Near space hypersonic speed target rapid increment identification method

Abstract

The invention discloses a near space hypersonic speed target rapid increment identification method. The method comprises the following steps: training a first hypersonic speed target identification network model based on a hypersonic speed target initial class data set to carry out target identification; the hypersonic speed target initial class data set comprises a plurality of classes of samples; the first hypersonic speed target recognition network model comprises a spectral feature extractor and a first target classifier; when there is a new target type, training a second hypersonic speed target identification network model based on the hypersonic speed target new class data set to carry out target identification; the second hypersonic speed target recognition network model comprises a spectral feature extractor and a second target classifier; the hypersonic speed target new class data set is constructed based on old class samples and new class samples; and when the second hypersonic speed target recognition network model is trained, the parameters of the spectral feature extractor are kept unchanged. According to the method, rapid updating and less disastrous forgetting of the actual hypersonic speed target new class identification application can be met.

Description

Technical Field

[0001] The present invention belongs to the field of hypersonic aircraft, and in particular relates to a method for rapid incremental recognition of near-space hypersonic targets. Background Art

[0002] Near-space hypersonic vehicles have the characteristics of being able to achieve ultra-high speed, long-range and highly maneuverable flight, and have broad application prospects in the military and civilian fields. Rapidly and accurately identifying the flight type, flight altitude and speed of near-space hypersonic vehicles is a prerequisite for tracking such targets.

[0003] Traditional ground-based radar detection is easily affected by the radius of curvature of the earth, while infrared-based detection methods cause detection signal degradation due to the high-speed flight of the aircraft. Both detection methods make it difficult to effectively identify hypersonic aircraft targets.

[0004] At present, the special application of hypersonic targets is still in the development stage, and the known target types are limited. Therefore, the recognition of hypersonic aircraft targets can be achieved by training conventional target recognition networks. However, with the emergence of new types of hypersonic targets in the future, how to quickly identify the newly added types of hypersonic targets has become a technical challenge. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a method for rapid incremental recognition of hypersonic targets in near space.

[0006] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] A method for rapid incremental recognition of near-space hypersonic targets, comprising:

[0008] Training a first hypersonic target recognition network model based on an initial class data set of hypersonic targets, and using the trained first hypersonic target recognition network model to perform target recognition; the first hypersonic target recognition network model includes a spectral feature extractor and a first target classifier;

[0009] Whenever there is a new target type, a second hypersonic target recognition network model is trained based on a new class data set of hypersonic targets, and the trained second hypersonic target recognition network model is used to perform target recognition; the second hypersonic target recognition network model includes the spectral feature extractor and the second target classifier;

[0010] Among them, the initial class data set of hypersonic targets includes samples of multiple categories; the newly added class data set of hypersonic targets is constructed based on old category samples and new category samples; when training the second hypersonic target recognition network model, the parameters of the spectral feature extractor are maintained unchanged at the state when the training of the first hypersonic target recognition network model is completed.

[0011] Optionally, both the first target classifier and the second target classifier are classifiers constructed based on spectral angle theory.

[0012] Optionally, the first target classifier is specifically used to: output a category scoring matrix according to the features extracted by the spectral feature extractor, so as to determine the target type according to the category scoring matrix;

[0013] The category scoring matrix output by the first target classifier is:

[0014] Score[i]=SA(f i ,w j ),j∈[1,N c ],

[0015]

[0016] Wherein, f is the feature extracted by the spectral feature extractor, and its subscript is used to identify different sample categories, i represents the i-th sample; w is the column in the parameter matrix, and its subscript is the column number, and different columns correspond to different target types. The parameter matrix of the first target classifier is N e ×N c Dimension; N e is the dimension of the features extracted by the spectral feature extractor, N c is the number of target types that the first target classifier can identify; the second target classifier has the same function as the first target classifier; the parameter matrix of the second target classifier is N e ×(N c +t) dimensions, where the 0th to Nth c The column parameters are the same as the 0th to Nth c The column parameters are the same, and t is the number of newly added target types.

[0017] Optionally, determining the target type according to the category scoring matrix is ​​implemented by the following formula:

[0018]

[0019] in, It means taking the category Label corresponding to the element with the smallest value in the category scoring matrix as the determined target type.

[0020] Optionally, the method of constructing the newly added hypersonic target data set includes:

[0021] Determine the sample size of new category samples;

[0022] For each class of samples in the old class samples, select samples from the class of samples with the same number as the samples;

[0023] The newly added category data set of hypersonic targets is constructed according to various types of samples selected from old category samples and newly added new category samples.

[0024] Optionally, for each class of samples in the old class samples, selecting samples with the same number as the samples from the class samples includes:

[0025] Calculate the feature mean of each category of samples in the old category samples;

[0026] For each category of samples in the old category samples, traverse each sample in the category. If any of the samples is included in the selected sample set corresponding to the category of samples, and the distance between the feature mean of the selected sample set and the feature mean of the category of samples is the smallest, then the sample is included in the selected sample set until the number of samples in the selected sample set reaches the said number of samples.

