044] h t = o t ⊙ tanh(c t )

[0045] where f t , i t , o t are the output states of the forget gate, input gate, and output gate of the long short-term memory network time series feature extraction layer at the current moment, c t is the internal unit state, h t is the external state of the hidden layer, e t is the candidate state obtained by concatenating the input at the current moment and the external state of the hidden layer at the previous moment and activating it with the hyperbolic tangent tanh. σ represents the sigmoid activation function, x t is the feature vector of the input network at the current moment, W i , W f , W o , W e are the first weight matrices corresponding to the long short-term memory network time series feature extraction layer, U i , U f , U o , U e are the second weight matrices corresponding to the long short-term memory network time series feature extraction layer, b i , b f , b o , b e are the bias vectors corresponding to the long short-term memory network time series feature extraction layer.

[0046] In an embodiment of the present invention, the network model of the Softmax layer is:

[0047]

[0048] where is the probability that the network outputs that this trajectory sample belongs to the k-th category, H i is composed of the hidden layer states h t output by the long short-term memory network time series feature extraction layer at each moment, w k , b k are the weight vector and bias term corresponding to the Softmax layer.

[0049] In one embodiment of the present invention, the training method of the trained maneuver conversion subsequence recognition network, the trained maneuver conversion point positioning network, and the trained single maneuver recognition network includes the steps:

[0050] Obtain the original target trajectory subsequence training samples, the original target maneuver conversion subsequence training samples, and the original target single maneuver training samples;

[0051] Perform oversampling and normalization processing on the original target trajectory subsequence training samples, and perform normalization processing on the original target maneuver conversion subsequence training samples and the original target single maneuver training samples to obtain the target trajectory subsequence training samples, the target maneuver conversion subsequence training samples, and the target single maneuver training samples;

[0052] Use the target trajectory subsequence training samples, the target maneuver conversion subsequence training samples, and the target single maneuver training samples to train the maneuver conversion subsequence recognition network, the maneuver conversion point positioning network, and the single maneuver recognition network connected in series in sequence, to obtain the trained maneuver conversion subsequence recognition network, the trained maneuver conversion point positioning network, and the trained single maneuver recognition network.

[0053] In one embodiment of the present invention, step S9 includes:

[0054] Unify the length of each single maneuver trajectory in the multi-segment single maneuver trajectory set to a fixed value to obtain a new multi-segment single maneuver trajectory set;

[0055] Input each single maneuver trajectory in the new multi-segment single maneuver trajectory set into the trained single maneuver recognition network for recognition to obtain the maneuver category prediction label sequence.

[0056] In one embodiment of the present invention, the maneuver categories in the maneuver category prediction label sequence include at least one of constant level flight, horizontal left spiral, horizontal right spiral, sharp climb, tactical dive, break-S, half loop, left break, right break, left combat turn, and right combat turn.

[0057] Compared with the prior art, the beneficial effects of the present invention:

[0058] The target maneuver online recognition method of the present invention uses a maneuver conversion subsequence recognition network to recognize maneuver conversion points, a maneuver conversion point positioning network to recognize the data bits of maneuver conversion points, and a single maneuver recognition network to recognize maneuver categories. It transforms the target maneuver online recognition problem into three classification problems that are easy to solve: maneuver conversion subsequence recognition, maneuver conversion point positioning, and single maneuver recognition, overcoming the problem of difficult recognition of continuous multi-segment unknown maneuver trajectories of the target. At the same time, it adopts a cascaded maneuver conversion subsequence recognition network, a maneuver conversion point positioning network, and a single maneuver recognition network to realize the mapping from the maneuver trajectory to the category label sequence, solving the convergence and timeliness problems of large-scale training samples. Therefore, this recognition method realizes the online recognition of target maneuvers in unmanned aerial vehicle autonomous air combat, improving the human-computer interaction and intelligent level of confrontation games and simulation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic flowchart of a method for online recognition of target maneuvers in an air combat simulation environment provided by an embodiment of the present invention;

[0060] Figure 2 It is a schematic diagram of a target motion coordinate system and related parameters provided by an embodiment of the present invention;

[0061] Figure 3 It is a schematic diagram of the true label of a subsequence provided by an embodiment of the present invention;

[0062] Figure 4 It is a schematic diagram of the true label of the data bits of a maneuver conversion subsequence provided by an embodiment of the present invention;

[0063] Figure 5 It is a schematic flowchart of the online recognition of target maneuvers provided by an embodiment of the present invention;

[0064] Figure 6 It is a schematic diagram of the confusion matrix of the recognition of maneuver conversion subsequences on a test set provided by an embodiment of the present invention;

[0065] Figure 7 It is a schematic diagram of the confusion matrix of the recognition of maneuver conversion subsequences on a test set provided by an embodiment of the present invention;

