Target identification method, device and equipment based on multi-platform radar RCS data

Abstract

The invention relates to a target identification method, device and equipment based on multi-platform radar RCS data, and the method comprises the steps: carrying out the coding processing of the RCS data obtained through the detection of a same target through a plurality of different platforms, and obtaining a coding sequence; a data fusion model is adopted to fuse local and global features in the RCS data through a coding sequence to obtain a first cooperative feature, and coding is performed on the basis of angle data corresponding to different platforms and the first cooperative feature based on angle information to obtain corresponding angle features; the method comprises the following steps: carrying out interaction on angle features, carrying out feature fusion by utilizing a data fusion model, carrying out interaction on data obtained after fusion again, carrying out feature fusion on interaction data, and carrying out target prediction classification according to the features obtained after fusion by adopting a classification head so as to realize target identification. By adopting the method, the targets in the RCS data of a plurality of different platforms and a plurality of angles can be fused and represented, and the recognition performance of a downstream recognition algorithm is effectively improved.

Description

Technical Field

[0001] The present application relates to the field of radar signal processing and computer intelligence technology, and in particular to a method, device and equipment for target recognition based on multi-platform radar RCS data. Background Art

[0002] Radar Auto Target Recognition (RATR) plays an important role in national security and aviation security. In RATR tasks, radar cross section (RCS) data is easy to obtain in actual combat, so it is widely used on different platforms to monitor enemy sensitive targets. However, the amount of information of target RCS data obtained by a single station is limited, and it cannot provide highly separable features for perception and recognition algorithms. Therefore, how to achieve multi-platform radar RCS collaborative characterization to obtain complete radar characteristic information of the target is an urgent problem to be solved in the field of multi-platform radar target recognition at this stage. Summary of the invention

[0003] Based on this, it is necessary to provide a target recognition method, device and equipment based on multi-platform radar RCS data, which can improve target recognition capability through multi-platform radar RCS data collaboration to address the above technical problems.

[0004] A target recognition method based on multi-platform radar RCS data, the method comprising:

[0005] Acquire RCS echo data at multiple different angles, each of the RCS echo data being obtained by detecting the same target from different radar platforms;

[0006] Preprocessing each of the RCS echo data respectively, using a signal coding model based on Markov Monte Carlo sampling to divide each of the preprocessed RCS echo data to obtain a plurality of weighted subsequences of the same length, and encoding each weighted subsequence to obtain a coding sequence;

[0007] Inputting the encoded sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature;

[0008] Inputting the angle data corresponding to each of the RCS echo data and the first cooperative feature into a coding model based on angle information to obtain an angle feature corresponding to each of the angle data;

[0009] Inputting each of the angle features into a first previous-level feature interaction model for feature interaction to obtain a first interaction feature corresponding to each of the angle features;

[0010] Inputting each of the first interactive features into a corresponding data fusion model based on an autonomous mechanism to obtain a corresponding plurality of second collaborative features;

[0011] Inputting each of the second collaborative features into a second front-level feature interaction model for feature interaction to obtain a second interaction feature corresponding to each of the second collaborative features;

[0012] A deep attention fusion model is used to fuse the second interaction features to obtain fused features, and a classification head is used to identify the target according to the fused features.

[0013] In one embodiment, when pre-processing each RCS echo data, normalization processing and center-of-gravity alignment processing are sequentially performed on each RCS echo data.

[0014] In one embodiment, encoding according to each weighted subsequence to obtain a coded sequence comprises:

[0015] Encode each weighted subsequence into a D-dimensional vector to obtain the corresponding weighted coded subsequence;

[0016] Processing each weighted coding subsequence, adding a category hint vector and a segmentation vector to the front end and the back end of each weighted coding subsequence, respectively, to obtain a processed weighted coding subsequence;

[0017] The coding sequence is obtained by constructing according to all processed weighted coding subsequences, sequence segmentation coding matrices and position coding matrices.

[0018] In one embodiment, the data fusion model based on the autonomous force mechanism includes multiple feature extraction layers connected in sequence;

[0019] Each feature extraction layer includes a first normalization layer, a multi-head attention layer, a second normalization layer, and a feedforward layer connected in sequence.

[0020] In one embodiment, the coding model based on angle information includes multiple angle guidance layers, and the input of each angle guidance layer is the angle data corresponding to the layer and the first collaborative feature. In each angle guidance layer:

[0021] Linearly encode the input angle data to obtain angle-encoded data;

[0022] Performing feature mapping on the input first collaborative feature to obtain a mapping feature;

[0023] The angle encoding data is multiplied by the mapping feature, and the multiplication result is summed with the first collaborative feature using a residual connection to obtain the angle feature output by the angle guidance layer.

