One-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization

Abstract

The invention provides a one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization, and belongs to the technical field of radar target recognition. The method comprises the following steps: preprocessing acquired sample one-dimensional range profile sequence data, marking real labels, and dividing according to a preset proportion to obtain a training set and a test set; constructing a one-dimensional range profile target recognition model; performing supervised iterative training on the identification model based on the training set, continuously realizing optimization of network parameters until the model converges, and obtaining an optimal model; and inputting target radar one-dimensional range profile data to be identified into the optimal model, and outputting an identification result of the radar target. The method can effectively improve the accuracy, convergence speed and robustness of target recognition.

Description

Technical Field

[0001] The present invention relates to the field of radar target recognition technology, and in particular to a one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization. Background Art

[0002] Identifying mobile target types is crucial in recognition systems. Based on the data type, mobile target recognition methods can be categorized into various approaches, including those based on one-dimensional range profiles, synthetic aperture radar images, infrared data, optical images, and multi-source data fusion. Because each type of data has unique characteristics and applicable scenarios, technicians can select the most appropriate recognition strategy based on target characteristics and actual application requirements.

[0003] A high-resolution range profile (HRRP) is a target echo signal generated by the superposition of scattering points of varying intensities in multiple independent range cells along the line of sight received by a wideband radar. It reflects the comprehensive echo characteristics of the target's key components. After the radar's transmitted signal reflects off the target, the receiver collects the echo, which is then processed through pulse compression and inverse Fourier transform to produce a one-dimensional range profile. This profile not only shows the distribution of the target's scattering centers but can also be used for target identification and classification, making it a crucial data format for feature extraction in radar target recognition systems.

[0004] As broadband radar echo data, one-dimensional range profiles exhibit amplitude, translation, and attitude sensitivity. Amplitude sensitivity reflects the influence of target reflection intensity on material and surface characteristics; translation sensitivity accounts for feature shifts caused by changes in target position; and attitude sensitivity results in significant feature differences at different angles. While these characteristics provide rich information, they can also affect feature extraction and classification stability, thereby reducing classification accuracy.

[0005] Commonly used one-dimensional range profile recognition methods can be categorized into two main categories: traditional algorithms and deep learning algorithms. Traditional algorithms, including statistical feature extraction, template matching, dimensionality reduction, and transformation methods, often rely on artificially designed features, struggle to fully capture the complexity of the target's scattering center, and are also sensitive to changes in position and posture. For example, template matching methods experience a significant drop in matching accuracy when the target's position or posture changes, while dimensionality reduction methods can lose key information during the data compression process. In contrast, deep learning algorithms utilize end-to-end learning, automatically extracting richer global and local features and reducing reliance on artificial features. Standard deep learning models, however, are limited in their ability to extract long-range temporal dependencies and local key information. Classic RNNs are prone to vanishing and exploding gradient problems during training, which in turn impacts model training and recognition performance. Furthermore, due to the increasing length and sample size of one-dimensional range profile sequences, single-layer LSTM networks are no longer sufficient for real-world tasks. They often fail to effectively exploit the long-range temporal dependencies of one-dimensional range profiles and are relatively weak in adapting to translation and posture changes. Summary of the Invention

[0006] In light of this, the present invention proposes a one-dimensional range profile target recognition method based on a deep stacked LSTM and attention optimization. This method improves the accuracy and robustness of moving target classification, effectively addressing the problem that traditional one-dimensional range profile recognition methods struggle to capture deep features in the data, resulting in insufficient representation of complex target scattering structures and neglect of key spatiotemporal correlations within the sequence. It also addresses the problem of poor model expressiveness and generalization performance when dealing with large-scale or highly complex data, which affects overall recognition accuracy and robustness.

[0007] The technical solution adopted in the present invention is:

[0008] The one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization includes the following steps:

[0009] Step 1: preprocess the collected one-dimensional range image sequence data to construct a training set;

[0010] Step 2: Build a one-dimensional range profile target recognition model, which includes an attention mechanism layer, a multi-layer stacked LSTM network, and a classifier.

