A deep learning method for radiation source individual identification based on complex spike neural network

Abstract

The present invention discloses a method for individual identification of radiation sources based on deep learning of complex pulse neurons. The method is identified by a recognition network based on complex pulse neurons. The recognition network based on complex pulse neurons includes a plurality of convolution modules connected end to end in sequence, and the convolution module includes a complex convolution layer and a pulse neural layer. The present invention adopts a complex convolution layer to extract the two-way channel features of the radiation source signal I and Q, encodes the features with a pulse neural layer, and maps the signal features into a sparse sequence, which greatly improves the expression ability, computational sparsity and computational efficiency of the radiation source signal, thereby improving the accuracy of individual identification of the radiation source; at the same time, a triangular gradient is used to circumvent the zero gradient problem of the ignition function during the training process, and solves the problem that the pulse neural layer cannot perform back propagation to update parameters.

Description

Technical Field

[0001] The invention belongs to the field of radiation source individual identification, and in particular relates to an individual identification method based on a complex pulse neural network. Background Art

[0002] Specific Emitter Identification (SEI), also known as emitter "fingerprinting", is a technique that extracts unique features from captured electromagnetic signals and uses existing knowledge to accurately identify the specific emitters that emit these signals. RF fingerprints are unique identifiers generated by inevitable variations in the hardware manufacturing process, so they are inherently unavoidable and difficult to replicate.

[0003] The extraction and classification of RF fingerprints are key parts of the core technology of radiation source individual identification. SEI is mainly divided into two methods:

[0004] 1) Manual radiation source identification techniques that require human intervention. Manual identification methods are further subdivided into: a) methods based on transient features; b) methods that rely on steady-state features. The development of RF fingerprinting technology can be traced back to the transient signal feature extraction technology proposed by Toonstra et al. in 1995. Factors such as channel noise and device hardware variation significantly increase the difficulty of extracting discernible features from on / off transients. Identification methods based on steady-state features require expert knowledge for manual feature extraction. The use of domain transformation techniques, such as wavelet domain features, envelope features, and transient phase analysis, has shown good performance in SEI. With the rapid development of IoT technology, the diversity and number of devices are increasing, making it increasingly difficult for traditional methods to meet the needs of extensive data processing and real-time performance.

[0005] 2) Automatic radiation source identification method based on deep neural network. With the continuous advancement of artificial intelligence technology, radiation source individual identification technology is shifting from traditional methods to deep learning methods. Researchers are actively exploring models such as convolutional neural networks, generative adversarial networks, and recurrent neural networks for the extraction of RF fingerprints and the identification of specific radiation sources, and have achieved satisfactory results. However, deep neural networks still face some challenges, mainly including long inference time, high power consumption, and low computational sparsity. Summary of the invention

[0006] The present invention proposes a deep learning radiation source individual identification method based on complex pulse neurons, the purpose of which is to solve the problems of long inference time, high power consumption and low computational sparsity in the existing deep learning-based radiation source individual identification method.

[0007] The technical solution of the present invention is as follows:

[0008] A method for identifying individual radiation sources by deep learning based on complex pulse neurons is disclosed. The identification is performed through a recognition network based on complex pulse neurons. The recognition network based on complex pulse neurons includes a plurality of convolution modules connected end to end in sequence. The convolution module includes a complex convolution layer and a pulse neural layer.

[0009] As a further improvement of the deep learning radiation source individual identification method based on complex pulse neurons, the processing method of the pulse neural layer is as follows: each channel in the input of the pulse neural layer is input into the neuron model as a membrane potential input sequence. When the cumulative potential of the membrane potential in the neuron model exceeds the threshold, the ignition function emits a spike, thereby generating a sparse sequence corresponding to the channel, and then the sparse sequences generated by all channels are merged into the features output by the pulse neural layer.

[0010] As a further improvement of the deep learning radiation source individual identification method based on complex spiking neurons, the neuron model in the spiking neuron layer is:

[0011]

[0012] Among them, τ m represents the membrane potential attenuation factor, u(t) represents the membrane potential, u rest represents the membrane reset voltage, u(t) represents the membrane potential, R m represents the membrane resistance and I(t) represents the input current.

