IFF signal identification method, system and device based on DRSN network anti-taikang interference and storage medium

Abstract

The method, system, device and storage medium for identifying IFF signals based on DRSN network anti-Tacan interference are as follows: training data is generated by simulating received data and performing overlapping sliding window processing on the received data; then, the DRSN network is trained after the generated training data is used to construct the DRSN network; the features are input into the trained multiple DRSN networks to perform pattern recognition and position recovery on the IFF data to be identified, and the presence or absence and position of each pattern in the signal segment are identified; in the process of constructing the network training data, the energy and position conflict relationship among different mode IFF signals, noise and Tacan signals is fully considered; the system, device and storage medium based on the method for identifying IFF signals based on DRSN network anti-Tacan interference perform IFF signal identification against Tacan interference; the present application solves the problems of difficult IFF signal feature extraction and poor IFF signal pattern recognition performance under low signal-to-noise ratio and serious interference from the same frequency band Tacan signal, and improves the ability to resist Tacan signal interference.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of IFF signal identification, and particularly relates to an IFF signal identification method, system and device based on a DRSN network and resisting tower jamming and a storage medium. BACKGROUND

[0002] IFF (Identification Friend or Foe) system is an important communication technology in joint operations. There are many different types of IFF systems, and typical IFF systems are Mark X and Mark XII systems. At present, there are few researches on IFF signal detection and identification at home and abroad, and most of them are brief introductions to signal characteristics of enemy and friend identification signals and working characteristics of enemy and friend identification systems, and few of them involve specific problems of enemy and friend identification signal detection and identification.

[0003] Identification of enemy and friend identification signals is actually identification of signal modulation types, and the main difference is that enemy and friend identification signals do not use the usual modulation method, but have their own unique system. The main identification methods are hypothesis testing theory and statistical pattern recognition method. Hypothesis testing theory uses probability theory to calculate a suitable threshold, and then obtains a suitable identification condition. From the perspective of probability, it is the most effective method, but the optimal classifier completely formula is very complex, and even for a simple signal, the degree of disorder is very high. Moreover, it also needs to analyze on the basis of correct hypothesis to determine the suitable threshold. The statistical pattern recognition method can be divided into two parts. The first part is the extraction of characteristic parameters, and the main content of this part is to extract the characteristic parameters that are quite different from other signals and quite obvious from the received signals. The second part is pattern recognition, and the main content of this part is to determine the signal modulation mode according to one or more characteristic parameters extracted in the characteristic parameter extraction through a certain algorithm.

[0004] The prior art mainly uses the second method, and different technologies use different parameter extraction methods.

[0005] The patent application number 202310450840.7 discloses an IFF signal recognition method based on a cross algorithm. The method processes the output signal through Gaussian filtering to obtain a to-be-identified Gaussian pulse, and obtains two Gaussian pulses through sampling and delay processing. The to-be-identified Gaussian pulse is marked according to the intersection points of the two Gaussian pulses, and the marked to-be-identified Gaussian pulse is identified according to the pulse framework search of different signal modes. In the presence of Tacan pulse interference, the IFF signal and the Tacan pulse will be mixed in the time-frequency domain. Because the IFF signal pulse duration is extremely short, the IFF signal is easy to be hidden in the interference signal, and it is difficult for the method to extract the available Gaussian pulse features, and the pulse features are also submerged in a low signal-to-noise ratio. In addition, the setting of the threshold is also easily affected by the signal-to-interference ratio and the signal-to-noise ratio, so the method has poor robustness.

[0006] The patent application number 202210545270.5 discloses an IFF signal recognition method based on deep learning. The method uses a deep learning model to extract signal features, inputs the to-be-identified IFF signal into the trained IFF signal recognition model, and obtains the classification of the radar IFF signal and the relative position relationship of the starting position of the radar IFF signal. The training data set is composed of signal segments with interference composed of different modes of signals with interference, including intermediate frequency jitter, sampling clock jitter, path loss, additive Alpha stable distribution noise, random frame loss, and TACAN navigation system signal. The model uses a convolutional neural network layer, a capsule network layer, a channel attention mechanism layer, and a fully connected layer. On the one hand, the network parameters are large, with up to 35177237 training parameters, which has high requirements for the hardware platform in actual application. On the other hand, the construction of the data set has certain defects, especially for the position and energy conflict between the Tacan pulse and the IFF pulse, which leads to insufficient robustness of the network to Tacan interference. SUMMARY

[0007] In order to overcome the defects existing in the prior art, the purpose of the present application is to provide an IFF signal recognition method, system, equipment and storage medium based on DRSN network anti-tacan interference, which can learn the global information of human action in the video while paying attention to the local fine-grained feature information, obtain the local detail information of the human body from the local video input obtained by local interception operation, and obtain more video dynamic local fine-grained information by combining the features of different local parts to obtain better global features; the present application fully considers the energy and position conflict relationship among different mode IFF signals, noise and tacan signals in the process of constructing network training data, solves the difficulty of IFF signal feature extraction, overcomes the poor recognition performance of the current recognition technology under low signal-to-noise ratio of IFF signal, and improves the anti-tacan signal interference capability.

[0008] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0009] An IFF signal recognition method based on DRSN network anti-tacan interference, first, the features of the entire data segment of the collected IFF signal are extracted in the form of variable length window overlap sliding window, then the features are input to the trained multiple DRSN networks for pattern recognition, and finally the presence and position of each mode in the signal segment are recognized by the method of sliding window combined with pattern recognition.