[0011] Step 3: Perform supervised iterative training on the one-dimensional range profile target recognition model based on the training set to optimize the network parameters of the one-dimensional range profile target recognition model until the one-dimensional range profile target recognition model converges and obtains the optimal one-dimensional range profile target recognition model;

[0012] Step 4: input the one-dimensional range profile data of the moving target to be identified into the optimal one-dimensional range profile target recognition model to obtain a recognition result of the one-dimensional range profile data of the moving target.

[0013] Furthermore, the specific method of preprocessing in step 1 is:

[0014] Performing denoising filtering on the acquired sample one-dimensional range image data to smooth the noise in the sample one-dimensional range image data;

[0015] Perform range migration correction and range unit alignment to eliminate range offset;

[0016] Normalize the amplitude and energy of the sample one-dimensional range image data;

[0017] Highlight target features and suppress irrelevant information through filtering and image processing;

[0018] The preprocessed one-dimensional range image data is obtained.

[0019] Furthermore, the multi-layer stacked LSTM network adopts a deep stacked LSTM network structure, wherein the feature extraction layer consists of a total of 6 layers of LSTM stacked along the same time step direction to perform layer-by-layer feature abstraction on the input sequence. Each layer contains multiple LSTM feature learning units, the low layer captures basic features, and the high layer extracts global features.

[0020] Furthermore, after each layer output of the multi-layer stacked LSTM network, an attention mechanism layer is introduced to dynamically assign the importance weight of each hidden state, and the LSTM layer and the attention mechanism layer are stacked alternately;

[0021] The attention mechanism layer is used to optimize the translation-invariant features extracted by each LSTM layer. The formula is expressed as:

[0022] = , ;

[0023] ;

[0024] ;

[0025] in, Indicates that the one-dimensional distance image data sample is regarded as a time series with a time step of t. Represents one-dimensional range image data The echo strength of each range unit, Indicates the total number of distance units; represents the lth LSTM layer, Indicates the The one-dimensional range image echo intensity data is extracted by the l-th layer LSTM for translation-invariant features. Indicates the The one-dimensional range image echo intensity data is processed by the The hidden layer output after the LSTM layer performs feature extraction, Represents the translation-invariant features of a one-dimensional distance image data sample extracted by the lth layer LSTM.

[0026] Furthermore, the attention mechanism parameter learning of the attention mechanism layer is divided into three steps:

[0027] (1) Solve the weight score. The weight is the contribution of the feature vector to the model recognition. The calculation method is:

[0028] ;

[0029] in, and are all hyperparameters, is the attention matrix, which is set by random initialization, b is the bias vector of the attention map; T is the total number of time steps, which is numerically equal to N; For the The input vector of time steps, specifically the LSTM layer at time step The input vector is the hidden state of LSTM or unit status , the dimension is , Indicates the number of samples in the model batch, is the number of neurons in the LSTM hidden layer; express Energy score for current attention;

[0030] Hidden State and cell status They are represented as follows:

[0031] ;

[0032] ;

[0033] in, and Represent the hidden layer state and unit state of the l-th layer LSTM unit at the t-th time step respectively;

[0034] (2) Solve the normalized weight, that is, the weighting coefficient:

[0035] ;

[0036] in, is the time step The feature attention weights, is an exponential function, Indicates that in one attention calculation The energy fraction, Indicates the first of multiple input vectors input vectors;

[0037] (3) Weighted summation:

[0038] ;

[0039] in, represents the context vector after attention calculation at the t-th time step, that is, the feature of the entire sequence after attention weighting; Indicates the The normalized weight of the j-th input vector in the attention calculation of the time step is, Represents the time step Among the multiple input vectors input vectors.

[0040] Furthermore, the LSTM feature learning unit consists of an input gate, a forget gate, an output gate, and a memory unit. The input and output functions of the input gate, forget gate, and output gate are:

[0041] ;

[0042] ;

[0043] ;

[0044] in, is the input of the input gate, forget gate and output gate, represents the output of the input gate, represents the output of the forget gate, represents the output of the output gate, represents the Elu activation function, 、 and is the weight matrix of the input gate, forget gate and output gate for the current input, 、 and is the weight matrix of the input gate, forget gate and output gate for the previous input, 、 and Bias for input gate, forget gate, and output gate;

[0045] Let the value stored in the memory cell at the current time step be , The calculation formula is:

[0046] ;

[0047] in, represents the value stored in the memory cell at the previous time step, represents the Tanh activation function, is the weight matrix of the memory unit for the current input, is the bias of the memory unit;

[0048] The output of the LSTM feature learning unit at the current moment is:

[0049] ;

[0050] in, Indicates that the LSTM feature learning unit is at the current time step Output.