[0013] As a further improvement of the deep learning radiation source individual identification method based on complex pulse neurons: in the pulse neural layer, the membrane potential input sequence is input into the above-mentioned neuron model as the initial value of the membrane potential, and the corresponding membrane potential u(t) is obtained after each sequence element value is input, and the membrane potential u(t) is accumulated to the cumulative potential; let the cumulative potential obtained after the membrane potential input sequence is input with n+1 points be u (n+1),pre If the value is greater than the preset trigger threshold U th , then the n+1th point in the sparse sequence u o The corresponding point u in o (n+1) The value is 1, and the accumulated potential is reset to u rest , otherwise the n+1th point is in the output sparse sequence u o The corresponding point u in o (n+1) The value is 0.

[0014] As a further improvement of the deep learning radiation source individual identification method based on complex pulse neurons: the trigger threshold is set to 0.5 and the attenuation factor is set to 0.75.

[0015] As a further improvement of the deep learning radiation source individual identification method based on complex pulse neurons: the recognition network based on complex pulse neurons is trained by using the gradient descent method combined with gradient transformation.

[0016] As a further improvement of the deep learning radiation source individual identification method based on complex pulse neural network: during training, the gradient of the loss function to the model parameters is calculated using the chain derivative rule;

[0017] At each iteration, for each convolutional module's spiking neural layer, the gradient of the weight parameters is calculated using the chain rule:

[0018]

[0019] Among them, L represents the loss function, W is the weight of the current pulse neural layer, and c (n) is the transition parameter of the chain derivative rule, T is the sampling length of the input of the spiking neural layer;

[0020] At the same time, the triangular gradient is used as the gradient of the ignition function part:

[0021]

[0022] When updating the weights, the following formula is used to calculate the new weight W of the spiking neural layer and the new weight W′ of the complex convolutional layer:

[0023]

[0024] Where W i+1 and W′ i+1 are the weights of the spike neural layer and the complex convolutional layer after the i+1th iteration, respectively, and W i and W′ i are the weights of the spike neural layer and the complex convolution layer after the i-th iteration, respectively, lr is the learning rate, is the calculated gradient.

[0025] As a further improvement of the deep learning radiation source individual identification method based on complex pulse neurons: the convolution module also includes a normalization layer and a maximum pooling layer, and the complex convolution layer, the normalization layer, the pulse neuron layer and the maximum pooling layer are connected in sequence.

[0026] As a further improvement of the deep learning radiation source individual identification method based on complex pulse neural network: the identification network based on complex pulse neural network also includes a data processing module and a classification head;

[0027] The data processing module processes the input signal data; several of the convolution modules are connected end to end in sequence, the complex signal vector output by the data processing module is the input of the first layer of convolution module, and the output of the previous layer of convolution module is the input of the next layer of convolution module; the output of the last layer of convolution module enters the classification head to obtain the classification recognition result.

[0028] As a further improvement of the deep learning radiation source individual identification method based on complex pulse neural network: the received radiation source signal data is first downsampled in the data processing module, and then normalized;

[0029] The classification head consists of an adaptive pooling layer, a linear layer, a PRelu activation layer, a Dropout layer, and a fully connected layer connected in sequence.

[0030] Compared with the prior art, the present invention has the following positive effects:

[0031] 1. The present invention adopts a complex convolutional layer to extract the I and Q channel features of the radiation source signal, encodes the features with a pulse neural layer, and maps the signal features into a sparse sequence, which greatly improves the expression ability, computational sparsity and computational efficiency of the radiation source signal, thereby improving the accuracy of individual radiation source identification.

[0032] 2. The present invention uses triangular gradient to avoid the zero gradient problem of the ignition function during training, solving the problem that the pulse neural layer cannot perform back propagation to update parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the overall architecture diagram of the identification network of the present invention;

[0034] Figure 2 It is a comparison chart of radiation source signal recognition results and model parameter quantities of different neural networks;

[0035] Figure 3 Comparison chart of inference time of different networks. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0037] A deep learning radiation source individual identification method based on complex spike neurons, the method performs identification through a recognition network based on complex spike neurons.

