Convolutional neural network variable-length radar signal identification method based on mercuric chloride AI platform

Abstract

The invention discloses a convolutional neural network variable-length radar signal recognition method based on a mercuric chloride AI platform, and the method comprises the steps: constructing a variable-length radar modulation recognition network based on a convolutional neural network, extracting the time domain and frequency domain features of a signal through one-dimensional convolution, carrying out the feature fusion based on an SE-ResNet module, and carrying out the recognition of a variable-length radar signal. The modulation type of the variable-length radar signal is effectively identified, network optimization and deployment are carried out based on the particularity of the mercuric chloride AI platform, and convolutional neural network variable-length radar signal identification based on the mercuric chloride AI platform is completed. According to the method, effective identification of various variable-length radar signals in a complex electromagnetic environment is realized, engineering deployment of a localized platform is carried out based on a mercuric chloride AI platform, the use condition of related hardware resources during model reasoning is measured, the problem that the accuracy of an existing processing method based on a fixed-length window is rapidly reduced is solved, and the accuracy of the variable-length radar signals is greatly improved. The method has the advantages of being excellent in recognition performance, high in processing efficiency, flexible, accurate, high in robustness and capable of being deployed in an engineering mode.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and particularly relates to a method for identifying variable-length radar signals of a convolutional neural network based on the Ascend AI platform. Background Technique

[0002] Radar emitter recognition is an important research content in electronic countermeasures. By analyzing radar signals, information such as the parameters, models, and intentions of the opposing radar can be obtained, thereby guiding the decision-making of one's own side. With the increase in the number of emitters and the application of new radar systems, the pulse density in the electromagnetic space is continuously increasing, and the overlapping probability of radar signal parameters intercepted by reconnaissance receivers is gradually increasing. In addition, the pulse widths emitted by different radars are not the same, and the pulse width of the same radar will also be adjusted with the change of the working mode. Therefore, the radar signals received by reconnaissance equipment usually have overlapping parameters and different lengths, and the recognition method based on pulse description words is difficult to apply. Since the in-pulse signals of radar contain a large amount of information compared with simple pulse signals, the in-pulse modulation recognition of radar signals has become the key to emitter recognition.

[0003] The existing methods for in-pulse signal modulation recognition of radar extract statistical features such as time domain and frequency domain, and use classifiers to identify signals. In 2019, the team of Professor Jin Weidong at Southwest Jiaotong University represented radar modulation signals in the form of wavelet ridges and reduced the dimension of the signals. The dynamic time warping average was used to extract the centroid of each type of signal, and the k-nearest neighbor algorithm was used to make a decision based on the Manhattan distance between the sample data and the centroid. The recognition accuracy rate could reach more than 40% at 6dB. In 2022, the team of Professor Liu Xinggao at Zhejiang University extracted five features including time domain information, frequency domain information, Hilbert transform, discrete wavelet transform, and empirical mode decomposition from the received signals. After setting weights for weighted processing, a multi-path aggregation classification method was used to realize signal modulation recognition. The documents "Method for extracting time-frequency distribution features of in-pulse modulation signals of radar. Application of Electronic Techniques, 2012, 38(08): 136-139" and "Radar signals classification using energy time-frequency distribution features. IET Radar, Sonar & Navigation, 2020, 14(5): 707-715" both used the method of combining the time-frequency features of signals with a support vector machine (SVM) classifier to realize the in-pulse modulation recognition of radar signals.

[0004] Compared with existing machine learning algorithms, due to its powerful data processing ability, deep learning can automatically learn important features helpful for signal recognition during the training process, and has gradually become a research hotspot in the field of modulation recognition. The research team led by Professor Guo Limin of Harbin Engineering University performed time-frequency transformation on the received radar signals, designed a convolutional neural network based on LeNet-5, and recognized the time-frequency diagrams of eight types of signals such as LFM, Frank code, and multi-time code, with the recognition accuracy reaching 90%. The research team led by Professor Li Yunjie of Beijing Institute of Technology proposed a new multi-instance multi-label learning framework based on deep convolutional neural network and achieved the recognition of overlapping low-intercept radar signals. The method based on deep learning has significantly improved the signal recognition accuracy, but the existing methods do not consider the situation of uncertain radar pulse length, which limits its practical application: when the pulse width of the received signal is large, the signal information extraction is incomplete, resulting in a decrease in recognition accuracy; when the pulse width is small, unnecessary computational resources will be wasted. Therefore, it is of great significance to study the recognition of variable-length radar signals. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a method for recognizing variable-length radar signals based on a convolutional neural network on the Ascend AI platform. By realizing the deployment on the domestic AI platform, the effective recognition of the modulation type of variable-length radar signals is achieved, and it has a certain robustness.

[0006] The technical solution adopted by the present invention is as follows: A method for recognizing variable-length radar signals based on a convolutional neural network on the Ascend AI platform, and the specific steps are as follows:

[0007] S1. Simulate the original radar signal and modulation parameters to generate original radar signals with different pulse widths;

[0008] Among them, the original radar signal is the time-domain signal, and the modulation parameters include: the serial number, bandwidth, pulse width, and signal-to-noise ratio.

