A friend-or-foe identification signal recognition method based on deep learning

Abstract

A deep learning-based IFF signal recognition method relates to the field of signal recognition. This invention addresses the difficulties of manually designing existing IFF signals and the high false alarm or missed alarm rates of existing IFF signal recognition methods, which make it difficult to meet IFF signal recognition requirements. The invention comprises the following steps: obtaining an IFF signal to be identified, and inputting the IFF signal to be identified into a trained IFF signal recognition model to obtain the classification of the radar IFF signal and the relative positional relationship of the radar IFF signal's starting position. The trained IFF signal recognition model is obtained by the following steps: obtaining an IFF signal training set and a test set; training the IFF signal recognition model using the training set to obtain a trained IFF signal recognition model; and testing the trained IFF signal recognition model using the test set to obtain a test IFF signal recognition model. This invention is used to identify IFF signals.

Description

Technical Field

[0001] The present invention relates to the field of signal recognition, and in particular to a friend-or-foe identification signal recognition method based on deep learning. Background Art

[0002] Identification Friend or Foe (IFF) technology is widely used in reconnaissance and defense systems for various long-range combat units, such as aircraft, warships, and air defense systems. Traditional IFF signal pattern recognition methods rely heavily on domain knowledge and artificial empirical features to extract specific signal characteristics. This makes signal feature extraction highly dependent on manual feature extraction and abundant prior information. Furthermore, the increasing number of radio signals and their rapid variability complicate the design of prior information. Therefore, research on IFF signal recognition algorithms in complex environments is crucial for ensuring the efficient and reliable interception and recognition of IFF signals in these environments.

[0003] Currently, the recognition of friend-or-foe (IFF) signals primarily relies on pulse frame correlation matching. This method further decodes a certain pattern signal after detecting it. However, this traditional sliding window method for IFF signals relies heavily on manual feature extraction and abundant prior information, as well as the increasing number of radios and the rapid variability of radio signals. For example, due to the relative position differences between different IFF devices, the discrimination threshold needs to be set based on the received signal's signal-to-noise ratio and signal density, making it difficult to manually design efficient feature representations. This makes the manual design of IFF signals difficult. Furthermore, this traditional method only works well in ideal or high-SNR environments. In complex electromagnetic environments, when the matching window is strictly designed, the received signal has a high false alarm rate. Conversely, many false pattern signals appear, resulting in an increased false alarm rate or false alarm rate, making it difficult to meet the recognition requirements for the IFF signal's operating mode. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems that the existing IFF signal is difficult to be manually designed and the false alarm rate or missed alarm rate of the existing IFF signal recognition method is too high, which makes it difficult to meet the IFF signal recognition needs. A deep learning-based IFF signal recognition method is proposed.

[0005] A deep learning-based friend-or-foe identification signal recognition method has the following specific process:

[0006] Obtaining an IFF signal to be identified, and inputting the IFF signal to be identified into a trained IFF signal recognition model to obtain the classification of the radar friend-or-foe identification signal and the relative position relationship of the starting position of the radar friend-or-foe identification signal;

[0007] The trained IFF signal recognition model is obtained by the following steps:

[0008] Step 1: Obtain the IFF signal training set and test set:

[0009] Step 1. Obtain the IFF signal dataset using the IFF signal model and radar receiver channel model:

[0010] First, the IFF signal model is used to randomly generate IFF signals of different modes in a blank signal segment, and the starting position of each mode signal in the blank signal segment is recorded;

[0011] The IFF signal includes: an inquiry IFF signal and a response IFF signal;

[0012] The interrogation IFF signal is divided into an interrogation IFF short signal and an interrogation IFF long signal by a first signal threshold;

[0013] The response IFF signal is divided into a response IFF short signal and a response IFF long signal according to a second signal threshold;

[0014] Then, the radar receiver channel model is used to simulate signal interference to obtain the interfered signal segments composed of signals with different interference modes;

[0015] The simulated signal interference includes: intermediate frequency jitter, sampling clock jitter, path loss, additive alpha stable distribution noise, random frame loss, and TACAN navigation system signal;

[0016] Then, several random position samplings are performed on the signal segment with interference, and each sampling result is saved as a data set sample;

[0017] Finally, a sample label is defined for each dataset sample and assigned a value to obtain a dataset sample with a label. All labeled dataset samples constitute the IFF signal dataset.

