Aircraft friend or foe attribute discrimination method based on expert experience and self-development neural network

By combining expert experience and self-development neural network, using Resnet to extract fine-grained features of the aircraft and conduct self-development training, the accuracy and robustness of traditional methods in the discrimination of enemy-friendly attributes in complex environments is solved, and efficiently versus unknown goals is achieved Identify and continue to learn.

CN119046745APending Publication Date: 2024-11-29BEIHANG UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202410878781.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-02
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

Traditional aircraft's enemy-me attribute discrimination methods are difficult to adapt to the update of enemy aircraft models and changes in electronic warfare environments in complex battlefield environments, resulting in a decrease in identification accuracy, especially when facing unknown targets, the misjudgment rate is high.

Method used

Combining expert experience and self-developed neural network, Resnet is used to extract fine-grained features of aircraft images, and through self-developed neural network development training, accurate judgment and continuous learning of aircraft's enemy-to-me attributes.

Benefits of technology

The onboard computer's ability to judge the attributes of unknown targets and enemies is improved, and the ability to accurately identify and continuously learn in complex environments is achieved, reducing the rate of misjudgment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046745A_ABST
    Figure CN119046745A_ABST
Patent Text Reader

Abstract

The invention discloses an airplane friend or foe attribute discrimination method based on expert experience and a self-development neural network, and the method is characterized in that the method comprises the steps: 1, building a database of airplane appearance features according to the expert experience; step 2, constructing a classification network of the Resnet101; 3, constructing an Adam optimizer, and optimizing model parameters in a neural network training process; step 4, establishing a normalized feature vector table; 5, establishing an autonomous development neural network model; 6, training an aircraft appearance geometric feature classification network according to the appearance feature data set; 7, according to the appearance characteristic data mapping relation established in the step 4, carrying out friend or foe attribute inference engine development; and 8, combining training results of the step 6 and the step 7 to realize end-to-end friend or foe attribute reasoning. According to the method, neural network training is facilitated; the components are convenient to deploy or replace; the network can continuously iterate along with updating of the expert knowledge base, and is more intelligent.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to an aircraft friend-or-foe attribute discrimination method based on expert experience and self-developing neural network, and belongs to the technical field of image recognition and fine-grained image classification. Background Art

[0002] In the aviation field, accurate identification of the enemy and friend attributes of aircraft is crucial to ensuring aviation safety and performing military missions. With the rapid development of modern science and technology, especially the remarkable progress in deep learning and neural network technology, the methods of aircraft enemy and friend attributes identification based on these technologies have become a hot topic of research. Traditional aircraft enemy and friend identification methods mainly rely on radar signal characteristics, aircraft shape recognition and other means. These methods can achieve the purpose of enemy and friend identification to a certain extent, but with the development of stealth technology and the advancement of electronic countermeasure technology, traditional methods are facing more and more challenges. Especially in complex electronic warfare environments and high-density airspace operations, it is particularly difficult to quickly and accurately identify the enemy and friend attributes of aircraft. The deep learning-based method can extract high-dimensional features of aircraft, including but not limited to radar cross-sectional area features, flight behavior patterns, electronic signal features, etc., by using a large amount of data for training, thereby achieving high-accuracy identification of aircraft enemy and friend attributes. In contrast, as a new type of neural network structure, the self-developing neural network is characterized by the ability to self-optimize the network structure according to task requirements, dynamically adjust the connection mode and the function of the processing unit, so as to better adapt to the needs of enemy and friend attribute identification tasks in complex and changing environments.

[0003] The self-developing neural network is an artificial neural network model that achieves intelligent learning and adaptation by simulating the self-developing process of the biological nervous system. It combines the research results of neuroscience, computer science and cognitive science, and aims to achieve machine learning and cognitive functions in a way similar to the biological brain. It has good self-organizing ability and plasticity, and can adapt to multimodal learning and online learning task scenarios. Existing deep learning methods often only directly identify enemy attributes through offline expert annotation, ignoring the intrinsic connection between the aircraft appearance characteristics and the camp; combining deep learning with self-developing neural network technology through expert experience can effectively improve the accuracy and robustness of aircraft enemy attribute discrimination. This method can not only learn the static characteristics of the aircraft, but also capture the dynamic changes of the aircraft, such as flight status, instant changes in electronic signals, etc., which provides a new idea for aircraft enemy attribute discrimination in complex environments. Although the research and application of this technology is still in the development stage, its potential application value and influence cannot be ignored, indicating that in the future aviation field, the aircraft enemy attribute discrimination method based on deep learning and self-developing neural network will become an important technical direction.

