A Radar Pulse Sequence Recognition Method Based on Self-Supervised Temporal Convolutional Network

Abstract

The present invention discloses a method for identifying radar pulse sequences based on a self-supervised temporal convolutional network. The method uses continuous pulse parameter features to identify radar emitters, and mines and models a large number of pulse parameter sequences through a data-driven deep learning method. First, a temporal convolutional deep neural network is designed according to the characteristics of the pulse sequence, then the self-supervised temporal convolutional network is used to extract features and train the network for the pulse sequence, and finally a trained pulse sequence classifier is obtained for identifying the type of emitter. The feature extraction structure in the temporal convolutional network model designed by the present invention can effectively extract the change features of the input pulse sequence parameters, and output the type of each pulse / pulse sequence according to the sequence relationship, improving the recognition effect of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of radar emitter pulse sequence recognition, and particularly to a radar pulse sequence recognition method and system based on a self-supervised temporal convolutional network. Background Art

[0002] Using the characteristic parameters of radar emitter signals obtained by reconnaissance to identify emitters plays an important role in electronic warfare.

[0003] Traditional radar emitter signal recognition methods often require building a template matching database, which not only has high requirements for professional knowledge, but also the manually established parameter database cannot guarantee applicability to the recognition of most types of signals, and the recognition accuracy and speed cannot be balanced. Traditional machine learning methods rely on a large amount of manual feature extraction and prior knowledge, and it is also difficult to handle time series problems. Summary of the Invention

[0004] In view of the above defects, the present invention designs a distributed multi-functional sensor resource management method based on the timing of function switching, and designs a deep convolutional network for pulse sequences to complete emitter recognition.

[0005] To achieve the above object, a radar pulse sequence recognition method based on a self-supervised temporal convolutional network adopted by the present invention includes the following steps:

[0006] S1: Sequence data acquisition;

[0007] S2: Abnormal pulse elimination;

[0008] S3: Design of a self-supervised temporal convolutional classification network; wherein, the network includes a temporal convolutional encoder, a temporal convolutional decoder, and a temporal target classifier

[0009] S4: Model training; wherein, the original pulse parameter sequence first passes through the temporal convolutional encoder, and high-dimensional temporal feature vectors are extracted through multiple layers of deep neural networks, and then these high-dimensional vectors are sent to the temporal convolutional decoder and the temporal target classifier

[0010] S5: Model recognition; wherein, the trained temporal convolutional encoder and the temporal target classifier are selected to form a temporal convolutional network for recognition.

[0011] Optionally, the step S1 specifically includes: receiving the pulse signals of a radar emitter for a period of time through a radar reconnaissance device, and constructing a data set using the parameter measurement data of these pulses.

[0012] Optionally, the data set includes pulse descriptor sequences and other target features.

[0013] Optionally, step S2 specifically includes: judging pulses with measurement error exceeding the tolerance or being abnormal according to system characteristics and requirements, and removing or normalizing the pulses that affect training and recognition.

[0014] Optionally, in the step S3, the encoder and the decoder have a symmetric network structure, which are respectively composed of an input layer, an output layer and a temporal convolutional layer. The multi-layer deep neural network structure is used to extract complex high-dimensional temporal features of the pulse parameter sequence.

[0015] Optionally, in the step S3, the temporal convolutional layer is stacked by multiple temporal convolutional residual modules. The internal of the residual module adopts a one-dimensional dilated convolutional structure design to model the dependency relationship before and after the temporal information. Different residual blocks are set with different dilation coefficients to accelerate the calculation between neural network layers.

[0016] Optionally, in step S4, after the high-dimensional vector is sent into the temporal target classifier, the target classification result is obtained, and the cross-entropy loss is calculated with the label information of the sequence. where n is the number of classifications, y i is the label value, and y i ′ is the predicted value output by the neural network; the high-dimensional vector is sent into the temporal convolutional decoding network to reconstruct the original pulse parameter sequence, and the reconstruction error loss is calculated with the original pulse sequence. where m is the number of batch samples, x i is the input pulse parameter sequence, and x i ′ is the reconstructed pulse parameter sequence.

