Hrrp multi-scale distribution learning and recognition method based on progressive probability model

Abstract

The application discloses a HRRP multi-scale distribution learning and recognition method based on a progressive probability model, and comprises the following steps: preprocessing a to-be-recognized HRRP data set, and converting the preprocessed to-be-recognized HRRP data set into a sequence form to obtain HRRP time sequence data; performing roughening processing on the HRRP time sequence data to obtain multi-layer sequence data; wherein the multi-layer sequence data comprises data of multiple time points; inputting the multi-layer sequence data into a trained progressive probability model to obtain a target hidden state corresponding to the multi-layer sequence data; and obtaining a recognition result corresponding to the to-be-recognized HRRP data set according to the target hidden state corresponding to the multi-layer sequence data. The application can fully extract multi-level feature information in data, and thus improves the recognition accuracy.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to a multi-scale distribution learning and recognition method based on an asymptotic probability model (HRRP). Background Technology

[0002] Automatic radar target identification (ARDI) is a technology that identifies targets by analyzing and processing the radar echo signals of captured targets. It is the most difficult of the four basic functions of radar, namely detection, tracking, imaging and identification, and it is also the one with a relatively low level of maturity in terms of application technology.

[0003] High-Resolution Range Profile (HRRP) signals are the echoes of target scattering points acquired through broadband radar signals. These signals contain a wealth of target information, including the distribution of scattering points, target size, and physical structure. Peak values ​​in the HRRP signal reflect the strength of the scattering structure, while fluctuations reflect the distribution of the scattering structure along the radar line of sight. Compared to SAR images, one-dimensional HRRP data is easier to acquire and requires less storage resources, leading to its increasingly widespread application in automatic radar target identification.

[0004] For radar target recognition tasks based on HRRP (High-Resolution Radar Programming), feature extraction and mining are the most critical steps. Extracting robust and separable signals from HRRP signals using a data-driven approach is particularly important for target recognition. However, existing technologies cannot fully capture the information in the HRRP sequence and lack sufficient learning of multi-level features, thus affecting recognition performance. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a multi-scale distribution learning and recognition method for HRRP based on an asymptotic probability model.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a multi-scale distribution learning and recognition method for HRRP based on an asymptotic probability model, the method comprising: The HRRP dataset to be identified is preprocessed, and the preprocessed HRRP dataset to be identified is converted into sequence form to obtain HRRP time series data; The HRRP time series data is coarsened to obtain multi-level series data; wherein, the multi-level series data includes data from multiple time points; The multi-layer sequence data is input into the trained asymptotic probability model to obtain the target hidden state corresponding to the multi-layer sequence data; The recognition result corresponding to the HRRP dataset to be identified is obtained based on the target hidden state corresponding to the multi-layer sequence data.

[0007] Optionally, the step of preprocessing the HRRP dataset to be identified and converting the preprocessed HRRP dataset to be identified into a sequence form to obtain HRRP time series data includes: The HRRP dataset to be identified is processed by modulo 2 norm normalization to obtain the preprocessed HRRP dataset to be identified; The preprocessed HRRP dataset to be identified is converted into a sequence form using a sliding window method to obtain the HRRP time series data.

[0008] Optionally, the HRRP time series data is represented as follows: ; in, This represents the first preprocessed HRRP dataset to be identified. One HRRP sample to be identified Indicates the first One HRRP sample to be identified The corresponding HRRP time series data at each time point, Indicates the duration of the sliding window. This indicates the number of overlap distance units between two windows.

[0009] Optionally, the coarsening process of the HRRP time series data to obtain multi-level series data includes: The HRRP time series data is mapped from the time domain to the frequency domain by using Fast Fourier Transform to obtain the HRRP time series data in the frequency domain. The HRRP time series data in the frequency domain is processed by different low-pass filters to obtain the processed HRRP time series data. The processed HRRP time series data is converted back to the time domain by inverse fast Fourier transform to obtain the multilayer sequence data.

[0010] Optionally, the training process of the asymptotic probability model includes: The training samples in the training dataset are input into the asymptotic probability model, and the reconstructed training samples are obtained based on the asymptotic generation module in the asymptotic probability model. The training samples before reconstruction are input into the autoencoder variational inference module in the asymptotic probability model to obtain the hidden state corresponding to the training samples before reconstruction, and then the sample recognition result is obtained based on the hidden state corresponding to the training samples before reconstruction. A joint optimization objective is constructed based on the reconstructed training samples and the sample recognition results. The progressive probability model is then optimized based on the optimization objective until a preset condition is met, resulting in the trained progressive probability model.