[0038] like Figure 1The recognition network based on complex pulse neural network includes a data processing module, a seven-layer convolution module and a classification head. The data processing module processes the input signal data; the seven-layer convolution modules are connected end to end in sequence, the complex signal vector output by the data processing module is the input of the first layer of convolution module, and the output of the previous layer of convolution module is the input of the next layer of convolution module; the output of the last layer of convolution module enters the classification head to obtain the classification recognition result.

[0039] Specifically:

[0040] 1. The processing process of the data processing module is as follows:

[0041] Step 1-1: down-sample the received radiation source signal data.

[0042] The signal data includes an I path representing the real part and a Q path representing the imaginary part. The two paths correspond to each other and the signal length (i.e., the number of sampling points) is consistent. Because the number of collected signal points is large, it is not conducive to feature extraction by the complex network, so downsampling is performed to extract data points from the original data at a certain ratio, skip some data points regularly, and only retain a part of the data, filtering out features with small effects and redundant information, and retaining key information.

[0043] Downsampling can be expressed as:

[0044] y[n]=x[n·M];

[0045] Wherein, x is the original signal data sequence, y is the signal data sequence after downsampling, n is the sampling point index of the signal data after downsampling, and M is the downsampling factor.

[0046] Step 1-2: Normalize the real and imaginary parts of the downsampled signal data to enhance learning efficiency and prevent gradient problems.

[0047] Suppose the sequence of the I path of the downsampled signal data is real(y), and the sequence of the Q path is img(y), then:

[0048]

[0049] Among them, max(real(y)) represents the maximum element value in real(y), max(img(y)) represents img(y), y I is the sequence of I-path after normalization, y Q is the sequence of Q paths after normalization. I and Q The complex signal vector that constitutes the output of the data processing module.

[0050] 2. The convolution module includes a complex convolution layer, a normalization layer, a spiking neural layer, and a maximum pooling layer connected in sequence.

[0051] The processing process of each layer of convolution module is as follows:

[0052] Step 2-1: Send the input of the convolution module to the complex convolution layer, and obtain the signal characteristics after convolution.

[0053] Suppose the input of the convolution module is s and the weight of the complex convolution kernel is W′, then the output of the complex convolution layer is W′*s, where * is the convolution operation.

[0054] Step 2-2: Send the output of the complex convolutional layer to the normalization layer in batches for complex batch normalization.

[0055] The batch normalization method can accelerate the learning of deep networks and effectively prevent the occurrence of gradient disappearance or explosion, so batch normalization is crucial for optimizing models. However, since the current batch normalization operation is only applicable to the processing of real data, the present invention will also complete the complex batch normalization layer calculation logic through equivalent real ordered pairs.

[0056] Let the input of the normalization layer be x in , then perform complex batch normalization as follows:

[0057] Step 2-2-1, calculate the expected value:

[0058]

[0059] Where l is the sampling length of the signal in the input batch data, and They are x in The mth sampling point on the I and Q paths, n bat is the batch size of the input data.

[0060] Step 2-2-3, calculate the covariance matrix:

[0061]

[0062] Among them, Cov(·) is the covariance function, x r Represents x in The real part, x i Represents x in The imaginary part of .

[0063] Step 2-2-4, calculate the intermediate value of complex batch normalization:

[0064]

[0065] Step 2-2-5, calculate the output of the normalization layer:

[0066]

[0067] Among them, the scaling factor matrix γ and the offset factor matrix β are trainable parameter matrices and parameter vectors respectively:

[0068]

[0069] Step 2-3: Send the output of the normalization layer to the spiking neural layer to obtain a sparse sequence.

[0070] The processing method of the spiking neural layer is as follows: each channel in the input of the spiking neural layer is input into the neuron model as a membrane potential input sequence. When the cumulative potential of the membrane potential in the neuron model exceeds the threshold, the ignition function emits a spike, thereby generating a sparse sequence corresponding to the channel, and then the sparse sequences generated by all channels are merged into the features output by the spiking neural layer.

[0071] The specific process is:

[0072] Step 2-3-1, build a neuron model:

[0073]

[0074] Among them, τ m represents the membrane potential attenuation factor, u(t) represents the membrane potential, u rest represents the membrane reset voltage, u(t) represents the membrane potential, R m represents the membrane resistance and I(t) represents the input current.