[0017] Optionally, in step S5, after the model training is completed, the trained temporal convolutional encoder and the temporal target classifier are selected to form a temporal convolutional network. During the recognition process, only the trained temporal convolutional encoder is used to calculate the high-dimensional features of the pulse sequence, and then the temporal target classifier is used to calculate the output probabilities of each type to obtain the final recognition result.

[0018] To achieve the above object, the present application also provides a radar pulse sequence recognition system based on a self-supervised temporal convolutional network. The system includes:

[0019] A sequence data acquisition module, used for sequence data acquisition;

[0020] An abnormal pulse removal module, used for abnormal pulse removal;

[0021] A self-supervised temporal convolutional classification network design module, used for self-supervised temporal convolutional classification network design; wherein, the network includes a temporal convolutional encoder, a temporal convolutional decoder and a temporal target classifier

[0022] A model training module for model training. In this module, the original pulse parameter sequence first passes through a temporal convolutional encoder, and high-dimensional temporal feature vectors are extracted through a multi-layer deep neural network. Then, these high-dimensional vectors are fed into a temporal convolutional decoder and a temporal target classifier.

[0023] A model recognition module for model recognition. In this module, a trained temporal convolutional encoder and a temporal target classifier are selected to form a temporal convolutional network for recognition.

[0024] Compared with the prior art, the present invention has the following beneficial effects: Compared with the single-pulse recognition method, this method can make full use of the temporal correlation information in the process of pulse parameter change, and the radiation source recognition model has higher discrimination. Compared with the traditional parameter matching method, the method of this patent has stronger robustness and better adaptability to noise and missing values in the sequence. Compared with traditional sequence deep learning models such as RNN / LSTM, this method effectively avoids gradient vanishing or gradient explosion during the training process through residual connections and dilated convolutions. In the case of long-time sequence input, the network parameters of the weight-sharing TCN of this patent are fewer, and the parallel processing mode also makes the memory usage lower, greatly reducing the computational amount and storage space. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the self-supervised temporal convolutional network recognition process;

[0026] Figure 2 It is a schematic diagram of the self-supervised temporal convolutional classification network;

[0027] Figure 3 It is a flowchart of pulse sequence recognition;

[0028] Figure 4 It is a graph of the change in the recognition accuracy of the self-supervised temporal convolutional network;

[0029] Figure 5 It is a graph of the change in the training loss function of the self-supervised temporal convolutional network.

[0030] The implementation, functional features, and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0032] In order to make the object, technical solution, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0033] The pulse parameter sequence contains rich timing features, such as pulse width, pulse repetition interval, frequency and other information. Compared with single pulse parameters, the features of timing changes can better reflect the change trends of various parameters over a period of time, avoiding the influence of dynamic errors caused by single-point parameter measurement. Therefore, using the pulse parameter sequence for identification is a feasible method.

[0034] In order to further extract the sequence features of pulse parameters, the present invention introduces an improved self-supervised deep time convolutional network into the identification of radiation source signals. The input of this model is the original pulse parameter time series. The high-dimensional feature extraction and classification identification processes are both carried out in the network, avoiding the incompleteness of manually selected features. At the same time, the parallel feature of the time convolutional network is used to solve the problem of excessive computational complexity, and it can process data sets with a large number of samples. The network can autonomously learn and extract pulse sequence features, can adapt to various types of signals, and the model training method is simple, without the need for too much domain knowledge modeling. In addition, the feature extraction structure in the time convolutional network model designed by the present invention can effectively extract the change features of the input pulse sequence parameters, and output the type of each pulse / pulse sequence according to the sequence relationship, improving the identification effect of the model.