[0004] The present invention proposes a method for distinguishing friendly or enemy attributes of aircraft based on expert experience and self-developing neural networks, so as to solve the problem that enemy aircraft models are constantly updated in complex battlefield environments and traditional neural network methods will misjudge new target categories as existing categories. This method can improve the ability of onboard computers to judge the friendly or enemy attributes of unknown targets; at the same time, it can also realize continuous learning of the network in actual engineering. Summary of the invention

[0005] The purpose of the present invention is to propose a method for distinguishing the friend-or-foe attributes of an aircraft based on expert experience and self-developing neural networks. The method uses Resnet (deep residual network) to extract fine-grained features of aircraft images, such as wing type, tail type, number of engines, etc.; the self-developing neural network is used to summarize the relationship between these fine-grained features and the friend-or-foe attributes of the aircraft, and the fine-grained features extracted by Resnet are normalized and sent to the self-developing neural network for development training, thereby realizing the friend-or-foe attribute discrimination of an unknown aircraft type by an onboard computer. The specific implementation steps are as follows:

[0006] Step 1: Based on expert experience, establish a database of aircraft appearance features.

[0007] The appearance features of general fighters include tail type, canard, wing type, engine type, number of engines, whether it carries missiles, whether it has machine guns, whether it has ventral fins, whether it has a hull structure (seaplane), and friend or foe attributes. Among them, tail types include: traditional tail, T-tail, cross tail, no tail, V-tail and double tail; wing types include: delta wing, swept wing, variable swept wing, high wing, tilt-rotor, straight wing, flying wing and diamond wing; engine types include: turbine engine, turbofan engine, turbojet engine and propeller engine; the number of engines is 1, 2, 4, 6, 8; attributes are only divided into friend or foe.

[0008] Step 2: Build a classification network of Resnet101.

[0009] ResNet performs well in many computer vision tasks, especially image classification and object detection. ResNet101 has 101 layers and is a deep model in the ResNet family. Deeper networks can learn more complex and abstract features, improving the performance of the model in image classification tasks. The basic module structure of Resnet is as follows: Figure 2 As shown. Assuming that the input of the convolutional neural network is x, the original expected output of a convolutional layer is H(x), and the target of the residual learning module in the Resnet101 network is the residual function F(x), then:

[0010] F(x)=H(x)-x (1)

[0011] Finally, relu is used as the activation function,

[0012] H(x)=W2relu(W1x)+W s x (2)

[0013]

[0014] In formula (2), W1 represents the first convolutional layer operation, W2 represents the second convolutional layer operation, and W s Represents the weight matrix of the identity mapping. Modify the last layer of the original Resnet101 into two fully connected layers, and the output dimensions of the two fully connected layers are 1×512 and 1×Num c , where Num c is the number of categories to be classified. Train the network, and train 50 rounds in each network training process. Use cross entropy loss as the loss function. For the binary classification problem in feature classification (whether to carry missiles, etc.), the cross entropy loss function formula is:

[0015]

[0016] Among them, L is the loss function, N is the number of samples, and y i is the true label of the i-th sample, which takes a value of 0 or 1, y i is the probability that the model predicts that the i-th sample is a positive class, and log represents the natural logarithm. For multi-classification problems in feature classification (such as tail type), the cross entropy loss function needs to use an independent formula for the predicted probability of each category, which can be expressed as:

[0017]

[0018] Where N is the number of samples, C is the total number of categories, and y i,c is the true label of sample i in category c (1 means it belongs to the category, 0 means it does not belong to the category), is the probability that the model predicts that sample i belongs to category c. For multi-classification tasks such as wing type and tail type, the multi-classification cross entropy is used to calculate the loss function; for binary classification tasks such as whether to carry missiles, the binary classification cross entropy is used to calculate the loss function.