[0051] Furthermore, the classifier consists of a first fully connected layer, a ReLU activation layer, a second fully connected layer, and a Softmax layer connected sequentially. The classifier performs category recognition based on translation-invariant features and outputs the recognition result. The formula is expressed as:

[0052] ;

[0053] ;

[0054] in, represents the classifier, represents the first fully connected layer, represents the ReLU activation layer, represents the second fully connected layer, represents the Softmax layer, Represents the output recognition result.

[0055] Furthermore, the specific method of step 3 is:

[0056] The sample one-dimensional range image data in the training set are input into the one-dimensional range image target recognition model in batches for supervised training, and the recognition results of the sample one-dimensional range image data are output by the one-dimensional range image target recognition model;

[0057] The cross entropy loss function is used to calculate the loss value between the true label of the sample one-dimensional range image data and the recognition result of the sample one-dimensional range image data output by the one-dimensional range image target recognition model; the cross entropy loss function is:

[0058] ;

[0059] in, Indicates the batch size, represents the total number of categories, represents the cross entropy loss function, Indicates the The sample category is the true label of the one-dimensional distance image data of the dth category, Represents the output of the one-dimensional range image target recognition model The recognition result of the one-dimensional range image data whose sample category is the dth category, Represents the calculated loss value;

[0060] Based on the loss value, the stochastic gradient descent method is used to continuously optimize and adjust the network parameters of the one-dimensional range profile target recognition model until the one-dimensional range profile target recognition model converges and the optimal one-dimensional range profile target recognition model is obtained.

[0061] The beneficial effects of the present invention are:

[0062] 1. The present invention preprocesses sample one-dimensional range profile data to reduce signal noise interference. Using the preprocessed samples to train a one-dimensional range profile target recognition model enables the model to fully exploit the translation-invariant characteristics of moving targets, thereby effectively improving the recognition accuracy of the one-dimensional range profile target recognition model.

[0063] 2. The one-dimensional range profile target recognition model constructed by the present invention adopts a combination of a stacked LSTM network and an attention mechanism as a feature extractor. The stacked LSTM network can deeply mine the global and local dependency information in time series data, while the attention mechanism uses the forward and backward information flows to adaptively give higher weights to key features, thereby significantly improving the model's robustness to translation and posture changes of moving targets, and further enhancing recognition performance, stability and reliability.

[0064] 3. The one-dimensional range profile target recognition model constructed by the present invention has a reasonable design and a simple structure. It does not rely on artificial feature design, but makes full use of the local and global characteristics of the data to achieve rapid feature extraction and target classification. It can identify mobile targets with high precision and high efficiency, has strong adaptability, and meets the needs of mobile target recognition in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the related technologies, the following briefly introduces the drawings required for use in the embodiments of the present application or the description of the related technologies. Obviously, the drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be derived from these drawings without inventive effort.

[0066] Figure 1 Schematic diagram of the structure of the attention mechanism in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present application more apparent, the present application will be described in detail below with reference to the accompanying drawings. In the embodiments of the present invention, numerous technical details are provided to help readers better understand the present invention. However, even without these technical details and the various variations and modifications based on the following embodiments, the technical solutions claimed in the present invention can still be implemented.

[0068] A one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization is applied to electronic devices, wherein the electronic device can be a terminal or a server. In the following embodiments, the electronic device is described using a server as an example.

[0069] The following describes the implementation details of the one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization. The following content is only provided for ease of understanding and is not required for the implementation of this solution.

[0070] The method comprises the following steps:

[0071] Step 101 : pre-process the acquired sample one-dimensional range image data, and then after marking the true labels, divide the data into training sets and test sets according to a preset ratio.

[0072] In the specific implementation, the server first collects one-dimensional range image samples of multiple targets, and then preprocesses these data to remove noise and achieve data normalization; during the preprocessing process, the server smoothes the noise to enhance the separability between data categories, and performs range migration correction and range unit alignment to eliminate distance offset and make the data comparable, thereby obtaining clean and standardized one-dimensional range image data.