[0024] In one embodiment, in the first front-stage feature interaction model and the second front-stage feature interaction model, each input data is weightedly added to all other input data to obtain the corresponding interaction feature.

[0025] In one embodiment, the target recognition method further includes:

[0026] Construct a multi-angle target recognition network according to the signal coding model based on Markov Monte Carlo sampling, the data fusion model based on the self-attention mechanism, the coding model based on the angle information, the first front-stage feature interaction model, the second front-stage feature interaction model, the deep attention fusion model and the classification head;

[0027] The pre-processed RCS echo data are input into the multi-angle target recognition network to obtain the target recognition result.

[0028] In one embodiment, when training the multi-angle target recognition network, the multi-angle target recognition network is trained using a loss function based on a cross entropy formula to construct an algorithm;

[0029] The signal coding model based on Markov Monte Carlo sampling includes adjustable weights, and the weights are used to weight each of the subsequences.

[0030] The present application also provides a target recognition device based on multi-platform radar RCS data, the device comprising:

[0031] A multi-platform data acquisition module is used to acquire RCS echo data at multiple different angles, each of which is obtained by detecting the same target from different radar platforms;

[0032] A coding sequence obtaining module is used to preprocess each of the RCS echo data respectively, divide each of the preprocessed RCS echo data using a signal coding model based on Markov Monte Carlo sampling to obtain a plurality of weighted subsequences of the same length, and encode each weighted subsequence to obtain a coding sequence;

[0033] A first collaborative feature obtaining module is used to input the coding sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature;

[0034] An angle feature acquisition module, used for inputting the angle data corresponding to each of the RCS echo data and the first collaborative feature into a coding model based on angle information to obtain an angle feature corresponding to each of the angle data;

[0035] A first interactive feature obtaining module, used for inputting each of the angle features into a first previous feature interaction model for feature interaction, to obtain a first interactive feature corresponding to each of the angle features;

[0036] A second collaborative feature obtaining module, used for inputting each of the first interactive features into a corresponding data fusion model based on an autonomous mechanism, to obtain a corresponding plurality of second collaborative features;

[0037] A second interaction feature obtaining module, used for inputting each of the second cooperative features into a second previous-level feature interaction model for feature interaction, to obtain a second interaction feature corresponding to each of the second cooperative features;

[0038] The target recognition module is used to use a deep attention fusion model to perform feature fusion on each of the second interaction features to obtain fusion features, and use a classification head to identify the target based on the fusion features.

[0039] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0040] Acquire RCS echo data at multiple different angles, each of the RCS echo data being obtained by detecting the same target from different radar platforms;

[0041] Preprocessing each of the RCS echo data respectively, using a signal coding model based on Markov Monte Carlo sampling to divide each of the preprocessed RCS echo data to obtain a plurality of weighted subsequences of the same length, and encoding each weighted subsequence to obtain a coding sequence;

[0042] Inputting the encoded sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature;

[0043] Inputting the angle data corresponding to each of the RCS echo data and the first cooperative feature into a coding model based on angle information to obtain an angle feature corresponding to each of the angle data;

[0044] Inputting each of the angle features into a first previous-level feature interaction model for feature interaction to obtain a first interaction feature corresponding to each of the angle features;

[0045] Inputting each of the first interactive features into a corresponding data fusion model based on an autonomous mechanism to obtain a corresponding plurality of second collaborative features;

[0046] Inputting each of the second collaborative features into a second front-level feature interaction model for feature interaction to obtain a second interaction feature corresponding to each of the second collaborative features;

[0047] A deep attention fusion model is used to fuse the second interaction features to obtain fused features, and a classification head is used to identify the target according to the fused features.

[0048] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0049] Acquire RCS echo data at multiple different angles, each of the RCS echo data being obtained by detecting the same target from different radar platforms;

[0050] Preprocessing each of the RCS echo data respectively, using a signal coding model based on Markov Monte Carlo sampling to divide each of the preprocessed RCS echo data to obtain a plurality of weighted subsequences of the same length, and encoding each weighted subsequence to obtain a coding sequence;

[0051] Inputting the encoded sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature;

[0052] Inputting the angle data corresponding to each of the RCS echo data and the first cooperative feature into a coding model based on angle information to obtain an angle feature corresponding to each of the angle data;

[0053] Inputting each of the angle features into a first previous-level feature interaction model for feature interaction to obtain a first interaction feature corresponding to each of the angle features;

[0054] Inputting each of the first interactive features into a corresponding data fusion model based on an autonomous mechanism to obtain a corresponding plurality of second collaborative features;

[0055] Inputting each of the second collaborative features into a second front-level feature interaction model for feature interaction to obtain a second interaction feature corresponding to each of the second collaborative features;

[0056] A deep attention fusion model is used to fuse the second interaction features to obtain fused features, and a classification head is used to identify the target according to the fused features.