[0027] Optionally, calculating the feature mean of each category of samples in the old category samples includes:

[0028] For each class of samples in the old class samples, the spectral feature extractor is used to extract the feature vector of each sample in the class of samples, and the mean of all feature vectors of the class of samples is calculated as the feature mean.

[0029] Optionally, the spectral feature extractor includes a convolution layer, a maximum pooling layer and a convolution network connected in sequence; the convolution network includes a first Dense module, a transition layer and a second Dense module connected in sequence; the first Dense module and the second Dense module are both Dense Blocks in a densely connected convolution network DenseNet, and the transition layer is a Transition Layer in the DenseNet; the convolution layer includes 64 convolution kernels of size 7×1, and the size of the maximum pooling layer is 3×1; the first Dense module includes 6 layers with a growth rate of 32; the second Dense module includes 12 layers with a growth rate of 32.

[0030] Optionally, when training the second hypersonic target recognition network model, the parameter updating method of the second target classifier is:

[0031]

[0032] Among them, W old is the N of the first target classifier e ×N c dimensional parameter matrix, for The lth row element in .

[0033] Optionally, when training the second hypersonic target recognition network model, a cross entropy loss function is used to evaluate the model loss during the training process.

[0034] The method for rapid incremental recognition of near-space hypersonic targets provided by the present invention, whenever there is a new target type, uses the spectral feature extractor of the trained hypersonic target recognition network model to combine with the new second target classifier to construct a new second hypersonic target recognition network model, and trains the second hypersonic target recognition network model by adding new category samples on the basis of the existing old category samples to form a new category data set of hypersonic targets; and in the incremental update process of the second hypersonic target recognition network, keeps the parameters of the spectral feature extractor unchanged, dynamically updates the incremental learning mode of the target classifier, protects the spectral feature extractor's knowledge of the learned near hypersonic targets, and dynamically adapts to the distribution of the new category of targets in the feature space. At the same time, since the parameters of the model update are reduced, the update time of the second hypersonic target recognition network model is effectively reduced, so that the newly added category of hypersonic targets can be quickly identified.

[0035] The present invention will be further described in detail below with reference to the accompanying drawings.

[0036] Figure 1 It is a flow chart of a method for rapid incremental recognition of hypersonic targets in near space provided by an embodiment of the present invention;

[0037] Figure 2 is a mean curve diagram of 16 types of spectral feature samples of simulated near-space hypersonic targets in an embodiment of the present invention;

[0038] Figure 3 is a structural diagram of a first hypersonic target recognition network model / a second hypersonic target recognition network model in an embodiment of the present invention;

[0039] Figure 4 is a recognition confusion matrix result of the first hypersonic target recognition network model in the embodiment of the present invention and the second hypersonic target recognition network model trained in multiple incremental learning stages thereafter;

[0040] Figure 5 The results of the comparison between the method of the embodiment of the present invention and several existing methods in terms of processing sample rate and time consumption are shown. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0042] In order to realize the rapid identification of newly added hypersonic targets, the embodiment of the present invention provides a method for rapid incremental identification of near-space hypersonic targets, which can meet the requirements of rapid update of new hypersonic target identification applications in practice and reduce catastrophic forgetting. Figure 1 As shown, the method comprises the following steps:

[0043] S10. Training a first hypersonic target recognition network model based on an initial class data set of hypersonic targets, and using the trained first hypersonic target recognition network model to perform target recognition; the first hypersonic target recognition network model includes a spectral feature extractor and a first target classifier.

[0044] Specifically, a hypersonic vehicle target will form a plasma sheath on the target surface during high-speed flight. The spectral radiation characteristics of the plasma sheath of the hypersonic target are closely related to the physical motion characteristics of the aircraft, which can provide richer information for the new generation of detection tasks and can be used as an important feature for fine identification of target type, altitude and speed. In an embodiment of the present invention, the measured near-space hypersonic targets are classified according to type, flight altitude and flight speed, and the hypersonic target spectra under different categories are obtained to construct a hypersonic target initial class data set to train the first hypersonic target recognition network model. The hypersonic target initial class data set includes samples of multiple categories.

[0045] In the above-mentioned first hypersonic target recognition network model, the spectral feature extractor is used to extract spectral features from the hypersonic target spectrum; and the first target classifier is used to predict the target type according to the spectral features.

[0046] In the embodiment of the present invention, there may be multiple specific structures of the spectral feature extractor. In a preferred implementation, the spectral feature extractor may be constructed based on a densely connected convolutional network DenseNet.

[0047] For example, when the dimension of the hypersonic target spectrum is low, for example, 128×1, Figure 2 As shown, the spectral feature extractor may include a convolutional layer, a maximum pooling layer, and a convolutional network connected in sequence. Figure 2The convolution layer and the maximum pooling layer are jointly displayed as Conv-Pool; the convolution network includes the first Dense module (DenseBlock1), the transition layer (Transition) and the second Dense module (Dense Block2) connected in sequence; the first Dense module and the second Dense module are both Dense Blocks in the densely connected convolution network DenseNet, and the transition layer is the Transition Layer in DenseNet. Among them, the convolution layer includes 64 convolution kernels of size 7×1, and the size of the maximum pooling layer is 3×1; the first Dense module includes 6 layers with a growth rate of 32; the second Dense module includes 12 layers with a growth rate of 32. In this implementation, considering the low dimensionality of the hypersonic target spectrum, the spectral feature extractor only uses two Dense Blocks in DenseNet.