[0035] In this embodiment, as Figure 1 shown, in order to fully extract the timing features of the pulse parameter sequence and reduce the influence of random errors and detection errors on the identification of single pulse parameters, the present invention designs a deep convolutional network for pulse sequences to complete radiation source identification. The specific implementation steps are as follows:

[0036] Obtaining sequence data: First, the pulse signals of a radar radiation source are received through radar reconnaissance equipment for a period of time, and a data set is constructed using the parameter measurement data of these pulses, which usually includes a pulse description word sequence (PDW), and other features can also be added according to needs;

[0037] Removing abnormal pulses: According to system characteristics and requirements, pulses with excessive measurement errors and abnormalities are judged, and the pulses that affect training and identification are removed or normalized, optimizing the quality of the pulse sequence and reducing the influence of abnormal pulses on the sequence.

[0038] Design of self-supervised time convolutional classification network: Design a self-supervised time convolutional classification network with the Figure 2 shown structure for training the classifier model.

[0039] The network includes a temporal convolutional encoder, a temporal convolutional decoder, and a temporal target classifier. The encoder and decoder have symmetric network structures, each consisting of an input layer, an output layer, and temporal convolutional layers. The multi-layer deep neural network structure is used to extract complex high-dimensional temporal features of the pulse parameter sequence. The temporal convolutional layer is composed of multiple stacked temporal convolutional residual modules, and the internal structure of the residual module adopts a one-dimensional dilated convolutional design to model the dependencies before and after temporal information. Different residual blocks are set with different dilation coefficients to accelerate the calculation between neural network layers. Dilated convolution samples the input of the convolution at intervals, which is equivalent to skipping some inputs to change the size of the convolution kernel. This not only enables the TCN to obtain a large receptive field with fewer layers to extract information, but also effectively reduces complex network parameters and computational volume. Using the changes in the PDW sequence over a period of time, jointly determine the recognition output of the temporal target classifier. The temporal target classifier consists of an input layer, a fully connected layer, and an output layer, and is used to reduce the dimension and classify the high-dimensional pulse sequence features, and output the final classification result.

[0040] Model training: The original pulse parameter sequence first passes through the temporal convolutional encoder, and high-dimensional temporal feature vectors are extracted through multiple layers of deep neural networks. Then, this high-dimensional vector is fed into two branch networks for supervised training. The first branch is passed into the temporal target classifier to obtain the target classification result, and the cross-entropy loss is calculated with the label information of the sequence. where n is the number of classifications, y i is the label value, and y i ′ is the predicted value output by the neural network. The other branch is passed into the temporal convolutional decoding network to reconstruct the original pulse parameter sequence, and the reconstruction error loss is calculated with the original pulse sequence. where m is the number of batch samples, x i is the input pulse parameter sequence, and x i ′ is the reconstructed pulse parameter sequence. This branch mainly further enhances the ability of the extracted high-dimensional features to depict the original temporal information through a self-supervised method. The loss functions of the two branches are combined into the loss function of the entire training network Loss = L CE +L MSE . Through the backpropagation of the neural network gradient, the overall loss function is minimized, and a convergent model is obtained through multiple iterations of training.

[0041] Model recognition: After the model training is completed, select the trained temporal convolutional encoder and temporal target classifier to form a temporal convolutional network. Figure 3 The process and flowchart for the recognition of the entire pulse sequence are as follows:

[0042] During the recognition process, only the trained temporal convolutional encoder is used to calculate the high-dimensional features of the pulse sequence, and then the temporal object classifier is used to calculate the output probabilities of each type to obtain the final recognition result.

[0043] The present invention adopts a pulse sequence recognition method based on a temporal convolutional network. This method designs a self-supervised temporal convolutional classification network, which can fully extract the variation characteristics of the pulse parameter (PDW) sequence in the time dimension, implicitly depicts the variation law of PDW in time in a data-driven manner, solves the over-reliance on parameter values in single-pulse recognition methods, makes up for the deficiency that single-pulse recognition methods cannot model the temporal variation law, and at the same time avoids the influence of single-point detection errors on the recognition result.