[0066] Figure 8 It is a schematic diagram of the confusion matrix of the recognition of single maneuvers on a test set provided by an embodiment of the present invention;

[0067] Figure 9 It is a schematic diagram of the three-dimensional trajectory of continuous multi-segment maneuvers of a target provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] The present invention will be further described in detail below in conjunction with specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0069] Embodiment 1

[0070] Please refer to Figure 1 , Figure 1 , which is a schematic flow chart of a method for on-line identification of target maneuvers in an air combat simulation environment provided by an embodiment of the present invention. The method for on-line identification of target maneuvers in the air combat simulation environment includes the following steps:

[0071] S1. Calculate the target trajectory feature information of the airborne sensor target trajectory data including multiple continuous maneuvers to obtain a target continuous maneuver trajectory feature information sequence. Specifically, it includes the following steps:

[0072] S11. Establish a geographic coordinate system, an inertial coordinate system, and a target track coordinate system. Among them, the inertial coordinate system moves with the target, and the direction from the origin of the target track coordinate system to the x-axis is the target velocity vector direction and is consistent with the longitudinal axis of the aircraft body.

[0073] Please refer to Figure 2 , Figure 2 , which is a schematic diagram of a target motion coordinate system and related parameters provided by an embodiment of the present invention. Figure 2 In it, Ox g y g z g is the geographic coordinate system; O t x gt y gt z gt is the inertial coordinate system, which moves with the target; O t x ht y ht z ht is the target track coordinate system. Ignoring the angle of attack and the sideslip angle, the direction of the O t x ht axis is the target velocity vector v direction and is consistent with the longitudinal axis of the aircraft body. θ represents the target track inclination angle, that is, the angle between the target velocity direction and the horizontal plane. When the component of the velocity vector on the O t y gt axis is positive, θ is positive, otherwise it is negative, and its value range is θ ∈ [-π / 2, π / 2]. φ represents the target track deviation angle, that is, the angle between the projection of the target velocity vector on the horizontal plane and O t x gt . When the projection of the target velocity vector on the horizontal plane rotates clockwise from O t x gt to the projection direction of the target velocity vector on the horizontal plane, φ is positive, otherwise it is negative, and its value range is φ ∈ [-π, π].

[0074] S12. Calculate the track dip angle, track deviation angle change rate, track dip angle change rate, altitude change rate, and horizontal displacement change rate of the airborne sensor target track data in the geographic coordinate system, inertial coordinate system, and target track coordinate system to obtain characteristic information.

[0075] Specifically, in this embodiment, the track dip angle θ, track dip angle change rate track deviation angle change rate altitude change rate and horizontal displacement change rate are used as the target track characteristic information.

[0076] Specifically, at time t, the coordinates of the target in the Ox g y g z g coordinate system are (x(t), y(t), z(t)), and the initial coordinates of the target are (x 0 , y 0 , z 0 ), which is the starting point of the track. The track deviation angle φ(t) of the target at time t is:

[0077]

[0078] where Δx(t), Δy(t), and Δz(t) are the position changes of the target in the geographic coordinate system Ox g y g z g at two consecutive times.

[0079] The track dip angle θ(t) of the target at time t is:

[0080]

[0081] where Δx(t), Δy(t), and Δz(t) are the position changes of the target in the geographic coordinate system Ox g y g z g at two consecutive times.

[0082] In the above two formulas, the calculation formulas for Δx(t), Δy(t), and Δz(t) are:

[0083] Δx(t) = x(t) - x(t - 1)

[0084] Δy(t) = y(t) - y(t - 1)

[0085] Δz(t) = z(t) - z(t - 1)

[0086] Among them, (x(t), y(t), z(t)) is the position coordinate of the target in the geographic coordinate system at time t, and (x(t - 1), y(t - 1), z(t - 1)) is the position coordinate of the target in the geographic coordinate system at time t - 1.

[0087] The horizontal displacement r(t) of the target at time t is:

[0088]

[0089] Among them, (x 0 , y 0 , z 0 ) is the initial coordinate of the target, and (x(t), y(t), z(t)) is the position of the target in the geographic coordinate system at time t.

[0090] The rate of change of the track deviation angle of the target at time t is:

[0091] The rate of change of the track inclination angle of the target at time t is:

[0092] The rate of change of the height of the target at time t is:

[0093] The rate of change of the horizontal displacement of the target at time t is:

[0094] Among them, h is the simulation step size.

[0095] S13. Normalize the feature information to obtain the target continuous maneuver trajectory feature information sequence.

[0096] Specifically, in order to avoid the influence of the feature parameter dimension, in this embodiment, the feature information is normalized by the maximum and minimum values of the feature parameters in each dimension for the sample data, that is Among them, x is a certain feature parameter value, x min and x max are respectively the minimum and maximum values in all samples, and x' is the normalized value, so as to obtain the target continuous maneuver trajectory feature information sequence.

