A method for electromagnetic signal type recognition based on a hybrid pulse neural network
Patent Information
- Application Number
- CN202410461502.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-17
- Publication Date
- 2026-10-20
- Estimated Expiration
- 2044-04-17
AI Technical Summary
[0004]鉴于上述的分析，本发明实施例旨在提供一种基于混合脉冲神经网络的电磁信号类型识别方法，用以解决现有无线电信号识别方法推理计算功耗要求高的技术问题
[0045]与现有技术相比，本发明至少可实现如下有益效果之一：
Smart Images
Figure CN118568537B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of signal processing technology, and in particular to a method for electromagnetic signal type identification based on a hybrid pulse neural network. Background Technology
[0002] Signal type identification refers to recognizing the system architecture or modulation scheme of received signals in wireless communication, classifying signals into different systems such as Wi-Fi, 4G (The 4th Generation Communication System), Bluetooth, etc., or different modulation types such as PSK (Phase Shift Keying) and FSK (Frequency Shift Keying). Signal type identification technology is crucial in wireless communication systems, helping the receiver understand the characteristics of the transmitted signal, thereby enabling correct signal demodulation and information reconstruction. Signal type identification technology has wide applications in wireless communication, such as radio spectrum monitoring, smart radio, the Internet of Things, and radar signal processing. Accurately identifying signal types can improve the efficiency and performance of wireless communication systems.
[0003] In recent years, with the development of deep learning technology, deep learning has gradually demonstrated its advantages in signal type recognition. Deep learning methods can automatically learn the feature representation and pattern recognition of signals without the need for manually designing feature extraction algorithms. Commonly used deep learning models include Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM), collectively known as Artificial Neural Networks (ANN). ANN has achieved good results in signal modulation type recognition; however, the performance of ANN depends on the depth and complexity of the model, requires a large amount of computational and storage resources, has high requirements for the deployment environment, and is not suitable for some small or micro devices with limited power consumption. Summary of the Invention
[0004] Based on the above analysis, the present invention aims to provide an electromagnetic signal type identification method based on a hybrid pulse neural network to solve the technical problem of high power consumption requirements for inference calculation in existing radio signal identification methods.
[0005] This invention provides a method for electromagnetic signal type identification based on a hybrid pulse neural network, comprising the following steps:
[0006] The in-phase and quadrature components of the collected radio signals are extracted and preprocessed to obtain the amplitude and phase of the normalized signal. The signal modulation type label is used to form a training sample set.
[0007] Initialize the shallow residual network and the deep residual network; train the shallow residual network and the deep residual network using the training sample set until the training converges, and obtain the trained shallow residual network and the deep residual network.
[0008] The trained deep residual network is transformed into an impulse residual network, and the trained shallow residual network and the impulse residual network together form a signal type recognition neural network model.
[0009] The amplitude and phase of the radio signals acquired in real time are calculated, normalized, and then input into the signal type recognition neural network model to obtain the prediction result of the signal modulation type of the radio signals.
[0010] Furthermore, the step of extracting the in-phase and quadrature components of the acquired radio signal and performing preprocessing includes:
[0011] Obtain the complex baseband sampling sequence z = [z1, z2, ..., z] of N radio signals. M ], where M is the number of signal sampling points and N is the total number of radio signal sample data acquired;
[0012] For each sampling point, the in-phase and quadrature components of the signal are extracted as follows:
[0013] z i =I i +jQ i
[0014] Where I and Q are the in-phase and quadrature components of the signal, respectively, j is the imaginary unit, and i = 1, 2, ..., M;
[0015] The amplitude 'a' of the signal is calculated based on its in-phase and quadrature components. i and phase p i The amplitude and phase of the signal are normalized respectively to obtain the normalized signal amplitude ai. ,nor and phase p i,nor .
[0016] Furthermore, the shallow residual network is connected in series with the input layer and one residual block;
[0017] The deep residual network consists of an input layer, first, second, third and fourth residual blocks, a fully connected layer and an output layer connected in series.
[0018] Each residual block consists of a first convolutional layer, a ReLU activation layer, a second convolutional layer, a BN batch normalization layer, a summing layer, and a ReLU activation layer in sequence; the input layer is skipped to the summing layer, and the convolutional kernel size of the convolutional layer is 1×1.
[0019] Furthermore, the shallow residual network is trained using the training sample set as follows:
[0020] The sample X in the training set is input into the shallow residual network. One X enters the first convolutional layer of the residual block of the shallow residual network to perform convolution operation, and then passes through the ReLU activation and the second convolutional layer, and then through the BN batch normalization layer to perform batch normalization processing to obtain F(X).
[0021] The other path X is summed with F(X) via a skip connection and then activated by ReLU to obtain the signal feature x, as follows:
[0022] x = ReLU(X + F(X)).