[0019] Step 3: Build the Adam optimizer

[0020] The Adam optimizer is constructed as an optimization algorithm for optimizing neural networks. Its principles are as follows:

[0021] First, calculate the gradient at time t. Gradient represents the gradient. represents the gradient of the loss function J with respect to the parameter θ,

[0022]

[0023] Then calculate the exponential moving average m of the gradient t , β1 is the decay rate 1, which is usually set to a value close to 1.

[0024] m t =β1m t-1 +(1-β1) gradient t (7)

[0025] Weighted square of the gradient, v t is the exponential moving average of the squared gradient, and β2 is the decay rate 2:

[0026] v t =β2v t-1 +(1-β1) gradient 2 (8)

[0027] The gradient mean is bias-corrected, This is the result after correction:

[0028]

[0029] The weighted squared gradient is bias-corrected. This is the result after bias correction:

[0030]

[0031] The general formula is as follows, α is the learning rate, and ε is used to prevent numerical instability caused by the denominator being 0:

[0032]

[0033] Step 4: Create a normalized feature vector table

[0034] In order to facilitate the conversion of appearance features into a language that can be understood by the machine, each feature is arranged in English in lexicographic order and mapped to the range [0, 1] using a maximum and minimum value normalization operation.

[0035]

[0036] Step 5: Establish an autonomous development neural network model

[0037] The autonomous development neural network consists of three layers of self-organizing networks: XYZ, where X is the sensory input layer, Y is the brain neural layer, and Z is the effector layer. Figure 1 shown.

[0038] (1) At time t = 0, for any region A in the network {X, Y, Z}, initialize the corresponding network layer N = (V, G) and the response vector r, where V is the total synaptic weights and G is the age of each neuron.

[0039] (2) At time t=1,2,..., for any region A in the network {X,Y,Z}, a unified algorithm function is used to update the network.

[0040] (r′,N′)=f(b,t,N) (13)

[0041] Where b is the bottom-to-top vector, t is the top-to-bottom vector, and r′ is the response vector. Each neuron in region A has a weight vector v=(v b ,v t ), its response vector is determined by the following formula:

[0042]

[0043] in, is the normalized neuron synaptic weight The unit vector of It is normalized

[0044] Input Vector The top-k competition mechanism is adopted, and k is 1. The winning neuron j is determined by the following formula:

[0045]

[0046] In the development stage, the activated neurons update the corresponding synaptic weights, and the bottom-top vector and top-bottom vector of the Y layer are updated with the bottom-top vector of the Z layer. The learning process of the entire developmental neural network neurons uses the leaf component analysis algorithm. The bottom-top vector and top-bottom vector of the synaptic weights of the Y layer neurons are updated with the bottom-top vector of the synaptic weights of the Z layer neurons. For the network layer L = (V, G, r), where V = (v1, v2, ..., v c ), contains c synaptic weight vectors, G = (g1, g2, ..., g c ), stores the neuron activation age, and r is the response vector of each neuron.

[0047] During the leaf component analysis algorithm, v is first randomly initialized. i , and set the activation age of each neuron to zero. In the updating process of leaf component analysis, calculate the pre-response vector of neuron synapse: Let the input p = ((1-k)x DN ,az DN ), x DNis the external input of the self-developing neural network layer X, z DN is the internal input of the Z layer of the self-developing neural network, k is the weight of the top-bottom input, and k=0.5. i The largest neuron is the winner. The network only updates the synaptic weight vector of the winner neuron, and does not update the others. The connection between neurons adjusts the strength of the connection due to the excitation behavior of the neuron itself, and then returns to affect the response of the subsequent neurons to receive stimulation. After each round of development, the excitation age of the winner neuron must be updated. i ←g i +1.

[0048]

[0049] Among them, w1 and w2 are learning factors that control the update rate of neuron synaptic weights. w1 and w2 are obtained by the forgetting average algorithm:

[0050]

[0051] w1(g i )=1-w2(g i ) (18)

[0052] In the formula, u(g i ) is the forgetting equation, which is defined as follows:

[0053]

[0054] After the developmental neural network completes learning, the prediction process only uses the synaptic weight vectors of the neurons in the Y and Z layers, based on the external input of the X layer, through the internal neuron synaptic connections, and finally selects the enemy and friend attributes of the Z layer for output.