[0048] In an embodiment of the present invention, a spectral feature extractor is constructed based on DenseNet, which can establish dense links between all previous layers and subsequent layers, and can achieve feature reuse through feature connections of different layers, thereby promoting the utilization of information of each spectral segment of the target spectral feature with fewer parameters and computational costs.

[0049] It should be noted that the specific structure of the spectral feature extractor in the embodiment of the present invention is not limited to Figure 2 For example, when the dimension of the hypersonic target spectrum is high, more Dense Blocks can be used in the spectral feature extractor, which can be optimized and adjusted.

[0050] The first target classifier may be a common classifier, such as a fully connected layer. Alternatively, in a preferred implementation of the embodiment of the present invention, the first target classifier may be a classifier constructed based on spectral angle theory.

[0051] In the spectral angle theory, the spectrum is represented as a vector in a multidimensional space, and the angle between the vectors of two different spectra can be used to describe the similarity between the two spectra. The calculation method of the spectral angle is highly sensitive to curves, and it can still maintain a high accuracy when processing different spectral amplitude changes. Even slight differences can distinguish the flight status of targets of different categories. Therefore, an embodiment of the present invention uses the spectral angle theory to construct a classifier. Specifically, the first target classifier projects the feature space onto the hypergeometric sphere by calculating the angle between the features of samples of different classes and the model parameters corresponding to different classes, so that even slight differences in the targets in the feature space can be effectively distinguished by the first target classifier.

[0052] Specifically, in the embodiment of the present invention, the first target classifier is specifically used to: output a category scoring matrix according to the features extracted by the spectral feature extractor, so as to determine the target type according to the category scoring matrix. The category scoring matrix output by the first target classifier is:

[0053] Score[i]=SA(f i ,w j ),j∈[1,N c ],

[0054]

[0055] In this formula, f is the feature extracted by the spectral feature extractor, and its subscript is used to identify different sample categories. i represents the i-th sample; w is the column in the parameter matrix of the classifier, and its subscript is the column number. Different columns correspond to different target types. The parameter matrix of the first target classifier is N e ×N c Dimension; N e is the dimension of the features extracted by the spectral feature extractor, N c is the number of target types that can be identified by the first target classifier. The parameter matrix of the first target classifier is represented as W old ∈R Ne×Nc , R is the set of real numbers,

[0056] In the embodiment of the present invention, the target type is determined according to the category scoring matrix, which can be achieved by the following formula:

[0057]

[0058] in, It means taking the category label corresponding to the element with the smallest value in the category scoring matrix as the determined target type.

[0059] When training the first hypersonic target recognition network model, the hypersonic target spectrum (here, the sample for training) is input into the spectral feature extractor to extract the spectral features, and the extracted spectral features are sent to the first target classifier for target recognition. In addition, in the process of training the first hypersonic target recognition network model, a loss function can be used to evaluate the model loss, so that the model parameters are gradually optimized to obtain a trained first hypersonic target recognition network model. Among them, the loss function is, for example, a cross entropy loss function, which is of course not limited to this.

[0060] Therefore, after the first hypersonic target recognition network model is trained, the hypersonic target spectrum of the target type to be identified (here, the spectrum of the actual hypersonic target) is input into the trained first hypersonic target recognition network model, and the target recognition result can be obtained at the output end of the first target classifier.

[0061] In addition, after the first hypersonic target recognition network model is trained, a certain number of samples can be selected and retained for each category from the initial class data set of the hypersonic target, and these samples can be used as old samples, waiting for the emergence of new class targets in the future to form an incremental data set again for the training of the incremental recognition network. Here, the incremental data set is named a hypersonic target newly added class data set in the embodiment of the present invention, and the incremental recognition network is named a second hypersonic target recognition network model in the embodiment of the present invention.

[0062] S20. Whenever there is a new target type, a second hypersonic target recognition network model is trained based on a new class data set of hypersonic targets, and the trained second hypersonic target recognition network model is used to perform target recognition; the second hypersonic target recognition network model includes a spectral feature extractor and a second target classifier; the new class data set of hypersonic targets is constructed based on old category samples and new category samples; when training the second hypersonic target recognition network model, the parameters of the spectral feature extractor are maintained unchanged at the state when the training of the first hypersonic target recognition network model is completed.

[0063] Specifically, the newly added class data set of hypersonic targets can be constructed by introducing the hypersonic target spectra (new class samples) of the newly added target types on the basis of the old class samples. Here, the old class samples refer to: whenever there is a newly added target type, the most recently trained supersonic target recognition network model. For example, after the first hypersonic target recognition network model is trained, when the newly added target type appears for the first time, the old class samples at this time are the samples in the initial class data set of the hypersonic target. If incremental training has been performed, that is, there is already a trained second hypersonic target recognition network model, then when the newly added target type appears at this time, the old class samples are the samples used in the most recent incremental training of the second hypersonic target recognition network model (which has been trained).