[0009] S2. Divide the original radar signals simulated in step S1 into training set data and test set data;

[0010] Among them, the division ratio is set according to the actual situation.

[0011] S3. Perform IQ sampling on the variable-length radar signals in the training set and test set obtained in step S2 and perform Fourier transform to obtain the frequency-domain signal, and splice the three-channel data obtained from the time-domain dual-channel signal and the frequency-domain signal obtained by IQ sampling to obtain the input feature matrix;

[0012] S4. Construct a convolutional variable-length radar modulation recognition network;

[0013] S5. Based on step S4, input the three-channel feature matrix of the time domain and frequency domain of the training set radar signals obtained in step S3 into the variable-length radar modulation recognition network for network training;

[0014] S6. Use the variable-length radar modulation recognition network trained in step S5 to identify the modulation types of the variable-length radar signals in the test set processed in step S3;

[0015] Input the test set processed in step S3 into the one-dimensional convolutional variable-length radar modulation recognition network trained in step S5 for forward propagation to obtain the classification results of the test set.

[0016] S7. Based on the Ascend AI platform, optimize the software and hardware of the variable-length radar modulation recognition model trained in step S5, make the optimized model offline in a fixed standard, and convert it into a model file supported by the Ascend AI platform. Finally, deploy the generated model file to the Ascend AI platform and modify the length of the input data based on the format restrictions of the input of the Ascend AI platform to complete the recognition of variable-length radar signals by the convolutional neural network based on the Ascend AI platform.

[0017] Further, the specific steps of step S3 are as follows:

[0018] For the training set and test set obtained in step S2, use IQ sampling to extract the time domain IQ two-channel feature data of the radar signals, and perform FFT on the extracted time domain digital signals to obtain the frequency domain single feature channel data. Then, the signal model obtained in step S2, the IQ sampling and FFT calculation expressions in step S3 are as follows:

[0019]

[0020] Among them, t represents time, z(t) represents the complex signal, x(t) represents the real part of the complex signal, a(t) represents the amplitude of the complex signal, θ(t) represents the signal phase, fc represents the signal carrier frequency, g(t) represents the complex envelope of x(t), and g(t) = a(t)e θ(t) ; represents the Hilbert transform of x(t). At the sampling rate f s , the sampling sequence of x(t) is x(i), and the number of sampling points is N S , then its complex analytical expression is as follows:

[0021]

[0022] Among them, x(i) represents the in-phase component of z(i), that is, the I component, represents the quadrature component of z(i), that is, the Q component.

[0023] Then perform discrete Fourier transform on z(i), and the expression is as follows:

[0024]

[0025] Among them, f(i) represents the frequency-domain data after Fourier transform, and i represents the serial number.

[0026] Then, signal features are extracted from the time domain and the frequency domain, that is, the three-channel data x(i) obtained by sampling and f(i) are concatenated to obtain the model input feature matrix X of a single signal. The expression is as follows:

[0027]

[0028] For n signals, they are cascaded as the batch input Y of the model. The expression is as follows:

[0029] Y = [X(0); X(2);......X(n - 1)]

[0030] Further, the specific steps of step S4 are as follows:

[0031] S41. Construct a feature dimension elevation module;

[0032] The feature dimension elevation module includes: a convolutional layer, a batch normalization layer, and a ReLU activation layer.

[0033] S42. Construct an SE-ResNet sub-module;

[0034] The SE-ResNet sub-module includes: 3 convolutional layers, 2 batch normalization layers, 2 ReLU activation layers, 1 residual block, and 1 attention mechanism layer. The attention mechanism layer sequentially includes: global average pooling, a fully connected layer, a ReLU activation layer, a fully connected layer, and a Sigmoid activation layer, which is used to extract independent features of different feature channels.

[0035] Among them, the input of the SE-ResNet sub-module is the feature map after the three feature channels pass through the feature dimension elevation module. Deep features are extracted after passing through 3 convolutional layers in sequence. The first 2 convolutional layers are followed by a batch normalization layer and a ReLU activation layer. The extracted deep features are fused with the original shallow features of the input through the residual block. Then, the obtained fused features are input into the attention mechanism layer. The output of the residual block is fused with the output of the original residual block after passing through the attention mechanism layer.

[0036] S43. Based on step S42, construct an SE-ResNet module;

[0037] The SE-ResNet module is composed of three SE-ResNet sub-modules connected in series to deeply extract independent features of different feature channels.

[0038] Among them, the output of each SE-RseNet sub-module is used as the input of the next one, and it is extracted layer by layer. Finally, deep and global features of three channels are obtained.

[0039] S44. Construct a feature fusion module;

[0040] The feature fusion module includes: a convolutional layer, a batch normalization layer, a ReLU activation function, and an attention mechanism layer to achieve feature fusion in the time domain and frequency domain.