[0055]

[0056] Step 6: Train the aircraft appearance feature classification network constructed in step 2 based on the appearance feature dataset established in step 1. Train a corresponding self-network for each sub-feature, for a total of 9 self-networks. Figure 3 The evolution of the network accuracy during the classification training of one of the sub-features is shown; Figure 4 The corresponding loss function evolution process is shown. The input of the development process is the normalized feature vector, and the correspondence of some features is shown in Table 6.

[0057] Step 7: According to step 4, a normalized feature vector table is established to develop the autonomous development neural network model constructed in step 5. During the development process, the input is a normalized feature vector with a shape of 1×9; the output size is 1×2; the number of initial Y layer neurons is 6, and the development round is 200. The split percentage is 75%, the split threshold is set to 10, the number of splits is set to 3, and the activation age of the split is set to 0. Figure 5 It shows the changes in the accuracy of friend-or-enemy attribute discrimination during the development process; Figure 6 Shows the changes in the number of neurons in layer Y during development.

[0058] Step 8: Combine the training results of steps 6 and 7 to achieve end-to-end friend-or-foe attribute reasoning. When the friend-or-foe attribute reasoning machine performs reasoning, the image is first sent to 9 Resnet101 sub-networks to extract the geometric appearance features of the aircraft, such as the number of engines and wing shapes. After obtaining the inferred geometric appearance features, they are normalized and converted into numerical vectors according to the expert knowledge base established previously, and sent to the self-developing neural network for friend-or-foe attribute reasoning. In actual combat, if the type of aircraft in the expert knowledge base is updated, its appearance features can be mapped to numerical vectors and sent to the self-developing neural network for development, thereby achieving sustainable learning. The entire reasoning process is as follows: Figure 7 shown.

[0059] The invention discloses an aircraft friend-or-foe attribute discrimination method based on expert experience and self-developing neural network, which has the advantages and effects of: constructing an appearance feature expert knowledge base based on an existing fighter data set, and mapping the appearance features in the semantic space into normalized vectors, which is convenient for neural network training; at the same time, establishing a modular end-to-end friend-or-foe attribute reasoning model, which is convenient for deploying or replacing components therein; and adopting a self-developing neural network model, so that the network can be continuously iterated with the update of the expert knowledge base, and become more intelligent. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of the self-developing neural network structure.

[0061] Figure 2 Schematic diagram of the basic residual block structure of Resnet.

[0062] Figure 3 Evolution curve of classification accuracy of Resnet basic image feature extraction network training.

[0063] Figure 4 Evolution curve of Resnet basic image feature extraction network training loss function.

[0064] Figure 5 The accuracy of friend-or-enemy attribute discrimination during the development process of self-developing neural network.

[0065] Figure 6 Changes in the number of neurons in layer Y during the development of the self-developing neural network.

[0066] Figure 7 The overall process of aircraft friend-or-foe attribute discrimination based on expert experience and self-developing neural network. DETAILED DESCRIPTION

[0067] The following is a specific example to verify the aircraft friend-or-foe attribute discrimination method based on expert experience and self-developing neural network proposed in the present invention. The simulation environment of this example is configured with an Intel i7-12700H processor, 2.30Ghz main frequency, 32G memory, 64-bit operating system, and an x64-based processor; the graphics card is NVIDIA GeForce RTX 3060LaptopGPU, and the programming language is python3.10. The configured environment dependencies are: torch=2.0.1, torchaudio=2.0.2, torchvision=0.15.2, tqdm=4.66.2, matplotlib=3.8.4, opencv-python=4.9.0.80, openpyxl=3.1.2, pandas=2.2.1, pyxl=3.1.2, pillow=10.2.0, pyarrow=15.0.0, scipy=1.12.0, seaborn=0.13.2 and numpy=1.26. The implementation flow chart is as follows Figure 3 As shown, the specific practical steps of this example are as follows:

[0068] Step 1: Build a database of aircraft appearance features based on expert experience.

[0069] Table 1 shows some of the aircraft appearance features based on tail type, canard presence, airfoil, engine type, number of engines, whether missiles are carried, machine guns, ventral fins, and hull structure (seaplane).