[0075] Step 2-3-2: Input the membrane potential input sequence as the initial value of the membrane potential into the above neuron model. After each sequence element value is input, the corresponding membrane potential u(t) is obtained, and the membrane potential u(t) is accumulated to the cumulative potential. Suppose the cumulative potential obtained after the membrane potential input sequence is input with n+1 points is u (n+1),pre If the value is greater than the preset trigger threshold U th , then the n+1th point in the sparse sequence u o The corresponding point u in o (n+1) The value is 1, and the accumulated potential is reset to u rest , otherwise the n+1th point is in the output sparse sequence u o The corresponding point u in o (n+1) The value is 0.

[0076] In this embodiment, the trigger threshold is set to 0.5, and the attenuation factor is set to 0.75. The above two parameter values ​​are obtained through ablation experiment results, and the experimental results are shown in Table 1.

[0077] Table 1 - Results of ablation experiments on key parameters of the spiking neural layer:

[0078] Value Recognition accuracy 0.25 75％ Trigger threshold 0.5 89％ 0.75 78％ 1 81％ 0 67％ 0.25 76％ Attenuation Factor 0.5 79％ 0.75 89％ 1 78％ .

[0079] Step 2-4: Send the features output by the pulse neural layer to the maximum pooling layer to obtain the processed feature map.

[0080] The pooling layer uses maximum pooling, which can extract the feature data of the specified window, significantly reduce the feature map, and also alleviate the sensitivity to the position of the features to be identified.

[0081] 3. The classification head includes an adaptive pooling layer, a linear layer, a PRelu activation layer, a Dropout layer, and a fully connected layer, which are connected in sequence to identify and classify the extracted high-dimensional features. The classification result is expressed as a probability vector [β1,β2,...,β cls ], cls represents the total number of categories to be identified, and the category corresponding to the maximum probability is the recognition result.

[0082] Through the adaptive pooling layer, the pooling size can be automatically adjusted according to the input size, reducing the number of parameters while retaining important information and improving the efficiency of the model. The introduction of the PRelu activation layer in the network has greater flexibility and can better adapt to different tasks and data sets. At the same time, in order to prevent the model from overfitting during training, the Dropout layer is added to the network. During the training process, the output of some neurons in the network is randomly set to 0 to reduce the dependency between neurons.

[0083] Before training, prepare the training set, validation set, and test set according to the following steps.

[0084] Step A: collecting signal data from different radiation sources, and processing each signal data distribution according to the data processing module to obtain a sample set.

[0085] In this embodiment, there are a total of 5 radiation source individuals, and the data of each radiation source individual has 400 samples, the total number of samples is 2000, each sample has 2 channels, I channel and Q channel, and each channel has 4800 sampling points.

[0086] Step B: label the samples in the sample set with labels in the form of one-hot encoding.

[0087] An N-bit status register is used to encode N radiation source devices, each radiation source device is represented by its own independent register bit, and only one bit is valid at any time, that is, set to 1, and the rest are 0. In this embodiment, there are 5 radiation sources, so the encoded label is a one-hot encoding of 1 row and 5 columns.

[0088] Step C: Randomly shuffle the sample data so that the model can be exposed to more data of different categories during training, which helps the model better generalize to unseen data. Shuffling the data also helps ensure that each subsequent small batch of data has a reasonable distribution of different categories, so that all categories can be effectively learned.

[0089] Step D: Divide the samples and their corresponding labels into training set, validation set and test set, with a split ratio of 8:1:1. The training set is used to train the model, that is, to determine the model parameters such as weight and bias; the validation set is used to select the model hyperparameters; the test set is used to evaluate the final model after the model training is completed.

[0090] Furthermore, the recognition network based on complex pulse neural network is trained by using the gradient descent method combined with gradient transformation. In this embodiment, the number of training rounds is set to 100, and the initial learning rate is set to 0.003. The training steps are:

[0091] Step T1: Initialize model parameters, randomly initialize model weights, biases and other parameters.