[0097] S2. Average the target continuous maneuver trajectory feature information sequence into several subsequences with a fixed length.

[0098] Specifically, there is a maneuver conversion point p s between every two consecutive maneuvers in the target continuous maneuver trajectory feature information sequence, and this point is the trajectory segmentation point. In this embodiment, the target continuous maneuver trajectory feature information sequence Seq is evenly divided into L subsequences with a fixed length lc, and there is at most one maneuver conversion point in each subsequence.

[0099] S3. Input several subsequences into the cascaded trained maneuver conversion subsequence recognition network, the trained maneuver conversion point positioning network, and the trained single maneuver recognition network in sequence for maneuver category recognition, and obtain a maneuver category prediction label sequence.

[0100] Specifically, in this embodiment, by designing an automatic target trajectory segmentation and maneuver recognition model, the maneuver conversion point recognition problem is first decomposed into a maneuver conversion subsequence recognition problem and a maneuver conversion point positioning problem, and then the online target maneuver recognition problem is transformed into three classification problems: maneuver conversion subsequence recognition, maneuver conversion point positioning, and single maneuver recognition.

[0101] For the maneuver conversion subsequence recognition problem, this embodiment transforms it into a binary classification problem of subsequences. The subsequence with a maneuver conversion point is a "maneuver conversion subsequence", and the subsequence without a maneuver conversion point is a "non - maneuver conversion subsequence". As Figure 3 shown, Figure 3 is a schematic diagram of the true label of a subsequence provided by an embodiment of the present invention. The maneuver conversion subsequence recognition problem can be described as: solving the mapping function MSS = f s (Seq) from the target continuous maneuver trajectory feature information sequence Seq to the set of maneuver conversion subsequences sw 1 , sw 2 , …} which is MSS = {sw S .

[0102] For the maneuver conversion point positioning problem, this embodiment transforms it into a multi - classification problem of maneuver conversion subsequences. The category corresponds to the data position where the maneuver conversion point is located. As Figure 4 shown, Figure 4 is a schematic diagram of the true label of the data position of a maneuver conversion subsequence provided by an embodiment of the present invention; the maneuver conversion point positioning problem can be described as: solving the mapping MSP = f s (MSS) from the set MSS = {sw 1 , sw 2 , …} of maneuver conversion subsequences sw to the set MSP = {p s , p 1 , …} of maneuver conversion points p 2 . The probability of each bit of data in the pre - recognition maneuver conversion subsequence as the maneuver conversion point is equal, but since the feature of the maneuver conversion point is a mutation in the time dimension; if the maneuver conversion point appears in sw P s ​The first digit cannot be recognized, so only the cases where the maneuver conversion point is in the second digit to the last digit are considered here. To solve the problem of the maneuver conversion point appearing in the first digit, in this embodiment, the maneuver trajectory is shifted one position backward and re-input into the maneuver conversion subsequence recognition network and the maneuver conversion point positioning network, and then the union of the two conversion point recognition results is taken. Thus, the case where the maneuver conversion point is in the first digit can be converted to the second digit for solution.

[0103] For the problem of single maneuver recognition, this embodiment converts it into a multi-classification problem, and the classes correspond to the 11 maneuvers in the set of target maneuver classes in step 1; the target single maneuver recognition problem can be described as: solving the mapping M s from the target single maneuver trajectory sample X k to the maneuver class M k = f M (X s ); After cascading MSS = f S (Seq), MSP = f P (MSS) and M k = f M (X s ), the target maneuver online recognition problem is converted into solving the mapping MK = f 1 (Seq) from the maneuver trajectory Seq to the maneuver class sequence MK = [M 2 , M MSP ,...].

[0104] For the above three classification problems, this embodiment designs a cascaded classification network structure for automatic segmentation of target trajectories and maneuver recognition. This cascaded classification network structure includes three classification networks, namely, the maneuver conversion subsequence recognition network, the maneuver conversion point positioning network, and the single maneuver recognition network, which are used to solve the above three classification problems.

[0105] Specifically, the three classification networks have the same structure, all including a mask layer, a long short-term memory network time series feature extraction layer, and a Softmax layer connected in sequence.