[0011] Optionally, the step of inputting the multi-layer sequence data into the trained asymptotic probability model to obtain the target hidden state corresponding to the multi-layer sequence data includes: The previous time step data from the multi-layer sequence data is input into the deep dynamic network module in the asymptotic probability model to obtain the latent state of the previous time step. The current time data in the multi-layer sequence data is input into the autoencoder variational inference module in the asymptotic probability model to obtain the processing result; The hidden state at the current moment is obtained based on the processing result and the latent state at the previous moment; The target hidden state corresponding to the multi-layer sequence data is obtained based on the hidden state at each time step.

[0012] Optionally, obtaining the recognition result corresponding to the HRRP dataset to be identified based on the target hidden state corresponding to the multi-layer sequence data includes: Based on the target hidden state, a hidden feature containing the time information of the multi-layer sequence data is obtained; The recognition result is obtained based on the multi-layer sequence data and the latent features.

[0013] Secondly, the present invention provides a multi-scale distribution learning and recognition device for HRRP based on an asymptotic probability model, the device comprising: The preprocessing module is used to preprocess the HRRP dataset to be identified and convert the preprocessed HRRP dataset into a sequence form to obtain HRRP time series data. The coarsening module is used to coarsen the HRRP time series data to obtain multi-level series data; wherein the multi-level series data includes data from multiple time points. The hidden state acquisition module is used to input the multi-layer sequence data into the trained asymptotic probability model to obtain the target hidden state corresponding to the multi-layer sequence data. The identification module is used to obtain the identification result corresponding to the HRRP dataset to be identified based on the target hidden state corresponding to the multi-layer sequence data.

[0014] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: In the above technical solution, multi-layer sequence data containing multiple time points at different scales are obtained from the HRRP dataset to be identified. The multi-layer sequence data is then processed by a deep dynamic network module and an autoencoder variational inference module to obtain the target hidden state of the HRRP dataset to be identified. This fully acquires the time information in the multi-layer sequence data, considers fine-grained temporal correlation, and fully extracts multi-level feature information from the HRRP dataset to be identified, thereby improving the accuracy of identification.

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart of a multi-scale distribution learning and recognition method for HRRP based on an asymptotic probability model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a HRRP sample processing procedure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of HRRP time series data coarsening processing provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a progressive generation process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a deep dynamic process provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a variational reasoning process provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of an identification process provided by an embodiment of the present invention; Figure 8 This is a schematic diagram of a 3D model simulating a Corsair II aircraft provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of a 3D model simulating a Hawker HT7 aircraft provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of a 3D model simulating a Mirage III aircraft provided in an embodiment of the present invention; Figure 11 This is a schematic diagram of a 3D model simulating a Zancudo T38 aircraft provided in an embodiment of the present invention; Figure 12 This is a block diagram of an HRRP multi-scale distribution learning and recognition device based on an asymptotic probability model, provided in an embodiment of the present invention. Detailed Implementation

[0017] To facilitate understanding of the present invention, a brief description of the relevant prior art and the inventive concept of the present invention will be given first.

[0018] Traditional HRRP recognition methods require prior knowledge and rely on manually extracting better classification features to achieve high recognition rates. Furthermore, the ability to extract feature representations is limited. Researchers have introduced deep learning to automatically extract target features.

[0019] To better utilize the time dependency in HRRP, some scholars proposed applying recurrent neural networks (RNNs) to HRRP recognition in the paper "Radar HRRP target recognition with target aware two-dimensional recurrent neural network; procedures of the 2019 IEEE International Conference on Signal Processing, Communications and Computing (ICSPCC), F, 2019 [C]." However, RNNs are prone to the vanishing gradient problem. To address this issue, other scholars proposed a method in "VGM-RNN: HRRP sequence extrapolation and recognition based on a novel optimized RNN[J]. 2020, 8: 70071-70081." to extrapolate missing samples in the sequence, which not only effectively solves the vanishing gradient problem but also exhibits excellent noise resistance. Another scholar, in “Region-factored recurrent attentional network with deep clustering for radar HRRP target recognition[J]. Signal Processing, 2021, 183: 108010.”, combined recurrent neural networks with deep clustering mechanisms, using region decomposition and attention mechanisms to accumulate the weights of each region.

[0020] The Transformer model stands out among numerous deep neural network models due to its powerful temporal modeling capabilities. The Transformer's self-attention mechanism can not only establish bidirectional correlations between HRRP local features but also capture the connections between local and global features. Therefore, in recent years, researchers have used attention modules and feedforward neural units to construct Transformer networks and extracted deep features by stacking attention modules and feedforward neural units.