[0057] The target recognition method, device and equipment based on multi-platform radar RCS data are obtained by encoding the RCS echo data obtained by detecting the same target on multiple different platforms based on Markov Monte Carlo sampling to obtain a coding sequence, and then a data fusion model based on a self-attention mechanism is used to fuse the local and global features in each RCS data through the coding sequence to obtain a first collaborative feature, and then the angle data of each platform is respectively encoded with the first collaborative feature based on the angle information to obtain the corresponding angle feature, and then the angle features are interacted and then the data fusion model is used to perform feature fusion, and the fused data are interacted again, and finally the interactive data is feature fused through a deep attention fusion model, and a classification head is used to predict and classify the target according to the fused features to achieve target recognition. The method can be used to fuse the targets in the RCS data of multiple angles of multiple different platforms, effectively improving the recognition performance of the downstream recognition algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 1 is a flow chart of a target recognition method based on multi-platform radar RCS data in one embodiment;

[0059] Figure 2 It is a flowchart of a target recognition method based on multi-platform radar RCS data in one embodiment;

[0060] Figure 3 is a flow chart of a signal encoding step based on Markov Monte Carlo sampling in one embodiment;

[0061] Figure 4 is a schematic structural diagram of an angle guiding layer in one embodiment;

[0062] Figure 5 A schematic diagram of the interaction of the front-end features in one embodiment;

[0063] Figure 6 The first is a schematic diagram of the experimental verification results of the separability of this method in the experimental simulation. Figure 6 (a) is a schematic diagram of the target echo data of the first radar station. Figure 6 (b) is a schematic diagram of the target echo data of the second radar station. Figure 6 (c) is a schematic diagram of the fusion of two radar stations;

[0064] Figure 7 It is a structural block diagram of a target recognition device based on multi-platform radar RCS data in one embodiment;

[0065] Figure 8 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0067] In the existing technology, the target RCS data obtained by single-platform radar detection is limited in the amount of target information, which cannot provide strong analytical features for the perception and recognition algorithm, resulting in unsatisfactory automatic recognition performance. Figure 1-2 As shown, a target recognition method based on multi-platform radar RCS data is provided, comprising the following steps:

[0068] Step S100, acquiring RCS echo data at multiple different angles, each RCS echo data being obtained by detecting the same target from different radar platforms;

[0069] Step S110, preprocessing each RCS echo data respectively, using a signal coding model based on Markov Monte Carlo sampling to divide each preprocessed RCS echo data into multiple weighted subsequences of the same length, and encoding each weighted subsequence to obtain a coded sequence;

[0070] Step S120, inputting the encoded sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature;

[0071] Step S130, inputting the angle data corresponding to each RCS echo data and the first cooperative feature into a coding model based on angle information respectively, to obtain an angle feature corresponding to each angle data;

[0072] Step S140, inputting each angle feature into a first previous feature interaction model for feature interaction, and obtaining a first interaction feature corresponding to each angle feature;

[0073] Step S150, inputting each first interactive feature into a corresponding data fusion model based on an autonomous mechanism to obtain a corresponding plurality of second collaborative features;

[0074] Step S160, inputting each second collaborative feature into a second previous-level feature interaction model for feature interaction, and obtaining a second interaction feature corresponding to each second collaborative feature;

[0075] Step S170, using a deep attention fusion model to fuse the second interaction features to obtain fused features, and using a classification head to identify the target based on the fused features.

[0076] In this method, based on the deep learning method, the input multi-platform RCS information is combined with the angle information, and the radar RCS data is fused at the data semantic level through the data encoding method based on the Markov Monte Carlo process, so as to achieve the collaborative representation of multi-platform radar RCS data. Experiments such as fusion feature dimensionality reduction visualization and downstream recognition task accuracy have proved that this technology can obtain the collaborative representation of RCS data sequences under multiple platforms and improve the recognition performance of downstream recognition algorithms.

[0077] In this method, a data fusion model based on the self-attention mechanism, a coding model based on angle information, and a signal coding model based on the Markov Monte Carlo process are mainly proposed. Combined with the existing advanced natural language processing model based on the self-attention mechanism, the angle information of radar RCS data under multiple platforms and the scattering characteristic information of the RCS data itself are introduced to realize the fusion representation of RCS data under multiple platforms.