[0041] Among them, the attention mechanism layer sequentially includes: global average pooling, a fully connected layer, a ReLU activation layer, a fully connected layer, and a Sigmoid activation layer; and the feature fusion module takes the output of the SE-ResNet module as the input, and the output result is a multi-dimensional feature fusion map after the fusion is completed.

[0042] S45. Construct a fully connected module;

[0043] The fully connected module includes: a pooling layer, a ReLU activation function, and a fully connected layer. Taking the output of the feature fusion module as the input, the posterior probability corresponding to each category is obtained.

[0044] S46. Combine the sub-modules constructed in steps S41-S45 to form a convolutional variable-length radar modulation recognition network;

[0045] After cascading the feature dimension increasing module, 3 parallel SE-ResNet modules, the feature fusion module, and the fully connected module in sequence, a one-dimensional convolutional variable-length radar modulation recognition network is formed.

[0046] The input of the recognition network is the feature matrix after the original radar signal is processed by step S3, and the output is the classification result.

[0047] Among them, the category corresponding to the data with a value greater than the pre-set threshold in the posterior probability is the classification result of this network structure.

[0048] Furthermore, the specific steps of step S5 are as follows:

[0049] S51. After splicing the IQ data and frequency domain data of each variable-length radar signal in the training set processed by step S3, input them into the one-dimensional convolutional variable-length radar modulation recognition network constructed in step S4 for forward propagation, and calculate the cost function value;

[0050] Among them, calculating the cost function value means using the binary cross-entropy loss function as the cost function, and the calculation expression is as follows:

[0051]

[0052] Among them, BCE represents the binary cross-entropy loss function, and prob ∈ R 1×mRepresents the output probability, lab ∈ R 1×m Represents the true value of the overlapping signal in one-hot code form, thre ∈ R 1×m Represents the threshold, m represents the total number of modulation types, and R represents the real number field.

[0053] S52. Use the backpropagation algorithm based on stochastic gradient descent to update the parameters of the variable-length radar modulation recognition network;

[0054] S53. Based on step S52, iterate the backpropagation process until the cost function converges to obtain a trained one-dimensional convolutional variable-length radar modulation recognition network.

[0055] Further, the specific steps of step S7 are as follows:

[0056] S71. Based on the characteristics of the Ascend AI platform, optimize the software and hardware of the model trained in step S5, that is, perform operator optimization and operator replacement;

[0057] First, optimize the model size based on the hardware conditions of the Ascend AI platform, and then perform operator optimization based on the software environment of the Ascend AI processor.

[0058] Among them, if there is no corresponding operator for the pooling layer of the fully connected module of the trained model in Ascend AI, then replace the pooling layer operator. Adopt the upsampling method and use linear interpolation to replace the pooling layer operator.

[0059] S72. Based on step S71, store the optimized variable-length radar modulation recognition network in the ONNX file format, perform model offlineization, and generate a model file supported by the Ascend AI platform;

[0060] First, perform model offlineization based on environmental standards, that is, use the official ONNX tool to offlineize the network to obtain the ONNX model file.

[0061] Then generate the model for the Ascend AI platform development environment, that is, generate the OM model according to the characteristics of the Ascend AI platform from the obtained ONNX model file to obtain a model file supported by the Ascend AI platform.

[0062] S73. Deploy the generated OM model file to the Ascend AI platform to achieve network engineering deployment, and process the length of each radar signal in the test set obtained in step S3 to a unified length to adapt to the model input format of the Ascend AI platform;

[0063] First, use the official ATC tool of Ascend AI, take the ONNX model as the input, and configure the relevant parameters of the OM model, including: input dimension, application platform, and convert the ONNX model into the OM model.

[0064] Then, truncate the input signal of the OM model. The truncation length is set to an integer multiple of 500 that is closest to the original length. The calculation formula is as follows:

[0065]

[0066] Among them, represents the floor function, and N s represents the original number of sampling points.

[0067] S74. Input the test set data processed in step S73 into the deployed convolutional variable-length radar modulation recognition network for forward propagation to obtain the classification result of the test set data;

[0068] Compare the classification result with the result of the conventional platform network. If the inference accuracy loss exceeds 5%, then, with the input format of the OM model as the standard, truncate the test set of the conventional platform model in the same format, repeat step S5 for retraining, and then redeploy on the Ascend AI platform according to steps S6 - S7.

[0069] Beneficial effects of the present invention: The method of the present invention constructs a variable-length radar modulation recognition network based on a convolutional neural network, extracts the time-domain and frequency-domain features of the signal using one-dimensional convolution, performs feature fusion based on the SE-ResNet module, realizes the effective recognition of the modulation types of variable-length radar signals, and, based on the particularity of the Ascend AI platform, conducts network optimization and deployment to complete the recognition of variable-length radar signals by the convolutional neural network based on the Ascend AI platform. The method of the present invention realizes the effective recognition of various variable-length radar signals in a complex electromagnetic environment, conducts engineering deployment of the domestic platform based on the Ascend AI platform, measures the usage of relevant hardware resources during model inference, solves the problem that the accuracy of the existing processing method based on a fixed-length window drops rapidly, and has the advantages of excellent recognition performance, high processing efficiency, flexibility, accuracy, strong robustness, and engineering deployability. Description of the Drawings

[0070] Figure 1 is a flowchart of a method for recognizing variable-length radar signals by a convolutional neural network based on the Ascend AI platform according to the present invention.