[0021] Some scholars have improved the encoding part of the Transformer model. In the paper "LightweightTransformer Network for Ship HRRP Target Recognition[J]. 2022, 12(19): 9728.", a lightweight Transformer model for ship HRRP target recognition was proposed. In the paper "Polarimetric HRRP recognition based on feature-guided Transformer model[J].2021, 57(18): 705-707.", a feature-guided Transformer model for radar polarization HRRP recognition was proposed. Artificial features were introduced into the attention module, allowing the model to focus on the range cells of HRRP with more target scattering information.

[0022] Although existing deep neural network-based HRRP sequence modeling and recognition methods have made significant improvements over traditional methods, and HRRP target recognition based on Transformer models has achieved good accuracy, the expressive power of Transformer models is limited when faced with complex HRRP time series. Because the implicit randomness in HRRP sequences is ignored, the ultimate potential of Transformer models is limited, which may result in Transformer models being unable to simulate complex HRRP sequences with complex distributions. At the same time, although probabilistic generation HRRP recognition methods based on single generation consider the implicit randomness in HRRP sequences, they are not sufficient for learning multi-level features.

[0023] To address the aforementioned technical problems, this invention proposes a multi-scale distribution learning and recognition method for HRRP based on an asymptotic probability model.

[0024] Figure 1 This is a flowchart of a multi-scale distribution learning and recognition method for HRRP based on an asymptotic probability model provided in an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps: S101. Preprocess the HRRP dataset to be identified, and convert the preprocessed HRRP dataset to be identified into a sequence form to obtain HRRP time series data.

[0025] Optionally, S101 may include: The HRRP dataset to be identified is processed by modulo 2 norm normalization to obtain the preprocessed HRRP dataset to be identified; A sliding window method was used to convert the preprocessed HRRP dataset to be identified into a sequence form, thus obtaining HRRP time series data.

[0026] Understandably, preprocessing requires normalizing the HRRP dataset to be identified using modulo 2 norm to eliminate the impact of amplitude sensitivity. Then, considering the temporal sequence of the data, a sliding window approach can be used to transform the preprocessed HRRP dataset into sequential data. Specifically, the first... A sample of HRRP to be identified can be defined as After modulo-2 norm normalization, the preprocessed HRRP samples to be identified are obtained. : ; in, This indicates the number of distance cells contained in the HRRP dataset to be identified. Indicates the first The amplitude within a distance unit; Figure 2 This is a schematic diagram of a HRRP sample processing procedure provided by an embodiment of the present invention, as shown below. Figure 2 As shown, the preprocessed HRRP dataset to be identified is transformed into HRRP time series data using the sliding time window method. It is caused by Sequential observations Composed of dimensional vectors, it can be further represented as , , Indicates the length of the time series data in the HRRP dataset to be identified. Represents a two-dimensional matrix composed of real numbers.

[0027] Therefore, HRRP time series data can be represented as follows: ; in, This represents the first preprocessed HRRP dataset to be identified. One HRRP sample to be identified Indicates the first One HRRP sample to be identified The corresponding HRRP time series data at each time point, Indicates the duration of the sliding window. This indicates the number of overlap distance units between two windows.

[0028] The time length of each HRRP time series data obtained after windowing is represented as follows: ; Where D represents a one-dimensional HRRP sample to be identified. The length.

[0029] S102. Coarsen the HRRP time series data to obtain multi-level series data; wherein, the multi-level series data includes data from multiple time points.

[0030] Optionally, S102 may include: The HRRP time series data is mapped from the time domain to the frequency domain by using the Fast Fourier Transform to obtain the HRRP time series data in the frequency domain. The HRRP time series data in the frequency domain is processed by different low-pass filters to obtain the processed HRRP time series data. The processed HRRP time series data is converted back to the time domain by inverse fast Fourier transform to obtain multi-level series data.

[0031] Figure 3 This is a schematic diagram of HRRP time series data coarsening processing provided by an embodiment of the present invention, as shown below. Figure 3 As shown, the HRRP time series data is first mapped from the time domain to the frequency domain using the Fast Fourier Transform (FFT): ,in This represents the frequency components after processing with a Fast Fourier Transform (FFT). Then, HRRP time series data in the frequency domain are used. By selecting different low-pass filters to generate multi-layer sequence data, it can be represented as follows: ; in, Represents multi-level sequence data. For multi-level input representation, In the above, the frequency sequence obtained after FFT processing is: When using a number of layers When processed by the filter, the resulting sequence is It can also be expressed in further detail as its subscript This indicates that the frequency components are retained from 1 to w / L, with the intermediate general representation term. In the middle, with The values ​​range from left to right (from 1 to 1). As the number of frequencies increases, the range of frequencies retained widens, indicating higher frequency resolution and more high-frequency information. The inverse Fast Fourier Transform and the asymptotic generative model are: Layers allow for the creation of multi-layered input representations. The number of network layers is layer, Indicates the first The input representation of the layer, Indicates from the 1st floor to the 2nd floor. The low-pass filter of the layer, Indicates the first The low-pass filter of the layer, The three here Taking the same number, the progressive generation model has multiple layers (a total of...). (Layer); Before model training or testing, it is necessary to obtain multi-layer HRRP time series data. According to the formula for obtaining multi-layer HRRP time series data, the number of layers in the multi-layer data is... Multi-layer HRRP time series data needs to go through multiple layers. Each low-pass filter is processed separately. Indicates after the first The HRRP time series data is represented in the frequency domain after low-pass filtering; then it is transformed back to the time domain by inverse fast Fourier transform to obtain multi-layer sequence data of different scales.