[0070] Table 1 Appearance characteristics of some aircraft

[0071]

[0072]

[0073] Step 2: Build a classification network based on Resnet101

[0074] The basic module structure of Resnet is as follows Figure 2As shown. Assume that the input of the convolutional neural network is x, the original expected output of a convolutional layer is H(x), the residual learning module in the Resnet101 network targets the residual function F(x) as shown in formula (1), and the activation function used is shown in formula (3).

[0075] Modify the last layer of the original Resnet101 into two fully connected layers, and the output dimensions of the two fully connected layers are 1×512 and 1×Num c , where Num c is the number of categories to be classified. The network is trained for 50 rounds in each network training process. The cross entropy loss is used as the loss function. For binary classification problems in feature classification (whether to carry missiles, etc.), the cross entropy loss function is as shown in formula (4). For multi-classification problems in feature classification (such as tail type), the cross entropy loss function needs to use an independent formula for the predicted probability of each category, and its loss function is as shown in formula (5).

[0076] Step 3: Build the Adam optimizer

[0077] Construct the Adam optimizer as the optimization algorithm for the neural network in the optimization paper. Calculate the gradient at time t according to formula (6); then calculate the exponential moving average of the gradient according to formula (7), where β1 = 0.9; weight the square of the gradient according to formula (8), where β2 = 0.999. Correct the mean of the gradient according to formula (9); correct the weight of the square of the gradient according to formula (10). The total formula is as follows (11), where α = 0.001, ε = 10 -8 .

[0078] Step 4: Create a normalized feature vector table

[0079] To facilitate the conversion of appearance features into a language that can be understood by the machine, each feature is arranged in English in lexicographic order and mapped to the range [0, 1] using the maximum and minimum value normalization operation. The mapping results are shown in Table 2-5.

[0080] Table 2 Tail type mapping

[0081] Feature Name Feature dictionary number Normalized Mapping Conventional 0 0 Cruciform 1 0.2 No-tail 2 0.4 T-tail 3 0.6 Twin Tail 4 0.8 V-tail 5 1.0

[0082] Table 3 Airfoil mapping

[0083] Feature Name Feature lexicographic number Normalized Mapping Delta Wing 0 0 Faceted 1 0.142857142 Flying Wing 2 0.285714285 High Wing 3 0.428571428 Straight 4 0.571428571 Swept Wing 5 0.714285714 Tiltrotor 6 0.857142857 Variable-sweep Wing 7 1.0

[0084] Table 4 Engine type mapping

[0085] Feature Name Feature dictionary number Normalized Mapping Tiltrotor 0 0 Turbofan 1 0.333333333 Turbojet 2 0.666666667 Turboprop 3 1.0

[0086] Table 5 Engine quantity mapping

[0087]

[0088]

[0089] Step 5: Establish an autonomous development neural network model

[0090] according to Figure 1 The structure shown in the figure constructs an autonomous development neural network. At time t = 0, for any region A in the network {X, Y, Z}, initialize the corresponding network layer N = (V, G) and the response vector r, where V is the total synaptic weight and G is the age of each neuron. At time t = 1, 2..., for any region A in the network {X, Y, Z}, use the algorithm function shown in formula (13) to update the network. Each neuron in region A has a weight vector v = (v b ,v t ), its response vector is determined by formula (14): The winning neuron j is determined by formula (15).

[0091] In the development stage, the activated neurons update the corresponding synaptic weights. The learning process of the entire developmental neural network neurons uses the leaf component analysis algorithm. During the operation of the leaf component analysis algorithm, v is first randomly initialized. i , and set the activation age of each neuron to zero. In the update process of leaf component analysis, calculate the pre-response vector of neuron synapse: let the input p = ((1-k)x,az), x is the external input of layer X, z is the internal input of layer Z, k is the weight of top-bottom input, take k = 0.5. i The largest neuron is the winner. The network only updates the synaptic weight vector of the winner neuron, and does not update the others. The connection between neurons adjusts the strength of the connection due to the excitation behavior of the neuron itself, and then returns to affect the response of the subsequent neurons to receive stimulation. After each round of development, the excitation age of the winner neuron must be updated. i ←g i +1.