[0010] An IFF signal recognition method based on DRSN network anti-tacan interference, comprising the following steps:

[0011] Step 1, generate training data by simulating received data and performing overlapping sliding window processing on the received data;

[0012] Step 2, construct a DRSN network according to the training data generated in step 1; first, construct the overall structure of the network, including the input layer, the convolution layer, the residual module group, the batch normalization, the nonlinear activation, the global pooling, the full connection layer and the output layer; second, construct the residual module group, build a general RSBU unit, each RSBU is composed of four parts, including two layers of convolution layer feature extraction part, attention mechanism module, soft thresholding module and cross-layer connection part, the output of the previous RSBU is also the input of the next RSBU unit;

[0013] Step 3: train the DRSN network constructed in step 2;

[0014] Step 4, use the trained DRSN network to identify the type of the IFF signal to be identified and recover the start and end positions.

[0015] The specific process of step 1 is as follows:

[0016] Step 1.1, using Matlab to simulate the received data; first fix the sampling frequency, set the interference to noise ratio (inr, 0 to 10 dB) and the signal to interference and noise ratio (sinr, -4 to 20 dB), generate more than 1 second of noise-free Tacan baseband real signal T under each signal to interference and noise ratio and interference to noise ratio, and do power normalization in the Tacan burst position section, as follows:

[0017]

[0018]

[0019] Wherein represents the total length of the burst part in the Tacan signal; add the noise corresponding to the required interference to noise ratio inr to the normalized Tacan signal to obtain X;

[0020] X is regarded as a template, and the IFF signal is added to the template X; define n signal modes; the start position start of each signal mode defined is random, there is no time domain overlap, and the number is equal; in the process of adding an IFF signal S with a length of L to X, first calculate the burst segment power of S , power normalize S, and then change the power of S according to the total power of the interference and noise in the start and end positions of S , the signal to interference and noise ratio sinr, and add S to X, as follows:

[0021]

[0022]

[0023]

[0024]

[0025] Wherein, represents the total length of the burst part in S; at this time, the template X contains the Tacan baseband signal, noise and IFF baseband signal, and the signal to interference and noise ratio and the signal to noise ratio are fixed; frequency conversion is performed on the template X, and the Hilbert transform is used to obtain the IQ data of the signal:

[0026] Random channel conditions are added to the template X signal, including Doppler shift (maximum 1KHz), sampling clock drift (maximum 4ppm) and receiver frequency offset (maximum 500Hz), to obtain the received signal Y; baseband frequency conversion and filtering are performed on Y to finally simulate the received data;

[0027] Step 1.2, the received data obtained in step 1.1 is subjected to overlapping sliding window processing: taking one sample, setting the current mode as t, fixing a window, and the window length win_len is twice the length of the IFF signal in the t mode, for the mode with information bits, only the synchronization bits are selected as the identification effective section, that is, the window size is twice the synchronization bits; the window is used for sliding, and the sliding uses overlapping sliding, one fifth of the window length is slid each time, and the IQ data frame in the window is recorded; if there is a complete IFF signal synchronization bit in the current sliding window, the label is set to 1, otherwise it is set to 0, and the label is recorded each time the window is slid; when the sliding window is slid, a sequence of 01 labels is obtained, which is the identification result of the sample in the t mode, 0 in the sequence represents no synchronization bit of the IFF signal in the mode, and 1 represents the synchronization bit of the IFF signal in the mode; if the 01 label is uneven, all 1 labels and their data are taken and an equal amount of data is randomly selected for the 0 label;

[0028] Step 1.3, the operation of step 1.2 is performed on all received data samples obtained in step 1.1, and the above overlapping sliding window processing is performed on each sample using the window corresponding to the n modes, n groups of IQ data and corresponding labels can be obtained, each group of data corresponds to a mode, and the number of frames and the number of labels in each group of data are equal to the number of overlapping sliding windows in the mode; the data and labels of the same mode of each received data sample are combined, and finally a data set and a label are obtained, the data set contains IQ data corresponding to n modes, which is divided into n training sets, validation sets and test sets.

[0029] Step 2.1, the overall structure of the network is constructed: the data set contains IQ data corresponding to n IFF signal modes, and the IQ data lengths are different between different modes, so the DRSN network is trained for the IQ data corresponding to each mode, a total of n DRSN networks; the network structure includes: input layer, convolutional layer, residual module group, batch normalization, nonlinear activation, global pooling, fully connected layer and output layer;

[0030] Among them, the input layer data is IQ data with a size of [2, win_len, 1], win_len is the window length of the current mode, the convolution kernel size of the convolutional layer is set to 3*3, and the same mode is used for padding, and the output dimension is [2, win, 16];

[0031] The output dimensions of the batch normalization layer and the nonlinear activation layer (ReLU function) are both [1, 1, 96]; the output dimension of the global pooling layer using average pooling is 96; the fully connected layer has 2 nodes; and the SoftMax activation layer is used as the output;

[0032] Step 2.2, constructing the residual module group; a general RSBU unit is built, each RSBU consisting of four parts, including a feature extraction part of two convolutional layers, an attention mechanism module, a soft thresholding module, and a cross-layer connection part;