[0106] Among them, the network model of the long short-term memory network time series feature extraction layer is:

[0107] i t = σ(W i x t + U i h t-1 + b i )

[0108] f t = σ(W f x t + U f h t-1 + bf )

[0109] o t = σ(W o x t + U o h t-1 + b o )

[0110] e t = tanh(W e x t + U e h t-1 + b e )

[0111] c t = f t ⊙ c t-1 + i t ⊙ e t

[0112] h t = o t ⊙ tanh(c t )

[0113] where f t , i t , o t are the output states of the forget gate, input gate, and output gate of the long short-term memory network's sequential feature extraction layer at the current moment, c t is the internal unit state, h t is the external state of the hidden layer, e t is the candidate state obtained by activating the concatenated vector of the current moment's input and the external state of the previous moment's hidden layer with the hyperbolic tangent tanh. σ represents the sigmoid activation function, x t is the feature vector of the current moment input to the network, W i , W f , W o , W e are the first weight matrices corresponding to the long short-term memory network's sequential feature extraction layer, U i , U f , U o , U e are the second weight matrices corresponding to the long short-term memory network's sequential feature extraction layer, b i , b f , b o , b e are the bias vectors corresponding to the long short-term memory network's sequential feature extraction layer.

[0114] The network model of the Softmax layer is:

[0115]

[0116] Among them, is the probability that the network outputs that this trajectory sample belongs to the k-th category, and H i The hidden layer states h at each moment are output by the long short-term memory network time series feature extraction layer t constitute, w k , b k are the weight vector and bias term corresponding to the Softmax layer.

[0117] Furthermore, the above cascaded classification network structure is trained, and the training method includes the steps:

[0118] 1) Obtain the original target trajectory subsequence training samples, the original target maneuver conversion subsequence training samples, and the original target single maneuver training samples.

[0119] 2) Oversample and normalize the original target trajectory subsequence training samples, and normalize the original target maneuver conversion subsequence training samples and the original target single maneuver training samples to obtain the target trajectory subsequence training samples, the target maneuver conversion subsequence training samples, and the target single maneuver training samples.

[0120] Specifically, first, to avoid the influence of the imbalance in the number of positive and negative class samples, this embodiment uses an oversampling method to resample the original target trajectory subsequence training sample data; then, to avoid the influence of the feature parameter dimension, this embodiment normalizes the original target trajectory subsequence training samples, the original target maneuver conversion subsequence training samples, and the original target single maneuver training samples through the maximum and minimum values of each dimension feature parameter, that is where x is a certain feature parameter value, x min and x max are respectively the minimum and maximum values in all training samples, and x' is the normalized value, so as to obtain the target trajectory subsequence training samples, the target maneuver conversion subsequence training samples, and the target single maneuver training samples.

[0121] Furthermore, obtain the training samples for target trajectory automatic segmentation and maneuver recognition. The target trajectory subsequence training samples are the mapping from the subsequence to its category of "maneuver conversion subsequence" or "non-maneuver conversion subsequence". The target maneuver conversion subsequence training samples are the mapping from the maneuver conversion subsequence to the data position of the maneuver conversion point. The target single maneuver training samples are the mapping from the single maneuver trajectory to the maneuver category.

[0122] 3) Use the training samples of the target trajectory subsequences, the training samples of the target maneuver conversion subsequences, and the training samples of the target single maneuver to train the maneuver conversion subsequence recognition network, the maneuver conversion point positioning network, and the single maneuver recognition network that are cascaded in sequence, respectively, to obtain the trained maneuver conversion subsequence recognition network, the trained maneuver conversion point positioning network, and the trained single maneuver recognition network.

[0123] Specifically, during the training process, due to the randomness of the air combat target maneuver trajectory input into the single maneuver recognition network, the length of a certain maneuver trajectory executed by the target is also very random. Therefore, before inputting into the single maneuver recognition network, the length of the input maneuver trajectory is unified to a fixed value. Specifically, the method of unifying the length of the input maneuver trajectory to a fixed value is: set a fixed value, fill the too short trajectory samples with a length less than the fixed value with the fixed value, and sample the too long trajectory samples with a length greater than the fixed value at equal intervals, so as to unify the length of the maneuver trajectory to the fixed value maxlen.

[0124] When the training meets the convergence condition, the trained maneuver conversion subsequence recognition network, the trained maneuver conversion point positioning network, and the trained single maneuver recognition network are obtained. Among them, the convergence condition can be that the training reaches the number of training times, for example, the number of training times reaches 800 times.

[0125] Please refer to Figure 5 , Figure 5 which is a schematic diagram of an online recognition process of target maneuvers provided by an embodiment of the present invention. Use the trained cascaded classification network structure to perform automatic segmentation and maneuver recognition of the target trajectory for several subsequences obtained in step S2, specifically including the steps:

[0126] S31. Input several subsequences into the trained maneuver conversion subsequence recognition network in sequence to identify whether there is a maneuver conversion point in each subsequence, obtain a set of maneuver conversion subsequences, and record the sequence number of each maneuver conversion subsequence in the target continuous maneuver trajectory.