[0044] At the same time, the dual-branch training process adopted by this method improves the robustness of feature extraction through the dual supervision training of the pulse sequence and data labels. The unique causal dilated convolution structure in the designed temporal convolutional network outputs rich long-term tracking information, can obtain longer sequence dependencies, and effectively models the variation of PDW. At the same time, compared with the sequential processing architecture of the recurrent neural network RNN, the weight sharing characteristic of the convolutional network can regard the PDW sequence data as a whole and process all the data in it in parallel, greatly reducing the calculation time of the model.

[0045] In some specific examples, the present invention proposes a pulse sequence recognition method based on a self-supervised temporal convolutional network, which can effectively and quickly identify the type of pulse signal by using the pulse parameter sequence data.

[0046] To prove the effectiveness of this method, in this embodiment, the proposed method is verified with measured data in an electronic reconnaissance project. First, the measured data of 10 different types of radars are used to make a data set according to the methods of steps (1) and (2) of the invention, and the total sample set is divided into a training set and a test set, where the training set accounts for 75% and the test set accounts for 25%. The training set contains 30,000 PDW pulse data, including main parameter information such as PRI, frequency, and pulse width.

[0047] Then, according to Figure 2 the shown network structure, a self-supervised temporal convolutional deep neural network is developed and designed. The number of residual blocks in the temporal convolutional layer is set to 4, the temporal feature vector is 64-dimensional, and the pulse sequence length is 25. The network model is trained with a large number of pulse data parameter features, and trained for 200 epochs until the network converges. The change of recognition accuracy and loss function during the training process of the self-supervised temporal convolutional network is as Figure 4 、 Figure 5 shown.

[0048] After training is completed, according to Figure 1The processing flow processes the test set and counts the recognition results. At the same time, comparative experiments are carried out on the same test set using a variety of mainstream machine learning and deep learning models. The recognition effects of various models are shown in Table 1.

[0049] Among them, the models based on traditional machine learning methods have poor recognition performance for pulse sequences and cannot depict the temporal relationship between sequences. The methods based on sequence deep learning models have certain recognition effects, and the pulse sequence recognition method based on self-supervised temporal convolutional network shows good recognition performance.

[0050] In summary, the method proposed in the present invention can fully extract the changing characteristics of the pulse parameter sequence in terms of time series, and accordingly achieve fast and accurate type recognition, proving the effectiveness of the invention.

[0051] Table 1 Comparison table of recognition accuracies of mainstream intelligent learning algorithms

[0052]

[0053] The present invention uses continuous pulse parameter features to identify radar emitters. Through a data-driven deep learning method, a large number of pulse parameter sequences are mined and modeled. First, a temporal convolutional deep neural network is designed according to the characteristics of the pulse sequence, and then a self-supervised temporal convolutional network is used to extract features and train the network for the pulse sequence. Finally, a trained pulse sequence classifier is obtained for the recognition of emitter types.

[0054] It should be noted that in this article, the terms "including", "comprising" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or system. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or system including that element.

[0055] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0056] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above and includes several instructions for causing a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in various embodiments of the present invention.

[0057] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for identifying radar pulse sequences based on a self-supervised temporal convolutional network, characterized in that, the method comprises the following steps: S1: Sequence data acquisition; S2: Abnormal pulse rejection; S3: Design of a self-supervised temporal convolutional classification network; wherein, the network includes a temporal convolutional encoder, a temporal convolutional decoder, and a temporal target classifier; S4: Model training; wherein, the original pulse parameter sequence first passes through a temporal convolutional encoder, and high-dimensional temporal feature vectors are extracted through a multi-layer deep neural network. Then, this high-dimensional vector is fed into a temporal convolutional decoder and a temporal target classifier, including: after the high-dimensional vector is fed into the temporal target classifier, a target classification result is obtained, and the cross-entropy loss is calculated with the label information of the sequence. where n is the number of classifications, y i is the label value, and y i ′ is the predicted value output by the neural network; the high-dimensional vector is fed into the temporal convolutional decoding network to reconstruct the original pulse parameter sequence, and the reconstruction error loss is calculated with the original pulse sequence. where m is the number of batch samples, x i is the input pulse parameter sequence, and x i ′ is the reconstructed pulse parameter sequence; the cross-entropy loss and the reconstruction error loss are combined into the loss function for the entire model training. S5: Model identification; wherein, the trained temporal convolutional encoder and the temporal target classifier are selected to form a temporal convolutional network for identification.