[0033] In the RSBU unit, the input is first divided into two paths: one of the input paths enters two convolutional layers for further feature refinement to obtain refined features, and each convolutional layer is accompanied by batch normalization and ReLU operations, and their convolution kernel numbers, sizes, and step lengths are the same; the refined features are processed by absolute value and then reduced to a one-dimensional vector by global average pooling; the one-dimensional vector is divided into two paths, one of the one-dimensional vectors is input to an attention mechanism module consisting of two fully connected layers, where the first fully connected layer contains BN and ReLU, and the second fully connected layer only has ReLU, and the node numbers of the two fully connected layers are the same as the length of the one-dimensional vector; the other path of the one-dimensional vector is calculated to obtain a threshold value after multiplication with the output of the attention mechanism module ; the soft thresholding module is used to process the refined features, and the features with absolute values less than the threshold value of the refined features are deleted, and the features with absolute values greater than of the refined features are shrunk towards zero, and the formula is as follows:

[0034]

[0035] wherein, denotes the refined features, denotes the result of the soft thresholding module operation on the refined features, and is added to the other input path to realize cross-layer identity connection, and this result is the output of the RSBU and the input of the next RSBU unit.

[0036] The specific process of step 3 is as follows:

[0037] The data set generated in step 1.2 is used to train the DRSN network corresponding to each IFF signal mode, a total of n DRSN networks.

[0038] The specific process of step 4 is as follows:

[0039] Step 4.1, multiple sliding windows are taken for the IFF data to be identified, and a window is fixed for different modes of the IFF signal in turn according to step 1.2, the window length win_len is twice the length of the IFF signal synchronization bit, and the sliding window is taken with a step of one fifth of the window length, the IQ data in the window is saved in the form of frames, a total of n groups of IQ data frames are obtained, the number of frames in each group is floor((L-win_len) / 5), L represents the total number of points of the IFF data to be identified, win_len represents the window length corresponding to different modes of IFF, and floor represents rounding down; the n groups of IQ data frames are sent into corresponding networks respectively, the label [1,0] is marked as 1, and the label [0,1] is marked as 0, then a total of n groups of 01 labels are obtained.

[0040] Step 4.2, the n groups of 01 labels obtained in step 4.1 are used to construct the original IFF mode type and sequence, in order to reduce the false alarm probability, the IFF different mode window length win_len is arranged in descending order, and the IFF signal length of n signal modes is sequentially numbered 1, 2, …, n in turn, and then the following steps are performed:

[0041] Initialize v_stpoint and v_edpoint to be empty as the starting position and ending position vectors of each predicted IFF signal; initialize v_midpos1 and v_midpos2 to be empty as the center position vectors of each predicted IFF signal; initialize v_predictlabel to be empty as the type of each predicted IFF signal; midpos1 represents the center position of the first window corresponding to the continuous 1 label; midpos2 represents the center position of the second window corresponding to the continuous 1 label; stpoint represents the starting position of the predicted IFF; edpoint represents the ending position of the predicted IFF, and i is initialized to 1,

[0042] Let the current signal mode number be i, the corresponding window length be win_len, and the corresponding original IFF length be L, and let the label index k be 0;

[0043] Search the 01 label sequence corresponding to the i mode, k=k+1, and let the kth label be searched, when the kth label and the k+1th label are both 1, record the window center point positions midpos1 and midpos2 corresponding to the two labels, set the two 1 labels to 0, record the starting position of the first label corresponding to the window as the starting position of the predicted IFF, that is, stpoint=midpos1-win_len / 4, and record the starting position of the second label corresponding to the window plus the original IFF signal length as the ending position of the predicted IFF, that is, edpoint=midpos2-win_len / 4+L;

[0044] find the position st of the value of v_stpoint which is smaller than stpoint and has the minimum distance to stpoint, judge whether the stth value of v_edpoint is greater than or equal to stpoint, if less than, enter e), otherwise judge whether the stth value of v_edpoint-stpoint is greater than 0.2 times of L, if less than, enter e), otherwise, let k=k+1 enter c);

[0045] find the position ed of the value of v_edpoint which is greater than edpoint and has the minimum distance to edpoint, judge whether the edth value of v_stpoint is less than or equal to stpoint, if less than, enter f), otherwise judge whether the edth value of v_stpoint-edpoint is greater than 0.2 times of L, if less than, enter f), otherwise, let k=k+1 enter c);

[0046] record stpoint, edpoint, midpos1, modpos2, i in v_stpoint, v_edpoint, v_midpos1, v_midpos2, v_predictlabel, and enter g);

[0047] judge whether k reaches the maximum value, if not, enter c), if yes, let i=i+1, then judge whether i reaches the maximum value, if not, enter b), otherwise end;

[0048] After step 4.2, v_predictlabel, v_stpoint, v_edpoint are obtained, wherein, v_predictlabel corresponds to the type of IFF signal, v_stpoint corresponds to the starting position of IFF signal, and v_edpoint corresponds to the termination position of IFF signal.

[0049] An IFF signal recognition system based on DRSN network anti-tacan interference, comprising an IFF signal generation module, a DRSN network feature extraction module, and an IFF signal recognition module based on an overlapping sliding window method;

[0050] The IFF signal generation module is used to generate data for training the deep learning network in step 1, and the signal-to-noise ratio and the signal-to-interference-and-noise ratio are used to control the power of the IFF signal with a burst segment, the Tacan interference with a burst segment, and the noise, so that the data set quality is controllable;

[0051] The DRSN network feature extraction module is used to extract features of the IFF signal in step 2, and the residual network and soft thresholding are used to improve the stability and anti-noise performance of the model;

[0052] An IFF signal identification module based on an overlapping sliding window method is used to identify multiple modes of the IFF signal and recover the position in step 4, simplifies the multi-task learning process, and effectively reduces the complexity of network training.