[0078] In step S110 , when pre-processing each RCS echo data, normalization processing and center of gravity alignment processing are sequentially performed on each RCS echo data.

[0079] Specifically, radar echo data has the problems of azimuth sensitivity, amplitude sensitivity and translation sensitivity, which will have a great impact on the model's extraction of robust target features. In order to reduce the radar echo data's sensitivity to amplitude, the modulo-2 norm normalization is used to process the RCS echo data to obtain the amplitude-normalized radar echo data. In order to reduce the translation sensitivity of radar echo data, the center alignment method is used to perform a cyclic shift operation on the amplitude-normalized RCS echo data so that its center is located at the center of the range window, and the preprocessed RCS echo data is obtained.

[0080] Furthermore, based on the preprocessed RCS echo data, according to the different local importance of different RCS echo data sequences, they are sampled based on the Markov Monte Carlo process to obtain weighted subsequences, which are then input into a feature extraction model based on the self-attention mechanism for feature extraction.

[0081] In this embodiment, a signal coding model based on Markov Monte Carlo sampling is used to process each pre-processed RCS echo data. First, each processed RCS echo data is divided according to a preset length to obtain multiple weighted subsequences of equal length, which are expressed as:

[0082]

[0083] Here, formula (1) and formula (2) respectively represent the RCS echo data obtained by different radar platforms. Here, these two formulas do not mean that the method is only applied to two platforms. They are only used as examples. The method can be applied to multiple radar platforms.

[0084] After that, for each subsequence we have:

[0085]

[0086] In formula (3) and formula (4), and Represent the learnable weights of the two RCS sequences. The weights will be iteratively updated according to the loss function during model training.

[0087] In this embodiment, encoding each weighted subsequence to obtain a coding sequence includes: encoding each weighted subsequence into a D-dimensional vector to obtain a corresponding weighted coding subsequence, processing each weighted coding subsequence, adding a category prompt vector and a segmentation vector to the front end and the back end of each weighted coding subsequence to obtain a processed weighted coding subsequence, and constructing according to all the processed weighted coding subsequences, a sequence segmentation coding matrix and a position coding matrix to obtain a coding sequence.

[0088] Specifically, for each subsequence after division and Encode the input and encode it into a D-dimensional vector to obtain the corresponding encoded sequence x MCMC，emb With y MCMC，emb ,Right now:

[0089]

[0090]

[0091] In formula (5) and formula (6), represents the encoder, represents weighted subsequence coding, Indicates the encoded RCS sequence.

[0092] In this embodiment, the signal coding model based on Markov Monte Carlo sampling uses a linear fully connected layer to achieve encoding of the input sequence, and all weighted subsequences share one input encoder.

[0093] At the same time, in order to aggregate the information on each subsequence in the radar echo data without bias, this model adds a category prompt vector before each weighted subsequence Implements the aggregation of RCS echo data features in the deep attention feature fusion module. Also adds a segmentation vector after each weighted subsequence Used to distinguish different input sequences. And add data sequence segmentation coding E seq , providing data sequence segmentation information for the model. Secondly, in order to make full use of the spatial position relationship of the subsequences in the radar echo data, in this embodiment, the category prompt vector, the segmentation vector, x MC，emb With y MC，emb Each weighted subsequence in pos ,Right now:

[0094] z in =[z cls ;x MCMC，emb ; z sep ;y MCMC.emb ; z sep ]+ Eseq +E pos , (7)

[0095] In formula (7), Respectively represent the sequence segmentation encoding matrix and the position encoding matrix, z in represents the output of the signal coding model based on Markov Monte Carlo sampling, and also serves as the input of the feature extraction layer, that is, the coding sequence. The entire data processing process of the signal coding model based on Markov Monte Carlo sampling is as follows: Figure 3 shown.

[0096] The self-attention mechanism has context perception capability and can fully utilize the local and global correlations between radar sequences. In step S120, the data fusion model based on the self-attention mechanism is used to extract relevant information containing target scattering characteristics from each RCS echo data based on the coding sequence.

[0097] In this embodiment, the data fusion model based on the autonomous mechanism includes multiple layers of feature extraction layers connected in sequence, and each feature extraction layer includes a first normalization layer, a multi-head attention layer, a second normalization layer and a feedforward layer connected in sequence.

[0098] Specifically, the data fusion model based on the autonomous mechanism is expressed as:

[0099] T n (Z in )=T(T(...T(Z in ))), (8)

[0100] In formula (8), n represents the number of feature extraction layers included in the data fusion model, z in represents the input encoding sequence, and T(·) represents a feature extraction layer based on the self-attention mechanism.