[0023] Furthermore, the deep residual network is trained as follows:
[0024] The signal features x extracted by the shallow residual network are input into the deep residual network and transformed by the first, second, third and fourth residual blocks in sequence to obtain the deep features of the signal. These features are then input into the fully connected layer for signal classification, and the fully connected layer outputs the prediction result of the signal modulation type.
[0025] Based on the prediction results and the loss of the signal modulation type label, backpropagation is used to update the connection weight parameters of the first and second convolutional layers and the BN batch normalized layer in each residual block.
[0026] Furthermore, x is input into the deep residual network, and after transformation by the first, second, third and fourth residual blocks, the deep features of the signal are obtained, as shown below;
[0027]
[0028] Where y1, y2, y3 and y4 are the outputs of the summation layer of the first, second, third and fourth residual blocks, respectively, and F1(x), F2(y1), F3(y2) and F4(y3) represent the results of the processing of the first, second, third and fourth residual blocks through the first convolutional layer, ReLU activation layer, second convolutional layer and BN batch normalization, respectively.
[0029] Inputting y4 into the fully connected layer yields a prediction result of the signal modulation type of the radio signal.
[0030] Furthermore, the pulse residual network includes an input layer and first, second, third and fourth pulse residual blocks connected in series, a fully connected layer and an output layer;
[0031] Each pulse residual block sequentially includes a first convolutional layer, a first spiking neuron, a second convolutional layer, and a second spiking neuron, with a skip connection from the input layer to the second spiking neuron.
[0032] Furthermore, converting the trained deep residual network into a pulse residual network includes:
[0033] The ReLU activation function in each residual block of the trained deep residual network is transformed into the first spiking neuron and the second spiking neuron in the corresponding residual block of the spiking residual network;
[0034] The weights and bias parameters of the first and second convolutional layers in the first, second, third, and fourth residual blocks of the trained deep residual network are normalized and then assigned to the weights and bias parameters of the first and second convolutional layers in the first, second, third, and fourth pulse residual blocks of the pulse residual network, respectively.
[0035] Remove the BN batch normalization layer and summation layer from each residual block in the deep residual network;
[0036] A scaling factor λ is added to the skip connection from the input of each residual block in the pulse residual network to the second spiking neuron to adjust the residual block.
[0037] Furthermore, the weights w1 and w2 and the bias parameters b1 and b2 of the first and second convolutional layers of each pulse residual block of the trained deep residual network are normalized as follows:
[0038]
[0039] in, , , represent the weights and biases of the first and second convolutional layers of each residual block in the normalized pulse residual network, respectively. max1, max2, and max3 represent the maximum input of the residual block in the deep residual network, the maximum output activation of the first convolutional layer of the residual block, and the maximum output activation of the second convolutional layer of the residual block, respectively.
[0040] Furthermore, after x is input into the pulse residual network, it undergoes transformation by the first, second, third, and fourth pulse residual blocks to obtain the deep-level features of the signal, as shown below;
[0041]
[0042] Where n1, n2, n3, and n4 are the outputs of the second spiking neurons in the first, second, third, and fourth spiking residual blocks, respectively; Q1(x), Q2(n1), Q3(n2), and Q4(n3) represent the processing results of the first, second, third, and fourth spiking residual blocks after passing through the first convolutional layer, the first spiking neuron, and the second convolutional layer, respectively; λx, λn1, λn2, and λn3 are the skip connection values in the first, second, third, and fourth spiking residual blocks, respectively.
[0043] Spike is an IF spiking neuron that converts data passing through the spiking neuron into a binary data sequence;
[0044] Inputting n4 into the fully connected layer yields a prediction result of the signal modulation type of the radio signal.
[0045] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:
[0046] 1. This invention utilizes skip connections in residual blocks, enabling the network to learn and represent complex features more effectively. Skip connections allow information to propagate directly between layers, avoiding problems such as vanishing or exploding gradients, thus helping to improve the efficiency and performance of feature learning;
[0047] 2. By employing a hybrid pulse neural network combined with deep learning, high-accuracy identification of radio signal modulation types is achieved, thereby improving the accuracy of signal modulation type identification;
[0048] 3. This invention is suitable for low-power environments. Due to the event-driven nature of the pulse residual network, it effectively saves computing and storage resources, making it suitable for small or micro devices with limited power consumption and reducing deployment costs.
[0049] 4. This invention is applicable to identification tasks of various digital communication modulation styles or types, including different systems and modulation methods, and has a wide range of applications and versatility, with broad application prospects and practicality;
[0050] 5. This invention combines shallow residual networks and deep residual networks to extract the same-direction and orthogonal components of radio signals, effectively extracting the time-domain and frequency-domain features of radio signals. It utilizes the skip connections of residual blocks to achieve direct information propagation, avoiding the gradient vanishing or gradient exploding problems, resulting in more stable gradients and improving the efficiency of feature extraction and signal modulation type identification.