[0018] Step 12: Divide the IFF signal data set obtained in step 11 into a training set and a test set to obtain an IFF signal training set and a test set;

[0019] Step 2: Using the training set to train the IFF signal recognition model to obtain a trained IFF signal recognition model;

[0020] The IFF signal recognition model includes: a convolutional neural network layer, a capsule network layer, a channel attention mechanism layer, and a fully connected layer;

[0021] The convolutional neural network layer is used to convert the input IFF signal into a scalar feature map, including: an input layer, a first convolution layer, a second convolution layer, a maximum pooling layer and a BN layer, a third convolution layer, a fourth convolution layer, and a splicing enhancement processing layer;

[0022] The capsule network layer includes: an initial capsule network unit and a digital capsule unit;

[0023] The initial capsule network unit is used to convert the scalar feature map output by the convolutional neural network layer into a vector feature map;

[0024] The digital capsule unit is used to compress the vector feature map into a two-dimensional feature vector;

[0025] The attention mechanism is used to compress the two-dimensional feature vector again to obtain a one-dimensional feature vector;

[0026] The number of layers of the fully connected layer is 3, which is used to output the classification of the signal and the relative position relationship of the signal starting position;

[0027] Step 3: Use the test set to test the trained IFF signal recognition model. If the accuracy is greater than the preset accuracy threshold, save the model. If the accuracy is lower than the preset accuracy threshold, adjust the model's hyperparameters and re-acquire the training set for training until the recognition accuracy is greater than the preset accuracy threshold.

[0028] The beneficial effects of the present invention are:

[0029] The present invention addresses the shortcomings of existing technologies by first extracting key features using a convolutional neural network. It then employs a capsule network and an attention mechanism to extract feature representations. IFF signals in both query and response modes are divided in detail using effective signal length ratios. A rapid long-short signal recognition model based on the spatial attention mechanism and capsule network is then designed. The present invention assigns long and short signal recognition and rapid recognition functions to the three output structures of the network, respectively. This introduces a neural network into IFF signal detection, alleviating the burden of manual design and selection. It also improves performance in terms of false alarm and false alarm rates, and expresses the final recognition focus (e.g., low false alarm or low false alarm rate) as a single threshold. The present invention proposes a method for constructing IFF signal dataset samples. Based on a superheterodyne receiver, the method combines an idealized signal model with a channel model for a complex, real-world electromagnetic environment to generate an IFF signal dataset comparable to a real-world environment. This dataset eliminates the complex workload and sample imbalance associated with manual labeling, providing guidance for improving the recognition of continuously collected IFF signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of the method of the present invention;

[0031] Figure 2 A schematic diagram of capsule dynamic routing calculation according to the present invention;

[0032] Figure 3 This is a schematic diagram of a simulation signal involved in the present invention;

[0033] Figure 4 Schematic diagram of the channel attention mechanism involved in the present invention;

[0034] Figure 5 is the confusion matrix involved in the present invention;

[0035] Figure 6 The recognition rate of each category of the data set involved in the present invention varies with the signal-to-noise ratio;

[0036] Figure 7 The present invention relates to a software interface. DETAILED DESCRIPTION

[0037] Specific implementation method 1: This implementation method is a friend-or-foe identification signal recognition method based on deep learning. The specific process is as follows: Figure 1 Specifically, the IFF signal to be identified (IFF signal) is obtained, and the IFF signal to be identified is input into a trained IFF signal recognition model to obtain the classification of the radar IFF signal and the relative position relationship of the starting position of the radar IFF signal.

[0038] The trained IFF signal recognition model is obtained by the following steps:

[0039] Step 1: Obtain the IFF signal training set and test set:

[0040] Step 11: Obtain an IFF signal dataset using the IFF signal model and radar receiver channel model, including the following steps:

[0041] First, the IFF signal model is used to randomly generate IFF signals of different modes in a blank signal segment, and the starting position of each mode signal in the blank signal segment is recorded;

[0042] The IFF signal includes: an inquiry IFF signal and a response IFF signal;

[0043] Then, the radar receiver channel model is used to simulate signal interference to obtain the interfered signal segments composed of signals with different interference modes;

[0044] The radar receiver channel model is a superheterodyne receiver;

[0045] The analog signal interference adopts multiple regularization methods including: intermediate frequency jitter, sampling clock jitter, path loss, additive alpha stable distribution noise, random frame loss, and TACAN navigation system signal;

[0046] Then, several random position samplings are performed on the signal segment with interference, and each sampling result is saved as a data set sample;

[0047] Finally, a sample label is defined for each dataset sample and assigned a value to obtain a dataset sample with a label. All labeled dataset samples constitute the IFF signal dataset.