[0064] The spectral feature extractor of the second hypersonic target recognition network model directly adopts the spectral feature extractor of the trained first hypersonic target recognition network model. The second target classifier is also a classifier constructed based on the spectral angle theory. Its function is the same as that of the first target classifier, which is to predict the target type based on the spectral features. However, the number of target types that the second target classifier can identify is different from that of the first target classifier. Specifically, assuming that the first hypersonic target recognition network model has been trained, its first target classifier has N e ×N c parameters, and the newly added target type appears for the first time, so the second target classifier constructed at this time has N e ×(N c+t) parameters, where 0th to Nth c The column parameters are related to the 0th to Nth order of the first target classifier. c The parameters of the second target classifier are the same, and t is the number of newly added target types. new ∈R Ne×(Nc+t) , more specifically, It can be seen that the number of target types that can be identified by the first target classifier is the number of types of old category samples, and the number of target types that can be identified by the second target classifier is the sum of the number of types of new category samples and old category samples.

[0065] It can be understood that whenever there is a new target type, in addition to constructing a new hypersonic target class data set based on old category samples and new category samples, the embodiment of the present invention will also adjust the structure of the second target classifier according to the number of categories of new category samples in the new hypersonic target class data set, so that the number of types of samples that can be identified can adapt to the increase, and correspondingly, its parameters will also be increased by one or more columns after the parameter matrix of the most recently trained first target classifier (currently to perform the first incremental training) or the second target classifier (currently to perform non-first incremental training).

[0066] In the embodiment of the present invention, considering that the generation mechanism of the spectral radiation characteristics of the hypersonic target is the same, the first hypersonic target recognition network model and the second hypersonic target recognition network model use the same spectral feature extractor to perform deep extraction of spectral features. In practice, when each new hypersonic target category appears, a column can be added to the parameter matrix of the old target classifier for incremental learning. Of course, if multiple hypersonic target categories are added at the same time, multiple columns can also be added to the parameter matrix of the old target classifier for incremental learning.

[0067] When training the second hypersonic target recognition network model, the hypersonic target spectrum (here, the sample used for training) is obtained from the newly added class data set of the hypersonic target and input into the spectral feature extractor to extract the spectral features, and the spectral features are sent to the second target classifier for target recognition. During the training process, the cross entropy loss function is used to evaluate the model loss during the training process, so that the parameters of the second target classifier are gradually optimized while keeping the parameters of the spectral feature extractor unchanged, and finally the trained second hypersonic target recognition network model is obtained. In this way, the old knowledge is separated from the new knowledge, which ensures that the old knowledge of the recognition model is not forgotten and the new knowledge is easy to learn during the incremental learning process. Therefore, it can solve the incremental recognition problem that with the development of hypersonic aircraft, the spectral features of the newly added class are similar to those of the original class, and the recognition model needs to be updated quickly.

[0068] Among them, the cross entropy loss function is:

[0069]

[0070] In this formula, x i represents the i-th class sample, p(x i ) represents the predicted data distribution for the i-th class of samples, q(x i ) represents the true data distribution of the i-th class sample, and Loss_add is the calculated model loss.

[0071] Therefore, after the second hypersonic target recognition network model is trained, the hypersonic target spectrum of the target type to be identified (here, the spectrum of the actual hypersonic target) is input into the trained second hypersonic target recognition network. Moreover, for new types of hypersonic targets, the second hypersonic target recognition network model can also quickly and effectively identify them.

[0072] The method for rapid incremental recognition of near-space hypersonic targets provided by an embodiment of the present invention, whenever there is a new target type, uses the spectral feature extractor of the trained hypersonic target recognition network model to combine with the new second target classifier to construct a new second hypersonic target recognition network model, and trains the second hypersonic target recognition network model by adding new category samples on the basis of the existing old category samples to form a new category data set of hypersonic targets; and in the incremental update process of the second hypersonic target recognition network, the parameters of the spectral feature extractor are kept inconvenient, the incremental learning method of the target classifier is dynamically updated, the spectral feature extractor is protected for the knowledge of the learned near hypersonic targets, catastrophic forgetting is avoided, and the distribution of the new category of targets in the feature space is dynamically adapted. At the same time, since the parameters of the model update are reduced, the update time of the second hypersonic target recognition network model is effectively reduced, so that the newly added category of hypersonic targets can be quickly identified.

[0073] It is worth mentioning that in traditional incremental learning methods, such as the common fine-tune training method, training directly on a new data set without restrictions often leads to catastrophic forgetting due to problems such as inconsistent distribution of new class target data sets, that is, the error rate of old class target data sets increases sharply. Incremental learning methods based on knowledge distillation generally require access to more old data sets or joint training with old data sets, but this is often limited in practical applications and consumes a lot of time and computational costs. In addition, when the newly added class has similar features to the initial class, the recognition model is difficult to learn the knowledge of the new class target and maintain the knowledge of the old class target due to feature confusion, resulting in catastrophic forgetting. Therefore, the existing incremental recognition methods are difficult to achieve rapid model updates, and the similarity in features between the new and old class targets will also cause catastrophic forgetting in the model.