[0032] The purpose of coarsening the HRRP time series data is to guide each layer in the successive generation process through the likelihood distribution of the multi-layer input representation, enabling it to obtain hierarchical representations that gradually increase in depth from highly specific to progressively general. On the other hand, research based on spectral bias shows that neural networks tend to learn low-frequency (simple and smooth) functions first during training, and then gradually learn high-frequency (complex) functions. Utilizing the FFT process to generate multiple time series ranging from simple / smooth to complex allows the network to encounter these different sequences, thereby learning multi-layered and multi-complex features more comprehensively.

[0033] S103. Input the multi-layer sequence data into the trained asymptotic probability model to obtain the target hidden state corresponding to the multi-layer sequence data.

[0034] Optionally, S103 may include: The previous time step data from the multi-layer sequence data is input into the deep dynamic network module in the asymptotic probabilistic model to obtain the latent state of the previous time step. The current time-in data from the multi-level sequence data is input into the autoencoder variational inference module of the asymptotic probability model to obtain the processing result; The hidden state at the current moment is obtained based on the processing result and the latent state at the previous moment; The target hidden state corresponding to the multi-layer sequence data is obtained based on the hidden state at each time step.

[0035] The asymptotic probabilistic model includes an asymptotic generation module, a deep dynamic network module, and an autoencoder variational inference module. Figure 4 This is a schematic diagram of a progressive generation process provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of a deep dynamic process provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of a variational reasoning process provided in an embodiment of the present invention, such as... Figure 4 , Figure 5 and Figure 6 As shown, the working process of the asymptotic probabilistic model mainly includes: an asymptotic generation process, a deep dynamics process, and a variational inference process, corresponding to the asymptotic generation module, the deep dynamics network module, and the autoencoder variational inference module, respectively. The three processes are explained in detail below: 1) Asymptotic probability generation process: To account for the randomness of the HRRP time series, this invention introduces a hierarchical Gaussian distributed hidden state at each time step. And the reconstructed hierarchical Gaussian distribution of hidden states The depth of the layer is , Indicates the first Layer in Hidden state at time The dimension, for the given multi-level series data obtained through the HRRP time series data coarsening process. , Indicates the first Layer in Multilevel HRRP time series data at time points The data dimension, the reconstructed multi-layer sequence data obtained through the progressive generation process, is denoted as... , can be defined in the first The progressive generation process of layers is as follows: ; in, , , This refers to a Gaussian distribution, with mean and covariance parameters of _____. , It is the set of all learnable parameters in the progressive generation module. It captures the dependencies between different layers and passes temporal information to the probabilistic hidden state. Indicates the first Layers are used to compute hidden states. mean The weight matrix; It is in the Layers are used to compute hidden states. covariance The weight matrix; Indicates the first Layers are used to calculate latent states. mean The weight matrix; Indicates the first Layers are used to calculate latent states. covariance Weight matrix; mean Covariance It can be used to represent hidden states. The probability distribution, This is the mathematical representation of a diagonal matrix; similarly, the mean... Covariance It can be used to represent reconstructed multi-level sequence data probability distribution: .by For example, it means from arrive of The transfer matrix, Indicates the first Layer, number One HRRP sample to be identified The latent state at any given moment. This invention will... and This is incorporated into the generation process to account for time dependence and randomness. It is a non-linear activation function. These are the weights of the neural network model, and their indices. and This indicates that these network weights are used for calculation. mean or covariance . , , , The difference lies in the fact that the mean and covariance without the horizontal line are used to represent the hidden state. probability distribution: , This represents the covariance matrix; similarly, the underlined mean and covariance are used to represent the multi-level HRRP time series data to be generated. probability distribution: This invention introduces multi-layer sequence data to insert more semi-supervised information, thereby improving the quality and efficiency of the generation process. Similarly, for the first layer, the reconstructed multi-layer sequence data... The asymptotic probability can be expressed as follows: ; in, and All of these are learnable parameters.