[0092] Step T2: Input the samples of the training set into the model to obtain the predicted labels, and then calculate the value of the loss function based on the predicted labels and the true labels of the samples.

[0093] In this embodiment, the cross entropy loss function is used to calculate the loss value:

[0094]

[0095] In the formula, is the predicted label and label is the actual label.

[0096] Step T3: Use the chain rule to calculate the gradient of the loss function with respect to the model parameters.

[0097] For each convolutional module's spiking neural layer, the gradient of the weight parameters is calculated using the chain rule:

[0098]

[0099] Among them, L represents the loss function, W is the weight of the current pulse neural layer, and c (n) is the transition parameter of the chain derivative rule, and T is the sampling length of the input of the spiking neural layer.

[0100] When calculating, since the output of the ignition function is discrete and non-differentiable, its derivative is zero almost everywhere, which makes the gradient-based optimization method unusable. To solve this problem, the triangular gradient is used as the gradient of the ignition function part:

[0101]

[0102] Step T4: Update the model parameters according to the gradient to reduce the loss function value.

[0103] The new weight W of the spiking neural layer and the new weight W′ of the complex convolutional layer are calculated using the following formula:

[0104]

[0105] Where W i+1 and W′ i+1 are the weights of the spike neural layer and the complex convolutional layer after the i+1th iteration, respectively, and W u and W′ i are the weights of the spike neural layer and the complex convolution layer after the i-th iteration, respectively, lr is the learning rate, The gradient calculated for step T3.

[0106] Repeat steps T2 to T4 until the preset number of training rounds is reached.

[0107] After training, save the network model with the best effect. Each time, select 64 data as a batch to input into the network for training. Each iteration only calculates the gradient of a small batch to reduce the amount of calculation. Since each iteration uses a different small batch of data, the model can converge to the optimal solution faster. Use the validation set to verify the network after each round of training. At the same time, set the network early stopping mechanism to avoid overfitting during network training. Finally, save the network model with the best effect.

[0108] Then use the test set data to test the model. Select the network model with the best training effect, input the test set data directly into the model, and compare the network output results with the one-hot encoding labels corresponding to the test set data after passing through the softmax layer. Calculate the model recognition accuracy based on the comparison results:

[0109]

[0110] In the formula, Acc is the recognition accuracy, Num Acc To identify the correct number of samples, Num is the number of samples in the test set.

[0111] The signal data was collected on-site using the SM200B software-defined radio platform produced by SignalHound. A high-quality radio signal data set was obtained by downsampling, header search, information decoding, CRC verification, ADS-B raw signal interception, and marking the ADS-B signal with the corresponding aircraft unique identifier. Traditional deep learning models require a large number of floating-point multiplication operations in electromagnetic signal recognition, high hardware requirements for computing resources and storage space, and long model inference time. The method based on complex neural networks can make full use of the information carried by complex data. At the same time, due to the binary transmission mechanism of the pulse neuron layer, it can be implemented using only adders and comparators, significantly reducing the computing cost and storage resources required for hardware implementation. Compared with traditional lightweight networks, this method achieves the best results in both recognition rate and inference speed.

[0112] Figure 2 It is a comparison chart of the radiation source signal recognition results and model parameter quantities of different neural networks. Considering only the recognition accuracy, ComplexNet (complex neural network) performs best. The recognition rate of SComplex (complex pulse neural network, i.e., the method of the present invention) reaches 89%, which is slightly lower than ComplexNet (complex neural network) and ResNet (residual neural network), but better than two mainstream lightweight networks: MobileNet (mobile network) and ShuffleNet (shuffle network). It can be seen that this method uses complex data. Compared with real data, complex data carries more information and therefore has stronger expression ability. On the other hand, the parameters of the remaining methods are more than SComplex (complex pulse neural network), especially ResNet (residual neural network). Although its recognition rate is slightly better than SComplex (complex pulse neural network) of this method, its parameter quantity is much higher than that of this method.

[0113] Figure 3 This is a comparison chart of the inference time of different networks. It can be seen that SComplex (complex spike neural network) has the shortest inference time.

[0114] Combination Figure 2 and Figure 3 It can be seen that the SComplex (complex spike neural network) proposed in the present invention has the shortest reasoning time under the premise of a higher recognition rate.

