Radar signal sorting method based on improved time convolution network

Abstract

The invention discloses a radar signal sorting method based on an improved time convolution network, and the method comprises the steps: receiving a radar pulse data flow through a current signal receiving window N, and obtaining N TOA, N PW and N RF according to the received radar pulse data flow; n is a positive integer greater than 1; generating a processed delta TOA sequence, a processed PW sequence and a processed RF sequence respectively based on the N TOAs, the N PWs and the N RFs; inputting the processed delta TOA sequence, the processed PW sequence and the processed RF sequence into a trained TCN-based radar signal sorting model for signal classification to obtain a classification result; wherein the trained TCN-based radar signal sorting model is obtained by training in a semi-supervised learning mode. According to the method, the calculation efficiency, accuracy and stability during classification can be improved, and the generalization ability of the model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal sorting, and in particular relates to a radar signal sorting method based on an improved temporal convolutional network. Background Art

[0002] Radar signal sorting has traditionally been designed based on a single parameter, the pulse repetition interval (PRI), using methods such as fixed threshold matching, histogram statistics, and pattern recognition. With the development of modern electronic countermeasures (ECM) technology, the complexity of electronic warfare environments has increased. This is reflected in the diverse pulse parameter designs of radar signals, such as non-uniform PRI, frequency hopping, and the interleaving of multiple PRIs. Furthermore, these methods are prone to misclassification in the presence of noise, making them difficult to meet the sorting requirements of electronic reconnaissance systems. Furthermore, rule-based methods have high computational complexity, making them difficult to meet the requirements of real-time processing.

[0003] To overcome the limitations of traditional PDW methods, researchers have begun employing multi-parameter analysis methods, such as cluster analysis and deep learning, for radar signal sorting. Clustering methods, such as the K-Means clustering algorithm (K-Means), group signals based on similarity in signal features. However, these methods require a preset number of clusters or similarity thresholds, making them difficult to process signals of unknown categories. They are also sensitive to noise and exhibit poor robustness. Deep learning methods, such as long-short-term memory networks (LSTMs), are suitable for sequence modeling due to their long-short-term memory capabilities. However, these methods suffer from low computational efficiency, vanishing gradients, and long training times, making them particularly difficult to meet the demands of efficient sorting when processing complex data streams. In other words, existing radar signal sorting methods suffer from low computational efficiency, limited sorting accuracy, and difficulty adapting to complex battlefield environments. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a radar signal sorting method based on an improved temporal convolutional network.

[0005] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] The present invention provides a radar signal sorting method based on an improved temporal convolutional network, comprising:

[0007] The radar pulse data stream is received using the current signal receiving window N, and N TOAs, N PWs, and N RFs are obtained according to the received radar pulse data stream; N is a positive integer greater than 1;

[0008] generating a processed ΔTOA sequence, a processed PW sequence, and a processed RF sequence based on the N TOAs, the N PWs, and the N RFs, respectively;

[0009] The processed ΔTOA sequence, the processed PW sequence, and the processed RF sequence are input into a trained TCN-based radar signal sorting model for signal classification to obtain a classification result; wherein the trained TCN-based radar signal sorting model is trained using a semi-supervised learning method.

[0010] Compared with the prior art, the present invention has the following beneficial effects:

[0011] 1) The present invention adopts TCN and dilated convolution technology, which can improve computational parallelism while maintaining timing information, and has higher computational efficiency than LSTM;

[0012] 2) The present invention performs radar signal sorting based on TOA (pulse arrival time), PW (pulse width), and RF (carrier frequency), enabling the network to more effectively distinguish different radar signals and improving sorting accuracy;

[0013] 3) The present invention combines a semi-supervised learning strategy, so that even with a small amount of labeled data, the TCN network can still be trained through signal reconstruction tasks, thereby improving the generalization ability of the model;

[0014] 4) The present invention adopts a dynamic window division strategy to adaptively adjust the window length according to the changing characteristics of the signal, so that TCN can learn complete timing information and extract more accurate signal features, thereby improving the accuracy and stability of classification.