[0051] 6. Deep residual networks are converted into spiking residual networks. Spiking residual networks (SNNs) have event-driven characteristics, utilizing pulse coding and temporal coding to ensure that neurons only become active when specific events (pulses) occur. This characteristic makes SNNs more computationally efficient, improving energy efficiency and reducing computational resource consumption. Furthermore, the conversion from deep residual networks to spiking residual networks does not add parameters and removes the batch normalization layer and summing layer. Through the conversion and real-time inference of spiking residual networks, instant recognition of real-time acquired radio signals can be achieved, meeting the application scenarios requiring fast response and real-time performance.
[0052] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description
[0053] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.
[0054] Figure 1 This is a flowchart illustrating an electromagnetic signal type identification method based on a hybrid pulse neural network.
[0055] Figure 2 This is a schematic diagram of a shallow residual network structure;
[0056] Figure 3 This is a schematic diagram of a deep residual network structure;
[0057] Figure 4 This is a schematic diagram of the pulse residual network structure;
[0058] Figure 5 This is a schematic diagram illustrating the conversion process between an artificial neural network and a spiking neural network constructed according to an embodiment of the present invention. Detailed Implementation
[0059] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.
[0060] To reduce the power consumption of traditional deep learning models, spiking residual networks (SNNs) were developed. SNNs are event-driven because they utilize pulse coding and temporal coding, meaning that neurons only become active when a specific event (spike) occurs. This characteristic makes SNNs more computationally efficient. For energy efficiency reasons, applying SNNs to signal type recognition is meaningful.
[0061] A specific embodiment of the present invention, such as Figure 1 As shown, a method for electromagnetic signal type identification based on a hybrid pulse neural network is disclosed, comprising the following steps:
[0062] Step S1: Extract the in-phase and quadrature components of the collected radio signal and preprocess them to obtain the amplitude and phase of the normalized signal. The signal modulation type label is used to form a training sample set.
[0063] Step S2: Initialize the shallow residual network and the deep residual network; train the shallow residual network and the deep residual network using the training sample set until the training converges, and obtain the trained shallow residual network and the deep residual network.
[0064] Step S3: The trained deep residual network is converted into an impulse residual network, and the trained shallow residual network and the impulse residual network form a signal type recognition neural network model.
[0065] Step S4: After normalizing the amplitude and phase of the radio signal acquired in real time, input the signal type recognition neural network model to obtain the prediction result of the signal modulation type of the radio signal.
[0066] Depend on Figure 1 As shown, the method in this embodiment utilizes the powerful feature extraction capabilities of shallow residual networks to extract the time-domain and frequency-domain features of radio signals; it uses deep residual networks to extract the deep-level features of the signals, and uses the fully connected layers of the deep residual networks to classify the deep-level features to obtain the prediction results of the signal modulation type of the radio signals.
[0067] The trained deep residual network transforms the ReLU activation function into the first and second spiking neurons of the spiking residual network; and assigns the weights and bias parameters of the first and second convolutional layers to the weights and bias parameters of the first, second, third, and fourth spiking residual blocks of the spiking residual network, respectively; removes the summation layer of the deep residual network, and improves the performance of the spiking network by assigning scaling factors to the edges of the jump connections of the residual blocks, thus transforming it into a spiking residual network; finally, while achieving signal modulation type recognition, the overall energy efficiency of the network is improved.
[0068] The method in this invention can be adapted to the identification of various digital communication modulation styles or types, and has versatility and practicality.
[0069] Step S1 includes steps S11-S12.
[0070] Step S11: Extract the in-phase and quadrature components of the acquired radio signal and perform preprocessing to obtain the amplitude and phase of the normalized signal.
[0071] The target radio signal is received and sampled, covering a variety of different modulation methods and signal types. These signal modulation types include, but are not limited to:
[0072] (1) Quadrature Phase Shift Keying (QPSK);
[0073] (2) Eight-phase phase shift keying (8PSK);
[0074] (3) Binary Phase Shift Keying (BPSK);
[0075] (4) Offset Quadrature Phase Shift Keying (OQPSK);
[0076] (4) Hexadecimal Quadrature Amplitude Modulation (16QAM);
[0077] (5) 32-Quadrature Amplitude Modulation (32QAM);
[0078] (5) 64-Quadrature Amplitude Modulation (64QAM);
[0079] (7) Binary Frequency Shift Keying (2FSK);
[0080] (8) Quad-Level Frequency Shift Keying (4FSK);
[0081] (9) Octal Frequency Shift Keying (8FSK);
[0082] (10) Quadruple Pulse Amplitude Modulation (4PAM);
[0083] (11) Signal categories such as octal pulse amplitude modulation (8PAM).