[0074] Taking into account the problems of insufficient completeness and diversity of hypersonic target data sets, the continuous growth of hypersonic target types, and the similarity between the spectral features of hypersonic targets under different target types, flight altitudes and flight speeds that easily lead to catastrophic forgetting, the embodiments of the present invention propose an incremental recognition scheme suitable for new categories of hypersonic targets in the future, which can add new samples of known classes to the existing knowledge system and enable the basic knowledge system mastered by the model to gradually evolve into a more complex and refined knowledge system, which is suitable for practical applications and can achieve rapid updates and less catastrophic forgetting.

[0075] In addition, the embodiments of the present invention also have the following beneficial effects:

[0076] 1. Design a spectral feature extractor based on DenseNet. Its dense connection characteristics promote the utilization of information in each spectral band of the target spectral features, and can extract a strong and robust feature space representation, making the target easier to classify and identify in the feature space.

[0077] 2. The classifier based on spectral angle quantity can dynamically adapt to the distribution of newly added near-space hypersonic targets in the feature space, and project the feature space onto the hypergeometric sphere, so that even the slightest differences in the targets in the feature space can be separated.

[0078] In one embodiment, the present invention manages the newly added category data set of hypersonic targets through a sample management mechanism, retains representative samples for the recognition network, can promote the fusion of new and old category target knowledge of near-space hypersonic targets, replays old category target knowledge during the update process of the recognition model, and further alleviates the catastrophic forgetting phenomenon.

[0079] Specifically, the construction method of the new hypersonic target data set includes:

[0080] (1) Determine the number of new category samples;

[0081] (2) For each category of samples in the old category samples, select the same number of samples from that category.

[0082] Here, for each category of samples in the old category samples, samples with the same number can be randomly selected from the samples of that category.

[0083] Alternatively, in a preferred implementation of the embodiment of the present invention, for each category of samples in the old category samples, selecting the same number of samples from the category of samples may include:

[0084] (2-1) Calculate the feature mean of each category of samples in the old category samples;

[0085] Specifically, for each class of samples in the old class samples, a spectral feature extractor is used to extract a feature vector of each sample in the class of samples, and a mean of all feature vectors of the class of samples is calculated as a feature mean.

[0086] (2-2) For each category of samples in the old category samples, traverse each sample in the category. If any of the samples is included in the selected sample set corresponding to the category, and the distance between the feature mean of the selected sample set and the feature mean of the category samples is the smallest, then the sample is included in the selected sample set until the number of samples in the selected sample set reaches the number.

[0087] For example, suppose the sample set consisting of the i-th type of samples is represented by x i ={x i1 ,x i2 ,…,x im}, the samples in this set all have the same category label y, the number of samples that need to be selected is K, and the set d represents the selected sample set corresponding to this type of sample.

[0088] First, the trained spectral feature extractor is used to extract each sample x ip The characteristics of {f i1 ,f i2 ,…,f im}, and calculate their class feature means

[0089] If a sample f ip When added to set d, the distance between the new set mean and the original class feature mean μ is the smallest, then the sample is added to set d, and the selected sample set can be obtained. This implementation process can be expressed as follows:

[0090]

[0091] Among them, f iq For any sample x in the selected sample set iq characteristics.

[0092] (3) Construct a new category dataset of hypersonic targets based on various samples selected from old category samples and newly added new category samples.

[0093] Specifically, various types of samples selected from the old category samples and the newly added new category samples are aggregated together to form a new category data set for hypersonic targets.

[0094] In the embodiment of the present invention, considering that the replay of old target knowledge is conducive to the integration of new and old target knowledge into the recognition model, and in order to realize the continuous incremental update of the recognition model, a mechanism for managing new samples is designed, a certain number of samples are reserved for each category, and according to the sample selection algorithm shown in the above step (2), the samples of the new target are evaluated and the samples closest to the category mean are reserved. Through such a sample management mechanism, the second hypersonic target recognition network model can take into account the review of old target knowledge and the learning of new target knowledge when updating, which is conducive to improving the recognition accuracy of the recognition model and saving the storage resources of the equipment.

[0095] The beneficial effects of the embodiments of the present invention are further illustrated by simulation experiments below.

[0096] In the first stage, the simulation data of the spectral characteristics of the existing typical blunt cone (RAMC-like configuration) and typical pointed cone (HTVⅡ-like configuration) targets in various flight states are used to simulate the embodiment of the present invention. The flight altitudes of these targets are 40km and 50km, and the flight speeds range from 10Mach to 25Mach. According to the type and flight state of hypersonic targets, 16 categories are selected as method verification, each of which contains 2000 spectral characteristic samples, and the mean spectral characteristic of each category is as follows: Figure 3 As shown. Among them, since the spectral radiation characteristics of hypersonic speed are based on the plasma sheath covering it, and the spectral radiation characteristics of the plasma sheath are closely related to the flight state, the spectral radiation characteristics of different models under the same flight state are extremely similar. The eight types of samples of HTVⅡ40km10Mach, HTVⅡ50km15Mach, RAMC40km20Mach, HTVⅡ50km25Mach, RAMC 50km15Mach, RAMC 40km10Mach, HTVⅡ40km20Mach, and RAMC50km10Mach in the above simulated near-space hypersonic target spectral characteristic samples are used as the initial class data set of hypersonic targets. 80% of the 2000 samples in each class are used as training sets, and 20% are used as test sets. The remaining 8 types of samples are used as new category samples, of which 1% are incremental learning training samples, and the rest are used as test sets.