[0092] The learning factors w1 and w2 that control the update rate of neuron synaptic weights are obtained by the forgetting average algorithm shown in formulas (17) and (18). After the developmental neural network completes learning, the prediction process only uses the synaptic weight vectors of neurons in the Y layer and the Z layer, according to the external input of the X layer, through the internal neuron synaptic connection, and finally selects the enemy and friend attributes of the Z layer for output according to formula (20).

[0093] Step 6: Train the aircraft appearance feature classification network constructed in step 2 based on the appearance feature dataset established in step 1. Train a corresponding self-network for each sub-feature, for a total of 9 self-networks. Figure 3The evolution of the network accuracy during the classification training of one of the sub-features is shown; Figure 4 The corresponding loss function evolution process is shown. The input of the development process is the normalized feature vector, and the correspondence of some features is shown in Table 6.

[0094] Table 6 Normalized appearance features of some aircraft

[0095] Aircraft Type Tail type Is there a duck wing? airfoil Engine Type …… A10 0 0 0.571428571 0.333333333 …… A400M 0.6 0 0.428571429 1 …… AG600 0.6 0 0.428571429 1 …… AV8B 0 0 0.714285714 0.333333333 …… …… …… …… …… …… ……

[0096] Step 7: According to step 4, the appearance feature data mapping relationship is established, and the enemy attribute inference machine constructed in step 5 is developed. During the development process, the input is a normalized feature vector with a shape of 1×9; the output size is 1×2; the number of initial Y layer neurons is 6, and the development round is 200. The split percentage is 75%, the split threshold is set to 10, the number of splits is set to 3, and the activation age of the split is set to 0. The change in the accuracy of enemy attribute discrimination during the development process is shown; Figure 6 Shows the changes in the number of neurons in layer Y during development.

[0097] Step 8: Combine the training results of steps 6 and 7 to achieve end-to-end friend-or-foe attribute reasoning. When the friend-or-foe attribute reasoning machine performs reasoning, the image is first sent to 9 Resnet101 sub-networks to extract the geometric appearance features of the aircraft, such as the number of engines and wing shapes. After obtaining the inferred geometric appearance features, they are normalized and converted into numerical vectors according to the expert knowledge base established previously, and sent to the self-developing neural network for friend-or-foe attribute reasoning. In actual combat, if the type of aircraft in the expert knowledge base is updated, its appearance features can be mapped to numerical vectors and sent to the self-developing neural network for development, thereby achieving sustainable learning. The entire reasoning process is as follows: Figure 7 shown.

[0098] This example uses pictures of aircraft models that have not been seen and have unknown attributes during the network training process to infer enemy attributes. The aircraft models used are H-6 (H-6) and JF-17 (Xiaolong fighter). In order to facilitate the comparison of network performance, Resnet is used alone to train an enemy attribute classification network for comparison. For H-6, the characteristic attributes inferred by the aircraft enemy attribute discrimination method based on expert experience and self-developed neural network are: V-shaped tail, no canard, swept wings, turbofan engine, twin engines, carrying missiles, with machine guns, no ventral fins, not a seaplane; the enemy attribute inference machine determines that the aircraft is a friendly aircraft. For JF-17, the characteristic attributes inferred by the aircraft enemy attribute discrimination method based on expert experience and self-developed neural network are: traditional tail, no canard, swept wings, turbofan engine, twin engines, carrying missiles, with machine guns, with ventral fins, not a seaplane; the enemy attribute inference machine determines that it is a friendly aircraft. Using only Resnet will misjudge both aircraft as enemy aircraft. Although the friend-or-foe attribute inference engine has never seen relevant data of these two types of aircraft during the training stage, the results show that the proposed aircraft friend-or-foe attribute discrimination method based on expert experience and self-developing neural network has good generalization ability in the friend-or-foe attribute discrimination of unknown types of aircraft, proving the effectiveness and efficiency of the present invention.