[0084] By receiving and sampling these signals with different modulation types, we can obtain the sample data required in the signal processing process, which can then be used to construct shallow residual network and deep residual network models to identify the signal modulation type.
[0085] The step of extracting the in-phase and quadrature components of the acquired radio signals and performing preprocessing includes:
[0086] Obtain the complex baseband sampling sequence z = [z1, z2, ..., z] of N radio signals. M ], where M is the number of signal sampling points and N is the total number of radio signal sample data acquired;
[0087] For each sampling point, the in-phase and quadrature components of the signal are extracted as shown in formula (1):
[0088] z i =I i +jQ i Formula (1)
[0089] Where I and Q are the in-phase and quadrature components of the signal, respectively, j is the imaginary unit, and i = 1, 2, ..., M;
[0090] The amplitude 'a' of the signal is calculated based on its in-phase and quadrature components. i and phase p i The amplitude and phase of the signal are normalized respectively to obtain the normalized signal amplitude ai. ,nor and phase p i,nor .
[0091] The amplitude 'a' of the signal is calculated based on its in-phase and quadrature components. i and phase p i As shown in formula (2):
[0092]
[0093] Then the amplitude and phase of the signal are normalized, as shown in formula (3):
[0094]
[0095] Among them, a min a max These are the minimum and maximum values of the amplitude, p. min p max These are the minimum and maximum values of the phase, respectively.
[0096] Step S12: The amplitude and phase of the normalized signal are combined with the signal modulation type label to form a training sample set.
[0097] The amplitude and phase data of the pre-processed normalized signal, along with the corresponding signal modulation type, are used as labels to form a training sample set.
[0098] 70% of the samples in the training set are used as the training set to train the shallow and deep residual networks in subsequent steps, and 30% of the samples are used as the validation set to validate the shallow and deep residual networks.
[0099] The training and validation sets are randomly divided to ensure that samples of each modulation type category are evenly distributed as much as possible.
[0100] Step S1 involves acquiring the radio signal, extracting its in-phase and quadrature components, and preprocessing them. It also calculates and normalizes the signal's amplitude and phase. Normalized amplitude and phase data allow for better extraction of key information from the signal, reducing redundancy and unnecessary information, thus improving the efficiency and accuracy of model training. Furthermore, amplitude and phase are easier to understand and more closely resemble the physical meaning of the signal. This representation is more intuitive and aids in the analysis and interpretation of signal characteristics. In addition, in some cases, the amplitude and phase representation is more robust and resistant to noise interference, improving the stability and robustness of the signal recognition method.
[0101] Step S2 includes steps S21-S22.
[0102] Residual networks extract signal features through the stacking of residual blocks. One of the main advantages of the residual network structure is its ability to better capture and learn features during network training. Residual connections within residual blocks allow the network to learn residuals rather than directly learning the original features. This mechanism enables residual networks to learn and represent complex features more effectively. Because residual network connections allow skip connections, information can propagate directly between layers without being affected by problems such as vanishing gradients, facilitating better feature learning and representation. This capability allows residual networks to extract features at deeper network layers, solving the gradient vanishing and model degradation problems encountered by traditional deep networks during training.
[0103] Step S21: Initialize the shallow residual network and the deep residual network, and train the shallow residual network using the training sample set.
[0104] like Figure 2 The structure of the shallow residual network is shown below:
[0105] The shallow residual network is connected in series with the input layer and one residual block.
[0106] like Figure 3 As shown, the structure of a deep residual network is as follows:
[0107] The deep residual network consists of an input layer, first, second, third and fourth residual blocks, a fully connected layer and an output layer connected in series.
[0108] The structure of each residual block in a deep residual network is the same as that in a shallow residual network.
[0109] Each residual block consists of a first convolutional layer, a ReLU activation layer, a second convolutional layer, a BN batch normalization layer, a summing layer, and a ReLU activation layer in sequence; the input layer is skipped to the summing layer, and the convolutional kernel size of the convolutional layer is 1×1.
[0110] The input of the first residual block is obtained from the input layer; the output of the first residual block is the input of the second residual block; the output of the second residual block is the input of the third residual block; and the output of the third residual block is the input of the fourth residual block.
[0111] The shallow residual network is trained using a training sample set. The samples in the training sample set are normalized amplitudes and phases, along with their corresponding signal modulation type labels.
[0112] like Figure 5 As shown, the shallow residual network is trained using the training sample set as follows:
[0113] The sample X in the training set is input into the shallow residual network. One X enters the first convolutional layer of the residual block of the shallow residual network to perform convolution operation, and then passes through the ReLU activation and the second convolutional layer, and then through the BN batch normalization layer to perform batch normalization processing to obtain F(X).