[0048] The sample labels include: long signal, short signal, and relative position of the signal;

[0049] Relative position = λ_loc·starting position / dataset sample length;

[0050] Among them, λ_loc is a coefficient factor introduced to prevent the values ​​of all relative position relationships from being too small, and can be set to 2-10;

[0051] The sample label is one-hot encoded;

[0052] Define sample labels for each dataset sample and assign values ​​to the labels as follows:

[0053] First, the signals in each dataset sample are divided into long signals and short signals according to the preset signal length threshold, and the signal category labels in the dataset samples are defined according to the length of the signals. Then, the relative positions of the signals are used as position labels.

[0054] The signal length threshold is set based on experience. The signal length threshold of the inquiry signal is set to 10us, and the signal length threshold of the response signal is set to 15us.

[0055] Then, the category labels are one-hot encoded and assigned to each signal category label according to the type of category label:

[0056] If the category label type is the long signal category, then determine whether all valid signals of a single mode are in the current dataset samples. If all are in the current dataset samples, label smoothing is used to assign a value to each signal category label. If part of the valid signal is in the current dataset samples, the category label is assigned according to the ratio of the valid signal length in the dataset samples to the total length of the valid signal (for example, the total length of the valid signal is [1,1,1,1,1], and the sample randomly extracted from the signal segment is [1,1,0,0,0,0,0,0,0]. At this time, there are only two valid signal points in the sample, so the label assignment for this sample is 2 (1) / 5 (1)=0.4, which is the number of valid signal points in the sample / total length of the valid signal). If all valid signals are not in the current dataset samples, the current dataset samples are considered to be noise, and the category label is assigned to 0.

[0057] If the category label type is the short signal category, then determine whether all valid signals of a single mode are within the top 45% of the current dataset samples. If so, label smoothing is used to assign a value to each signal category label. If not all valid signals are within the top 45%, the current dataset samples are noise, and the category label is assigned a value of 0.

[0058] The effective signal is the friend-or-foe identification signal excluding the information data pulse block;

[0059] The label smoothing is used to assign each signal category label as follows:

[0060]

[0061] Among them, K_ls is the total number of multi-classification categories, λ_ls is 0.1, i_ls is the valid signal of a single mode, target is the current target;

[0062] Finally, assign the position label:

[0063] For the position label, its value is a floating point number. When the current dataset sample contains a complete category signal, the ratio of the starting position of the first signal to the position corresponding to 10us is assigned to the position label; when the sample contains partial pattern category signals or does not contain pattern category signals, the position label value is 1;

[0064] Step 12: Divide the IFF signal data set obtained in step 11 into a training set and a test set to obtain an IFF signal training set and a test set;

[0065] The training set:test set=3:1;

[0066] Step 2: Using the training set to train the IFF signal recognition model to obtain a trained IFF signal recognition model;

[0067] The IFF signal recognition model is trained using the Adam optimizer; the Adam optimizer has a learning rate of 0.05, a learning rate decay of 0.9, and an early stopping mechanism of 3 times;

[0068] The IFF signal recognition model includes: a convolutional neural network layer, a capsule network layer, a channel attention mechanism layer, and a fully connected layer;

[0069] The convolutional neural network layer is used to convert the input IFF signal into a scalar feature map, including: an input layer, a first convolution layer, a second convolution layer, a maximum pooling layer and a BN layer, a third convolution layer, a fourth convolution layer, and a splicing enhancement processing layer;

[0070] The capsule network layer includes: an initial capsule network unit and a digital capsule unit;

[0071] The initial capsule network unit is used to convert the scalar feature map output by the convolutional neural network layer into a vector feature map;

[0072] The initial capsule layer dimension is 8, the number of channels is 32, the convolution kernel size is 9, the stride is 2, and the padding is the same;

[0073] The digital capsule unit is used to compress the vector feature map into a two-dimensional feature vector;

[0074] The attention mechanism is used to compress the two-dimensional feature vector again to obtain a one-dimensional feature vector;

[0075] The number of layers of the fully connected layer is 3, which is used to output the classification of the signal and the relative position relationship of the signal starting position.

[0076] Step 3: Use the test set to test the trained IFF signal recognition model. If the accuracy is greater than the preset accuracy threshold, save the model. If the accuracy is lower than the preset accuracy threshold, adjust the model's hyperparameters and re-acquire the training set for training until the recognition accuracy is greater than the preset accuracy threshold.