[0073] Ultimately, the training and test sets formed after labeling and division provide high-quality input data for the subsequent target recognition model using stacked LSTM and attention mechanism, effectively improving the model's training effect and recognition ability.

[0074] It can be understood that the preprocessing operation can make the one-dimensional range image data of the samples in the training set and the test set cleaner and more reliable. Using such a training set to train the one-dimensional range image target recognition model can improve the training quality of the one-dimensional range image target recognition model and enable the one-dimensional range image target recognition model to quickly acquire recognition capabilities.

[0075] Step 102 constructs a one-dimensional range profile object recognition model, consisting of a stacked LSTM network with an attention mechanism layer as a feature extractor and a classifier. The stacked LSTM network automatically extracts global and local features from the input data through multi-layer temporal processing, capturing the temporal dependencies between different range units. The attention mechanism layer adaptively weights key features in the hidden states of each layer based on previous and subsequent information, thereby extracting translation-invariant features. Finally, the classifier identifies the target category based on these attention-enhanced features and outputs the final recognition result.

[0076] In one example, a one-dimensional range profile recognition network specifically adopts a deep stacked LSTM network structure. The feature extraction layer consists of a total of 6 layers of LSTM stacked along the same time step direction to perform layer-by-layer feature abstraction on the input sequence. Each layer contains multiple LSTM feature learning units. The low layer captures basic features, and the high layer extracts more abstract global features. After stacking the output of each layer of the LSTM network, an attention mechanism is introduced to dynamically assign the importance weights of each hidden state to highlight the feature representation of key scattered information. LSTM layers and attention layers are stacked alternately to learn the conditional distribution of more complex abstract feature levels, better represent the structural relationship between different feature levels, and thus obtain translation-invariant features with stronger nonlinear expression capabilities.

[0077] In an example, the translation-invariant features extracted by each LSTM layer of a deep stacked LSTM attention optimization network are expressed as:

[0078] = , ;

[0079] ;

[0080] ;

[0081] in, Indicates that the one-dimensional distance image data sample is regarded as a time series with a time step of t. Represents one-dimensional range image data The echo strength of each range unit, Indicates the total number of distance units; represents the lth LSTM layer, Indicates the The one-dimensional range image echo intensity data is extracted by the l-th layer LSTM for translation-invariant features. Indicates the The one-dimensional range image echo intensity data is processed by the The hidden layer output after the LSTM layer performs feature extraction, Represents the translation-invariant features of a one-dimensional distance image data sample extracted by the lth layer LSTM.

[0082] In one example, Figure 1 As shown in Figure 2, the attention mechanism parameter learning of the attention mechanism layer can be divided into three steps:

[0083] (1) First, solve the weight score. The weight is the contribution of the feature vector to the model recognition. The calculation method is:

[0084] ;

[0085] in, and are all hyperparameters, is the attention matrix, which is set by random initialization, b is the bias vector of the attention map; T is the total number of time steps, which is numerically equal to N; For the The input vector of time steps is usually the LSTM layer at time step The input vector is the hidden state of LSTM or unit status , the dimension is , Indicates the number of samples in the model batch, is the number of neurons in the LSTM hidden layer. express The energy score of the current attention. The above hidden state and cell status They are represented as follows:

[0086] ;

[0087] ;

[0088] in, and They represent the hidden layer state and unit state of the l-th layer LSTM unit at the t-th time step respectively.

[0089] (2) Solve the normalized weight, that is, the weighting coefficient:

[0090] ;

[0091] in, is the time step The feature attention weight of is obtained by normalizing all energy scores according to the above Softmax formula; is an exponential function; Indicates that in one attention calculation energy fraction. Indicates the first of multiple input vectors input vectors.

[0092] (3) Weighted summation:

[0093] ;

[0094] in, Represents the context vector after attention calculation at the t-th time step, that is, the feature of the entire sequence after attention weighting. Indicates the The normalized weight of the j-th input vector in the attention calculation of the time step is, Represents the time step Among the multiple input vectors input vectors.

[0095] It can be understood that the model adds an attention mechanism layer after each LSTM layer outputs the feature vector, takes the hidden state and unit state of the previous LSTM layer as input, calculates weights for each region of the hidden layer feature vector, and the obtained weighted LSTM hidden state can learn the conditional distribution of more complex abstract feature levels, better represent the structural relationship between different feature levels, and thus enhance the nonlinear expression ability of features.