[0114] The other path X is summed with F(X) through a skip connection and then activated by ReLU to obtain the signal feature x, as shown in formula (4):
[0115]
[0116] Residual blocks are the basic building blocks of shallow residual networks. Shallow residual networks extract features, specifically as follows:
[0117] (1) Input layer: Normalized signal amplitude and phase data;
[0118] (2) Convolution operation: The input data passes through two convolutional layers, the first convolutional layer and the second convolutional layer. The convolution operation is used to extract features from the input data. The kernel size of the first and second convolutional layers is 1×1.
[0119] (3) ReLU activation layer: After the first convolutional layer, the ReLU (Modified Linear Unit) activation function is used to perform nonlinear mapping on the convolutional output; and after the summation layer, the summation result is activated by the ReLU activation function;
[0120] (4) Batch Normalization (BN): Normalizes the features output by the second convolutional layer in batches; accelerates training convergence speed, reduces internal covariate shift, and improves the generalization ability of shallow residual networks.
[0121] (5) Skip connection: One of the key features of shallow residual networks is skip connection. Skip connection allows the network to skip the learning of one or more layers, thereby learning the residual more effectively. In shallow residual networks, skip connection is between the input layer and the summation layer.
[0122] (6) Summation layer: The training sample X in the training sample set is input into the shallow residual network. One path X enters the first convolutional layer of the residual block of the shallow residual network for convolution operation, then passes through ReLU activation and then through the second convolutional layer, and then through the BN batch normalization layer for batch normalization to obtain F(X); the other path X is summed with F(X) through skip connections and then activated by ReLU to obtain signal feature x;
[0123] (7) Output: After the summation layer, the shallow residual network is activated by ReLU to output x.
[0124] The purpose of step S21 is to extract the time-domain and frequency-domain features of the radio signal based on the normalized amplitude and phase data and using a shallow residual network.
[0125] Step S22: Train the deep residual network.
[0126] A deep residual network is constructed as a classifier to classify the time-domain and frequency-domain features extracted by the shallow residual network, thereby obtaining the prediction results of the signal modulation type.
[0127] like Figure 5 As shown, the deep residual network is trained as follows:
[0128] The signal features x extracted by the shallow residual network are input into the deep residual network and transformed by the first, second, third and fourth residual blocks in sequence to obtain the deep features of the signal. These features are then input into the fully connected layer for signal classification, and the fully connected layer outputs the prediction result of the signal modulation type.
[0129] x is input into the deep residual network, and after transformation by the first, second, third and fourth residual blocks, the deep features of the signal are obtained, as shown in formula (5);
[0130]
[0131] Where y1, y2, y3 and y4 are the outputs of the summation layer of the first, second, third and fourth residual blocks, respectively, and F1(x), F2(y1), F3(y2) and F4(y3) represent the results of the processing of the first, second, third and fourth residual blocks through the first convolutional layer, ReLU activation layer, second convolutional layer and BN batch normalization, respectively.
[0132] Inputting y4 into the fully connected layer yields a prediction result of the signal modulation type of the radio signal.
[0133] like Figure 3 As shown, the deep residual network acts as a classifier to classify the features extracted by the shallow residual network, including:
[0134] (1) Input layer: shallow neural network extracts x, including the time domain features and frequency domain features of the signal;
[0135] (2) Processing: The deep residual network's processing includes four residual blocks, each of which performs feature extraction and representation learning. Within each residual block, the input features are processed through a first convolutional layer, a ReLU activation layer, a second convolutional layer, and a BN batch normalization layer. Skip connections between the input layer and the summing layer are used to facilitate gradient flow and information transfer. The convolution, ReLU activation functions, and BN batch normalization operations in each residual block are the same as in the shallow residual network.
[0136] (3) Output: The output of the deep residual network is the prediction result of the signal modulation type. The output of the fourth residual block is activated by ReLU and then input into the fully connected layer for classification. The final output is the prediction result of the signal modulation type.
[0137] The main purpose of using a deep residual network with 4 residual blocks is to increase the depth of the network in order to better capture and learn the deep features of the input data.
[0138] Specifically, increasing the number of residual blocks helps deep residual networks gradually build more complex feature representations. Each residual block can be viewed as a feature extractor, which can extract features from the input data through convolutional layers, batch normalization (BN) layers, and activation functions. As the number of residual blocks increases, the network can gradually learn more abstract and complex deep features, thereby improving its classification performance for signal modulation types.
[0139] In addition, increasing the number of residual blocks can help the network better solve problems such as vanishing and exploding gradients, thus enabling more stable training and making it easier for the network to learn effective feature representations during training.
[0140] Therefore, using four residual blocks can make the network deeper and more expressive, thereby improving the performance of signal modulation type recognition tasks.