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 1 is a flow chart of a radar signal sorting method based on an improved temporal convolutional network provided by an embodiment of the present invention;

[0017] Figure 2 Schematic diagram of the extended convolution structure of TCN provided by an embodiment of the present invention;

[0018] Figure 3 3. It is a schematic diagram comparing the confusion matrices of radar signal sorting performed by the method proposed in the present invention and the prior art method. DETAILED DESCRIPTION

[0019] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0020] In modern electronic warfare and complex electromagnetic environments, radar signals are often subject to multiple parameter agility, pulse overlap, and noise interference. Existing pulse descriptor word (PDW)-based or clustering methods struggle to maintain high sorting accuracy. To address this issue, this paper proposes a semi-supervised sorting method that combines dynamic window partitioning with a multi-channel TCN to adapt to diverse radar signal characteristics, improving sorting accuracy and computational efficiency.

[0021] Figure 1 is a flow chart of a radar signal sorting method based on an improved time convolutional network provided by an embodiment of the present invention, such as Figure 1 As shown, the method includes:

[0022] S101 , using a current signal receiving window N to receive a radar pulse data stream, and obtaining N TOAs, N PWs, and N RFs according to the received radar pulse data stream; N is a positive integer greater than 1.

[0023] S102 : Generate a processed ΔTOA sequence, a processed PW sequence, and a processed RF sequence based on N TOAs, N PWs, and N RFs, respectively.

[0024] S103 . Input the processed ΔTOA sequence, the processed PW sequence, and the processed RF sequence into a trained TCN-based radar signal sorting model for signal classification to obtain a classification result. The trained TCN-based radar signal sorting model is trained using a semi-supervised learning method.

[0025] In some embodiments, the above S102 can be implemented by the following steps:

[0026] S1021 : Generate an original ΔTOA sequence including N ΔTOAs based on N TOAs.

[0027] Specifically, the calculation formula for the i-th ΔTOA in the original ΔTOA sequence is: ΔTOA i =TOA i -TOA i-1 ,i=1,2,…,N, where ΔTOA i It represents the i-th ΔTOA in the original ΔTOA sequence. The first ΔTOA in the original ΔTOA sequence is TOA1.

[0028] S1022: Use N PWs to form an original PW sequence including N PWs, and use N RFs to form an original RF sequence including N RFs.

[0029] S1023 : Normalize each value in the original ΔTOA sequence, the original PW sequence, and the original RF sequence to obtain a processed ΔTOA sequence, a processed PW sequence, and a processed RF sequence.

[0030] Specifically, linear function normalization processing may be performed on each value in the original ΔTOA sequence, the original PW sequence, and the original RF sequence. The expression for performing linear function normalization processing on a value X is as follows: X represents each value of the original ΔTOA sequence, the original PW sequence, and the original RF sequence.

[0031] In some embodiments, after the above S102, step S104 may be further included, wherein S104 may be executed simultaneously with S103 or may be executed successively with S103, which is not limited in the present invention.

[0032] S104 : Based on the processed ΔTOA sequence, dynamically adjust the current signal receiving window N to obtain an updated signal receiving window N′ that matches the radar signal and the TCN.

[0033] Here, the i-th change rate R can be calculated based on the i-th ΔTOA and the i-1-th ΔTOA in the processed ΔTOA sequence. i , obtaining N change rates R corresponding to the N ΔTOAs one by one; based on the relationship between the N change rates R and the preset threshold τ, an updated signal receiving window N′ that matches the radar signal and TCN is obtained, so that the radar pulse data stream is subsequently received based on the updated signal receiving window N′.

[0034] Window partitioning can better leverage the advantages of the TCN method and improve the ability to model the timing characteristics of complex radar signals. This paper adopts a dynamic window partitioning strategy, adaptively adjusting the window length according to the changing characteristics of the signal. This enables the TCN to learn complete timing information, thereby extracting more accurate signal features, thereby improving classification accuracy and stability.

[0035] Here, by calculating the rate of change R, we can mine different PRI models based on the ΔTOA of the radar signal. Specifically, R i is the ratio between the i-th ΔTOA and the i-1-th ΔTOA in the processed ΔTOA sequence, that is, Among them, ΔTOA' i represents the i-th ΔTOA in the processed ΔTOA sequence.