[0071] Figure 2 is a structural diagram of the SE-ResNet sub-module in an embodiment of the present invention.

[0072] Figure 3 is a structural diagram of the SE-ResNet module (attention-residual module) in an embodiment of the present invention.

[0073] Figure 4 is a structural diagram of the feature fusion module in an embodiment of the present invention.

[0074] Figure 5 This is the structural diagram of the convolutional variable-length radar modulation recognition network in the embodiment of the present invention.

[0075] Figure 6 This is the deployment flow chart of the Ascend AI platform in the embodiment of the present invention.

[0076] Figure 7 This is the recognition result diagram of the variable-length radar signal modulation type based on the Ascend AI platform in the embodiment of the present invention. Detailed implementation manners

[0077] The method of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0078] As Figure 1 shown, the flow chart of the method for recognizing variable-length radar signals by a convolutional neural network based on the Ascend AI platform of the present invention is as follows:

[0079] S1. Simulate the original radar signal and modulation parameters to generate original radar signals with different pulse widths;

[0080] Among them, the original radar signal is the time-domain signal, and the modulation parameters include: serial number, bandwidth, pulse width, and signal-to-noise ratio.

[0081] S2. Divide the original radar signals simulated in step S1 into training set data and test set data;

[0082] Among them, the division ratio is set according to the actual situation.

[0083] S3. Perform IQ sampling on the variable-length radar signals in the training set and test set obtained in step S2 and perform Fourier transform to obtain the frequency-domain signal, and splice the three-channel data obtained from the time-domain dual-channel signal and the frequency-domain signal obtained by IQ sampling to obtain the input feature matrix;

[0084] S4. Construct a convolutional variable-length radar modulation recognition network;

[0085] S5. Based on step S4, input the three-channel feature matrix of the time domain and frequency domain of the training set radar signals obtained in step S3 into the variable-length radar modulation recognition network for network training;

[0086] S6. Use the variable-length radar modulation recognition network trained in step S5 to recognize the modulation types of the variable-length radar signals in the test set processed in step S3;

[0087] Input the test set processed in step S3 into the one-dimensional convolutional variable-length radar modulation recognition network trained in step S5 for forward propagation to obtain the classification result of the test set. Thus, the effectiveness of the method on a conventional platform is verified.

[0088] S7. Based on the Ascend AI platform, optimize the variable-length radar modulation recognition model completed in step S5 in terms of software and hardware, make the optimized model offline in a fixed standard, and convert it into a model file supported by the Ascend AI platform. Finally, deploy the generated model file to the Ascend AI platform, and modify the length of the input data based on the format restrictions of the input on the Ascend AI platform to complete the recognition of variable-length radar signals based on the convolutional neural network of the Ascend AI platform.

[0089] In this embodiment, step S2 is specifically as follows:

[0090] The division ratio in this embodiment is 8:2. The specific dataset parameters of this embodiment are shown in Tables 1 and 2.

[0091] Table 1

[0092] Parameter Range Pulse width 10~100 us Sampling rate 200 MHz SNR 15 dB

[0093] Table 2

[0094]

[0095] Among them, Table 1 is the original radar signal parameters. In this embodiment, seven typical radar signal modulation types are simulated, including: linear frequency modulation (LFM), non-linear frequency modulation (NLFM), continuous wave (CW) signal, frequency agility within pulse (FAM) signal, binary frequency shift keying (Costas), phase coding (Barker), and frequency modulation and phase modulation combined coding (Costas + Barker). The simulation signals are all intermediate frequency signals, the sampling frequency is 200 MHz, the pulse width range is 1 - 10 us, and the signal-to-noise ratio is 15 dB. Table 2 is the corresponding table of the dataset parameters of the in-pulse modulation signal, that is, the parameters of radar signals of different modulation types, including: bandwidth, pulse width, and the corresponding number.

[0096] In this embodiment, step S3 is specifically as follows:

[0097] For the training set and test set obtained in step S2, use IQ sampling to extract the time-domain IQ dual-channel feature data of the radar signal, and perform FFT on the extracted time-domain digital signal to obtain the frequency-domain single-feature channel data. Then, the signal model obtained in step S2, the IQ sampling in step S3, and the FFT calculation expressions are as follows:

[0098]

[0099] Among them, t represents time, z(t) represents the complex signal, x(t) represents the real part of the complex signal, a(t) represents the amplitude of the complex signal, θ(t) represents the signal phase, f c represents the signal carrier frequency, g(t) represents the complex envelope of x(t), and g(t) = a(t)eθ(t) ; denotes the Hilbert transform of x(t). At the sampling rate f s , the sampling sequence of x(t) is x(i), and the number of sampling points is N S , then its complex analytical expression is as follows:

[0100]

[0101] where x(i) represents the in-phase component of z(i), i.e., the I component, denotes the quadrature component of z(i), i.e., the Q component.