[0141] Deep residual networks take the time-domain and frequency-domain output features of shallow residual networks as input, perform feature extraction and representation learning through four residual blocks, and finally output the prediction result of the signal modulation type.
[0142] Using the stochastic gradient descent method in the backpropagation algorithm, the deep residual network is trained using the features extracted from the shallow residual network in step S21 until the training converges. The trained deep residual network is then used.
[0143] The purpose of step S22 is to extract deeper features to accelerate the training of the deep residual network to be transformed and the inference process of the spiking residual network.
[0144] The shallow residual network and the deep residual network are validated using the validation set in the training sample set until training converges, resulting in the trained shallow residual network and the deep residual network.
[0145] Step S3, specifically.
[0146] The deep residual network is transformed into a spiking neural network, and the features extracted by the shallow residual network are used for inference to output the prediction result of the signal modulation type.
[0147] like Figure 4 As shown, the pulse residual network includes an input layer and first, second, third and fourth pulse residual blocks connected in series, a fully connected layer and an output layer;
[0148] Each pulse residual block sequentially includes a first convolutional layer, a first spiking neuron, a second convolutional layer, and a second spiking neuron, with a skip connection from the input layer to the second spiking neuron.
[0149] Converting the trained deep residual network into a pulse residual network includes:
[0150] The ReLU activation function in each residual block of the trained deep residual network is transformed into the first spiking neuron and the second spiking neuron in the corresponding residual block of the spiking residual network;
[0151] The weights and bias parameters of the first and second convolutional layers in the first, second, third, and fourth residual blocks of the trained deep residual network are normalized and then assigned to the weights and bias parameters of the first and second convolutional layers in the first, second, third, and fourth pulse residual blocks of the pulse residual network, respectively.
[0152] Remove the BN batch normalization layer and summation layer from each residual block in the deep residual network;
[0153] A scaling factor λ is added to the skip connection from the input of each residual block in the pulse residual network to the second spiking neuron to adjust the residual block.
[0154] The trained deep residual network is converted into an impulse residual network; the trained shallow residual network and the impulse residual network together form a signal type recognition neural network model.
[0155] like Figure 5 As shown, the trained deep residual network is transformed into a spiking neural network, and then the spiking neural network is used to perform inference on the features extracted by the shallow residual network, including:
[0156] (1) Transform the ReLU activation functions in the residual blocks of the deep residual network into the first and second spiking neurons of the spiking residual network. This matches the radio frequency range of the spiking neurons in the spiking residual network, since spiking neurons are not allowed to fire multiple pulses within a single time step. Otherwise, the information carried by high activation values will be lost during the transformation process. For stacked layers, the weights and biases of each layer are normalized by the maximum possible activation value calculated during training;
[0157] (2) Transform weights and biases: The parameters of the first and second convolutional layers in the first, second, third and fourth residual blocks of the trained deep residual network, including weights and biases, are normalized and then assigned to the parameters of the first and second convolutional layers of the first, second, third and fourth residual blocks of the spiking residual network to ensure the operation of the network on the spiking neurons.
[0158] (3) Discard BN batching to a single layer: Since BN batching to a single layer is not required in the impulse residual network, it is necessary to discard the BN batching to a single layer in the deep residual network;
[0159] (4) Modify the structure: Since the spiking neurons automatically perform the superposition operation, the summation layer of the deep residual network is not required. The network structure of the spiking residual network is modified accordingly during the conversion process.
[0160] (5) When making skip connections from the input layer to the second spiking neuron, add a scaling factor to adjust the output of the residual block in order to improve the effect of the converted spiking residual network.
[0161] The weights w1 and w2 and the bias parameters b1 and b2 of the first and second convolutional layers of each pulse residual block of the trained deep residual network are normalized, as shown in formula (6):
[0162]
[0163] in, , , represent the weights and biases of the first and second convolutional layers of each residual block in the normalized pulse residual network, respectively. max1, max2, and max3 represent the maximum input of the residual block in the deep residual network, the maximum output activation of the first convolutional layer of the residual block, and the maximum output activation of the second convolutional layer of the residual block, respectively.
[0164] The conversion of artificial neural networks (deep residual networks) to spiking residual networks involves transforming the weights and biases of the artificial neural network into the appropriate form in spiking residual networks.
[0165] Finally, the transformed pulse residual network is used to infer the time-domain and frequency-domain features extracted by the shallow residual network.
[0166] The inference process of a Spike Residual Network (SNN) involves accumulating impulses over multiple time steps to identify the signal modulation type. Each neuron receives impulse inputs from the previous layer over multiple time steps and accumulates these inputs. At each time step, the neuron decides whether to output an impulse based on the received impulse inputs and its own state. This process is performed step-by-step, with a classification result at each time step, until the network's maximum number of time steps is reached or a preset stopping condition is met.