[0036] 2) Deep dynamic process: To uncover the multi-layered time dependence of HRRP, this invention may construct a deep dynamic process based on two typical structures: recurrent neural networks (RNN) and Transformer structures, and combine it with an asymptotic probabilistic generative model.

[0037] Recurrent Neural Networks: For a given hidden state The hidden dependencies between data are inferred through the following process: ; in, , [·] indicates a join operation, Indicates will The latent state obtained from the input RNN neural network, Indicates will The latent state obtained from the input RNN neural network, subscript Indicates time, and Similarly, For learnable feedforward neural networks, Representation layer normalization utilizes multi-layer recurrent neural network blocks to summarize the multi-layer temporal information of multi-layer sequence data into latent states from both input and latent state perspectives. .

[0038] Stacked Transformers: Given the Transformer's strong ability to capture long-range dependencies and interactions, making it particularly attractive for time series modeling, this invention introduces a stacked structure into the Transformer framework, constructing stacked Transformer modules. Specifically, a multi-head self-attention mechanism block is first introduced to capture the temporal dependencies of each layer. The operation is defined as follows: ; in, express Dimensional input, This indicates the self-attention mechanism. This indicates a multi-head attention mechanism. Indicates a connection operation. Represents the vector dimension. Indicates the first An attention mechanism head, Indicates the first The query vector of the attention mechanism head. Indicates the first The key vector of the attention mechanism head. Indicates the first The value vector of the attention mechanism head. , It refers to the number of heads in a multi-head attention mechanism. This is the output of the self-attention mechanism block. For each layer, a stacked Transformer can be set for the latent state. Its structure is as follows: ; Among them, superscript Indicates the first layer, Indicates will Enter the first The latent states are obtained through a multi-head self-attention mechanism in each layer. Each layer employs a multi-head self-attention mechanism and a feedforward network block to explore the hierarchical temporal dependencies, learn expressive representations, and thus obtain the latent states of multiple layers.

[0039] 3) Variational reasoning: The purpose of inference is to map multi-level sequence data to hidden states to obtain a target hidden state containing information about the multi-level sequence data. To avoid the statistical hidden states collapsing to their priors at higher levels due to the hierarchical structure, this invention makes two modifications: i) a deterministic upward path is constructed to connect the hidden states with... arrive The multi-level representations are connected: ii) Using the parameterized mapping function of the input multi-level sequence data to supplement the hidden state information, it can be represented as follows: ; in, Indicates the first layer, The intermediate hidden state at time step. Indicates the first The mean of the hidden states of the layer. Indicates the first Standard deviation of the hidden state; and These are all learnable parameters of the inference network. For the index, the position after the comma indicates... The vector interacting with it, and what it generates after the interaction, are indicated by the position before the comma. To interact with a vector (interaction refers to the multiplication of a matrix and a vector); express Dimensions express The dimension. By combining latent features with prior knowledge of random downward paths, the model transforms the latent features from... arrive The variational posterior contraction is: ; in, Indicates the first Layer in Hidden state at time The variational distribution of is a multidimensional normal distribution, whose mean and diagonal covariance matrix are respectively derived from . and definition; As weight; This represents the transformation layer of a linear neural network. The activation function is represented by this. Based on the hierarchical structure of the inference network, the posterior of the hidden state is approximated by combining bottom-up likelihood information and bottom-up prior information, giving the proposed model a richer representation and yielding the target hidden state containing information from multi-layer sequence data.

[0040] Optionally, the training process of the asymptotic probabilistic model includes: The training samples in the training dataset are input into the asymptotic probability model, and the reconstructed training samples are obtained based on the asymptotic generation module in the asymptotic probability model. The training samples before reconstruction are input into the autoencoder variational inference module in the asymptotic probability model to obtain the hidden state corresponding to the training samples before reconstruction, and then the sample recognition result is obtained based on the hidden state corresponding to the training samples before reconstruction. A joint optimization objective is constructed based on the reconstructed training samples and the sample recognition results. The asymptotic probability model is then optimized based on the optimization objective until the preset conditions are met, resulting in a fully trained asymptotic probability model.

[0041] Understandably, referring to the incremental generation process, the reconstruction part of the optimization objective can be constructed based on the reconstructed training samples and sample recognition results, and the incremental probability model can be optimized according to the final optimization objective; firstly, the marginal likelihood is defined based on the prediction loss and reconstruction loss, as follows: ; in, Indicates marginal likelihood. The first part represents the reconstruction loss, indicating the state given the hidden state. Under the condition that the upper right corner index indicates that the hidden state is a hidden state of the first layer, the reconstructed multi-layer sequence data The probability distribution; This is the second part of the reconstruction loss, representing the state given the hidden state. Under the condition of reconstruction, the first Layer sequence data representation The probability distribution from layer 1 to Layers can be obtained Let there be such probability distributions, and multiply them together, denoted as [a_n] = [a_n] *b_n*. ; This section defines the prediction loss, which is expressed as the loss given all hidden states. Under these conditions, the predicted reconstructed multilayer sequence data The probability distribution of the hidden state. In This indicates taking time from the 1st moment to the 2nd moment. All moments of time, This indicates taking the first to the last layer. All layers of a layer; the integral symbol represents all hidden states. The points.