[0096] In an example, the LSTM feature learning unit consists of an input gate, a forget gate, an output gate, and a memory unit. The input and output functions of the input gate, forget gate, and output gate are expressed as follows:

[0097] ;

[0098] ;

[0099] ;

[0100] in, is the input of the input gate, forget gate and output gate, represents the output of the input gate, represents the output of the forget gate, represents the output of the output gate, represents the Elu activation function, 、 and is the weight matrix of the input gate, forget gate and output gate for the current input, 、 and is the weight matrix of the input gate, forget gate and output gate for the previous input, 、 and Bias for input gate, forget gate, and output gate;

[0101] Let the value stored in the memory cell at the current time step be , The calculation formula is:

[0102] ;

[0103] in, represents the value stored in the memory cell at the previous time step, represents the Tanh activation function, is the weight matrix of the memory unit for the current input, is the bias of the memory unit;

[0104] The output of the LSTM feature learning unit at the current moment is expressed by the formula:

[0105] ;

[0106] in, Indicates that the LSTM feature learning unit is at the current time step Output.

[0107] In an example, a classifier consists of a first fully connected layer, a ReLU activation layer, a second fully connected layer, and a Softmax layer connected sequentially. The classifier performs category recognition based on translation-invariant features and outputs the recognition result. The formula is expressed as:

[0108] ;

[0109] ;

[0110] in, represents the classifier, represents the first fully connected layer, represents the ReLU activation layer, represents the second fully connected layer, represents the Softmax layer, Represents the output recognition result.

[0111] Step 103 , performing supervised iterative training on the one-dimensional range profile target recognition model based on the training set to continuously optimize the network parameters of the one-dimensional range profile target recognition model until the one-dimensional range profile target recognition model converges to obtain the optimal one-dimensional range profile target recognition model.

[0112] In the specific implementation, after the server completes the construction of the one-dimensional range image target recognition model, it can perform supervised iterative training on the one-dimensional range image target recognition model based on the training set to continuously optimize the network parameters of the one-dimensional range image target recognition model until the one-dimensional range image target recognition model converges and obtains the optimal one-dimensional range image target recognition model.

[0113] In one example, the server needs to input the sample one-dimensional range image data from the training set into the one-dimensional range image target recognition model in batches for supervised training, obtaining the recognition results of the sample one-dimensional range image data output by the one-dimensional range image target recognition model. The cross-entropy loss function is then used to calculate the loss between the true label of the sample one-dimensional range image data and the recognition results of the sample one-dimensional range image data output by the one-dimensional range image target recognition model. Finally, based on the loss value, the stochastic gradient descent method is used to continuously optimize and adjust the network parameters of the one-dimensional range image target recognition model until the one-dimensional range image target recognition model converges, resulting in the optimal one-dimensional range image target recognition model.

[0114] In one example, the server uses the cross entropy loss function to calculate the loss value between the true label of the sample one-dimensional range image data and the recognition result of the sample one-dimensional range image data output by the one-dimensional range image target recognition model, which can be achieved by the following formula:

[0115]

[0116] in, Indicates the batch size, represents the total number of categories, represents the cross entropy loss function, Indicates the The sample category is the true label of the one-dimensional distance image data of the dth category, Represents the output of the one-dimensional range image target recognition model The recognition result of the one-dimensional range image data whose sample category is the dth category, Represents the calculated loss value.

[0117] In one example, a test set is used to perform various performance tests on an optimal one-dimensional range profile target recognition model. The server inputs the sample one-dimensional range profile data in the test set into the optimal one-dimensional range profile target recognition model. Based on the recognition results of the sample one-dimensional range profile data in the test set output by the optimal one-dimensional range profile target recognition model and the true labels of the sample one-dimensional range profile data in the test set, the server obtains various test indicators. If one or more tests fail, the server returns the optimal one-dimensional range profile target recognition model for retraining and network parameter tuning.

[0118] Step 104 : inputting the one-dimensional range profile data of the moving target to be identified into the optimal one-dimensional range profile target recognition model to obtain a recognition result of the one-dimensional range profile data of the moving target outputted by the optimal one-dimensional range profile target recognition model.