[0167] After x is input into the pulse residual network, it is transformed by the first, second, third and fourth pulse residual blocks to obtain the deep features of the signal, as shown in formula (7);
[0168]
[0169] Where n1, n2, n3, and n4 are the outputs of the second spiking neurons in the first, second, third, and fourth spiking residual blocks, respectively; Q1(x), Q2(n1), Q3(n2), and Q4(n3) represent the processing results of the first, second, third, and fourth spiking residual blocks after passing through the first convolutional layer, the first spiking neuron, and the second convolutional layer, respectively; λx, λn1, λn2, and λn3 are the skip connection values in the first, second, third, and fourth spiking residual blocks, respectively.
[0170] Spike is an IF spiking neuron that converts data passing through the spiking neuron into a binary data sequence;
[0171] Inputting n4 into the fully connected layer yields a prediction result of the signal modulation type of the radio signal.
[0172] Spike is an IF spiking neuron. The neuron integrates the incoming weighted spurious signals in the time domain. The equation for the input weighted spurious signal at a certain time step t is: s(t) = wx(t) + b, where w is the spurious weight, b is the neuron bias, and x(t) is the input from the previous layer of spiking neurons. After input integration, the signal enters the spiking neuron to accumulate the membrane potential v. When the membrane potential reaches the threshold voltage v... th At this time, the neuron fires an output pulse, and then the membrane potential is reduced by a value. th The following formula describes this: v(t) = v(t-1) + s(t) - v th δ(t). Here, δ(t) is a step function used to determine whether to fire a pulse. Thus, the data passing through the spiking neurons is converted into a binary data sequence of 0s or 1s and output.
[0173] The purpose of step S3 is to transform the trained deep residual network into an impulse residual network, and then use the impulse residual network to infer the features extracted by the shallow residual network to obtain the prediction result of the signal type, thereby improving energy efficiency.
[0174] In summary, the electromagnetic signal type identification method based on a hybrid pulse neural network in this embodiment of the invention has the following beneficial effects:
[0175] 1. This invention utilizes skip connections in residual blocks, enabling the network to learn and represent complex features more effectively. Skip connections allow information to propagate directly between layers, avoiding problems such as vanishing or exploding gradients, thus helping to improve the efficiency and performance of feature learning;
[0176] 2. By employing a hybrid pulse neural network combined with deep learning, high-accuracy identification of radio signal modulation types is achieved, thereby improving the accuracy of signal modulation type identification;
[0177] 3. This invention is suitable for low-power environments. Due to the event-driven nature of the pulse residual network, it effectively saves computing and storage resources, making it suitable for small or micro devices with limited power consumption and reducing deployment costs.
[0178] 4. This invention is applicable to identification tasks of various digital communication modulation styles or types, including different systems and modulation methods, and has a wide range of applications and versatility, with broad application prospects and practicality;
[0179] 5. This invention combines shallow residual networks and deep residual networks to extract the same-direction and orthogonal components of radio signals, effectively extracting the time-domain and frequency-domain features of radio signals. It utilizes the skip connections of residual blocks to achieve direct information propagation, avoiding the gradient vanishing or gradient exploding problems, resulting in more stable gradients and improving the efficiency of feature extraction and signal modulation type identification.
[0180] 6. Deep residual networks are converted into spiking residual networks. Spiking residual networks (SNNs) have event-driven characteristics, utilizing pulse coding and temporal coding to ensure that neurons only become active when specific events (pulses) occur. This characteristic makes SNNs more computationally efficient, improving energy efficiency and reducing computational resource consumption. Furthermore, the conversion from deep residual networks to spiking residual networks does not add parameters and removes the batch normalization layer and summing layer. Through the conversion and real-time inference of spiking residual networks, instant recognition of real-time acquired radio signals can be achieved, meeting the application scenarios requiring fast response and real-time performance.
[0181] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for electromagnetic signal type identification based on a hybrid pulse neural network, characterized in that, Includes the following steps: The in-phase and quadrature components of the collected radio signals are extracted and preprocessed to obtain the amplitude and phase of the normalized signal. The signal modulation type label is used to form a training sample set. Initialize the shallow residual network and the deep residual network; The shallow residual network and the deep residual network are trained using the training sample set until the training converges, resulting in the trained shallow residual network and the deep residual network. The trained deep residual network is transformed into an impulse residual network, and the trained shallow residual network and the impulse residual network together form a signal type recognition neural network model. Converting the trained deep residual network into a pulse residual network includes: The ReLU activation function in each residual block of the trained deep residual network is transformed into the first spiking neuron and the second spiking neuron in the corresponding residual block of the spiking residual network; The weights and bias parameters of the first and second convolutional layers in the first, second, third, and fourth residual blocks of the trained deep residual network are normalized and then assigned to the weights and bias parameters of the first and second convolutional layers in the first, second, third, and fourth pulse residual blocks of the pulse residual network, respectively. Remove the BN batch normalization layer and summation layer from each residual block in the deep residual network; A scaling factor is added to the skip connections from the input of each residual block in the pulse residual network to the second spiking neuron. To adjust the residual block; The amplitude and phase of the radio signals acquired in real time are calculated, normalized, and then input into the signal type recognition neural network model to obtain the prediction result of the signal modulation type of the radio signals.