[0053] An IFF signal identification device based on a DRSN network anti-Tacan interference, comprising:

[0054] A memory for storing a computer program;

[0055] A processor for implementing the IFF signal identification method based on the DRSN network anti-Tacan interference in steps 1 to 4 when executing the computer program.

[0056] A computer readable storage medium storing a computer program, which can perform the IFF signal identification method based on the DRSN network anti-Tacan interference in steps 1 to 4 when executed by a processor.

[0057] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0058] 1. The present application introduces the DRSN network into the IFF signal identification, fully utilizes the characteristics of the residual network that can solve the gradient explosion or disappearance, and improves the stability of the model; the soft threshold operation is used to enable the samples under different signal-to-interference-and-noise ratios to independently adjust the threshold according to the content of the interference and noise in the samples; and the anti-noise performance of the model is improved.

[0059] 2. The IFF signal identification method adopted by the present application is to design different length windows and use the overlapping sliding window method to preprocess the IFF signal. For each IFF mode, the window length and the sliding length are set, the IFF signal is divided into a series of frames under the mode window length, and the 01 label sequence is extracted by inputting the IFF signal into the corresponding DRSN network. Finally, the 01 label sequence of each mode is designed to recover the IFF mode and its position. This method avoids the multi-task learning process of IFF signal identification and position recovery, effectively reducing the complexity of network training

[0060] 3. In the data set construction, the power problem of IFF and Tacan signals in the burst segment is fully considered, the signal-to-interference-and-noise ratio and the interference-to-noise ratio are used, and the power ratio relationship among IFF, Tacan and noise is more effectively controlled, and the signal modeling process is improved.

[0061] 4. From the identification method, the present application effectively simplifies the recovery type and position task in the form of overlapping sliding window, which also makes the network model sharply reduce in parameter amount, and the parameter amount of the present application using DRSN is only 6.64% of the parameter amount of the prior art MTCNAM.

[0062] In summary, the present application uses a deep learning method instead of a complex artificial feature extraction step, which overcomes the defect that artificial feature extraction is sensitive to noise and interference, and avoids the instability of the artificial feature extraction step. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 Flowchart of the present application.

[0064] Figure 2 Design form of the sliding window of the present application.

[0065] Figure 3 DRSN network structure used in the present application.

[0066] Figure 4 Specific structure of each RSBU inside the present application, where S represents soft threshold operation.

[0067] Figure 5 Flowchart of the type identification and position recovery of the IFF signal of the present application.

[0068] Figure 6 Simulated data generated by the embodiment of the present application, the influence of signal-to-interference-and-noise ratio on identification performance, including false alarm probability and identification accuracy.

[0069] Figure 7 Simulated data generated by the embodiment of the present application, the influence of signal-to-interference-and-noise ratio on identification performance, including false alarm probability and identification accuracy. DETAILED DESCRIPTION

[0070] In order to enable personnel in the art to better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0071] Embodiment:

[0072] A DRSN network anti-Tacan interference based IFF signal identification method, specifically comprising the following steps:

[0073] Step 1, generate training data by simulating received data and performing overlapping sliding window processing on the received data;

[0074] Step 1.1, use Matlab to simulate received data; first fix the sampling frequency, set the interference-to-noise ratio (inr, 0 to 10 dB) and the signal-to-interference-and-noise ratio (sinr, -4 to 20 dB), generate 1 second or more of Tacan baseband real signal T without noise under each signal-to-interference-and-noise ratio and interference-to-noise ratio, and perform power normalization on the Tacan burst position segment, the formula is as follows:

[0075]

[0076]

[0077] wherein represents the total length of the burst part in Tacan signal; X is obtained by adding noise corresponding to the inr to the normalized Tacan signal;

[0078] X is regarded as a template, and the IFF signal is added to the template X; the IFF signal includes six modes, which are IFF Mark X system mode 1, mode 2, mode 3 / A, mode C four modes and IFF Mark X II system mode 4, mode S two modes in turn; the start positions of each mode on the template X are random, have no time domain overlap and are equal in number; in the process of adding an IFF signal S with a length of L to X, the burst segment power of S is calculated first , S is power-normalized, and then S is power-varied according to the total power of the interference and noise in the start and end positions of S and the signal-to-interference-and-noise ratio sinr, and is added to X, and the formula is as follows:

[0079]

[0080]

[0081]

[0082]

[0083] wherein represents the total length of the burst part in S; at this time, the template X contains the Tacan baseband signal, noise and IFF baseband signal, and the signal-to-interference-and-noise ratio and the signal-to-noise ratio are fixed; the template X is frequency-converted, and the Hilbert transform is used to obtain the IQ data of the signal:

[0084] Random channel conditions are added to the template X signal, including Doppler shift (maximum 1KHz), sampling clock drift (maximum 4ppm) and receiver frequency offset (maximum 500Hz), to obtain a received signal Y; the Y is baseband-converted and filtered, and finally the received data is simulated;