[0101] For each feature extraction layer, the input sequence after feature extraction can be expressed as:

[0102] T(z in ) = f FFN +f MSA , (9)

[0103] In formula (9), f MSA represents the output of the multi-head attention mechanism layer in the feature extraction layer, f FFN Represents the output of the feed-forward layer in the feature extraction layer.

[0104] Among them, the multi-head attention mechanism layer f MSA It can be described as:

[0105] f MSA =Concat(head1, head2,..., head M )W O , (10)

[0106] In formula (10), M represents the number of attention heads, Concat represents the concatenation operation, Represents the recovery matrix, which is used to ensure the consistency of input and output dimensions.

[0107] Among them, the feedforward layer of the feature extraction layer includes a multi-layer perceptron, which can map the input features to a high-dimensional latent space, and then map it back to the original space from the high-dimensional latent space to achieve the extraction and screening of input features. The specific process of the feedforward layer can be defined as:

[0108]

[0109] In formula (11), is the input of the feed-forward layer, and MLP(·) represents a multilayer perceptron with two fully connected layers.

[0110] Therefore, the encoded radar RCS data is finally obtained through the data fusion model based on the autonomous mechanism to obtain the collaborative representation under multi-platform data, that is, the first collaborative feature.

[0111] In step S130, in order to introduce angle information under different platforms, an encoding model based on angle information is constructed in the present method, and the angle information is added to the output of the fusion model in the form of a feature vector. The fusion of multi-stage features is then realized through the previous feature interaction model in step S140, and finally the fused features are input into the downstream classification and recognition task.

[0112] In this embodiment, the coding model based on angle information includes multiple angle guidance layers, and the input of each angle guidance layer is the angle data corresponding to the layer and the first collaborative feature. In each angle guidance layer: linear encoding is performed on the input angle data to obtain angle coding data, feature mapping is performed on the input first collaborative feature to obtain mapping features, the angle coding data is multiplied by the mapping features, and the multiplication result is summed with the first collaborative feature using a residual connection to obtain the angle feature output by the angle guidance layer, such as Figure 4 shown.

[0113] Specifically, the angle guidance layer first linearly encodes a through the angle encoding module. The encoding module consists of two fully connected layers, and the output encoding information γ(a) can be expressed as:

[0114] γ(a)=(aW1+b1)W2+b2, (12)

[0115] In formula (12), W1, W2 and b1, b2 represent the weights and biases of the two linear transformations respectively.

[0116] Specifically, the mapping function based on the convolution module converts the first collaborative feature T output by the data fusion model based on the self-attention mechanism n (z in ) for feature mapping. The mapping function R includes two convolutional layers, a batch normalization layer and a GELU activation layer. The feature mapping process can be expressed as:

[0117] R(T n (z in ))=Conv(GELU(BN(Conv(T n (z in ))))), (13)

[0118] In formula (13), BN(·) represents batch normalization, Conv(·) represents a one-dimensional convolution operation, and GELU(·) represents a nonlinear activation layer.

[0119] Finally, the mapped features are multiplied with the angle-encoded data and combined with the input features T using a residual connection n (z in ) and sum.

[0120] Furthermore, the angle-guided layer outputs the feature f out It can be expressed as:

[0121] f out =T n (z in )+R(T n (z in )).γ(a). (14)

[0122] In order to realize the multi-stage feature fusion of the model, a front-end feature interaction model is also designed in this method to realize the multi-stage hierarchical fusion of multi-platform features in the feature extraction process. By fusing the data features of each platform, the model can associate and learn with the shallow features from other radars during shallow feature extraction.

[0123] In step S140, a multi-stage hierarchical fusion is implemented through a first front-stage feature interaction model, such as Figure 5 shown.

[0124] In this embodiment, in the first front-stage feature interaction model, each angle feature is weightedly added to all other angle features to obtain a corresponding interaction feature.

[0125] Specifically, the input data of the first front-end feature interaction model is the output of the angle guidance layer of each platform The superscripts represent the serial numbers of different platforms. For the sake of illustration, three radar platforms are taken as examples. The output of the previous feature interaction module can be expressed as:

[0126]

[0127] In formula (15), They represent the three interactive features output by the previous feature interaction module, ω 11 ,ω 12 ,ω 13 ,ω 21 ,ω 22 ,ω 23 ,ω 31 ,ω 32 ,ω 33 is the weight of each platform feature interaction.

[0128] After the first front-stage feature interaction model, the first interaction features consistent with the number of radar platforms are obtained. In step S150, the data fusion model based on the autonomous mechanism is used to fuse the first interaction features again to obtain multiple second collaborative features corresponding to the first interaction features. The data processing process in the data fusion model based on the autonomous mechanism is consistent with the post-processing process of inputting the coding sequence into the data fusion model based on the self-attention mechanism, so it will not be repeated.