[0115] In this embodiment, the program is mainly implemented in python, and the main libraries used include torch (deep learning library), numpy (scientific mathematical operation library) and matplotlib (mainly used for drawing).

[0116] It should be noted that it is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. The scope of the present invention is defined by the claims rather than the above description.

Claims

1. A method for identifying individual radiation sources based on deep learning of complex pulse neurons, characterized in that: Recognition is performed by a recognition network based on complex spiking neurons, wherein the recognition network based on complex spiking neurons includes a plurality of convolution modules connected end to end in sequence, wherein the convolution module includes a complex convolution layer and a spiking neural layer; The processing method of the spiking neural layer is as follows: each channel in the input of the spiking neural layer is input into the neuron model as a membrane potential input sequence. When the accumulated potential of the membrane potential in the neuron model exceeds the threshold, the ignition function emits a spike, thereby generating a sparse sequence corresponding to the channel. The sparse sequences generated by all channels are then merged into the features output by the spiking neural layer. The neuron model in the spiking neural layer is: Among them, τ m represents the membrane potential attenuation factor, u(t) represents the membrane potential, u rest represents the membrane reset voltage, u(t) represents the membrane potential, R m represents the membrane resistance, I(t) represents the input current; In the pulse neural layer, the membrane potential input sequence is input into the above neuron model as the initial value of the membrane potential. After each sequence element value is input, the corresponding membrane potential u(t) is obtained, and the membrane potential u(t) is accumulated to the cumulative potential. Let the cumulative potential obtained after the membrane potential input sequence is input to n+1 points be u (n+1),pre If the value is greater than the preset trigger threshold U t h, then the n+1th point is in the sparse sequence u o The corresponding point u in o (n+1) The value is 1, and the accumulated potential is reset to u rest , otherwise the n+1th point is in the output sparse sequence u o The corresponding point u in o (n+1) The value is 0; The recognition network based on complex pulse neural network is trained by using the gradient descent method combined with gradient transformation; During training, the chain rule is used to calculate the gradient of the loss function with respect to the model parameters; At each iteration, for each convolutional module's spiking neural layer, the gradient of the weight parameters is calculated using the chain rule: Among them, L represents the loss function, W is the weight of the current pulse neural layer, and c (n) is the transition parameter of the chain derivative rule, T is the sampling length of the input of the spiking neural layer; At the same time, the triangular gradient is used as the gradient of the ignition function part: When updating the weights, the following formula is used to calculate the new weight W of the spiking neural layer and the new weight W′ of the complex convolutional layer: Where W i+1 and W′ i+1 are the weights of the spike neural layer and the complex convolutional layer after the i+1th iteration, respectively, and W i and W′ i are the weights of the spike neural layer and the complex convolution layer after the i-th iteration, respectively, lr is the learning rate, is the calculated gradient.

2. The method for individual radiation source identification based on deep learning of complex pulse neural networks according to claim 1, characterized in that: The trigger threshold is set to 0.5 and the decay factor is set to 0.

75.

3. The method for individual radiation source identification based on deep learning of complex pulse neural networks according to claim 1, characterized in that: The convolution module also includes a normalization layer and a maximum pooling layer, and the complex convolution layer, the normalization layer, the pulse neural layer and the maximum pooling layer are connected in sequence.

4. The method for individual radiation source identification based on deep learning of complex pulse neural network according to any one of claims 1 to 3, characterized in that: The complex spike neural based recognition network also includes a data processing module and a classification head; The data processing module processes the input signal data; several of the convolution modules are connected end to end in sequence, the complex signal vector output by the data processing module is the input of the first layer of convolution module, and the output of the previous layer of convolution module is the input of the next layer of convolution module; the output of the last layer of convolution module enters the classification head to obtain the classification recognition result.

5. The method for individual radiation source identification based on deep learning of complex pulse neural networks according to claim 4, characterized in that: In the data processing module, the received radiation source signal data is first downsampled and then normalized; The classification head consists of an adaptive pooling layer, a linear layer, a PRelu activation layer, a Dropout layer, and a fully connected layer connected in sequence.

Images