[0097] The dimension of the simulated near-space hypersonic target spectral feature is 128-dimensional, that is, the feature dimension of the input first hypersonic target recognition network model is 128×1, and the structure of the spectral feature extractor is the same as Figure 2, so a total of 64 large-scale convolution kernels are defined, and the size of each convolution kernel is 7×1. The size of the maximum pooling layer is defined as 3×1. After the features input to the model pass through the large-scale convolution kernel, the features change from the input 1 dimension to 64 dimensions. The convolutional network includes the Dense Block and Transition Layer in Densenet. The specific sequential structure includes Dense Block 1, Transition Layer, and DenseBlock 2. Dense Block 1 has a total of 6 layers with a growth rate of 32. Dense Block 2 has a total of 12 layers with a growth rate of 32. Therefore, the number of features input to Dense Block 1 is 64, and the number of features output from Dense Block 1 is 64+32×6=256. 256 features are input into the Transition layer, and the output is 128 features. Then the number of features input into Dense Block 2 is 128, and the feature dimension of the samples after Dense Block 2 is 128+32×12=512. Define the parameters of the first target classifier as W old ∈R Ne×Nc , N e =512,N c =8, that is, the input feature number of the first target classifier is 512 dimensions, but there are 8 categories of other target types: HTVⅡ40km10Mach, HTVⅡ50km15Mach, RAMC40km20Mach, HTVⅡ50km25Mach, RAMC50km15Mach, RAM40km10Mach, HTVⅡ40km20Mach, RAMC50km10Mach.

[0098] The first hypersonic target recognition network model is trained using the cross entropy function. The gradient descent rate l during the training process r =0.01, training epoch=150, and batch_size of each training iteration is 20. Here, the training batch_size is set to 20 because the observation scene of hypersonic targets varies greatly and the observation noise has a significant impact. Therefore, the spectral features of hypersonic targets of the same type have the characteristics of "same object, different spectrum", and the intra-class differences are much greater than those of hyperspectral images in the remote sensing field. Therefore, a larger batch_size value can suppress the intra-class feature differences, so that the new classifier has universal recognition accuracy for different observation samples.

[0099] In the second stage, after training the first hypersonic target recognition network model, two categories are added: RAMC50km20Mach and RAMC40km15Mach. The structure and parameters of the spectral feature extractor remain unchanged, and the parameters of the second target classifier become W new∈R 512×(8+2) , which adds two columns compared to the parameters of the first target classifier. That is, the second target classifier can convert the input 512-dimensional features into 10 classification probabilities. Except for the parameters of the newly added columns, the remaining parameters of the second target classifier remain the same as those of the first target classifier.

[0100] Since the samples of the newly added target classes are often small in actual situations, there is a problem of unbalanced training samples. Therefore, when constructing a newly added class data set for hypersonic targets, samples with the same number of new class samples are selected from each class of samples in the old class samples to jointly constitute a newly added class data set for hypersonic targets. Among them, the two new class samples are RAMC50km20Mach and RAMC40km15Mach, and the number of their samples is 20. According to this number of samples, the sample management mechanism mentioned in the above method embodiment is used to select old class samples from the initial class data set of hypersonic targets to respectively constitute a newly added class data set for hypersonic targets together with the two new class samples.

[0101] During the training process, pytorch is used to perform incremental learning model training. In the optimizer optim.Adadelta(), only the parameters of the second target classifier are defined, and the training parameters l r =0.01, epoch=100, and the batch_size of each training iteration is 20. During the training process, the parameter update method of the second target classifier is the same as that given in the previous method embodiment. In addition, as mentioned in the previous method embodiment, when training the second hypersonic target recognition network model, the cross entropy loss function is used to evaluate the model loss during the training process.

[0102] After the second stage of incremental learning, three stages of incremental learning are performed again. Two new category targets are learned in each stage, namely {HTVⅡ40km15Mach, HTVⅡ50km10Mach}, {RAMC40km25Mach and RAMC50km25Mach}, and {HTVⅡ50km20Mach, HTVⅡ40km25Mach}.

[0103] Figure 4The target recognition results obtained in the initial stage of the above four stages, as well as the recognition results of two additional categories of targets each time, are shown. 1-16 in the confusion matrix represent: HTVⅡ40km10Mach, HTVⅡ50km15Mach, HTVⅡ40km20Mach, HTVⅡ50km25Mach, RAMC50km15Mach, RAMC40km10Mach, HTVⅡ40km20Mach, RAMC50km10Mach, HTVⅡ40km15Mach, HTVⅡ50km10Mach, RAMC40km25Mach, RAMC50km25Mach, HTVⅡ50km20Mach, HTVⅡ40km25Mach.