[0077] In this embodiment, before the network model is trained, in order to prevent the model from overfitting and gradient vanishing, a variety of regularization methods are used when generating the data set, such as negative samples, interference noise, and label smoothing. In order to facilitate debugging, the Adam optimizer with faster convergence speed is adopted, and then optimization measures such as learning rate decay and early stopping are set. At the same time, a continuous learning strategy is adopted during training, and the regularization method in the data set is dynamically adjusted according to the number of training iterations. That is, as the number of iterations increases, the range of random generation of the corresponding regularization parameters is also constantly increasing. The ultimate goal is to ensure that the model is more adaptable while ensuring accuracy; the final output confusion matrix and the changes in each type of recognition accuracy with the signal-to-noise ratio can be seen Figure 5 and Figure 6 .

[0078] Specific implementation method two; the IFF interrogation signal works under the interrogation machine, including: M1, M2, M3 / A, MB, MC, MD, M4, MS, M5;

[0079] Inquiry short signals include: M1, M2, M3, M4, MS;

[0080] The answer IFF signal works under the transponder, including: MarkX, M4, MS, M5_L1, M5_L2;

[0081] Among them, the short response signals include: M4, MS, and M5_L1.

[0082] Since the interrogator and transponder operate in different frequency bands and the duration of different mode signals also varies greatly, the longest valid identification data segment of all mode signals is used to determine that the number of input points of the network model for inquiry and response under different working modes is the product of 50us and 35us and the sampling rate respectively.

[0083] Specific implementation method three: using the radar receiver channel model to simulate signal interference, obtaining interference signal segments composed of signals with different interference modes, as follows:

[0084] S1. At a set IF of 70 MHz and a sampling rate of 90 MHz, add ±2.5% IF random jitter and ±5% sampling frequency jitter to randomly generate a series of ideal IF friend-or-foe identification signals. Then simulate the additive Gaussian thermal noise inside the receiver to obtain the IF IFF signal that has passed through the Gaussian thermal noise.

[0085] When generating the M5 signal, because this mode uses MSK modulation, that is, IQ quadrature modulation has two outputs, the superheterodyne receiver model can be used to simulate the integration of the received signals into one. After the MSK signal modulator inputs the binary data sequence, it undergoes differential encoding and serial / parallel conversion, and the I branch signal is converted by cos(πt / 2Ts ) weighted modulation and co-phase carrier cosw c t multiplication output in-phase classification x I (t), the Q branch signal is delayed by T s , by sin(πt / 2T s ) weighted modulation and quadrature carrier sinw c t multiply output quadrature component x Q (t).x I (t) and x Q The modulated MSK signal can be obtained by subtracting (t).

[0086] The simulated Gaussian thermal noise is between 10 and 20 dB;

[0087] S2. Using a radar equation formula, randomly multiply the intermediate frequency IFF signal that has passed through the thermal Gaussian noise by an amplitude, to obtain an intermediate frequency IFF signal after the random amplitude multiplication.

[0088] The amplitude random range is 0.6-1;

[0089] S3. Since the IFF system operates in the 1030±4MHz and 1090±4MHz frequency bands, all signals existing in this band, namely the L-band, are considered, as shown in Table 1. In addition to the signals related to the IFF system, the TACAN navigation system signal is regarded as an interference signal. Therefore, the TACAN navigation system signal is simulated:

[0090]

[0091] Get the radio frequency pulses transmitted by TACAN navigation system beacons and airborne equipment:

[0092] The TACAN system beacon and airborne equipment both transmit RF pulse pairs of the same shape as the basic pulse waveform. In order to compress the signal spectrum width and reduce adjacent channel interference, the pulse is selected as a Gaussian bell signal. The Gaussian modulation pulse function expression is:

[0093]

[0094] Among them, τ is the half-amplitude value point width of the pulse, which is specified as 3.5us. It is also specified that the rising edge time of the pulse is 2±0.25us, the falling edge time should be 2.5±0.5us, b=0.5, t is time, and A is the Gaussian modulation pulse amplitude, which is randomly selected between 1-1.5;

[0095] S4. Using the additive alpha stable distribution noise model to simulate the environmental clutter in the receiving channel;

[0096] The characteristic function of the input additive alpha stable distribution noise model is:

[0097]

[0098]

[0099] Where i is the imaginary unit, α≤2 is the characteristic exponent, -1≤β≤1 is the distribution slope parameter, σ is the dispersion coefficient, μ is the location parameter, α, β, σ, μ are input parameters, sign(t) is the intermediate variable, and t is any random variable with this distribution;