[0042] Under inference networks and variational distributions, the optimization objective of the model can be achieved by maximizing the lower bound of evidence on the log-marginal likelihood, which can be calculated using the following formula: ; in, It is the first item in the reconstruction part, representing the state given the hidden state. Under these conditions, observation data The log-likelihood; The second term in the reconstruction section represents the hidden state at a given level. Under the condition, the middle represents The log-likelihood; This corresponds to the prediction part. This term corresponds to the KL divergence term, representing the hidden state distribution calculated by this invention. With prior distribution KL divergence between them; It is a balancing parameter used to balance the intensity distribution of the prediction and reconstruction parts in the optimization.

[0043] S104. Obtain the recognition result corresponding to the HRRP dataset to be identified based on the target hidden state corresponding to the multi-level sequence data.

[0044] Optionally, S104 may include: Based on the target's hidden state, obtain the hidden features containing time information of multi-level sequence data; The identification results are obtained based on multi-level sequence data and latent features.

[0045] Understandable Figure 7 This is a schematic diagram of an identification process provided by an embodiment of the present invention, such as... Figure 7 As shown, the target hidden state is the hidden state at all times. This invention connects all times (i.e., from the first time to the second time) to achieve this hidden state. Hidden state at time (time) Construct a latent feature that contains information from all time points and all layers of multi-level sequence data. Specifically, the present invention can utilize hidden features The calculation is as follows: ; in, This invention integrates the hidden state information of all time points in the HRRP time series data. Since the asymptotic probabilistic model can discover the hierarchical characteristics of the HRRP time series data and learns multi-layer representations through the asymptotic generation module, the hidden states of different layers reveal different levels of abstraction. Therefore, this invention connects all hidden layers... To fully utilize multi-layer representations and construct latent features, as shown below: ; The softmax function is used to predict the label vector of the input data, as shown below: ; Using true value label distribution With predicted label distribution The cross-entropy loss between them is used as the training loss, and is expressed as follows: ; in, For a length of The vector, It represents the number of categories in the classification task. have Each of the following values ​​represents the probability of belonging to each class.

[0046] Considering the generation and classification processes, the model's classifier can be optimized by minimizing the modified loss function, which is expressed as follows: ; in, This represents the lower bound of negative evidence for the model, i.e. , It is a regularization hyperparameter.

[0047] During the training phase of the model's classifier, the predicted label distribution is used for the training data. and the distribution of true value labels Train the classifier; after training, for the test data, the predicted label distribution of the output can be obtained. The category corresponding to the category with the highest probability is taken as the recognition result.

[0048] The effects of this invention can be further illustrated by the following simulation experiments.

[0049] 1. Simulation conditions: Figure 8 This is a schematic diagram of a 3D model simulating a Corsair II aircraft provided in an embodiment of the present invention. Figure 9 This is a schematic diagram of a 3D model simulating a Hawker HT7 aircraft provided in an embodiment of the present invention. Figure 10 This is a schematic diagram of a 3D model simulating a Mirage III aircraft provided in an embodiment of the present invention. Figure 11 This is a schematic diagram of a 3D model simulating a Zancudo T38 aircraft provided in an embodiment of the present invention, as shown below. Figure 8 , Figure 9 , Figure 10 and Figure 11 As shown, the right subplot represents the HRRP datasets collected by various simulated aircraft. The vertical axis represents the amplitude, and the horizontal axis represents the length of each data point in the HRRP dataset. During the training phase, each simulated aircraft has a total of 3200 HRRP data points. This experimental dataset uniformly covers the field of view from 0° to 180°, with a fixed pitch angle of 84°. In the testing phase, the pitch angle is fixed at 86°, and the size of each HRRP data point is 256. This invention assumes that the size of the simulated aircraft and their signal regions are similar, meaning the model may be able to easily distinguish targets simply by their size. Furthermore, this invention adds additional white noise and sets the signal-to-noise ratio to 10dB to verify the robustness of the model; then, automatic HRRP radar target identification is performed on the experimental dataset.