[0127] Specifically, input L subsequences into the maneuver conversion subsequence recognition network in sequence to identify whether each input subsequence is a maneuver conversion subsequence. The maneuver conversion subsequence sw s will form a set MSS = {sw 1 , sw 2 ,...}, and record that the sequence number corresponding to the maneuver conversion subsequence sw s in the target continuous maneuver trajectory feature information sequence Seq is ns s .

[0128] S32. Input each maneuver conversion subsequence into the trained maneuver conversion point positioning network to identify the data bits of the maneuver conversion points in the maneuver conversion subsequence.

[0129] Specifically, take the identified maneuver conversion subsequence sw s as the input sequence for the identification by the maneuver conversion point positioning network, and identify the data bit np s of the maneuver conversion point p s in the maneuver conversion subsequence sw s .

[0130] S33. Calculate each maneuver conversion point using the sequence number and the data bits of the maneuver conversion points to obtain an ordered set of maneuver conversion points.

[0131] Specifically, calculate the maneuver conversion point p s using the sequence number ns s and the data bit np s = ns s ·lc + np s , then all the maneuver conversion points p s in the target continuous maneuver trajectory feature information sequence Seq form an ordered set of maneuver conversion points MSP = {p 1 , p 2 ,...}.

[0132] S34. After removing the feature parameters of the first moment point in the target continuous maneuver trajectory, form a new target continuous maneuver trajectory, and repeat steps S2, S31, S32, and S33 to obtain several new maneuver conversion points.

[0133] Specifically, after removing the feature parameters of the first moment point from the target continuous maneuver trajectory feature information sequence Seq, form a new target continuous maneuver trajectory feature information sequence Seq′, and input Seq′ as another target trajectory into the maneuver conversion subsequence recognition network and the maneuver conversion point positioning network to obtain a series of maneuver conversion points p s ′.

[0134] S35. Shift each new maneuver conversion point one position backward to return it to the real data bit in the target continuous maneuver trajectory, obtain a new ordered set of maneuver conversion points, and take the union of the new ordered set of maneuver conversion points and the ordered set of maneuver conversion points to obtain the recognition result of the maneuver conversion points of the target trajectory.

[0135] Specifically, shift each maneuver conversion point p s ′ one position backward to return it to the real data bit in the target continuous maneuver trajectory feature information sequence Seq, and obtain a new ordered set of conversion points MSP′ = {p 1 ′, p2 ', …}, take the union of MSP and MSP' and reassign it to MSP as the recognition result of the maneuver conversion points of the final target trajectory.

[0136] S36. Use the recognition result of the maneuver conversion points of the target trajectory to segment the continuous maneuver trajectory of the target, and obtain a set of multi-segment single maneuver trajectories.

[0137] Specifically, use the set of MSP of the recognition result of the maneuver conversion points of the target trajectory to segment the characteristic information sequence Seq of the continuous maneuver trajectory of the target, and obtain a set of multi-segment single maneuver trajectories XS = {X 1 , X 2 , …}.

[0138] S37. Input each single maneuver trajectory in the set of multi-segment single maneuver trajectories into the trained single maneuver recognition network for maneuver category recognition, and obtain a maneuver category prediction label sequence. Specifically, it includes the steps:

[0139] S371. Unify the length of each single maneuver trajectory in the set of multi-segment single maneuver trajectories to a fixed value, and obtain a new set of multi-segment single maneuver trajectories.

[0140] Specifically, set a fixed value, fill the too short trajectory samples with a length less than the fixed value with the fixed value, and use the method of equidistant sampling for the too long trajectory samples with a length greater than the fixed value, so as to unify the length of the single maneuver trajectory to the fixed value maxlen, and obtain a new set of multi-segment single maneuver trajectories XS'.

[0141] S372. Input each single maneuver trajectory in the new set of multi-segment single maneuver trajectories into the trained single maneuver recognition network for recognition, and obtain a maneuver category prediction label sequence.

[0142] Specifically, input each single maneuver trajectory in the new set of multi-segment single maneuver trajectories XS' into the single maneuver recognition network one by one for recognition, and finally obtain a maneuver category prediction label sequence MK = [M 1 , M 2 , …].

[0143] Specifically, in this embodiment, the maneuver category set M = {M 1 , M 2 ,..., M n} designed for the autonomous air combat of unmanned aerial vehicles includes 11 kinds of maneuvers such as constant level flight, horizontal left spiral, horizontal right spiral, sharp climb, tactical dive, break-S, half loop, left break, right break, left combat turn, and right combat turn. Therefore, the maneuver categories in the maneuver category prediction label sequence include at least one of the above 11 maneuvers.