[0036] Specifically, when the N change rates R are all greater than the preset threshold τ, an updated signal receiving window N′ that is smaller than the current signal receiving window N is generated; when at least one of the N change rates R is less than or equal to the preset threshold τ, the current signal receiving window N is used as the updated signal receiving window N′. If the N change rates R all exceed the set threshold τ, it means that the PRI of the signal may change drastically, and the window needs to be narrowed to prevent different types of signals from mixing. In some embodiments, when the N change rates R are all greater than the preset threshold τ, the current signal receiving window N can be halved to obtain the updated signal receiving window N′; in some embodiments, when the N change rates R are all greater than the preset threshold τ, a preset receiving window that is smaller than the current signal receiving window N can also be selected as the updated signal receiving window N′, where the preset receiving window is set according to an empirical value.

[0037] In some embodiments, a trained TCN-based radar signal sorting model includes: a trained TCN, a concatenation layer, and a fully connected layer; wherein the trained TCN includes four dilated convolutional layers, and the dilation rate of the four dilated convolutional layers gradually increases. TCN has good time series modeling capabilities, supports parallel computing, and can model long-term dependencies through dilated convolution. Its calculation formula is: Among them, y(t) is the output after dilated convolution, x(t) is the input signal, W i' is the weight of the convolution kernel, d is the expansion rate, and k is the size of the convolution kernel. For example, the expansion rates of the 4-layer expanded convolution layer are 1, 2, 4, and 8, respectively. By increasing the expansion rate layer by layer, the network can capture the temporal relationship of long-term dependence, thereby realizing feature extraction at multiple time scales, which is conducive to improving the accuracy of radar signal recognition; the number of channels used in each layer is set to 16, 32, 64, and 128 respectively. For example, Figure 2 This is the extended convolution structure diagram of TCN with the expansion rate of the extended convolution layer set to 1, 2, 4, and 8 respectively. Figure 2 In the “x0 x1 x2…x k-2 x k-1 x k " indicates input, "y0 y1y2…y k-2 y k-1 y k ” indicates the output corresponding to the input.

[0038] For a TCN-based radar signal sorting model, the present invention incorporates a semi-supervised learning approach and employs a hybrid loss function when training the model. This hybrid loss function combines classification loss and self-supervised loss to optimize the model's classification accuracy and signal reconstruction capabilities, thereby improving the model's robustness in complex electromagnetic scenarios. In the present invention, each sample in the training dataset is a matrix composed of a processed ΔTOA sequence, a processed PW sequence, and a processed RF sequence. Furthermore, a small number of samples in the training dataset are labeled with categories, and the category label of a sample represents the actual radar signal category corresponding to each ΔTOA, PW, and RF in the sample.

[0039] In the present invention, the classification loss is the cross entropy loss L CE , which calculates the difference between the predicted probability and the true category label. The formula is: Among them, y i is the actual category label corresponding to the i-th ΔTOA, i-th PW and i-th RF in a sample, p i is the predicted probability that the i-th ΔTOA, i-th PW, and i-th RF output by the model belong to category i, and N is the total number of categories.

[0040] Because only a small part of the training data set has labels, self-supervised loss is introduced to learn the temporal structure of unlabeled radar signals. The self-supervised loss is used to calculate the mean square error between the predicted value and the true value. The calculation formula is: Among them, x i is the original ΔTOA value of the input (i.e., the i-th ΔTOA in a sample input to the model), is the ΔTOA value after model reconstruction (i.e., it is the ΔTOA output by the TCN in the model according to the i-th ΔTOA in a sample input into the model), and N is the length of the ΔTOA sequence.

[0041] Finally, the expression of the comprehensive loss L used in the present invention when training the above model is: L = L CE +λL MSE , where λ is a coefficient. By adjusting λ, the model's classification ability and time series modeling ability can be balanced. At the same time, the present invention can use the Adam optimizer to train the model and use Dropout regularization to prevent overfitting, thereby improving the model's generalization ability.

[0042] In some embodiments, when the trained TCN-based radar signal sorting model includes: a trained TCN, a concatenation layer, and a fully connected layer, the above S103 is implemented by the following steps:

[0043] S1031. Input the processed ΔTOA sequence, the processed PW sequence, and the processed RF sequence into a trained TCN. The trained TCN extracts features from the processed ΔTOA sequence, the processed PW sequence, and the processed RF sequence, respectively, to obtain a pulse timing feature sequence and two different inter-pulse feature sequences.