[0102] Then, perform the discrete Fourier transform on z(i), and the expression is as follows:

[0103]

[0104] where f(i) represents the frequency-domain data after the Fourier transform, and i represents the sequence number.

[0105] In order to effectively modulate and identify the in-pulse radar signal, in this embodiment, signal features are extracted from the time domain and the frequency domain, that is, the three-channel data x(i) obtained by sampling, and f(i) are concatenated to obtain the model input feature matrix X of a single signal, and the expression is as follows:

[0106]

[0107] For n signals, they are cascaded as the batch input Y of the model, and the expression is as follows:

[0108] Y = [X(0); X(2);... X(n - 1)]

[0109] In this embodiment, the step S4 is specifically as follows:

[0110] S41. Construct a feature dimension elevation module;

[0111] The feature dimension elevation module includes: a convolutional layer, a batch normalization layer, and a ReLU activation layer.

[0112] Among them, since the signal is affected by noise during transmission, the modulation characteristics of the received signal are submerged in the noise. Therefore, in this embodiment, the feature dimension elevation module is first used to map the signal to a high-dimensional space to learn the essential features of the signal and reduce the influence of noise. Then, through the batch normalization layer, the possibility of gradient explosion in subsequent training can be reduced. Finally, the ReLU activation layer introduces non-linear factors to enhance the expression ability of the network.

[0113] S42. Construct an SE-ResNet sub-module;

[0114] As shown Figure 2 in the figure, the SE-ResNet sub-module includes: 3 convolutional layers, 2 batch normalization (BN) layers, 2 ReLU activation layers, 1 residual block, and 1 attention mechanism layer. The attention mechanism layer sequentially includes: global average pooling, fully connected layer, ReLU activation layer, fully connected layer, and Sigmoid activation layer, which is used to extract independent features of different feature channels.

[0115] Among them, the input of the SE-ResNet sub-module is the feature map after the three feature channels (input feature matrix X) pass through the feature dimension elevation module. Deep features are extracted after passing through 3 convolutional layers in sequence. The first 2 convolutional layers are followed by a batch normalization layer and a ReLU activation layer. The extracted deep features are fused with the original shallow features through the residual block. Then, the obtained fused features are input into the attention mechanism layer. The output of the residual block is fused with the output of the original residual block after passing through the attention mechanism layer. This can take into account the advantages of both network modules, suppress information useless for the recognition task, and extract features that can better represent the signal type.

[0116] S43. Based on step S42, construct the SE-ResNet module;

[0117] As shown Figure 3 in the figure, the SE-ResNet module is composed of three SE-ResNet sub-modules connected in series to deeply extract independent features of different feature channels.

[0118] Among them, the output of each SE-RseNet sub-module is used as the input of the next one, and the extraction goes deeper layer by layer. Finally, deep and global features of three channels are obtained.

[0119] S44. Construct the feature fusion module;

[0120] As shown Figure 4 in the figure, the feature fusion module includes: convolutional layer, batch normalization layer, ReLU activation function, and attention mechanism layer to achieve feature fusion in the time domain and frequency domain.

[0121] Among them, the attention mechanism layer sequentially includes: global average pooling, fully connected layer, ReLU activation layer, fully connected layer, and Sigmoid activation layer; and the feature fusion module takes the output of the SE-ResNet module as the input, and the output result is the multi-dimensional feature fusion map after the fusion is completed.

[0122] The feature fusion module first cascades the deep features obtained by the feature extraction module for each channel of the signal to obtain the final multi-dimensional feature fusion map, and then uses the attention mechanism layer to obtain the weight of each feature for the classification result, and fuses the feature fusion map with the weight to obtain the result of multi-faceted and multi-level feature fusion.

[0123] S45. Construct a fully connected module;

[0124] The fully connected module includes: a pooling layer, a ReLU activation function, and a fully connected layer. Taking the output of the feature fusion module as the input, the posterior probability corresponding to each category is obtained.

[0125] S46. Combine the sub-modules constructed in steps S41 - S45 to form a convolutional variable-length radar modulation recognition network;

[0126] As Figure 5 shown, after cascading the feature dimension increasing module, 3 parallel SE-ResNet modules, the feature fusion module, and the fully connected module in sequence, a one-dimensional convolutional variable-length radar modulation recognition network is formed.

[0127] The input of the recognition network is the feature matrix after the original radar signal is processed by step S3, and the output is the classification result.

[0128] Among them, after the original radar signal is processed by step S3, a three-channel feature matrix in the time domain and frequency domain is obtained. After the feature matrix is input into the variable-length radar modulation recognition network, the posterior probability corresponding to each category can be obtained through the fully connected module. The category corresponding to the data whose value in the posterior probability is greater than the preset threshold is the classification result of this network structure.