[0085] Step 1.2, the received data obtained in step 1.1 is subjected to overlapping sliding window processing: taking one sample, setting the current mode as t, fixing a window, and the window length win len is twice the length of the IFF signal in the t mode, for the mode with information bits, only the synchronization bits are selected as the identification effective section, that is, the window size is twice the synchronization bits; the window is used for sliding, and the sliding uses overlapping sliding, one fifth of the window length is slid each time, and the IQ data frame in the window is recorded; if there is a complete t mode IFF signal synchronization bit in the current sliding window, the label is set to 1, otherwise it is set to 0, and the label is recorded each time the window is slid; when the sliding window is slid, a sequence of 01 labels is obtained, which is the identification result of the t mode for the sample, and 0 in the sequence indicates that there is no t mode IFF signal synchronization bit, and 1 indicates that there is a t mode IFF signal synchronization bit; for example, a window with a length of twice the t mode length is used for sliding on a certain sample, and it is assumed that the current dashed window contains the t mode synchronization bit (shadow part), and the gray dashed window is obtained after three times of sliding, a total of 4 frames of IQ data frames and the corresponding labels 1110 are obtained, wherein 1 indicates that the t mode synchronization bit is inside the window, and 0 indicates that the t mode synchronization bit is not inside the window. Note that at this time, due to the unevenness of the 01 label, all 1 labels and their data are taken, and an equal amount of data is randomly selected for the 0 label. Figure 2

[0086] Step 1.3, the above step 1.2 operation is performed on all received data samples obtained in step 1.1, and the above overlapping sliding window processing is performed on each sample using the windows corresponding to the 6 modes, and 6 groups of IQ data and corresponding labels are obtained, each group of data corresponds to a mode, and the number of frames and the number of labels in each group of data are equal to the number of overlapping windows in the mode. The data and labels of the same mode of each received data sample are combined to finally obtain a data set and a label, and the data set contains IQ data corresponding to 6 modes, which is divided into 6 groups of training set, validation set and test set.

[0087] Step 2, constructing a DRSN network;

[0088] Step 2.1, constructing the overall structure of the network: the data set contains IQ data corresponding to 6 IFF signal modes, and the IQ data lengths of different modes are different, so the DRSN network is trained for the IQ data corresponding to each mode, a total of 6 DRSN networks. The network structure includes: input layer, convolutional layer, residual module group, batch normalization, nonlinear activation, global pooling, fully connected layer and output layer;

[0089] Among them, the input layer data is IQ data with a size of [2, win len, 1], win len is the window length of the current mode, the convolution kernel size of the convolutional layer is set to 3*3, the number is 16, the step is 1, and the output dimension is [2, win, 16] using the same mode padding.​

[0090] The residual module group consists of 13 modified deep residual shrinkage building units (RSBUs), which are introduced in step 2.2;

[0091] The batch normalization layer and the nonlinear activation layer (ReLU function) have an output dimension of [1, 1, 96]; the global pooling layer uses average pooling with an output dimension of 96; the fully connected layer has 2 nodes; and the SoftMax activation layer is used as the output;

[0092] Step 2.2, build the residual module group; build a general RSBU unit, each RSBU consists of four parts, including two layers of convolutional layers for feature extraction, an attention mechanism module, a soft thresholding module, and a cross-layer connection part;

[0093] Within the RSBU unit, the input is first divided into two paths: one path of the input enters two layers of convolutional layers for further feature refinement to obtain refined features, and each layer of convolution is accompanied by batch normalization and ReLU operations, which have the same number of convolutional kernels, size, and step length settings; the refined features are processed by absolute value and then reduced to a one-dimensional vector by global average pooling; this one-dimensional vector is divided into two paths, one path of the one-dimensional vector is input to an attention mechanism module consisting of two fully connected layers, where the first fully connected layer contains BN and ReLU, and the second fully connected layer only has ReLU, and the number of nodes of the two fully connected layers is the same as the length of the one-dimensional vector; the other path of the one-dimensional vector is calculated by averaging and multiplied by the output of the attention mechanism module to obtain the threshold ; the soft thresholding module is used to process the refined features, and the features with absolute values less than the threshold are deleted, and the features with absolute values greater than are shrunk towards zero, as follows:

[0094]

[0095] wherein represents the refined features, represents the result of the soft thresholding module operation on the refined features, and is added to the other path of the input to realize the cross-layer identity connection, and this result is the output of the RSBU and the input of the next RSBU unit. The convolutional kernel size, number, and step length settings of the convolutional layers within the 13 RSBU units are shown in Table 1.

[0096] Table 1 Convolutional parameter settings within each RSBN

[0097]

[0098] Step 3: training the DRSN network.

[0099] According to the description in step 2.1, the DRSN network corresponding to each IFF signal mode is trained using the data set generated in step 1.2, a total of 6 DRSN networks; each IQ sample is zero-centered and normalized, and the formula is as follows:

[0100]

[0101] wherein is the I or Q data, is the mean, is the variance.

[0102] Set the training batch size batchsize, and in the training process, the network is trained once using batchsize size data each time, the validation set is used to validate the network, the validation set accuracy is obtained, if the validation set accuracy is greater than the validation set accuracy obtained after the last training network, the training model parameters are saved, otherwise the training network parameters are not saved.

[0103] The training process uses the Adam optimizer, the label is one-hot encoding of [1, 0] and [0, 1], the objective function is cross-entropy function, the learning rate is set to 0.1, if the validation set loss does not decrease during 8 rounds of learning, the learning rate is reduced by 50%, if the validation set loss does not decrease during 16 rounds of learning, the iteration is stopped.

[0104] Step 4: mode recognition and position recovery of the IFF data to be identified, which is subjected to analog-to-digital conversion, frequency conversion, filtering, sampling rate adaptation, etc.