[0104] from Figure 4 It can be seen that the confusion matrices in the initial first stage and each subsequent incremental stage show good recognition effects, the predicted values ​​and the true values ​​are basically consistent, and the degree of confusion is extremely low, which proves the superiority of the present invention in the task of superb target recognition.

[0105] Figure 5 It is a comparison between the method of the embodiment of the present invention and several existing methods in terms of processing sample rate and time consumption, wherein the numbers 1-4 marked on the horizontal axis represent four incremental stages respectively. Figure 5 In the figure, “the proposed method” represents the method of the embodiment of the present invention, EWC is the existing Elastic Weight Consolidation method, iCaRL is the existing Incremental Classifier and Representation Learning method, LWF is the existing Learning without Forgetting method, and Funetune is a fine-tuning method, which is a baseline method that only uses the current new class data to fine-tune the model, such as adjusting the parameters of certain layers of the model or changing the learning rate. The specific implementation of this method can be found in the document “Zhou Da-Wei, Wang Qi-Wei, Qi Zhi-Hong, Ye Han-Jia, Zhan De-Chuan, Liu Ziwei. Class-Incremental Learning: A Survey. IEEE Trans Pattern Anal Mach Intell2024: 1-20. Doi: 10.1109 / TPAMI.2024.3429383”. The non-incremental update method is a method for target recognition using the model trained in the first stage of the embodiment of the present invention.

[0106] exist Figure 5In the left sub-figure (a) of , since the feature extractor parameters are kept unchanged during the incremental learning process and only the classifier parameters are updated, fewer parameters are updated during incremental learning, and the sample processing rate is faster than that of existing similar methods. Figure 5 In the right sub-figure (b) of , the time consumed for model update when incrementally learning a new target is much less than that of the non-incremental method and the iCaRL method, indicating that the embodiment of the present invention has significant advantages in sample processing efficiency and model update time, and realizes rapid update of the model in the task of super-target recognition.

[0107] Table 1 shows the recognition results of the hypersonic target recognition network model trained in the above stages, where the AA (Average Accuracy) value and Kappa value are used to evaluate the indicators of this method. The AA value and Kappa value reflect the accuracy performance of the network in identifying targets at each stage.

[0108] Table 1

[0109]

[0110] From the data described in Table 1, it can be seen that the network trained in the first stage performed well and maintained high performance as the subsequent incremental stages progressed. Although the AA value dropped from 0.9818 to 0.9422, the Kappa value remained above 0.9, which shows that the classification performance of the model is still reliable after multiple incremental stages. In addition, among the initial categories of the entire incremental update stage, HTVⅡ40km10Mach, RAMC50km15Mach, and RAMC40km10Mach maintained high recognition accuracy throughout the incremental stage, and each incremental stage did not produce large fluctuations in accuracy due to learning new target knowledge. However, after four incremental stages, the accuracy of HTVⅡ50km15Mach, RAMC40km20Mach, HTVⅡ50km25Mach, HTVⅡ40km20Mach, and RAMC50km10Mach can still be maintained at a high level, but compared with the initial stage, they have decreased by 0.0293, 0.0555, 0.0384, and 0.0283 respectively, and their accuracy fluctuations are relatively large, indicating that with the deepening of the recognition incremental update, the model has a slight degree of forgetting for these categories. This is because the inter-class differences of these spectral features are particularly small, while the intra-class differences between them are extremely large, which leads to large fluctuations in classification accuracy.

[0111] In summary, the embodiments of the present invention make full use of the way that old knowledge is fixed and new knowledge is added in columns in the classifier, the category affiliation is calculated according to the spectral angle, and the knowledge playback of some representative old class samples is realized through the sample management mechanism, and a fast incremental recognition method for near-space hypersonic targets that can meet the actual application of new class recognition of hypersonic targets, with fast updates and less catastrophic forgetting is proposed. The embodiments of the present invention can quickly adapt to the typical flight state of new near-space hypersonic target categories on the premise of contacting some old data sets and almost forget the knowledge of old types of targets, realize high-precision recognition of newly added hypersonic targets, and greatly save time and computing costs.

[0112] The embodiments of the present invention can be applied to future new hypersonic target recognition tasks. It makes up for the shortcoming that the recognition model can only process the analogy of learned targets and cannot process the knowledge of newly added target classes. It meets the need for rapid update of the recognition network and lays the foundation for applications such as near-space hypersonic target detection and recognition based on spectral features.

[0113] The method provided in the embodiment of the present invention can be applied to electronic devices. Specifically, the electronic device can be: a desktop computer, a portable computer, an intelligent mobile terminal, a server, etc. This is not limited here, and any electronic device that can implement the present invention belongs to the protection scope of the present invention.

[0114] It should be noted that the terms "first", "second", etc. are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention.