[0129] In this embodiment, step S5 is specifically as follows:

[0130] S51. After splicing the IQ data and frequency domain data of each variable-length radar signal in the training set processed by step S3, input them into the one-dimensional convolutional variable-length radar modulation recognition network constructed in step S4 for forward propagation, and calculate the cost function value;

[0131] Among them, calculating the cost function value means using the binary cross-entropy loss function as the cost function, and the calculation expression is as follows:

[0132]

[0133] Among them, BCE represents the binary cross-entropy loss function, prob ∈ R 1×m represents the output probability, lab ∈ R 1×m represents the ground truth of the overlapping signal in the form of a one-hot code, thre ∈ R 1×m represents the threshold, m represents the total number of modulation types, and R represents the real number field.

[0134] S52. Use the backpropagation algorithm based on stochastic gradient descent to update the parameters of the variable-length radar modulation recognition network;

[0135] S53. Based on step S52, iterate the backpropagation process until the cost function converges to obtain a trained one-dimensional variable-length radar modulation recognition network.

[0136] As Figure 6 shown, in this embodiment, the step S7 is specifically as follows:

[0137] S71. Based on the characteristics of the Ascend AI platform (comparing the differences between the Ascend AI platform and the conventional AI platform), optimize the model trained in step S5 in terms of software and hardware, that is, perform operator optimization and operator replacement;

[0138] First, optimize the model size based on the hardware conditions of the Ascend AI platform, and then perform operator optimization based on the software environment of the Ascend AI processor.

[0139] Among them, by comparing the operator list of the Ascend AI platform, it is found that there is no corresponding operator for the pooling layer in the fully connected module of the trained model in the Ascend AI. Therefore, method replacement is performed on the pooling layer operator. According to the principle of the pooling layer in the fully connected module, the upsampling method is used to perform operator replacement on the pooling layer by means of linear interpolation.

[0140] S72. Based on step S71, store the optimized variable-length radar modulation recognition network in the ONNX file format, perform model offline processing, and generate a model file supported by the Ascend AI platform;

[0141] First, perform model offline processing based on environmental standards, that is, use the official ONNX tool to make the network offline to obtain an ONNX model file.

[0142] Then, generate a model for the Ascend AI platform development environment, that is, generate an OM model according to the characteristics of the Ascend AI platform from the obtained ONNX model file to obtain a model file supported by the Ascend AI platform.

[0143] S73. Deploy the generated OM model file to the Ascend AI platform to achieve network engineering deployment, and process the length of each radar signal in the test set obtained in step S3 to a unified length to adapt to the model input format of the Ascend AI platform;

[0144] First, use the official ATC tool of the Ascend AI to take the ONNX model as the input and configure the relevant parameters of the OM model, including: input dimension, application platform, and convert the ONNX model into an OM model.

[0145] Since the data input of the OM model only supports fixed-length data input, the input signal of the OM model is truncated, and the truncation length is set to an integer multiple of 500 closest to the original length. The calculation expression is as follows:

[0146]

[0147] Among them, represents the floor function, and N s represents the original number of sampling points.

[0148] S74. Input the test set data processed in step S73 into the deployed convolutional variable-length radar modulation recognition network for forward propagation to obtain the classification result of the test set data;

[0149] Compare the classification result with the result of the conventional platform network. If the inference accuracy loss exceeds 5%, then, with the input format of the OM model as the standard, truncate the test set of the conventional platform model in the same format, and repeat step S5 for retraining, and then redeploy on the Ascend AI platform according to steps S6 - S7.

[0150] To demonstrate the recognition performance of the variable-length radar modulation recognition network in the method of the present invention on the Ascend AI platform, this embodiment will test the recognition accuracy of different modulation types and obtain a confusion matrix. The confusion matrix of the recognition result is as Figure 7 shown. The recognition accuracy of each type of modulation signal is above 95%. The specific test result data is shown in Table 3, which shows the recognition accuracy of different modulation types. It can be seen from Table 3 that the overall recognition accuracy of the radar signal modulation type is 96.5%.

[0151] Table 3

[0152]

[0153] And to test the engineering feasibility of the method of the present invention, this embodiment also tests the inference time of each piece of data on the Ascend AI platform. The specific data is shown in Table 4, which shows the recognition time of different length pulse modulations.

[0154] Table 4

[0155]

[0156] It can be seen from Table 4 that the inference time of each piece of data is within 1 ms, meeting the engineering usage scenario.

[0157] In summary, the method of the present invention addresses the recognition problem in the case of uncertain in-pulse signal length of radar, comprehensively considers the application requirements of accuracy and real-time performance, takes the time-domain waveform and spectrum of the in-pulse signal as inputs, constructs a feature dimension elevation module to achieve shallow feature extraction of radar signals with different pulse widths, uses a one-dimensional convolutional neural network to extract multi-channel features of the signal, combines the attention-residual mechanism to deeply extract independent features of different channels, constructs a feature fusion module to achieve the fusion of independent features of different feature channels, completes the in-pulse modulation recognition of radar signals, and realizes the effective recognition of various variable-length radar signals in a complex electromagnetic environment. In addition, the method of the present invention has completed the engineering deployment of the method based on the Ascend AI platform and measured the usage of relevant hardware resources during model inference. The above experimental results show that the recognition performance of the method of the present invention is excellent, the processing efficiency is high, and it has the advantages of flexibility, accuracy, strong robustness, and being engineering-deployable.