[0044] Specifically, TCN can learn the pulse timing features based on the processed ΔTOA sequence to obtain a pulse timing feature sequence, TCN can learn the inter-pulse features based on the processed PW sequence to obtain an inter-pulse feature sequence, and TCN can learn the inter-pulse features based on the processed RF sequence to obtain another inter-pulse feature sequence. By learning the pulse timing features and two different inter-pulse feature sequences, the model's signal discrimination can be improved.

[0045] S1032. Input the pulse timing feature sequence and two different inter-pulse feature sequences into the splicing layer, which splices them together to obtain a three-channel feature matrix.

[0046] S1033. Input the three-channel feature matrix into the fully connected layer, and the fully connected layer classifies the radar signal to obtain a classification result.

[0047] Specifically, the expression for classification in the fully connected layer is: P = Softmax(WF + b), where P is the category probability of the radar signal, F is the three-channel feature matrix obtained after feature concatenation, W is the weight of the fully connected layer, and b is the bias term.

[0048] Specifically, the classification results indicate the probability that each ΔTOA in the processed ΔTOA sequence, each PW in the processed PW sequence, and each RF in the processed RF sequence belongs to each radar signal.

[0049] The present invention introduces TCN network and dynamic window partitioning technology for multi-channel input parameters, which solves the limitations of single-parameter modeling and sorting methods and multi-parameter sorting methods in complex environments such as complex modulation, and has higher accuracy, adaptability and robustness. At the same time, combined with semi-supervised learning and self-supervised loss technology, it not only reduces the dependence on large-scale labeled data, but also improves training efficiency and reduces computing costs, providing a feasible and efficient technical solution for radar signal sorting in complex electromagnetic environments.

[0050] The feasibility of the present invention is further illustrated below in conjunction with simulation experiments.

[0051] Experimental conditions:

[0052] The hardware platform for the simulation experiment of the present invention is: Intel(R) Core(TM) i7-14700HX@2.10GHz;

[0053] The software platform for the simulation experiment of the present invention is: on the Windows 11 Home Chinese version 64-bit operating system, MATLAB R2024a is used as the main simulation software for signal simulation, data analysis, etc.; the experimental development environment uses Python 3.9 version in Anaconda, combined with deep learning frameworks such as TensorFlow, Keras and PyTorch to realize the construction and training of TCN and classification models.

[0054] Simulation Parameter Settings: This simulation simulates the PDW data streams from 10 radars received by a radar reconnaissance receiver within 0.1 seconds. These data streams exhibit various modulation patterns, with different modulation schemes reflected in each radar's PRI, PW, and RF parameters. The specific data are shown in Table 1. This simulates the complex electromagnetic environment encountered by the sorting method through radar signal variations. The dataset is divided into training and validation sets in an 8:2 ratio, with 90% of the training set and 90% of the validation set unlabeled data, enabling semi-supervised network learning. The network employs a TCN for temporal modeling, consisting of four dilated convolutional layers. The number of channels used in each layer is set to 16, 32, 64, and 128, respectively, with the dilation rate increasing layer by layer to 1, 2, 4, and 8, respectively. This increasing dilation rate enables the network to capture long-term dependent temporal relationships.

[0055] Table 1. Test dataset of radar signal TCN sorting method

[0056]

[0057]

[0058] Figure 3 This is a comparison diagram of the confusion matrix of radar signal sorting by the method proposed in the present invention and the prior art method, where: Figure 3Figure (a) shows the radar signal sorting confusion matrix for the LSTM method, and Figure (b) shows the radar signal sorting confusion matrix for the method proposed in this paper. Both figures (a) and (b) were obtained after 200 network training runs based on the dataset in Table 1. Because the test dataset was randomly partitioned from the total dataset, the number of individual radars used in each training run fluctuated. The results show that both the LSTM method and the proposed method can perform radar signal sorting in complex environments with pulse overlap and multi-parameter agility by extracting PDW parameter features. However, the proposed method achieves a sorting success rate of 91.97%, significantly exceeding the LSTM method's 81.74%. This demonstrates the effectiveness and superiority of the proposed method in complex electromagnetic environments.