[0100] When α = 2 and β = 0, it is a Gaussian distribution, when α = 1, it is a Cauchy distribution, and when α = 1 / 2, it is a Pearson distribution;

[0101] The output noise is generated in the standard parameter system as follows:

[0102] When α≠1,

[0103]

[0104] When α=1,

[0105]

[0106]

[0107]

[0108] Among them, V obeys the uniform distribution of (-π / 2,π / 2), W obeys the exponential distribution with mean 1, and N a,β and M a,β is an intermediate variable;

[0109] S5. Use S2 and S3 to obtain an interference signal segment composed of signals with different interference modes:

[0110] Create a blank signal of about 1ms in length, randomly fill the blank segment with the signal processed in step ②, record the position of each segment, and then randomly fill the blank signal with the TACAN system signal simulated in step ③ after passing 10-20dB additive Gaussian noise. Then pass the signal segment through additive Alpha distribution noise. At this time, the signal-to-noise ratio can be 0-15dB. Finally, use low-pass filtering and maximum normalization to extract samples from random positions in the signal segment several times. Taking the response mode as an example, the sample length is 90Mhz×35us=3150 and is placed in the data set. The sample diagram generated in this step is shown as follows: Figure 3 shown.

[0111] Specific embodiment 4: The convolutional neural network includes in sequence: an input layer, a first convolutional layer, a second convolutional layer, a maximum pooling layer and a BN layer, a third convolutional layer, a fourth convolutional layer, and a splicing enhancement processing layer;

[0112] The input layer dimension is (4500,1);

[0113] The number of convolution kernels in the first convolution layer is 64, the convolution kernel size is 11, the stride is 1, the dilation rate is 4, the activation function is selu, and the output dimension is (4500, 64);

[0114] The second convolutional layer has 64 convolution kernels, a kernel size of 11, a stride of 4, a dilation rate of 1, an activation function of selu, and an output dimension of (1125, 64).

[0115] The kernel size and stride of the maximum pooling layer and BN layer are both 2, and the output dimension is (562, 64);

[0116] The third convolutional layer has 128 convolution kernels, a kernel size of 5, a stride of 1, a dilation rate of 4, an activation function of selu, and an output dimension of (562, 128).

[0117] The fourth convolutional layer has 128 convolution kernels, a kernel size of 5, a stride of 4, a dilation rate of 1, an activation function of selu, and an output dimension of (141, 128).

[0118] The splicing enhancement processing layer is used to copy and splice the first 1 / 4 of the feature map output by the fourth convolutional layer to the feature map output by the first convolutional layer, and the output dimension is (204, 128).

[0119] Specific embodiment 5: The initial capsule unit is used to convert the scalar feature map output by the convolutional neural network layer into a vector feature map, including:

[0120] The original feature map dimension is expanded from (204, 128) to (102, 256) through a one-dimensional convolution operation, and then converted into a vector of dimension 8 by row through a Reshape operation, with an output dimension of (3264, 8).

[0121] Specific implementation method six: The digital capsule unit is used to compress the vector feature map into a two-dimensional feature vector using a dynamic routing iterative algorithm such as Figure 2 , specifically:

[0122] The output vector of the initial capsule unit is used to obtain the input vector of the digital capsule layer, and the input vector of the digital capsule unit is input into the "Squashing" nonlinear function to obtain the output vector of the digital capsule unit;

[0123] The "Squashing" nonlinear function is:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129] Where i_pri is the initial capsule unit, j_dig is the digital capsule unit, and v j_dig is the output vector of the digital capsule unit, s j_dig is the input vector of the digital capsule unit and the weighted sum of the vectors output from i_pri to j_dig, s j_dig is the total predicted vector in the initial capsule unit The weighted sum of i_pri,j_dig is the coupling coefficient between i_pri and all j_dig, and its value is the initial coupling coefficient b between capsule i_pri and j_dig i_pri,j_dig The logarithmic prior probability, c i_pri,j_dig is the coupling coefficient determined by the iterative dynamic routing consensus algorithm, W i_pri,j_dig is the weight matrix, u j_dig is the output of the initial capsule unit, k_dig is the dimension of the output capsule of the digital capsule unit, b i_pri,k_dig It is b i_pri,j_dig The k_digth initial coupling coefficient in , where k_dig is an integer.