[0129] Next, each second collaborative feature is passed through the second front-stage feature interaction model, and the process is consistent with the data processing process in the first front-stage feature interaction model, so it will not be repeated here.

[0130] In step S160, the second previous feature interaction model is used to perform feature interaction on each second collaborative feature to obtain a corresponding second interaction feature.

[0131] In step S170, finally in this step, the second interaction features corresponding to multiple angles obtained by processing the signal coding model based on Markov Monte Carlo sampling, the data fusion model based on the self-attention mechanism, the coding model based on angle information, the first front-level feature interaction model, and the second front-level feature interaction model are handed over to the target recognition algorithm downstream of the task for target recognition.

[0132] In this embodiment, a deep attention fusion model is first used to fuse the second interaction features to obtain fused features, and then a classification head is used to classify and predict the target based on the fused features to achieve target recognition.

[0133] In this embodiment, each of the above model structures is implemented using a learning network, so a multi-angle target recognition network can be constructed according to a signal coding model based on Markov Monte Carlo sampling, a data fusion model based on a self-attention mechanism, a coding model based on angle information, a first front-stage feature interaction model, a second front-stage feature interaction model, a deep attention fusion model, and a classification head. In this way, when identifying a target, it is only necessary to input the pre-processed RCS echo data into the multi-angle target recognition network to obtain a target recognition result.

[0134] In this embodiment, when training the multi-angle target recognition network, the multi-angle target recognition network is trained using a loss function based on a cross entropy formula to construct an algorithm. The signal coding model based on Markov Monte Carlo sampling includes adjustable weights, which are used to weight each of the subsequences.

[0135] In order to realize the training of multi-angle target recognition network, in this method, the specific downstream target recognition and classification tasks and the corresponding loss function are combined to realize the iterative update of model parameters.

[0136] Specifically, the algorithm inputs a multi-platform RCS signal sequence with a random target category combination during the training phase, and then extracts and fuses it through multiple layers of feature extraction layers based on the self-attention mechanism. Finally, the fused feature combined with the signal encoding is Input the classification head to obtain the final prediction of the RCS target category.

[0137] Assume that the target category label of the RCS sequence is Then in each batch, the sequence labels are set as:

[0138]

[0139] In formula (16), m represents the total number of sequences during training, and C represents the total number of targets. Then, the loss function of the algorithm is constructed based on the cross entropy formula, namely:

[0140]

[0141] In formula (17), C(•) represents the classification head, which is used to output the predicted probability.

[0142] Furthermore, the training process of the multi-angle target recognition network can be divided into four steps. First, the input multi-platform RCS signal sequence is encoded based on the radar echo data of Markov Monte Carlo sampling to obtain the corresponding encoding vector Secondly, the obtained encoding vector is input into the cross-domain data fusion model to obtain the final feature vector Then the obtained fusion feature vector The final fusion feature output is obtained after angle information encoding; finally, the fusion feature obtained by the model is input into the classification head of the downstream task to calculate the loss function of target recognition and classification, and the learnable parameters in the overall algorithm are gradient updated accordingly.

[0143] The test process of the multi-angle target recognition network is divided into two steps. First, the trained data fusion model is connected to the recognition and classification head of the downstream task and the parameters are adjusted in the downstream data set (fixed fusion model parameters). After the parameter adjustment is completed, the parameters of the fusion model and the recognition and classification head are fixed, and the recognition test is performed in the downstream data set.

[0144] In the above-mentioned target recognition method based on multi-platform radar RCS data, the multi-platform radar RCS data collaborative characterization technology based on Markov Monte Carlo process can realize the feature-level unified characterization of RCS signal sequences under multiple platforms, and the obtained feature characterization has good separability. The obtained multi-platform RCS signal sequence fusion features can improve the performance of the classification and recognition algorithm.