[0105] Step 4.1: multiple sliding window value taking is performed on the IFF data to be identified, and according to the description in step 1.2, a window is fixed in turn for different modes of IFF signal, the window length win_len is twice the length of the IFF signal synchronization bit, and the sliding window value taking is performed with a step of one-fifth of the window length, the IQ data in the window is saved in the form of frames, a total of 6 groups of IQ data frames are obtained, the number of frames in each group is floor((L-win_len) / 5), L represents the total number of points of the IFF data to be identified, win_len represents the window length corresponding to different modes of IFF, and floor represents rounding down; the 6 groups of IQ data frames are sent into the corresponding network, the label [1, 0] is marked as 1, and the label [0, 1] is marked as 0, then a total of 6 groups of 01 labels are obtained.

[0106] Step 4.2, using the 6 groups of 01 labels, the original IFF mode type and order of construction, in order to reduce the false alarm probability, according to the IFF different mode window length win_len descending order, corresponding to the mode C, mode 4, mode S, mode 3, mode 2, mode 1 IFF signal length, according to the order, the mode is numbered 1, 2, …, 6, then the following steps are carried out:

[0107] Initialize v_stpoint, v_edpoint to empty as the starting position and ending position vector of each predicted IFF signal; initialize v_midpos1, v_midpos2 to empty as the center position vector of each predicted IFF signal; initialize v_predictlabel to empty as the type of each predicted IFF signal; use midpos1 to represent the center position of the corresponding first window of the continuous 1 label; use midpos2 to represent the center position of the corresponding second window of the continuous 1 label; use stpoint to represent the starting position of the predicted IFF; use edpoint to represent the ending position of the predicted IFF, i is initialized to 1,

[0108] Let the current mode number be i, the corresponding window length be win_len, and the corresponding original IFF length be L, let the label index k be 0;

[0109] Search the 01 label sequence corresponding to the i mode, k=k+1, and let the current search be the kth label. When the kth label and the k+1th label are both 1, record the window center point positions midpos1 and midpos2 corresponding to the two labels, set the two 1 labels to 0, record the starting position of the first label corresponding window as the starting position of the predicted IFF, i.e. stpoint=midpos1-win_len / 4, and record the starting position of the second label corresponding window plus the original IFF signal length as the ending position of the predicted IFF, i.e. edpoint=midpos2-win_len / 4+L;

[0110] Find the position st of the value of v_stpoint smaller than stpoint and the smallest distance from stpoint, judge whether the stth value of v_edpoint is greater than or equal to stpoint, if less than, enter e), otherwise, judge whether the stth value of v_edpoint-stpoint is greater than 0.2 times of L, if less than, enter e), otherwise, let k=k+1 enter c);

[0111] find the position ed of the value of v_edpoint which is greater than edpoint and has the minimum distance to edpoint, judge whether the edth value of v_stpoint is less than or equal to stpoint, if less, enter f), otherwise, judge whether the edth value of v_stpoint minus edpoint is greater than 0.2 times of L, if less, enter f), otherwise, let k=k+1 enter c);

[0112] record stpoint, edpoint, midpos1, modpos2, i in v_stpoint, v_edpoint, v_midpos1, v_midpos2, v_predictlabel, and enter g);

[0113] judge whether k reaches the maximum value, if not, enter c), if yes, let i=i+1, then judge whether i reaches the maximum value, if not, enter b), otherwise, end;

[0114] After step 4.2, v_predictlabel, v_stpoint, v_edpoint are obtained, wherein, v_predictlabel corresponds to the type of IFF signal, v_stpoint corresponds to the starting position of IFF signal, and v_edpoint corresponds to the ending position of IFF signal.

[0115] Monte Carlo experiments are performed 1000 times using Matlab according to the embodiment, and the false alarm probability (blue line) and recognition accuracy (red line) can be obtained, as shown in Figure 6 、 7 It can be seen from Figure 6 that when the signal-to-interference ratio is greater than 4dB, the false alarm probability of the present application is less than 10%, and the recognition probability is greater than 90%; when the signal-to-interference ratio is greater than 8dB, the false alarm probability of the present application is less than 2%, and the recognition probability is greater than 95%; it can be seen from Figure 7 that the influence of the interference-to-noise ratio on the recognition performance is small, and whether the interference-to-noise ratio is large or small, the recognition accuracy is more than 85%, and the false alarm probability is less than 15%. The present application still has good recognition performance under poor signal-to-interference ratio conditions, and the recognition performance is continuously improved as the signal-to-interference ratio rises; at the same time, the present application is robust to Tacan interference, and the recognition performance is not affected by the strength change of Tacan signal interference.

[0116] The embodiments of the present application are not limited by the above-mentioned embodiments, and any changes, modifications, substitutions, combinations and simplifications made without departing from the spirit and principles of the present application shall be equivalent replacement modes and shall be included in the protection scope of the present application.