[0130] Specific implementation method seven: The attention mechanism is used to compress the two-dimensional feature vector again to obtain a one-dimensional feature vector, such as Figure 4 , specifically:

[0131] I. Input the output of the digital capsule unit into the g(x) function to obtain the intermediate weight of Attention:

[0132] e at =g(x)=xW at

[0133] Among them, x refers to the output vector of the digital capsule unit, x∈R 10*128 , W at ∈R 128*1 is the last element in the output of the fully connected layer;

[0134] II. Probability vector using the intermediate weights of Attention:

[0135]

[0136] Among them, a at ∈R 10*1 is the intermediate weight e at The probability vector of at,j_at is a at The j_atth element in e k_at It is e at The k_atth element in e j_at It is e at The j_atth element of .

[0137] III. Input the output of the digital capsule unit into the spatial attention mechanism through the probability vector of the intermediate weights to obtain the weighted features captured by the intermediate attention mechanism:

[0138]

[0139] Among them, v j_dig|j_at It is v j_dig The j_atth element of .

[0140] IV. Finally, the last element W in I at With the output vector (c at ∈R 128*1 ) is concatenated into a 256*1 vector, and then outputted through a fully connected layer of size 128 to obtain a feature vector of dimension 128*1.

[0141] Specific implementation method eight: The fully connected layer consists of three layers, and the output feature vectors of the attention mechanism layer are divided into three tasks, namely:

[0142] The first task obtains the output of the long signal classification task through three layers of fully connected layers with sizes of 256, 128, and 10. The second task obtains the output of the short signal classification task through two layers of fully connected layers with sizes of 64 and 6. The third task first flattens the feature map output by the deep convolutional network, splices it with the output of the attention mechanism layer, and then obtains the output of the position prediction through two layers of fully connected layers with sizes of 64 and 1 respectively.

[0143] Example: The GUI interface designed according to the method described in the specific embodiment is as follows Figure 7 As shown, specifically:

[0144] First, select the recognition mode as query or response, then import the trained recognition model, and then set the recognition parameters, such as the number of interval points for point detection, data storage type, sampling rate, etc. Finally, record and display the timestamp of each pattern category signal in a single sampling file.

Claims

1. A friend-or-foe identification signal recognition method based on deep learning, characterized in that The method specifically comprises the following steps: obtaining an IFF signal to be identified, and inputting the IFF signal to be identified into a trained IFF signal recognition model to obtain the classification of the radar friend-or-foe identification signal and the relative position relationship of the starting position of the radar friend-or-foe identification signal; The trained IFF signal recognition model is obtained by the following steps: Step 1: Obtain the IFF signal training set and test set: Step 1. Obtain the IFF signal dataset using the IFF signal model and radar receiver channel model: First, the IFF signal model is used to randomly generate IFF signals of different modes in a blank signal segment, and the starting position of each mode signal in the blank signal segment is recorded; The IFF signal includes: an inquiry IFF signal and a response IFF signal; The interrogation IFF signal is divided into an interrogation IFF short signal and an interrogation IFF long signal by a first signal threshold; The response IFF signal is divided into a response IFF short signal and a response IFF long signal according to a second signal threshold; Then, the radar receiver channel model is used to simulate signal interference to obtain the interfered signal segments composed of signals with different interference modes; The simulated signal interference includes: intermediate frequency jitter, sampling clock jitter, path loss, additive alpha stable distribution noise, random frame loss, and TACAN navigation system signal; Then, several random position samplings are performed on the signal segment with interference, and each sampling result is saved as a data set sample; Finally, a sample label is defined for each dataset sample and assigned a value to obtain a dataset sample with a label. All labeled dataset samples constitute the IFF signal dataset. Step 12: Divide the IFF signal data set obtained in step 11 into a training set and a test set to obtain an IFF signal training set and a test set; Step 2: Using the training set to train the IFF signal recognition model to obtain a trained IFF signal recognition model; The IFF signal recognition model includes: a convolutional neural network layer, a capsule network layer, a channel attention mechanism layer, and a fully connected layer; The convolutional neural network layer is used to convert the input IFF signal into a scalar feature map; The capsule network layer includes: an initial capsule network unit and a digital capsule unit; The initial capsule network unit is used to convert the scalar feature map output by the convolutional neural network layer into a vector feature map; The digital capsule unit is used to compress the vector feature map into a two-dimensional feature vector; The attention mechanism is used to compress the two-dimensional feature vector again to obtain a one-dimensional feature vector; The number of layers of the fully connected layer is 3, which is used to output the classification of the signal and the relative position relationship of the signal starting position; Step 3: Use the test set to test the trained IFF signal recognition model. If the accuracy is greater than the preset accuracy threshold, save the model. If the accuracy is lower than the preset accuracy threshold, adjust the model's hyperparameters and re-acquire the training set for training until the recognition accuracy is greater than the preset accuracy threshold.