[0144] In summary, the target maneuver online recognition method of this embodiment uses a maneuver conversion subsequence recognition network to recognize maneuver conversion points, a maneuver conversion point positioning network to recognize the data positions of maneuver conversion points, and a single maneuver recognition network to recognize maneuver categories. It transforms the target maneuver online recognition problem into three classification problems that are easy to solve: maneuver conversion subsequence recognition, maneuver conversion point positioning, and single maneuver recognition, overcoming the problem of difficult recognition of continuous multi-segment unknown maneuver trajectories of the target. At the same time, a cascaded maneuver conversion subsequence recognition network, a maneuver conversion point positioning network, and a single maneuver recognition network are adopted. Each cascaded network uses a long short-term memory classification network and an algorithm recognition logic to realize the mapping from the maneuver trajectory to the category label sequence, solving the convergence and timeliness problems of large-scale training samples. Therefore, this method realizes the online recognition of target maneuvers in UAV autonomous air combat, meets the timeliness requirements of target maneuver online recognition, and improves the human-computer interaction and intelligent level of the confrontation game and simulation system.

[0145] Embodiment 2

[0146] On the basis of Embodiment 1, this embodiment further illustrates the target maneuver online recognition method in the air combat simulation environment through the following simulation.

[0147] The computer environment for running the program in this embodiment is: an Inter Core i7 2.9GHZ processor, 8G of memory, and a Windows10 64-bit operating system. The programming language of the program is Python3, and the keras third-party library is used for network construction. The simulation step size is 0.01s.

[0148] First, randomly select the initial track deflection angle, track inclination angle, speed, altitude, and process control amount within a certain range, and use a three-degree-of-freedom flight simulation program to generate a large number of target trajectory samples. The target trajectories are segmented with a fixed length of 20, and category labels are assigned according to whether there are maneuver conversion points in the segmented subsequences, forming a trajectory feature parameter subsequence sample set. This embodiment generates a total of 11,486 pieces of data to form a network training set, 3,633 pieces of data to form a network validation set, and 3,575 pieces of data to form a network test set.

[0149] There is only one point as the maneuver conversion point in every two maneuver trajectories. After the randomly generated trajectory sequence is segmented into multiple subsequences, the maneuver conversion subsequences are in the minority and the non-maneuver conversion subsequences are in the majority. Therefore, there will be a problem of imbalance in the number of samples in the two categories. This embodiment uses the oversampling method to process the maneuver subsequence category training sample set to expand the minority category samples and monitor the network training process through the validation set at all times.

[0150] In the maneuver conversion subsequence recognition network designed in this embodiment, the dimension of the network input sample is [20×5]; the depth of the long short-term memory network time series feature extraction layer is 2, and the number of neurons is 8 and 4 respectively; the Softmax layer serves as the network output layer, and the number of neurons corresponds to the number of categories 2; the number of iterations is 800 times, and the Adam optimizer is used to optimize the network parameters. The maneuver conversion point recognition network is trained, and the training effect of the network is tested by the test set. Please refer to Figure 6 , Figure 6 which is a schematic diagram of the confusion matrix for maneuver conversion subsequence recognition on the test set provided by the embodiment of the present invention. The values on the horizontal axis are the true sample labels, and the values on the vertical axis are the predicted sample labels. The label "1" represents that the sample is a maneuver conversion subsequence, and the label "0" represents that the sample is a non-maneuver conversion subsequence. In the problem of maneuver conversion subsequence recognition, the recognition accuracy of the trained network for the test set is 99.1%, and the recall rate is 98.7%.

[0151] In the maneuver conversion point positioning network designed in this embodiment, the dimension of the network input sample is [20×5]; the depth of the long short-term memory network time series feature extraction layer is 2, and the number of neurons is 16 and 8 respectively; the Softmax layer serves as the network output layer, and the number of neurons corresponds to the number of categories 19; the number of iterations is 300 times, and the Adam optimizer is used to optimize the network parameters. The number of samples in the training set of this network is 2564, the number of samples in the validation set is 848, and the number of samples in the test set is 851. Please refer to Figure 7 , Figure 7 which is a schematic diagram of the confusion matrix for maneuver conversion subsequence recognition on the test set provided by the embodiment of the present invention. The values on the horizontal axis are the true data positions of the maneuver conversion points, and the values on the vertical axis are the data positions where the algorithm predicts the maneuver conversion points. In the problem of maneuver conversion point positioning, the recognition accuracy of the test set is 97.3%, and the recall rate of the test set is 97.3%; it can be seen from the confusion matrix that the number of misclassified samples is small, and the algorithm can accurately identify the location of the maneuver conversion point.

[0152] In the single maneuver recognition network designed in this embodiment, the dimension of the network input sample is [100×5]; the depth of the long short-term memory network time series feature extraction layer is 2, and the number of neurons is 16 and 8 respectively; the Softmax layer serves as the network output layer, and the number of neurons corresponds to the number of categories, which is 11; the number of iterations is 300 times, and the Adam optimizer is used to optimize the network parameters. In this embodiment, a total of 3473 single maneuver samples are generated to form the training set, 1240 samples form the validation set, and 1136 samples form the test set. The test set is used to test the training effect. Please refer to Figure 8 , Figure 8A schematic diagram of the confusion matrix for single maneuver recognition on the test set provided by the embodiments of the present invention. In the single maneuver recognition problem, the recognition accuracy rate of the test set is 98.9%, and the recall rate of the test set is 98.9%.