Claims

1. A method for IFF signal identification based on DRSN network anti-TACAN interference, characterized in that, Specifically, the following steps are included: Step 1: By simulating the reception of data and performing overlapping sliding window processing on the received data, training data is generated. The received data effectively controls the power ratio relationship between IFF, Tacan, and noise. The overlapping sliding window processing includes setting the window length and sliding length for each IFF mode, dividing a segment of IFF signal into a series of frames under the window length of the mode, and extracting the 01 label sequence. Step 2: Construct the DRSN network based on the training data generated in Step 1. First, construct the overall network structure, including the input layer, convolutional layer, residual module group, batch normalization, non-linear activation, global pooling, fully connected layer, and output layer. Second, construct the residual module group and build a general RSBU unit. Each RSBU consists of four parts, including a feature extraction part of two convolutional layers, an attention mechanism module, a soft thresholding module, and a cross-layer connection part. The output of the previous RSBU is also the input of the next RSBU unit. Step 3: Train the DRSN network constructed in Step 2; Step 4: Use the trained DRSN network to perform type identification and start / end position recovery on the IFF signal to be identified.

2. The IFF signal identification method based on DRSN network anti-TACAN interference according to claim 1, characterized in that, The specific process of step 1 is as follows; Step 1.1: Simulate data reception using Matlab. First, fix the sampling frequency, set the interference-to-noise ratio (INR) to 0-10dB and the signal-to-interference-to-noise ratio (SNR) to sinR to -4-20dB. For each SNR and INR, generate a Tacan baseband real signal T with noise-free signal lasting more than 1 second, and normalize the power within the burst position segment of the Tacan signal. Add the noise corresponding to the required INR to the normalized Tacan signal to obtain X. Treating X as a template, IFF signals are added to template X; n custom signal modes are defined; the start positions of each custom signal mode on template X are random, with no temporal overlap, and the number is equal; when adding an IFF signal S of length L to X, first calculate the burst power P of S, normalize the power of S, and then calculate the total power P of interference and noise within the start and end positions of S. x The signal-to-interference-plus-noise ratio (SIR) is calculated using sinr, a power transformation is applied to S, and this power transformation is added to X. The template X is then frequency-converted, and a Hilbert transform is used to obtain the IQ data of the signal. Random channel conditions are added to the template signal X, including a maximum Doppler frequency shift of 1 kHz, a maximum sampling clock drift of 4 ppm, and a maximum receiver frequency offset of 500 Hz, to obtain the received signal Y; Baseband frequency conversion and filtering are performed on Y to finally simulate the received data; Step 1.2: Perform overlapping sliding window processing on the received data obtained in Step 1.1: Take one sample, let the current mode be t, fix a window, the window length win_len is twice the length of the IFF signal in mode t. For modes with information bits, only the synchronization bit is selected as the valid segment for recognition, that is, the window size is twice the synchronization bit. Use this window to slide the sample, using overlapping sliding, sliding one-fifth of the window length each time, and record the IQ data frame in the window. If there is a complete synchronization bit of the IFF signal in mode t in the current sliding window, the tag is set to 1, otherwise it is set to 0. Record the tag each time it slides. When the sliding window is finished, a set of 01 tag sequence is obtained as the recognition result of the sample in mode t. In the sequence, 0 indicates that there is no synchronization bit of the IFF signal in this mode, and 1 indicates that there is a synchronization bit of the IFF signal in this mode. If the 01 tags are not uniform, take all 1 tags and their data and randomly select an equal amount of data for the 0 tags. Step 1.3: Perform the operation of step 1.2 on all received data samples obtained in step 1.

1. For each sample, use n different modes to perform the above overlapping sliding window processing in sequence to obtain n sets of IQ data and corresponding tags. Each set of data corresponds to a certain mode. The number of frames and tags in each set of data is equal to the number of overlapping sliding windows in that mode. The data and labels of the same pattern from each received data sample are merged to obtain the final dataset and labels. The dataset contains IQ data corresponding to n patterns, which are then divided into n training sets, validation sets, and test sets.

3. The IFF signal identification method based on DRSN network anti-TACAN interference according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1, Construct the overall network structure: The dataset contains IQ data corresponding to n IFF signal modes, and the IQ data lengths are different between different modes. Therefore, DRSN networks are trained on the IQ data corresponding to each mode, resulting in a total of n DRSN networks. The network structure includes: input layer, convolutional layer, residual module group, batch normalization, nonlinear activation, global pooling, fully connected layer, and output layer. The batch normalization layer and the non-linear activation layer both have ReLU function output dimensions of [1, 1, 96]; the global pooling layer uses average pooling with an output dimension of 96; the fully connected layer has 2 nodes; and finally, a SoftMax activation layer is used as the output. Step 2.2, construct residual module group; build a general RSBU unit, each RSBU consists of four parts, including a feature extraction part of two convolutional layers, an attention mechanism module, a soft thresholding module, and a cross-layer connection part; Within the RSBU unit, the input is first split into two paths: one path enters two convolutional layers for further feature refinement, resulting in refined features. Each convolutional layer includes batch normalization and ReLU operations, with identical kernel numbers, sizes, and strides. The refined features undergo absolute value processing and are then reduced to a one-dimensional vector using global average pooling. This one-dimensional vector is split into two paths: one path inputs to an attention mechanism module consisting of two fully connected layers. The first fully connected layer includes batch normalization and ReLU, while the second fully connected layer only performs ReLU. The number of nodes in both fully connected layers is the same as the length of the one-dimensional vector. The other path calculates the average value of the one-dimensional vector and multiplies it by the output of the attention mechanism module to obtain a threshold τ. A soft thresholding module is then used to process the refined features, deleting features with absolute values ​​less than the threshold τ and shrinking features with absolute values ​​greater than τ towards zero, as shown in the following formula: Here, x represents the refined feature, and y represents the result of the refined feature after the soft thresholding module operation. Adding y to the other input path achieves cross-layer identity connection. This result is the output of RSBU and also the input of the next RSBU unit.