2. The method for identifying friend-or-foe (IFF) signals based on deep learning according to claim 1, characterized in that: The first signal threshold is 10 us; the second signal threshold is 15 us.

3. The method for identifying friend-or-foe (IFF) signals based on deep learning according to claim 2, wherein: The radar receiver channel model is used to simulate signal interference to obtain the interfered signal segments composed of signals of different interfered modes: ① First, at a set intermediate frequency of 70MHz and a sampling rate of 90MHz, a series of intermediate frequency identification friend-or-foe (IFFF) signals are randomly generated by adding ±2.5% IF random jitter and ±5% sampling frequency jitter. This is then simulated by passing additive Gaussian thermal noise through the receiver to obtain an IF IFF signal that has passed through the Gaussian thermal noise. Then, using the radar equation, the IF IFF signal that has passed through the thermal Gaussian noise is randomly amplitude-multiplied to obtain the IF IFF signal after random amplitude multiplication. The simulated Gaussian thermal noise is between 10 and 20 dB; the amplitude is random in the range of 0.6-1; ②Simulate the signal transmitted by the TACAN system as follows: Where τ is the pulse half-amplitude value point width, b = 0.5, t is the time independent variable, and A is the Gaussian modulation pulse amplitude; ③Use the additive alpha stable distribution noise model to simulate the environmental clutter in the receiving channel, as shown in the following formula: When α≠1, the environmental clutter is: When α=1, the environmental clutter is: Among them, α≤2 is the characteristic index, -1≤β≤1 is the distribution slope parameter, σ is the dispersion coefficient, μ is the location parameter, V obeys the uniform distribution of (-π / 2,π / 2), W obeys the exponential distribution with mean 1, N α,β and M α,β is an intermediate variable; When α = 2, it is Gaussian distribution, when α = 1, it is Cauchy distribution, and when α = 1 / 2, it is Pearson distribution; ④. Use ② and ③ to obtain the interference signal segment composed of signals with different interference modes: First, create a blank signal and randomly fill the blank segment with the signal processed in step ②, and record the position of each signal in the segment; Then, the TACAN system signal simulated in step ③ is randomly filled in the blank signal after being subjected to 10-20dB additive Gaussian noise; Then, the signal segment obtained after filling is subjected to additive alpha distribution noise to obtain an interfered signal segment composed of signals with different interference modes.

4. The method for identifying friend-or-foe (IFF) signals based on deep learning according to claim 3, characterized in that: The steps of defining sample labels for each dataset sample and assigning values ​​to the labels are as follows: First, the signals in each dataset sample are divided into long signals and short signals according to the preset signal length threshold, and the signal category label in the dataset sample is defined according to the length of the signal. Then, the relative position of the signal is used as the position label; Among them, relative position = λ_loc·starting position / dataset sample length; λ_loc∈[2,10] is the coefficient factor; Then, the category labels are one-hot encoded and assigned to each signal category label according to the type of category label: If the category label type is the long signal category, then determine whether all valid signals of a single mode are in the current dataset samples. If all are in the current dataset samples, label smoothing is used to assign a value to each signal category label; if part of the valid signal is in the current dataset samples, the ratio of the valid signal length in the dataset samples to the total valid signal length is assigned to the category label; if all valid signals are not in the current dataset samples, then the current dataset samples are considered to be noise, and the category label is assigned to 0; If the category label type is the short signal category, then determine whether all valid signals of a single mode are within the top 45% of the current dataset samples. If so, label smoothing is used to assign a value to each signal category label. If not all valid signals are within the top 45%, the current dataset samples are noise, and the category label is assigned a value of 0. The effective signal is the friend-or-foe identification signal excluding the information data pulse block; The label smoothing is used to assign each signal category label as follows: Among them, K_ls is the total number of multi-classification categories, λ_ls is 0.1, i_ls is the valid signal of a single mode, target is the current target; Finally, assign the position to the position label: If the current dataset sample contains a complete valid signal, the ratio of the starting position of the first signal to the position corresponding to 10us is assigned to the position label; if the current dataset sample contains a partial valid signal or no valid signal, the position label value is 1.