[0153] In this embodiment, 100 continuous maneuver trajectories of 2 to 8 segments are randomly generated to form a test set, and this test set is used to test the cascaded classification network; the sliding similarity is used to calculate the similarity between the predicted label sequence of the maneuver category and the true label sequence of the maneuver category to evaluate the recognition effect of the above cascaded network. The experimental results are shown in Table 1.

[0154] Table 1 Recognition effect of the cascaded network

[0155]

[0156] It can be seen from Table 1 that in the test of 2 to 8 continuous maneuver trajectories, the number of segmentation errors and recognition errors of the cascaded network is small, and the average similarity between the predicted label sequence and the true label sequence is relatively high, all above 93%.

[0157] Please refer to Figure 9 , Figure 9 A schematic diagram of the three-dimensional trajectory of the target's continuous multi-segment maneuver provided by the embodiments of the present invention. According to the target trajectory automatic segmentation and maneuver recognition model and algorithm designed in this embodiment, and at the same time combined with the classification network models at all levels obtained by offline training in the above experiment, the Figure 9 The target maneuver trajectory in is segmented online and the maneuver is recognized, and the maneuver trajectory data with 10 simulation steps added per second is used as the network input. Table 2 records the target maneuver time periods and the true maneuver categories.

[0158] Table 2 Target maneuver time periods and true maneuver categories

[0159]

[0160] It can be seen from Table 2 that the method proposed in this embodiment can segment the previous trajectory before the end of each maneuver and accurately identify the corresponding maneuver action, verifying the feasibility and effectiveness of the proposed target maneuver online recognition method.

[0161] The above content is a further detailed description of the present invention in combination with specific preferred implementation manners, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. An online recognition method for target maneuvers in an air combat simulation environment, characterized in that, it includes the steps: S1. Calculate the target trajectory feature information of the airborne sensor target trajectory data containing multiple continuous maneuvers to obtain a target continuous maneuver trajectory feature information sequence; S2. Evenly divide the target continuous maneuver trajectory feature information sequence into several subsequences with a fixed length; S3. Input the several subsequences into the trained maneuver conversion subsequence recognition network in sequence to identify whether there is a maneuver conversion point in each subsequence, obtain a set of maneuver conversion subsequences, and record the sequence number of each maneuver conversion subsequence in the target continuous maneuver trajectory; S4. Input each maneuver conversion subsequence into the trained maneuver conversion point positioning network to identify the maneuver conversion point data bits in the maneuver conversion subsequence; S5. Calculate each maneuver conversion point using the sequence number and the maneuver conversion point data bits to obtain an ordered set of maneuver conversion points; S6. Remove the feature parameters of the first moment point in the target continuous maneuver trajectory to form a new target continuous maneuver trajectory, and repeat steps S2 - S5 to obtain several new maneuver conversion points; S7. Move each new maneuver conversion point one position backward to make it return to the real data bit of the target continuous maneuver trajectory, obtain a new ordered set of maneuver conversion points, and take the union of the new ordered set of maneuver conversion points and the ordered set of maneuver conversion points to obtain the recognition result of the target trajectory maneuver conversion points; S8. Segment the target continuous maneuver trajectory using the recognition result of the target trajectory maneuver conversion points to obtain a set of multiple single maneuver trajectories; S9. Input each single maneuver trajectory in the set of multiple single maneuver trajectories into the trained single maneuver recognition network for maneuver category recognition to obtain a maneuver category prediction label sequence.

2. The online recognition method for target maneuvers in an air combat simulation environment according to claim 1, characterized in that, step S1 includes: Establish a geographic coordinate system, an inertial coordinate system, and a target track coordinate system, where the inertial coordinate system moves with the target, and the direction from the origin to the x - axis of the target track coordinate system is the target velocity vector direction and is consistent with the longitudinal axis of the aircraft body; Calculate the track inclination angle, track deviation angle change rate, track inclination angle change rate, altitude change rate, and horizontal displacement change rate of the airborne sensor target trajectory data in the geographic coordinate system, the inertial coordinate system, and the target track coordinate system to obtain feature information; Perform normalization processing on the feature information to obtain the target continuous maneuver trajectory feature information sequence.