[0119] In a specific implementation, after obtaining the optimal one-dimensional range image target recognition model, the server can deploy the optimal one-dimensional range image target recognition model to an application scenario in need. When using the one-dimensional range image target recognition model, the one-dimensional range image data of the mobile target to be identified is input into the optimal one-dimensional range image target recognition model to obtain the recognition result of the one-dimensional range image data of the mobile target output by the optimal one-dimensional range image target recognition model.

[0120] The one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization proposed in this paper has the following positive effects:

[0121] First, after collecting the sample one-dimensional range image data, the present invention needs to perform preprocessing to reduce the interference of signal noise. A training set is constructed based on the preprocessed sample one-dimensional range image data. Using such a training set to train the one-dimensional range image target recognition model can enable the one-dimensional range image target recognition model to fully exploit the translation-invariant features of the moving target, thereby effectively improving the recognition accuracy of the one-dimensional range image target recognition model.

[0122] Second, the one-dimensional range profile target recognition model constructed by the present invention adopts a combination of stacked LSTM networks and attention mechanisms as feature extractors. The stacked LSTM networks can deeply mine the global and local dependency information in time series data, while the attention mechanism uses the forward and backward information flows to adaptively give higher weights to key features, thereby significantly improving the model's robustness to translation and posture changes of moving targets, and further enhancing recognition performance, stability and reliability.

[0123] Third, the one-dimensional range profile target recognition model constructed by the present invention has a reasonable design and a simple structure. It does not rely on artificial feature design, but makes full use of the local and global characteristics of the data to achieve rapid feature extraction and target classification. It can identify mobile targets with high precision and high efficiency, has strong adaptability, and meets the needs of mobile target recognition in complex scenarios. Therefore, it has important application value in both military and civilian fields.

[0124] The step division of the above various methods is only for the purpose of clear description. During implementation, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this application; adding insignificant modifications or introducing insignificant designs to the algorithm or process without changing the core design of the algorithm and process are all within the scope of protection of this application.

[0125] Those skilled in the art will understand that all or part of the steps in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a program. The program is stored in a storage medium and includes a number of instructions for causing a device (such as a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a USB flash drive, a mobile hard drive, ROM (Read-Only Memory), RAM (Random Access Memory), a magnetic disk, or an optical disk, etc., various media that can store program code.

[0126] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application, and that in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present application.

Claims

1. One-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization, characterized by: The following steps are involved: Step 1: preprocess the collected one-dimensional range image sequence data to construct a training set; Step 2: Build a one-dimensional range profile target recognition model, which includes an attention mechanism layer, a multi-layer stacked LSTM network, and a classifier. Step 3: Perform supervised iterative training on the one-dimensional range profile target recognition model based on the training set to optimize the network parameters of the one-dimensional range profile target recognition model until the one-dimensional range profile target recognition model converges and obtains the optimal one-dimensional range profile target recognition model; Step 4: input the one-dimensional range profile data of the moving target to be identified into the optimal one-dimensional range profile target recognition model to obtain a recognition result of the one-dimensional range profile data of the moving target.

2. The one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization according to claim 1 is characterized in that The specific method of preprocessing in step 1 is: Performing denoising filtering on the acquired sample one-dimensional range image data to smooth the noise in the sample one-dimensional range image data; Perform range migration correction and range unit alignment to eliminate range offset; Normalize the amplitude and energy of the sample one-dimensional range image data; Highlight target features and suppress irrelevant information through filtering and image processing; The preprocessed one-dimensional range image data is obtained.

3. The one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization according to claim 1 is characterized in that The multi-layer stacked LSTM network adopts a deep stacked LSTM network structure, in which the feature extraction layer consists of a total of 6 layers of LSTM stacked along the same time step direction to perform layer-by-layer feature abstraction of the input sequence. Each layer contains multiple LSTM feature learning units, the low layer captures basic features, and the high layer extracts global features.