2. The method according to claim 1, characterized in that, The step of extracting the in-phase and quadrature components of the acquired radio signals and performing preprocessing includes: Obtain the complex baseband sampling sequence of N radio signals Where M is the number of signal sampling points and N is the total number of radio signal sample data acquired; For each sampling point, the in-phase and quadrature components of the signal are extracted as follows: Where I and Q are the in-phase and quadrature components of the signal, respectively. Let i be the imaginary unit, i = 1, 2, ..., M; The amplitude of the signal is calculated based on its in-phase and quadrature components. and phase The amplitude and phase of the signal are normalized respectively to obtain the amplitude of the normalized signal. and phase .
3. The method according to any one of claims 1-2, characterized in that, The shallow residual network is connected in series with the input layer and one residual block; The deep residual network consists of an input layer, first, second, third and fourth residual blocks, a fully connected layer and an output layer connected in series. Each residual block sequentially comprises a first convolutional layer, a ReLU activation layer, a second convolutional layer, a batch normalization layer (BN), a summation layer, and another ReLU activation layer; the input layer is skipped to the summation layer, where the convolutional kernel size is... .
4. The method according to claim 3, characterized in that, The shallow residual network is trained using the training sample set as follows: The samples in the training set of the training sample set Input the shallow residual network, one path The residual blocks entering the shallow residual network undergo convolution operations in the first convolutional layer, followed by ReLU activation, then a second convolutional layer, and finally batch normalization processing via a BN layer to obtain the final result. ; Another route Connect via jumps The signal characteristics are obtained after summation and ReLU activation. ,as follows: 。 5. The method according to claim 4, characterized in that, The deep residual network is trained as follows: The signal features extracted by the shallow residual network The deep residual network is input and transformed through the first, second, third and fourth residual blocks in sequence to obtain the deep features of the signal. These features are then input to the fully connected layer for signal classification. The fully connected layer outputs the prediction result of the signal modulation type. Based on the prediction results and the loss of the signal modulation type label, backpropagation is used to update the connection weight parameters of the first and second convolutional layers and the BN batch normalized layer in each residual block.
6. The method according to claim 5, characterized in that, The deep residual network is input and transformed by the first, second, third and fourth residual blocks to obtain the deep features of the signal, as shown below; in, , , and These are the outputs of the summation layer for the first, second, third, and fourth residual blocks, respectively. , , and These represent the results of the processing of the first, second, third, and fourth residual blocks through the first convolutional layer, ReLU activation layer, second convolutional layer, and BN batch normalization, respectively. Will The signal is input to the fully connected layer to obtain a prediction result of the signal modulation type of the radio signal.
7. The method according to claim 6, characterized in that, The pulse residual network includes an input layer and first, second, third and fourth pulse residual blocks connected in series, a fully connected layer and an output layer; Each pulse residual block sequentially includes a first convolutional layer, a first spiking neuron, a second convolutional layer, and a second spiking neuron, with a skip connection from the input layer to the second spiking neuron.
8. The method according to claim 1, characterized in that, The weights of the first and second convolutional layers of each residual block in the trained deep residual network and and bias parameters and Normalization, as follows: in, , , , These represent the weights and biases of the first and second convolutional layers for each residual block of the normalized pulse residual network. , and These represent the maximum input of the residual block in the deep residual network, the maximum output activation of the first convolutional layer of the residual block, and the maximum output activation of the second convolutional layer of the residual block, respectively.
9. The method according to claim 1, characterized in that, After being input into the pulse residual network, the signal undergoes transformation through the first, second, third, and fourth pulse residual blocks to obtain its deep-level features, as shown below; in, , , and These are the outputs of the second spiking neurons in the first, second, third, and fourth spiking residual blocks, respectively. , , and These represent the processing results of the first, second, third, and fourth pulse residual blocks after passing through the first convolutional layer, the first spiking neuron, and the second convolutional layer, respectively. , , and These are the jump connection values in the first, second, third, and fourth pulse residual blocks, respectively; Spike is an IF spiking neuron that converts data passing through the spiking neuron into a binary data sequence; Will The signal is input to the fully connected layer to obtain a prediction result of the signal modulation type of the radio signal.