[0059] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0060] In the specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple situations. Certain measures are recorded in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0061] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A radar signal sorting method based on an improved temporal convolutional network, characterized in that: include: The radar pulse data stream is received using the current signal receiving window N, and N TOAs, N PWs, and N RFs are obtained according to the received radar pulse data stream. N is a positive integer greater than 1; generating a processed ΔTOA sequence, a processed PW sequence, and a processed RF sequence based on the N TOAs, the N PWs, and the N RFs, respectively; The processed ΔTOA sequence, the processed PW sequence, and the processed RF sequence are input into a trained TCN-based radar signal sorting model for signal classification to obtain a classification result; wherein the trained TCN-based radar signal sorting model is trained using a semi-supervised learning method.

2. The radar signal sorting method based on the improved temporal convolutional network according to claim 1 is characterized in that: The method further comprises: Based on the processed ΔTOA sequence, the current signal receiving window N is dynamically adjusted to obtain an updated signal receiving window N′ that matches the radar signal and the TCN.

3. The radar signal sorting method based on the improved temporal convolutional network according to claim 2 is characterized in that: The processed ΔTOA sequence includes N ΔTOAs; and dynamically adjusting the current signal receiving window N based on the processed ΔTOA sequence to obtain an updated signal receiving window N′ that matches the radar signal and the TCN, including: Calculate the i-th change rate R according to the i-th ΔTOA and the i-1-th ΔTOA in the processed ΔTOA sequence. i , obtaining N change rates R corresponding to the N ΔTOAs; Based on the magnitude relationship between the N change rates R and the preset threshold, an updated signal receiving window N′ matching the radar signal and the TCN is obtained.

4. The radar signal sorting method based on the improved temporal convolutional network according to claim 3 is characterized in that: The step of obtaining an updated signal receiving window N′ that matches the radar signal and the TCN based on a magnitude relationship between the N change rates R and a preset threshold comprises: When the N change rates R are all greater than the preset threshold, generating the updated signal receiving window N′ that is smaller than the current signal receiving window N; When at least one change rate R among the N change rates R is less than or equal to the preset threshold, the current signal receiving window N is used as the updated signal receiving window N′.

5. The radar signal sorting method based on the improved temporal convolutional network according to claim 1 is characterized in that: The generating a processed ΔTOA sequence, a processed PW sequence, and a processed RF sequence based on the N TOAs, the N PWs, and the N RFs, respectively, includes: generating an original ΔTOA sequence including N ΔTOAs based on the N TOAs; Using the N PWs to form an original PW sequence including N PWs, and using the N RFs to form an original RF sequence including N RFs; Normalization processing is performed on each value of the original ΔTOA sequence, the original PW sequence, and the original RF sequence to obtain a processed ΔTOA sequence, a processed PW sequence, and a processed RF sequence.

6. The radar signal sorting method based on the improved temporal convolutional network according to claim 1 is characterized in that: The trained TCN-based radar signal sorting model includes: a trained TCN, a splicing layer, and a fully connected layer; wherein the trained TCN includes four layers of extended convolutional layers, and the expansion rates of the four layers of extended convolutional layers gradually increase.

7. The radar signal sorting method based on the improved temporal convolutional network according to claim 6 is characterized in that: The expansion rates of the 4-layer dilated convolutional layers are 1, 2, 4, and 8 respectively.

8. The radar signal sorting method based on the improved temporal convolutional network according to claim 6 is characterized in that: The processed ΔTOA sequence, the processed PW sequence, and the processed RF sequence are input into a trained TCN-based radar signal sorting model for signal classification to obtain a classification result, including: Inputting the processed ΔTOA sequence, the processed PW sequence, and the processed RF sequence into the trained TCN, and having the trained TCN perform feature extraction on the processed ΔTOA sequence, the processed PW sequence, and the processed RF sequence, respectively, to obtain a pulse timing feature sequence and two different inter-pulse feature sequences; Inputting the pulse timing feature sequence and the two different inter-pulse feature sequences into the splicing layer, and splicing them together to obtain a three-channel feature matrix; The three-channel feature matrix is input into the fully connected layer, and the fully connected layer classifies the radar signal to obtain a classification result.

9. The radar signal sorting method based on the improved temporal convolutional network according to claim 3 is characterized in that: The i-th change rate R i is the ratio between the i-th ΔTOA and the i-1-th ΔTOA in the processed ΔTOA sequence.

Images