[0158] Those skilled in the art will realize that the examples described herein are for helping readers understand the principles of the present invention and should be understood as the protection scope of the present invention but not limited to such specific statements and implementation cases. For those skilled in the art, various changes and modifications can be made to the present invention. Any changes, equivalent substitutions, and improvements made within the spirit and principle of the present invention are included within the scope of the claims of the present invention.

Claims

1. A method for identifying variable-length radar signals of a convolutional neural network based on the Ascend AI platform, the specific steps are as follows: S1. Simulate the original radar signal and modulation parameters to generate original radar signals with different pulse widths; Among them, The original radar signal is the time-domain signal, and the modulation parameters include: serial number, bandwidth, pulse width, and signal-to-noise ratio; S2. Divide the original radar signals simulated in step S1 into training set data and test set data; Among them, the division ratio is set according to the actual situation; S3. Perform IQ sampling on the variable-length radar signals in the training set and test set obtained in step S2 and perform Fourier transform to obtain frequency-domain signals, and splice the three-channel data obtained from the time-domain dual-channel signal and frequency-domain signal obtained by IQ sampling to obtain an input feature matrix; S4. Construct a convolutional variable-length radar modulation recognition network; S5. Based on step S4, input the three-channel feature matrix of the time domain and frequency domain of the training set radar signals obtained in step S3 into the variable-length radar modulation recognition network for network training; S6. Use the variable-length radar modulation recognition network trained in step S5 to identify the modulation types of the variable-length radar signals in the test set processed in step S3; Input the test set processed in step S3 into the one-dimensional convolutional variable-length radar modulation recognition network trained in step S5 for forward propagation to obtain the classification result of the test set; S7. Based on the Ascend AI platform, perform software and hardware optimization on the variable-length radar modulation recognition model trained in step S5, offline the optimized model in a fixed standard, and convert it into a model file supported by the Ascend AI platform. Finally, deploy the generated model file to the Ascend AI platform and modify the length of the input data based on the format restrictions of the input of the Ascend AI platform to complete the identification of variable-length radar signals of the convolutional neural network based on the Ascend AI platform.

2. The method for identifying variable-length radar signals of a convolutional neural network based on the Ascend AI platform according to claim 1, wherein The specific steps of step S3 are as follows: For the training set and test set obtained in step S2, use IQ sampling to extract the time-domain IQ dual-channel feature data of the radar signal, and perform FFT on the extracted time-domain digital signal to obtain the frequency-domain single-feature channel data. Then, the signal model obtained in step S2, the IQ sampling and FFT calculation expressions in step S3 are as follows: where t represents time, z(t) represents a complex signal, x(t) represents the real part of the complex signal, a(t) represents the amplitude of the complex signal, θ(t) represents the signal phase, and f c represents the signal carrier frequency, g(t) represents the complex envelope of x(t), and g(t) = a(t)e θ(t) ; represents the Hilbert transform of x(t); at the sampling rate f s the sampling sequence of x(t) is x(i), and the number of sampling points is N S , then its complex analytical expression is as follows: Among them, x(i) represents the in-phase component of z(i), that is, the I component. represents the quadrature component of z(i), that is, the Q component. Perform discrete Fourier transform on z(i) again, and the expression is as follows: Among them, f(i) represents the frequency-domain data after Fourier transform, and i represents the serial number; Then, signal features are extracted from the time domain and the frequency domain, that is, the three-channel data x(i) obtained by sampling, and f(i) are concatenated to obtain the model input feature matrix X of a single signal, and the expression is as follows: For n signals, cascade them as the batch input Y of the model, and the expression is as follows: Y = [X(0); X(2);... X(n - 1)].