4. The one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization according to claim 3 is characterized in that After each layer output of the multi-layer stacked LSTM network, an attention mechanism layer is introduced to dynamically assign the importance weight of each hidden state, and the LSTM layer and the attention mechanism layer are stacked alternately; The attention mechanism layer is used to optimize the translation-invariant features extracted by each LSTM layer. The formula is expressed as: = , ； ； ； in, Indicates that the one-dimensional distance image data sample is regarded as a time series with a time step of t. Represents one-dimensional range image data The echo strength of each range unit, Indicates the total number of distance units; represents the lth LSTM layer, Indicates the The one-dimensional range image echo intensity data is extracted by the l-th layer LSTM for translation-invariant features. Indicates the The one-dimensional range image echo intensity data is processed by the The hidden layer output after the LSTM layer performs feature extraction, Represents the translation-invariant features of a one-dimensional distance image data sample extracted by the l-th layer LSTM.

5. The one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization according to claim 4 is characterized in that The attention mechanism parameter learning of the attention mechanism layer is divided into three steps: (1) Solve the weight score. The weight is the contribution of the feature vector to the model recognition. The calculation method is: ； in, and are all hyperparameters, is the attention matrix, which is set by random initialization, b is the bias vector of the attention map; T is the total number of time steps, which is numerically equal to N; For the The input vector of time steps, specifically the LSTM layer at time step The input vector is the hidden state of LSTM or unit status , the dimension is , Indicates the number of samples in the model batch, is the number of neurons in the LSTM hidden layer; express Energy score for current attention; Hidden State and cell status They are represented as follows: ； ； in, and Represent the hidden layer state and unit state of the l-th layer LSTM unit at the t-th time step respectively; (2) Solve the normalized weight, that is, the weighting coefficient: ； in, is the time step The feature attention weights, is an exponential function, Indicates that in one attention calculation The energy fraction, Indicates the first of multiple input vectors input vectors; (3) Weighted summation: ； in, represents the context vector after attention calculation at the t-th time step, that is, the feature of the entire sequence after attention weighting; Indicates the The normalized weight of the j-th input vector in the attention calculation of the time step is, Represents the time step Among the multiple input vectors input vectors.

6. The one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization according to claim 4 is characterized in that The LSTM feature learning unit consists of an input gate, a forget gate, an output gate, and a memory unit. The input and output functions of the input gate, forget gate, and output gate are: ； ； ； in, is the input of the input gate, forget gate and output gate, represents the output of the input gate, represents the output of the forget gate, represents the output of the output gate, represents the Elu activation function, 、 and is the weight matrix of the input gate, forget gate and output gate for the current input, 、 and is the weight matrix of the input gate, forget gate and output gate for the previous input, 、 and Bias for input gate, forget gate, and output gate; Let the value stored in the memory cell at the current time step be , The calculation formula is: ； in, represents the value stored in the memory cell at the previous time step, represents the Tanh activation function, is the weight matrix of the memory unit for the current input, is the bias of the memory unit; The output of the LSTM feature learning unit at the current moment is: ； in, Indicates that the LSTM feature learning unit is at the current time step Output.

7. The one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization according to claim 1 is characterized in that The classifier consists of a first fully connected layer, a ReLU activation layer, a second fully connected layer, and a Softmax layer connected sequentially. The classifier performs category recognition based on translation-invariant features and outputs the recognition result. The formula is expressed as: ； ； in, represents a classifier, represents the first fully connected layer, represents the ReLU activation layer, represents the second fully connected layer, represents the Softmax layer, Represents the output recognition result.

8. The one-dimensional range profile target recognition method based on deep stacked LSTM and attention optimization according to claim 1 is characterized in that The specific method of step 3 is: The sample one-dimensional range image data in the training set are input into the one-dimensional range image target recognition model in batches for supervised training, and the recognition results of the sample one-dimensional range image data are output by the one-dimensional range image target recognition model; The cross entropy loss function is used to calculate the loss value between the true label of the sample one-dimensional range image data and the recognition result of the sample one-dimensional range image data output by the one-dimensional range image target recognition model; the cross entropy loss function is: ； in, Indicates the batch size, represents the total number of categories, represents the cross entropy loss function, Indicates the The sample category is the true label of the one-dimensional distance image data of the dth category, Represents the output of the one-dimensional range image target recognition model The recognition result of the one-dimensional range image data whose sample category is the dth category, Represents the calculated loss value; Based on the loss value, the stochastic gradient descent method is used to continuously optimize and adjust the network parameters of the one-dimensional range profile target recognition model until the one-dimensional range profile target recognition model converges and the optimal one-dimensional range profile target recognition model is obtained.

Images