2. The method for identifying radar pulse sequences based on a self-supervised temporal convolutional network according to claim 1, characterized in that, the step S1 specifically includes: obtaining the pulse signals of a radar radiation source for a period of time through radar reconnaissance equipment, and constructing a data set by using the parameter measurement data of these pulses.

3. The method for identifying radar pulse sequences based on a self-supervised temporal convolutional network according to claim 2, characterized in that, the data set contains pulse descriptor sequences and other target features.

4. The method for identifying radar pulse sequences based on a self-supervised temporal convolutional network according to claim 1, characterized in that, the step S2 specifically includes: judging the pulses with measurement error exceeding the tolerance or being abnormal according to the system characteristics and requirements, and rejecting or normalizing the pulses that affect training and identification.

5. The method for identifying radar pulse sequences based on a self-supervised temporal convolutional network according to claim 1, characterized in that, in the step S3, the encoder and the decoder have a symmetric network structure, and are respectively composed of an input layer, an output layer, and a temporal convolutional layer. The multi-level deep neural network structure is used to extract the complex high-dimensional temporal features of the pulse parameter sequence.

6. The method for identifying radar pulse sequences based on a self-supervised temporal convolutional network according to claim 5, characterized in that, in the step S3, the temporal convolutional layer is composed of multiple stacked temporal convolutional residual modules. The internal structure of the residual module adopts a one-dimensional dilated convolutional structure design to model the dependency relationship before and after the temporal information. Different residual blocks are set with different dilation coefficients to accelerate the calculation between neural network layers.

7. The method for identifying radar pulse sequences based on a self-supervised temporal convolutional network according to claim 1, characterized in that, in the step S5, after the model training is completed, the trained temporal convolutional encoder and the temporal target classifier are selected to form a temporal convolutional network. During the identification process, only the trained temporal convolutional encoder is used to calculate the high-dimensional features of the pulse sequence, and then the temporal target classifier is used to calculate the output probabilities of each type to obtain the final identification result.

8. A system for identifying radar pulse sequences based on a self-supervised temporal convolutional network, characterized in that, the system includes: a sequence data acquisition module for sequence data acquisition; an abnormal pulse rejection module for abnormal pulse rejection; a self-supervised temporal convolutional classification network design module for designing a self-supervised temporal convolutional classification network; wherein, the network includes a temporal convolutional encoder, a temporal convolutional decoder, and a temporal target classifier; The model training module is used for model training. Among them, the original pulse parameter sequence first passes through a temporal convolutional encoder, and high-dimensional temporal feature vectors are extracted through a multi-layer deep neural network. Then, this high-dimensional vector is sent to a temporal convolutional decoder and a temporal target classifier, including: after the high-dimensional vector is sent to the temporal target classifier, the target classification result is obtained, and the cross-entropy loss is calculated with the label information of this sequence. where n is the number of classifications, y i is the label value, y i ′ is the predicted value output by the neural network; the high-dimensional vector is sent to the temporal convolutional decoding network to reconstruct the original pulse parameter sequence, and the reconstruction error loss is calculated with the original pulse sequence. where m is the number of batch samples, x i is the input pulse parameter sequence, x i ′ is the reconstructed pulse parameter sequence; the cross-entropy loss and the reconstruction error loss are combined into the loss function for the entire model training. A model recognition module for model recognition. Among them, a trained temporal convolutional encoder and a temporal target classifier are selected to form a temporal convolutional network for recognition.

Images