3. The online recognition method for target maneuvers in an air combat simulation environment according to claim 2, characterized in that, The track deviation angle of the target at time t is: where Δx(t), Δy(t), Δz(t) are the position changes of the target in the geographic coordinate system at two consecutive times; The track inclination angle of the target at time t is: where Δx(t), Δy(t), Δz(t) are the position changes of the target in the geographic coordinate system at two consecutive times; Δx(t) = x(t) - x(t - 1), Δy(t) = y(t) - y(t - 1), Δz(t) = z(t) - z(t - 1), where (x(t), y(t), z(t)) is the position coordinates of the target in the geographical coordinate system at time t, and (x(t - 1), y(t - 1), z(t - 1)) is the position coordinates of the target in the geographical coordinate system at time t - 1; The horizontal displacement of the target at time t is: Among them, (x 0 , y 0 , z 0 ) is the initial coordinate of the target, and (x(t), y(t), z(t)) is the position of the target in the geographical coordinate system at time t; The rate of change of the track deflection angle of the target at time t is: The rate of change of the track inclination angle of the target at time t is as follows: The height change rate of the target at time t is as follows: The horizontal displacement change rate of the target at time t is as follows: where h is the simulation step size.

4. The method for online recognition of target maneuvers in an air combat simulation environment according to claim 1, characterized in that the trained maneuver conversion subsequence recognition network, the trained maneuver conversion point positioning network, and the trained single maneuver recognition network all include a masking layer, a long short-term memory network time series feature extraction layer, and a Softmax layer connected in sequence.

5. The method for online recognition of target maneuvers in an air combat simulation environment according to claim 4, characterized in that the network model of the long short-term memory network time series feature extraction layer is: i t = σ(W i x t + U i h t-1 + b i ) f t = σ(W f x t + U f h t-1 + b f ) o t = σ(W o x t + U o h t-1 + b o ) e t = tanh(W e x t + U e h t-1 + b e ) c t = f t ⊙ c t-1 + i t ⊙ e t h t = o t ⊙tanh(c t ) Among them, f t , i t , o t are the output states of the forgetting gate, input gate, and output gate of the long short-term memory network time series feature extraction layer at the current moment, respectively. c t is the internal unit state, h t is the external state of the hidden layer. e t is the candidate state obtained by activating the concatenated vector of the current moment input and the external state of the hidden layer at the previous moment by the hyperbolic tangent tanh. σ represents the sigmoid activation function. x t is the feature vector input to the network at the current moment. W i , W f , W o , W e is the first weight matrix corresponding to the long short-term memory network time series feature extraction layer. U i , U f , U o , U e is the second weight matrix corresponding to the long short-term memory network time series feature extraction layer. b i , b f , b o , b e is the bias vector corresponding to the long short-term memory network time series feature extraction layer.

6. The method for online recognition of target maneuvers in an air combat simulation environment according to claim 4, characterized in that the network model of the Softmax layer is: Among them, is the probability that the network outputs that this trajectory sample belongs to the k-th category, H i The hidden layer state h at each moment is output by the long short-term memory network time series feature extraction layer t constitutes, w k , b k are the weight vector and bias term corresponding to the Softmax layer.

7. The method for online recognition of target maneuvers in an air combat simulation environment according to claim 1, characterized in that the training methods of the trained maneuver conversion subsequence recognition network, the trained maneuver conversion point positioning network, and the trained single maneuver recognition network include the steps of: Obtain the original target trajectory subsequence training samples, the original target maneuver conversion subsequence training samples, and the original target single maneuver training samples; Perform oversampling and normalization processing on the original target trajectory subsequence training samples, and perform normalization processing on the original target maneuver conversion subsequence training samples and the original target single maneuver training samples to obtain target trajectory subsequence training samples, target maneuver conversion subsequence training samples, and target single maneuver training samples; Use the target trajectory subsequence training samples, the target maneuver conversion subsequence training samples, and the target single maneuver training samples to train the maneuver conversion subsequence recognition network, the maneuver conversion point positioning network, and the single maneuver recognition network connected in cascade in sequence, to obtain the trained maneuver conversion subsequence recognition network, the trained maneuver conversion point positioning network, and the trained single maneuver recognition network.

8. The method for online recognition of target maneuvers in an air combat simulation environment according to claim 1, characterized in that Step S9 includes: Unify the length of each single maneuver trajectory in the multi-segment single maneuver trajectory set to a fixed value to obtain a new multi-segment single maneuver trajectory set; Input each single maneuver trajectory in the new multi-segment single maneuver trajectory set into the trained single maneuver recognition network for recognition to obtain the maneuver category prediction label sequence.

9. The method for online recognition of target maneuvers in an air combat simulation environment according to claim 1, characterized in that The maneuver categories in the maneuver category prediction tag sequence include at least one of level constant flight, horizontal left spiral, horizontal right spiral, sharp climb, tactical dive, break-S, half loop, left break, right break, left combat turn, and right combat turn.

Images