Claims

1. A method for distinguishing friendly or enemy attributes of aircraft based on expert experience and self-developed neural network, characterized by: The steps of this method are as follows: Step 1: Build a database of aircraft appearance features based on expert experience; Step 2: Build a classification network of Resnet101; Assuming that the input of the convolutional neural network is x, the original expected output of a convolutional layer is H(x), and the target of the residual learning module in the Resnet101 network is the residual function F(x), then: F(x)=H(x)-x (1) Finally, relu is used as the activation function, H(x)=W2relu(W1x)+W s x (2) In formula (2), W1 represents the first convolutional layer operation, W2 represents the second convolutional layer operation, and W s The weight matrix representing the identity mapping; Step 3: Build the Adam optimizer The Adam optimizer is constructed as an optimization algorithm for optimizing neural networks. The general formula is as follows, where α is the learning rate and ε is used to prevent numerical instability caused by the denominator being 0: Step 4: Create a normalized feature vector table In order to convert the appearance features into a language that the machine can understand, each feature is arranged in lexicographic order and mapped to the range [0, 1] using the maximum and minimum value normalization operation; Step 5: Establish an autonomous development neural network model The autonomous development neural network consists of three layers of self-organizing networks: XYZ, where X is the sensory input layer, Y is the brain neural layer, and Z is the effector layer; Step 6: Train the aircraft appearance feature classification network constructed in step 2 according to the appearance feature database established in step 1; train a corresponding self-network for each sub-feature; Step 7: According to step 4, a normalized feature vector table is established, and the enemy and friend attributes autonomous development neural network model constructed in step 5 is carried out; Step 8: Combine the training results of steps 6 and 7 to achieve end-to-end friend-or-enemy attribute reasoning.

2. The method according to claim 1, characterized in that: The database of aircraft appearance features includes: tail type, whether there is a canard, wing shape, engine type, number of engines, whether it carries missiles, whether it has a machine gun, whether it has a ventral fin, whether it has a hull structure, and friend or foe attributes.

3. The method according to claim 2, characterized in that: The tail types include: traditional tail, T-tail, cross tail, tailless, V-tail and double tail; the wing shapes include: delta wing, swept wing, variable swept wing, high wing, tilt-rotor, straight wing, flying wing and diamond wing; the engine types include: turbine engine, turbofan engine, turbojet engine and propeller engine; the number of engines includes 1, 2, 4, 6 and 8.

4. The method according to claim 1, characterized in that: Modify the last layer of the original Resnet101 into two fully connected layers, and the output dimensions of the two fully connected layers are 1×512 and 1×Num c , where Num c is the number of categories to be classified.

5. The method according to claim 1, characterized in that: The cross entropy loss is used as the loss function. For the binary classification problem in feature classification, the cross entropy loss function formula is: Among them, L is the loss function, N is the number of samples, and y i is the true label of the i-th sample, which takes a value of 0 or 1. is the probability that the model predicts that the i-th sample is a positive class, and log represents the natural logarithm; For multi-classification problems in feature classification, the cross entropy loss function requires a separate formula for the predicted probability of each category, expressed as: Where N is the number of samples, C is the total number of categories, and y i,c is the true label of sample i in category c, where 1 means it belongs to the category and 0 means it does not belong to the category. is the probability that the model predicts that sample i belongs to category c; For multi-classification tasks such as airfoil and tail type, the cross entropy of multi-classification is used to calculate the loss function; For binary classification tasks such as whether or not to carry missiles, binary cross entropy is used to calculate the loss function.

6. The method according to claim 1, characterized in that: In the step eight, friend-or-foe attribute reasoning, the specific process is to first send the image to the Resnet101 sub-network to extract the geometric appearance features of the aircraft, including the number of engines and the wing shape; after obtaining the inferred geometric appearance features, they are normalized and converted into numerical vectors according to the expert knowledge base established previously, and sent to the autonomous development neural network for friend-or-foe attribute reasoning.

7. The method according to claim 6, characterized in that: If the types of aircraft in the expert knowledge base are updated, their appearance features are mapped into numerical vectors and sent to the self-developing neural network for development to achieve sustainable learning.

Citation Information

Patent Citations

  • Chaotic autonomous development neural network-based scene identification method, device and system

    CN107832833A

  • Out-of-library target identification method and device

    CN115718897A

  • Battlefield combat trend prediction method based on deep neural network

    CN115730743A

  • Intelligent data and knowledge-driven method for modulation recognition

    US11700156B1