[0050] 2. Simulation content: To evaluate the performance of this invention, the model was compared with seven baseline models using the average accuracy on the test data as the evaluation metric. Simulations were performed simultaneously on methods including Maximum Correlation Coefficient (MCC), Adaptive Gaussian Classifier (AGC), Hidden Markov Model (HMM), Fully Convolutional Network (FCN), Recurrent Neural Network Encoder (RNNenc), Attention Recurrent Neural Network (TARAN), and Transformer. WinSize was set to 16, and overlap to 8. Because the experimental dataset was much smaller than previous datasets, the training epochs were reduced from 30 to 10.

[0051] 3. Simulation results: The average accuracy of various identification methods is shown in Table 1: Table 1

[0052] The average recognition accuracy rates of MCC and AGC methods are 65.01% and 85.34%, respectively; the average accuracy rate of HMM is 87.25%; the average recognition accuracy rates of RNNenc and its two variants TARAN and RFRAN are 88.13%, 90.56% and 92.68%, respectively, while the average recognition accuracy rate of the present invention is 94.75%.

[0053] As shown in Table 1, traditional HRRP identification methods such as maximum correlation coefficient, adaptive Gaussian classifier, and hidden Markov model outperform recurrent neural networks and their variants based on deep learning, highlighting the importance of deep neural networks in automatic target identification for radar HRRP. Furthermore, methods based on hidden Markov models and fully convolutional networks outperform non-sequence pair methods, indicating that the introduction of temporal correlation contributes to HRRP target identification. The model of this invention achieves optimal performance by a significant margin, with an average target identification accuracy of 94.75%.

[0054] In the above technical solution, multi-level sequence data of different scales are obtained based on the HRRP dataset to be identified, and the multi-level sequence data is processed by a deep dynamic network module and an autoencoder variational inference module to obtain the target hidden state of the HRRP dataset to be identified. The time information in the multi-level sequence data is fully obtained, and the fine-grained temporal correlation can be considered. The multi-level feature information in the HRRP dataset to be identified is fully extracted, thereby improving the accuracy of identification.

[0055] Figure 12 This is a block diagram of an HRRP multi-scale distribution learning and recognition device based on an asymptotic probability model provided in an embodiment of the present invention, as shown below. Figure 12 As shown, the device 1200 may include: Preprocessing module 1201 is used to preprocess the HRRP dataset to be identified and convert the preprocessed HRRP dataset to be identified into a sequence form to obtain HRRP time series data; The coarsening module 1202 is used to coarsen the HRRP time series data to obtain multi-level series data; wherein, the multi-level series data includes data from multiple time points; The hidden state acquisition module 1203 is used to input multi-layer sequence data into the trained asymptotic probability model to obtain the target hidden state corresponding to the multi-layer sequence data. The recognition module 1204 is used to obtain the recognition result corresponding to the HRRP dataset to be recognized based on the target hidden state corresponding to the multi-layer sequence data.

[0056] For details on the specific processing procedures of each module of the device, please refer to the relevant content in the first section, which will not be repeated here.

[0057] 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 such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders 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. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0058] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0059] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0060] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0061] It should be noted that the apparatus of the present invention is an apparatus that applies the above-mentioned HRRP multi-scale distribution learning and recognition method based on the progressive probability model. Therefore, all embodiments of the above-mentioned HRRP multi-scale distribution learning and recognition method based on the progressive probability model are applicable to this apparatus and can achieve the same or similar beneficial effects.

[0062] This invention is described with reference to flowchart illustrations and / or block diagrams of methods and apparatus according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0065] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A multi-scale distribution learning and recognition method for HRRP based on an asymptotic probability model, characterized in that, The method includes: The HRRP dataset to be identified is preprocessed, and the preprocessed HRRP dataset to be identified is converted into sequence form to obtain HRRP time series data; The HRRP time series data is coarsened to obtain multi-level series data; wherein, the multi-level series data includes data from multiple time points; The multi-layer sequence data is input into the trained progressive probability model to obtain the target hidden state corresponding to the multi-layer sequence data; wherein, the progressive probability model includes a progressive generation module, a deep dynamic network module, and an autoencoder variational inference module; the deep dynamic network module is a recurrent neural network or a stacked Transformer; the progressive generation module reconstructs the multi-layer sequence data by introducing a hierarchical Gaussian distribution hidden state and a reconstructed hierarchical Gaussian distribution hidden state at each time step; The recognition result corresponding to the HRRP dataset to be identified is obtained based on the target hidden state corresponding to the multi-layer sequence data. The step of coarsening the HRRP time series data to obtain multi-level series data includes: The HRRP time series data is mapped from the time domain to the frequency domain by using Fast Fourier Transform to obtain the HRRP time series data in the frequency domain. The HRRP time series data in the frequency domain is processed by different low-pass filters to obtain the processed HRRP time series data; wherein, the different low-pass filters are low-pass filters that gradually increase the frequency of retention as the level increases. The processed HRRP time series data is converted back to the time domain by inverse fast Fourier transform to obtain the multilayer sequence data; The step of inputting the multi-layer sequence data into the trained asymptotic probability model to obtain the target hidden state corresponding to the multi-layer sequence data includes: The previous time step data from the multi-layer sequence data is input into the deep dynamic network module in the asymptotic probability model to obtain the latent state of the previous time step. The current time data in the multi-layer sequence data is input into the autoencoder variational inference module in the asymptotic probability model to obtain the processing result; The hidden state at the current moment is obtained based on the processing result and the latent state at the previous moment; The target hidden state corresponding to the multi-layer sequence data is obtained based on the hidden state at each time step.