[0145] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0146] In one embodiment, Figure 7 As shown, a target recognition device based on multi-platform radar RCS data is provided, including: a multi-platform data acquisition module 200, a coding sequence acquisition module 210, a first cooperative feature acquisition module 220, an angle feature acquisition module 230, a first interactive feature acquisition module 240, a second cooperative feature acquisition module 250, a second interactive feature acquisition module 260 and a target recognition module 270, wherein:

[0147] The multi-platform data acquisition module 200 is used to acquire RCS echo data at multiple different angles, each of which is obtained by detecting the same target from different radar platforms;

[0148] The coding sequence obtaining module 210 is used to pre-process each of the RCS echo data respectively, divide each of the pre-processed RCS echo data using a signal coding model based on Markov Monte Carlo sampling to obtain a plurality of weighted subsequences of the same length, and encode each weighted subsequence to obtain a coding sequence;

[0149] A first collaborative feature obtaining module 220 is used to input the encoding sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature;

[0150] An angle feature acquisition module 230, used to input the angle data corresponding to each of the RCS echo data and the first collaborative feature into a coding model based on angle information, to obtain an angle feature corresponding to each of the angle data;

[0151] A first interaction feature obtaining module 240 is used to input each of the angle features into a first previous feature interaction model for feature interaction to obtain a first interaction feature corresponding to each of the angle features;

[0152] A second collaborative feature obtaining module 250 is used to input each of the first interactive features into a corresponding data fusion model based on an autonomous mechanism to obtain a plurality of corresponding second collaborative features;

[0153] A second interaction feature obtaining module 260 is used to input each of the second cooperative features into a second previous feature interaction model for feature interaction, and obtain a second interaction feature corresponding to each of the second cooperative features;

[0154] The target recognition module 270 is used to use a deep attention fusion model to perform feature fusion on each of the second interaction features to obtain fusion features, and use a classification head to identify the target according to the fusion features.

[0155] For the specific definition of the target recognition device based on multi-platform radar RCS data, please refer to the definition of the target recognition method based on multi-platform radar RCS data above, which will not be repeated here. Each module in the above-mentioned target recognition device based on multi-platform radar RCS data can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0156] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a target recognition method based on multi-platform radar RCS data is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0157] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0158] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0159] Acquire RCS echo data at multiple different angles, each of the RCS echo data being obtained by detecting the same target from different radar platforms;

[0160] Preprocessing each of the RCS echo data respectively, using a signal coding model based on Markov Monte Carlo sampling to divide each of the preprocessed RCS echo data to obtain a plurality of weighted subsequences of the same length, and encoding each weighted subsequence to obtain a coding sequence;

[0161] Inputting the encoded sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature;

[0162] Inputting the angle data corresponding to each of the RCS echo data and the first cooperative feature into a coding model based on angle information to obtain an angle feature corresponding to each of the angle data;

[0163] Inputting each of the angle features into a first previous-level feature interaction model for feature interaction to obtain a first interaction feature corresponding to each of the angle features;

[0164] Inputting each of the first interactive features into a corresponding data fusion model based on an autonomous mechanism to obtain a corresponding plurality of second collaborative features;

[0165] Inputting each of the second collaborative features into a second front-level feature interaction model for feature interaction to obtain a second interaction feature corresponding to each of the second collaborative features;

[0166] A deep attention fusion model is used to fuse the second interaction features to obtain fused features, and a classification head is used to identify the target according to the fused features.

[0167] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0168] Acquire RCS echo data at multiple different angles, each of the RCS echo data being obtained by detecting the same target from different radar platforms;

[0169] Preprocessing each of the RCS echo data respectively, using a signal coding model based on Markov Monte Carlo sampling to divide each of the preprocessed RCS echo data to obtain a plurality of weighted subsequences of the same length, and encoding each weighted subsequence to obtain a coding sequence;

[0170] Inputting the encoded sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature;

[0171] Inputting the angle data corresponding to each of the RCS echo data and the first cooperative feature into a coding model based on angle information to obtain an angle feature corresponding to each of the angle data;

[0172] Inputting each of the angle features into a first previous-level feature interaction model for feature interaction to obtain a first interaction feature corresponding to each of the angle features;

[0173] Inputting each of the first interactive features into a corresponding data fusion model based on an autonomous mechanism to obtain a corresponding plurality of second collaborative features;

[0174] Inputting each of the second collaborative features into a second front-level feature interaction model for feature interaction to obtain a second interaction feature corresponding to each of the second collaborative features;

[0175] A deep attention fusion model is used to fuse the second interaction features to obtain fused features, and a classification head is used to identify the target according to the fused features.