[0115] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0116] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art can understand and implement other changes to the disclosed embodiments by viewing the drawings and the disclosed content. In the description of the present invention, the term "comprising" does not exclude other components or steps, "one" or "an" does not exclude multiple situations, and "multiple" means two or more, unless otherwise clearly and specifically limited. In addition, certain measures are recorded in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0117] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, devices (equipment), or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware, which are collectively referred to as "modules" or "systems" herein. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The computer program is stored / distributed in a suitable medium, provided together with other hardware or as part of the hardware, or may take other distribution forms, such as through the Internet or other wired or wireless telecommunications systems.

[0118] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, apparatuses (devices) and computer program products of the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0119] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0120] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0121] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A method for rapid incremental recognition of near-space hypersonic targets, characterized in that: include: Training a first hypersonic target recognition network model based on an initial class data set of hypersonic targets, and using the trained first hypersonic target recognition network model to perform target recognition; the first hypersonic target recognition network model includes a spectral feature extractor and a first target classifier; Whenever there is a new target type, a second hypersonic target recognition network model is trained based on a new class data set of hypersonic targets, and the trained second hypersonic target recognition network model is used to perform target recognition; the second hypersonic target recognition network model includes the spectral feature extractor and the second target classifier; Among them, the initial class data set of hypersonic targets includes samples of multiple categories; the newly added class data set of hypersonic targets is constructed based on old category samples and new category samples; when training the second hypersonic target recognition network model, the parameters of the spectral feature extractor are maintained unchanged at the state when the training of the first hypersonic target recognition network model is completed.

2. The method for rapid incremental recognition of near-space hypersonic targets according to claim 1, characterized in that: Both the first target classifier and the second target classifier are classifiers constructed based on spectral angle theory.

3. The method for rapid incremental recognition of near-space hypersonic targets according to claim 2, characterized in that: The first target classifier is specifically used to: output a category scoring matrix according to the features extracted by the spectral feature extractor, so as to determine the target type according to the category scoring matrix; The category scoring matrix output by the first target classifier is: Score[i]=SA(f i ,w j ),j∈[1,N c ], Wherein, f is the feature extracted by the spectral feature extractor, and its subscript is used to identify different sample categories, i represents the i-th sample; w is the column in the parameter matrix, and its subscript is the column number, and different columns correspond to different target types. The parameter matrix of the first target classifier is N e ×N c Dimension; N e is the dimension of the features extracted by the spectral feature extractor, N c is the number of target types that the first target classifier can identify; the second target classifier has the same function as the first target classifier; the parameter matrix of the second target classifier is N e ×(N c +t) dimensions, where 0~N c The column parameters are the same as the 0th to Nth c The column parameters are the same, and t is the number of newly added target types.

4. The near-space hypersonic target rapid incremental recognition method according to claim 3 is characterized in that: Determining the target type according to the category scoring matrix is ​​achieved by the following formula: in, It means taking the category Label corresponding to the element with the smallest value in the category scoring matrix as the determined target type.

5. The method for rapid incremental recognition of near-space hypersonic targets according to claim 1, characterized in that: The method for constructing the new hypersonic target data set includes: Determine the sample size of new category samples; For each class of samples in the old class samples, select samples from the class of samples with the same number as the samples; The newly added category data set of hypersonic targets is constructed according to various types of samples selected from old category samples and newly added new category samples.

6. The near-space hypersonic target rapid incremental recognition method according to claim 5, characterized in that: For each category of samples in the old category samples, select samples with the same number of samples from the category samples, including: Calculate the feature mean of each category of samples in the old category samples; For each category of samples in the old category samples, traverse each sample in the category. If any of the samples is included in the selected sample set corresponding to the category of samples, and the distance between the feature mean of the selected sample set and the feature mean of the category of samples is the smallest, then the sample is included in the selected sample set until the number of samples in the selected sample set reaches the said number of samples.

7. The method for rapid incremental recognition of near-space hypersonic targets according to claim 6, characterized in that: The calculating of the feature mean of each category of samples in the old category samples includes: For each class of samples in the old class samples, the spectral feature extractor is used to extract the feature vector of each sample in the class of samples, and the mean of all feature vectors of the class of samples is calculated as the feature mean.

8. The method for rapid incremental recognition of near-space hypersonic targets according to claim 7, characterized in that: The spectral feature extractor includes a convolution layer, a maximum pooling layer and a convolution network connected in sequence; the convolution network includes a first Dense module, a transition layer and a second Dense module connected in sequence; the first Dense module and the second Dense module are both Dense Blocks in a densely connected convolution network DenseNet, and the transition layer is a Transition Layer in the DenseNet; the convolution layer includes 64 convolution kernels of a size of 7×1, and the size of the maximum pooling layer is 3×1; the first Dense module includes 6 layers with a growth rate of 32; the second Dense module includes 12 layers with a growth rate of 32.

9. The method for rapid incremental recognition of near-space hypersonic targets according to claim 3, characterized in that: When training the second hypersonic target recognition network model, the parameter update method of the second target classifier is: Among them, W old is the N of the first target classifier e ×N c dimensional parameter matrix, for The lth row element in .

10. The near-space hypersonic target rapid incremental recognition method according to claim 1, characterized in that: When training the second hypersonic target recognition network model, a cross entropy loss function is used to evaluate the model loss during the training process.

Images