2. The HRRP multi-scale distribution learning and recognition method based on an asymptotic probability model according to claim 1, characterized in that, The process involves preprocessing the HRRP dataset to be identified and converting the preprocessed HRRP dataset into a sequence format to obtain HRRP time series data, including: The HRRP dataset to be identified is processed by modulo 2 norm normalization to obtain the preprocessed HRRP dataset to be identified; The preprocessed HRRP dataset to be identified is converted into a sequence form using a sliding window method to obtain the HRRP time series data.

3. The HRRP multi-scale distribution learning and recognition method based on an asymptotic probability model according to claim 2, characterized in that, The HRRP time series data is represented as follows: ； in, This represents the first preprocessed HRRP dataset to be identified. One HRRP sample to be identified Indicates the first One HRRP sample to be identified The corresponding HRRP time series data at each time point, Indicates the duration of the sliding window. This indicates the number of overlap distance units between two windows.

4. The HRRP multi-scale distribution learning and recognition method based on an asymptotic probability model according to claim 1, characterized in that, The training process of the asymptotic probability model includes: The training samples in the training dataset are input into the asymptotic probability model, and the reconstructed training samples are obtained based on the asymptotic generation module in the asymptotic probability model. The training samples before reconstruction are input into the autoencoder variational inference module in the asymptotic probability model to obtain the hidden state corresponding to the training samples before reconstruction, and then the sample recognition result is obtained based on the hidden state corresponding to the training samples before reconstruction. A joint optimization objective is constructed based on the reconstructed training samples and the sample recognition results. The progressive probability model is then optimized based on the optimization objective until a preset condition is met, resulting in the trained progressive probability model.

5. The HRRP multi-scale distribution learning and recognition method based on an asymptotic probability model according to claim 1, characterized in that, The step of obtaining the recognition result corresponding to the HRRP dataset to be identified based on the target hidden state corresponding to the multi-layer sequence data includes: Based on the target hidden state, a hidden feature containing the time information of the multi-layer sequence data is obtained; The recognition result is obtained based on the multi-layer sequence data and the latent features.

6. A multi-scale distribution learning and recognition device for HRRP based on an asymptotic probability model, characterized in that, The device includes: The preprocessing module is used to preprocess the HRRP dataset to be identified and convert the preprocessed HRRP dataset into a sequence form to obtain HRRP time series data. The coarsening module is used to coarsen the HRRP time series data to obtain multi-layer sequence data; wherein the multi-layer sequence data includes data from multiple time points; wherein the asymptotic probability model includes an asymptotic generation module, a deep dynamic network module, and an autoencoder variational inference module. The hidden state acquisition module is used to input the multi-layer sequence data into the trained progressive probability model to obtain the target hidden state corresponding to the multi-layer sequence data; the deep dynamic network module is a recurrent neural network or a stacked Transformer; the progressive generation module reconstructs the multi-layer sequence data by introducing hierarchical Gaussian distributed hidden states and reconstructed hierarchical Gaussian distributed hidden states at each time step. The identification module is used to obtain the identification result corresponding to the HRRP dataset to be identified based on the target hidden state corresponding to the multi-layer sequence data; The coarsening module is further configured to map the HRRP time series data from the time domain to the frequency domain using a fast Fourier transform to obtain the HRRP time series data in the frequency domain. The HRRP time series data in the frequency domain is processed by different low-pass filters to obtain the processed HRRP time series data; wherein, the different low-pass filters are low-pass filters that gradually increase the frequency of retention as the level increases. The processed HRRP time series data is converted back to the time domain by inverse fast Fourier transform to obtain the multilayer sequence data; The hidden state acquisition module is also used to input the previous time step data in the multi-layer sequence data into the deep dynamic network module in the asymptotic probability model to obtain the previous time step's latent state. The current time data in the multi-layer sequence data is input into the autoencoder variational inference module in the asymptotic probability model to obtain the processing result; The hidden state at the current moment is obtained based on the processing result and the latent state at the previous moment; The target hidden state corresponding to the multi-layer sequence data is obtained based on the hidden state at each time step.

Images