4. The IFF signal identification method based on DRSN network anti-TACAN interference according to claim 1, characterized in that, The specific process of step 3 is as follows: Using the dataset generated in step 1.2, train DRSN networks on the IQ data corresponding to each IFF signal mode, for a total of n DRSN networks.

5. The IFF signal identification method based on DRSN network anti-TACAN interference according to claim 1, characterized in that, The specific process of step 4 is as follows: Step 4.1: Perform multiple sliding window values ​​on the IFF data to be identified. According to Step 1.2, fix a window for different IFF signal modes in sequence. The window length win_len is twice the length of the IFF signal synchronization bit. Slide window values ​​in increments of one-fifth of the window length. Save the IQ data in the window as frames. A total of n sets of IQ data frames are obtained. The number of frames in each set is floor((L-win_len) / 5), where L represents the total number of IFF data points to be identified, win_len represents the window length corresponding to different IFF modes, and floor means rounding down. Send the n sets of IQ data frames into the corresponding networks respectively. Mark the label [1,0] as 1 and the label [0,1] as 0. Then, a total of n sets of 01 labels are obtained. Step 4.2: Using the n sets of 01 tags obtained in Step 4.1, construct the original IFF mode type and order. To reduce the false alarm probability, sort them in descending order according to the window length win_len of different IFF modes, corresponding to the IFF signal lengths of the n signal modes. Number each signal mode sequentially as 1, 2, ..., n according to this order, and then perform the following steps: a) Initialize v_stpoint and v_edpoint to empty, serving as the start and end position vectors for each predicted IFF signal; initialize v_midpos1 and v_midpos2 to empty, serving as the center position vector for each predicted IFF signal; initialize v_predictlabel to empty, serving as the type of each predicted IFF signal; use midpos1 to represent the center position of the first window corresponding to consecutive 1 labels; use midpos2 to represent the center position of the second window corresponding to consecutive 1 labels; use stpoint to represent the start position of the predicted IFF; use edpoint to represent the end position of the predicted IFF, and initialize i to 1. b) Let the current signal mode number be i, the corresponding window length be win_len, the corresponding original IFF length be L, and let the label index k be 0; c) Search the 01 label sequence corresponding to mode i, k = k + 1. Suppose that the current search finds the kth label. When the kth label and the (k+1)th label are both 1, record the window center point positions midpos1 and midpos2 corresponding to these two labels. Set these two 1 labels to 0. Record the starting position of the window corresponding to the first label as the starting position of the predicted IFF, i.e., stpoint = midpos1 - win_len / 4. Record the starting position of the window corresponding to the second label plus the original IFF signal length as the ending position of the predicted IFF, i.e., edpoint = midpos2 - win_len / 4 + L. d) Find the position st of the value that is smaller than stpoint and is closest to stpoint. Determine if the st-th value of v_edpoint is greater than or equal to stpoint. If it is less than stpoint, proceed to e). Otherwise, determine if the st-th value of v_edpoint minus stpoint is greater than 0.2 times L. If it is less than stpoint, proceed to e). Otherwise, set k = k + 1 and proceed to c). e) Find the position ed of the value that v_edpoint is larger than edpoint and is closest to edpoint. Determine whether the ed-th value of v_stpoint is less than or equal to stpoint. If it is less, proceed to f). Otherwise, determine whether the ed-th value of v_stpoint minus edpoint is greater than 0.2 times L. If it is less, proceed to f). Otherwise, set k = k + 1 and proceed to c). f) Record stpoint, edpoint, midpos1, midpos2, and i in v_stpoint, v_edpoint, v_midpos1, v_midpos2, and v_predictlabel, and proceed to g); g) Determine if k has reached its maximum value. If not, proceed to c). If it has, set i = i + 1. Then determine if i has reached its maximum value. If not, proceed to b). Otherwise, end. After step 4.2, we obtain v_predictlabel, v_stpoint, and v_edpoint. v_predictlabel corresponds to the type of the IFF signal, v_stpoint corresponds to the start position of the IFF signal, and v_edpoint corresponds to the end position of the IFF signal.

6. An IFF signal identification system based on DRSN network to resist TACAN interference, used to implement the method according to any one of claims 1 to 5, characterized in that, It includes an IFF signal generation module, a DRSN network feature extraction module, and an IFF signal recognition module based on the overlapping sliding window method; The IFF signal generation module is used to generate data for training the deep learning network in step 1. It uses the interference-to-noise ratio and the signal-to-interference-to-noise ratio to perform power control on the IFF signal with burst segments, the Tacan interference with burst segments, and noise, so that the quality of the dataset is controllable. The DRSN network feature extraction module is used to extract features from the IFF signal in step 2, and uses residual networks and soft thresholding to improve the stability and noise resistance of the model. The IFF signal recognition module based on the overlapping sliding window method is used to identify and recover the position of various IFF signals in step 4, which simplifies the multi-task learning process and effectively reduces the complexity of network training.

7. An IFF signal identification device based on DRSN network anti-TACAN interference, used to implement the method described in any one of claims 1 to 5, characterized in that, include: Memory, used to store computer programs; A processor is used to implement the IFF signal identification method based on DRSN network anti-TACAN interference described in steps 1 to 4 when executing the computer program.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, is capable of identifying IFF signals that resist TACAN interference according to the IFF signal identification method based on DRSN network anti-TACAN interference as described in any one of claims 1 to 5.

Images