5. The method for identifying friend-or-foe (IFF) signals based on deep learning according to claim 4, characterized in that: The training IFF signal recognition model in step 2 adopts Adam optimizer; The Adam optimizer has a learning rate of 0.05, a learning rate decay of 0.9, and an early stopping mechanism of 3 times.

6. The method for identifying friend-or-foe (IFF) signals based on deep learning according to claim 5, characterized in that: The convolutional neural network includes: an input layer, a first convolutional layer, a second convolutional layer, a maximum pooling layer and a batch normalization layer, a third convolutional layer, a fourth convolutional layer, and a splicing enhancement processing layer, specifically: The input layer dimension is (4500,1); The number of convolution kernels in the first convolution layer is 64, the convolution kernel size is 11, the stride is 1, the dilation rate is 4, the activation function is selu, and the output dimension is (4500, 64); The second convolutional layer has 64 convolution kernels, a kernel size of 11, a stride of 4, a dilation rate of 1, an activation function of selu, and an output dimension of (1125, 64). The kernel size and stride of the maximum pooling layer and BN layer are both 2, and the output dimension is (562, 64); The third convolutional layer has 128 convolution kernels, a kernel size of 5, a stride of 1, a dilation rate of 4, an activation function of selu, and an output dimension of (562, 128). The fourth convolutional layer has 128 convolution kernels, a kernel size of 5, a stride of 4, a dilation rate of 1, an activation function of selu, and an output dimension of (141, 128). The splicing enhancement processing layer is used to copy and splice the first 1 / 4 of the feature map output by the fourth convolutional layer to the feature map output by the first convolutional layer, and the output dimension is (204, 128).

7. The method for identifying friend-or-foe (IFF) signals based on deep learning according to claim 6, characterized in that: The initial capsule layer dimension is 8, the number of channels is 32, the convolution kernel size is 9, the stride is 2, and the padding is the same.

8. The method for identifying friend-or-foe (IFF) signals based on deep learning according to claim 7, characterized in that: The digital capsule unit is used to compress the vector feature map into a two-dimensional feature vector, specifically: Where i_pri is the initial capsule unit, j_dig is the digital capsule unit, and v j_dig is the output vector of the digital capsule unit, s j_dig is the input vector of the digital capsule unit, is the prediction vector of the initial capsule unit, c i_pri,j_dig is the coupling coefficient between i_pri and all j_dig, b i_pri,j_dig is the initial coupling coefficient of i_pri and j_dig, W i_pri,j_dig is the weight matrix, u j_dig is the output of the initial capsule unit, k_dig is the dimension of the output capsule of the digital capsule unit, b i_pri,k_dig It is b i_pri,j_dig The k_digth initial coupling coefficient in , where k_dig is an integer.

9. The method for identifying friend-or-foe (IFF) signals based on deep learning according to claim 8, characterized in that: The attention mechanism is used to compress the two-dimensional feature vector again to obtain a one-dimensional feature vector, specifically: I. Input the output of the digital capsule unit into the g(x) function to obtain the intermediate weight of Attention: e at =g(x)=xW at Among them, W at ∈R 128*1 is the last element in the output of the fully connected layer, x∈R 10*128 is the output vector of the digital capsule unit; II. Probability vector using the intermediate weights of Attention: Among them, a at ∈R 10*1 is the intermediate weight e at The probability vector of at,j_at is a at The j_atth element in e k_at It is e at The k_atth element in e j_at It is e at The j_atth element of ; III. Input the output of the digital capsule unit into the spatial attention mechanism through the probability vector of the intermediate weights to obtain the weighted features captured by the intermediate attention mechanism: Among them, v j_dig|j_at It is v j_dig The j_atth element of ; IV. The last element W that will be used at With the output vector c at The vectors are concatenated into 256*1 and then output through a fully connected layer of size 128 to obtain a feature vector of dimension 128*1.

10. The method for identifying friend or foe signals based on deep learning according to claim 9, characterized in that: The fully connected layer is divided into: a first fully connected unit, a second fully connected unit, and a third fully connected unit; The first fully connected unit includes three fully connected layers with sizes of 256, 128 and 10 respectively, for outputting the IFF long signal classification task; The second fully connected unit includes two fully connected layers with sizes of 64 and 6 respectively, for outputting the IFF short signal classification task; The third fully connected unit includes two fully connected layers with sizes of 64 and 1 respectively, which are used to output the relative position relationship of the starting position of the radar friend-or-foe identification signal.

Images