[0176] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0177] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0178] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A target recognition method based on multi-platform radar RCS data, characterized in that: The method comprises: Acquire RCS echo data at multiple different angles, each of the RCS echo data being obtained by detecting the same target from different radar platforms; Preprocessing each of the RCS echo data respectively, using a signal coding model based on Markov Monte Carlo sampling to divide each of the preprocessed RCS echo data to obtain a plurality of weighted subsequences of the same length, and encoding each weighted subsequence to obtain a coding sequence; Inputting the encoded sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature; Inputting the angle data corresponding to each of the RCS echo data and the first cooperative feature into a coding model based on angle information to obtain an angle feature corresponding to each of the angle data; Inputting each of the angle features into a first previous-level feature interaction model for feature interaction to obtain a first interaction feature corresponding to each of the angle features; Inputting each of the first interactive features into a corresponding data fusion model based on an autonomous mechanism to obtain a corresponding plurality of second collaborative features; Inputting each of the second collaborative features into a second front-level feature interaction model for feature interaction to obtain a second interaction feature corresponding to each of the second collaborative features; A deep attention fusion model is used to fuse the second interaction features to obtain fused features, and a classification head is used to identify the target according to the fused features.

2. The target recognition method according to claim 1, characterized in that: When preprocessing each of the RCS echo data, normalization processing and center of gravity alignment processing are performed on each of the RCS echo data in turn.

3. The target recognition method according to claim 2, characterized in that: The step of encoding the weighted subsequences to obtain the encoded sequence comprises: Encode each weighted subsequence into a D-dimensional vector to obtain the corresponding weighted coded subsequence; Processing each weighted coding subsequence, adding a category hint vector and a segmentation vector to the front end and the back end of each weighted coding subsequence, respectively, to obtain a processed weighted coding subsequence; The coding sequence is obtained by constructing according to all processed weighted coding subsequences, sequence segmentation coding matrices and position coding matrices.

4. The target recognition method according to claim 1, characterized in that: The data fusion model based on the autonomous mechanism includes multiple feature extraction layers connected in sequence; Each feature extraction layer includes a first normalization layer, a multi-head attention layer, a second normalization layer, and a feedforward layer connected in sequence.

5. The target recognition method according to claim 1, characterized in that: The coding model based on angle information includes a plurality of angle guidance layers, the input of each angle guidance layer is the angle data corresponding to the layer and the first collaborative feature, and in each angle guidance layer: Linearly encode the input angle data to obtain angle-encoded data; Performing feature mapping on the input first collaborative feature to obtain a mapping feature; The angle encoding data is multiplied by the mapping feature, and the multiplication result is summed with the first collaborative feature using a residual connection to obtain the angle feature output by the angle guidance layer.

6. The target recognition method according to claim 1, characterized in that: In the first front-stage feature interaction model and the second front-stage feature interaction model, each input data is weightedly added to all other input data to obtain the corresponding interaction feature.

7. The target recognition method according to any one of claims 1 to 6, characterized in that: The target recognition method further comprises: Construct a multi-angle target recognition network according to the signal coding model based on Markov Monte Carlo sampling, the data fusion model based on the self-attention mechanism, the coding model based on the angle information, the first front-stage feature interaction model, the second front-stage feature interaction model, the deep attention fusion model and the classification head; The pre-processed RCS echo data are input into the multi-angle target recognition network to obtain the target recognition result.

8. The target recognition method according to claim 7, characterized in that: When training the multi-angle target recognition network, the multi-angle target recognition network is trained using a loss function based on a cross entropy formula to construct an algorithm; The signal coding model based on Markov Monte Carlo sampling includes adjustable weights, and the weights are used to weight each of the subsequences.

9. A target recognition device based on multi-platform radar RCS data, characterized in that: The device comprises: A multi-platform data acquisition module is used to acquire RCS echo data at multiple different angles, each of which is obtained by detecting the same target from different radar platforms; A coding sequence obtaining module is used to preprocess each of the RCS echo data respectively, divide each of the preprocessed RCS echo data using a signal coding model based on Markov Monte Carlo sampling to obtain a plurality of weighted subsequences of the same length, and encode each weighted subsequence to obtain a coding sequence; A first collaborative feature obtaining module is used to input the coding sequence into a data fusion model based on a self-attention mechanism to perform feature fusion to obtain a first collaborative feature; An angle feature acquisition module, used for inputting the angle data corresponding to each of the RCS echo data and the first collaborative feature into a coding model based on angle information to obtain an angle feature corresponding to each of the angle data; A first interactive feature obtaining module, used for inputting each of the angle features into a first previous feature interaction model for feature interaction, to obtain a first interactive feature corresponding to each of the angle features; A second collaborative feature obtaining module, used for inputting each of the first interactive features into a corresponding data fusion model based on an autonomous mechanism, to obtain a corresponding plurality of second collaborative features; A second interaction feature obtaining module, used for inputting each of the second cooperative features into a second previous-level feature interaction model for feature interaction, to obtain a second interaction feature corresponding to each of the second cooperative features; The target recognition module is used to use a deep attention fusion model to perform feature fusion on each of the second interaction features to obtain fusion features, and use a classification head to identify the target based on the fusion features.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images