3. The variable-length radar signal recognition method of a convolutional neural network based on the Ascend AI platform according to claim 1, wherein The specific steps of step S4 are as follows: S41. Construct a feature dimension elevation module; The feature dimension elevation module includes: a convolutional layer, a batch normalization layer, and a ReLU activation layer; S42. Construct an SE-ResNet sub-module; The SE-ResNet sub-module includes: 3 convolutional layers, 2 batch normalization layers, 2 ReLU activation layers, 1 residual block, and 1 attention mechanism layer; the attention mechanism layer sequentially includes: global average pooling, a fully connected layer, a ReLU activation layer, a fully connected layer, and a Sigmoid activation layer, which are used to extract independent features of different feature channels; Among them, the input of the SE-ResNet sub-module is the feature map after the three feature channels pass through the feature dimension elevation module. After passing through 3 convolutional layers in sequence, deep features are extracted. The first 2 convolutional layers are followed by a batch normalization layer and a ReLU activation layer. The extracted deep features are fused with the original shallow features input through a residual block. Then, the obtained fused features are input into the attention mechanism layer. The output of the residual block is fused with the output of the original residual block after passing through the attention mechanism layer; S43. Based on step S42, construct an SE-ResNet module; The SE-ResNet module is composed of three SE-ResNet sub-modules in series to deeply extract independent features of different feature channels; Among them, the output of each SE-RseNet sub-module is used as the input of the next one, and extraction is carried out layer by layer. Finally, deep and global features of three channels are obtained; S44. Construct a feature fusion module; The feature fusion module includes: a convolutional layer, a batch normalization layer, a ReLU activation function, and an attention mechanism layer to achieve feature fusion in the time domain and frequency domain; Among them, the attention mechanism layer includes in sequence: global average pooling, a fully connected layer, a ReLU activation layer, a fully connected layer, and a Sigmoid activation layer; and the feature fusion module takes the output of the SE-ResNet module as the input, and the output result is the multi-dimensional feature fusion map after fusion completion; S45. Construct a fully connected module; The fully connected module includes: a pooling layer, a ReLU activation function, and a fully connected layer; taking the output of the feature fusion module as the input to obtain the posterior probability corresponding to each category; S46. Combine the sub-modules constructed in steps S41 - S45 to form a convolutional variable-length radar modulation recognition network; After cascading the feature dimension elevation module, 3 parallel SE-ResNet modules, the feature fusion module, and the fully connected module in sequence, a one-dimensional convolutional variable-length radar modulation recognition network is formed; The input of the recognition network is the feature matrix after the original radar signal is processed by step S3, and the output is the classification result; Among them, the category corresponding to the data with a value greater than the pre-set threshold in the posterior probability is the classification result of this network structure.

4. A method for identifying variable-length radar signals of a convolutional neural network based on the Ascend AI platform according to claim 1, characterized in that The specific steps of step S5 are as follows: S51. After splicing the IQ data and frequency domain data of each variable-length radar signal in the training set processed by step S3, input them into the one-dimensional convolutional variable-length radar modulation recognition network constructed in step S4 for forward propagation, and calculate the cost function value; Among them, calculating the cost function value is to use the binary cross-entropy loss function as the cost function, and the calculation expression is as follows: Among them, BCE represents the binary cross-entropy loss function, prob ∈ R 1×m represents the output probability, lab ∈ R 1×m represents the ground truth of the overlapping signal in one-hot code form, thre ∈ R 1×m represents the threshold, m represents the total number of modulation types, and R represents the real number field; S52. Use the backpropagation algorithm based on stochastic gradient descent to update the parameters of the variable-length radar modulation recognition network; S53. Based on step S52, iterate the backpropagation process until the cost function converges to obtain the trained one-dimensional convolutional variable-length radar modulation recognition network.

5. A variable-length radar signal recognition method for a convolutional neural network based on the Ascend AI platform according to claim 1, characterized in that The specific steps of step S7 are as follows: S71. Based on the characteristics of the Ascend AI platform, optimize the model trained in step S5 in terms of software and hardware, that is, perform operator optimization and operator replacement; First, optimize the model size based on the hardware conditions of the Ascend AI platform, and then optimize the operators based on the software environment of the Ascend AI processor; Among them, if there is no corresponding operator for the pooling layer of the fully connected module of the trained model in Ascend AI, the method of the pooling layer operator is replaced. The upsampling method is adopted, and the linear interpolation method is used to replace the operator of the pooling layer; S72. Based on step S71, store the optimized variable-length radar modulation recognition network in the ONNX file format, perform model offline conversion, and generate a model file supported by the Ascend AI platform; First, perform model offline conversion based on environmental standards, that is, use the official ONNX tool to offline the network and obtain the ONNX model file; Then generate the model for the Ascend AI platform development environment, that is, generate the OM model according to the characteristics of the Ascend AI platform from the obtained ONNX model file, and obtain the model file supported by the Ascend AI platform; S73. Deploy the generated OM model file to the Ascend AI platform to achieve network engineering deployment, and process the length of each radar signal in the test set obtained in step S3 into a unified length to adapt to the model input format of the Ascend AI platform; First, use the official ATC tool of Ascend AI, take the ONNX model as the input, and configure the relevant parameters of the OM model, including: input dimension, application platform, and convert the ONNX model into the OM model; Then truncate the input signal of the OM model, and set the truncation length to an integer multiple of 500 closest to the original length. The calculation formula is as follows: Among them, represents the floor function, and N s represents the original number of sampling points; S74. Input the test set data processed in step S73 into the deployed convolutional variable-length radar modulation recognition network for forward propagation to obtain the classification results of the test set data; Compare the classification results with the results of the conventional platform network. If the inference accuracy loss exceeds 5%, then, taking the OM model input format as the standard, truncate the test set of the conventional platform model in the same format, repeat step S5 for retraining, and then redeploy the Ascend AI platform according to steps S6 - S7.

Images