Intelligent classification method for IFF signals

Abstract

The application discloses an IFF signal intelligent classification and recognition method, first, a training set containing positive and negative samples under different sampling rates is simulated by using signals of known modes, then the simulated training set is trained by using a hierarchical capsule network, finally, a deep learning detection model is built by using the trained network, so that continuous detection and recognition of actual signals are realized. The problem of intelligent classification and recognition of continuous signals in a complex electromagnetic environment is solved, that is, the working modes of each position of the measured IFF signal are recognized and output.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of signal sorting, and particularly relates to an intelligent classification and identification method for IFF signals. BACKGROUND

[0002] IFF refers to identification of the friend-or-foe attribute of a target on a battlefield through various technical means. Accurate and efficient identification of the friend-or-foe attribute of an aerial target can greatly enhance the accuracy of combat command and control, and plays a very important role in the victory of joint operations. With the development of various modern high-tech technologies, the concealment of weapon equipment has been greatly improved, especially the development of electromagnetic interference capability, which makes the air combat environment extremely complex, which poses a severe challenge to the accuracy of the friend-or-foe identification capability of aerial targets.

[0003] At present, the identification of IFF signals is still in the use of traditional methods, that is, the framework pulse detection based on the sliding window method as a general detection means. In the paper "Research on Real-time Signal Processing System of Generalized Secondary Radar" published by Zhu Liang et al., the processing idea of the method is described: constructing a radar communication signal knowledge base, matching the real-time input signal stream with the knowledge base, then estimating the characteristics and parameters of the matched signal, and classifying based on this. This method strictly depends on the length, pulse width, pulse number, pulse sequence and other subtle features of each mode signal. However, in the real electromagnetic environment, the received signal pulse width will be narrowed, tailing and other phenomena, which will lead to the failure of the traditional method detection, resulting in signal missed detection. At the same time, because the carrier of each mode signal under inquiry or inquiry is in the same frequency band, at this time the real-time received signal exists the difficulty of normalization caused by the large difference between the amplitudes of adjacent mode signals, which makes the traditional method exist false detection problem. The above are all the problems that cannot be avoided by the traditional detection method.

[0004] While the multi-layer convolutional neural network CNN can also use one-dimensional convolutional layers to extract signal features, but because of the insufficient feature processing capability, it will give a strong detection probability as soon as a pulse appears, and it is also unable to detect when the signal amplitude is too low, so the detection capability also has defects. SUMMARY

[0005] The purpose of the present application is to provide an intelligent classification and identification method for IFF signals, which uses a hierarchical capsule neural network to fully exploit the limited time-domain features of signals, to make up for the shortcomings of traditional methods and multi-layer convolutional neural network CNN methods.

[0006] The technical solution of the application is as follows: an IFF signal intelligent classification and identification method based on a hierarchical capsule neural network fully mines limited signal time domain features, first, a sample training set with different signal-to-noise ratios is simulated by using signals of known modes, then the hierarchical capsule network is trained by using the simulated training set, finally, a deep learning detection model is built by using the trained network, so that continuous detection and identification of actual signals are realized. The steps are as follows:

[0007] Step 1, the working mode is selected as a response signal, and a training sample set is generated.

[0008] Step 2, the hierarchical capsule neural network is trained by using the training sample set, and learning parameters are generated.

[0009] Step 3, the neural network is trained and the neural network output vector v j is generated, and the data attention distribution probability is obtained by using the attention mechanism algorithm.

[0010] Step 4, the data mode is perceived and the data is fitted by using the full connection layer processing, and finally the mode recognition result is output.

[0011] Compared with the prior art, the application has the following advantages:

[0012] First, the sliding window method based on frame information has a fixed threshold, and when the amplitude of the signal to be detected is low and the signal-to-noise ratio is low, the detection will be missed and the detection will be wrong, and the signal must be completely in the frame to be detected, and the detection performance is low. The method proposed in the application can use the hollow convolution layer to extract features and use the hierarchical capsule network to further strengthen and weight the features, overcoming the problem of complex pulse judgment in the prior art which needs to use a large amount of prior knowledge and consider special cases. The method has good recognition effect within a certain range of signal-to-noise ratio.

[0013] Second, when setting the capsule network parameters, different numbers and sizes of convolution and capsule parameters are set according to the signal characteristics of different working modes, and the parameters of each layer of the network are fully used, overcoming the shortcomings of the prior art which relies too much on manual feature extraction. DETAILED DESCRIPTION

[0014] Figure 1 The flowchart of the IFF signal intelligent classification and identification method of the application.

[0015] Figure 2 The response signal sample simulation diagram under a 240MHz sampling rate of the application is shown in the figure. Figure 2 (a) is a MarkX series response signal waveform diagram, Figure 2 (b) is a mode 4 signal waveform diagram, Figure 2(c) is the waveform chart of the first half of the mode S signal, because the data chain of the mode S signal is too long and does not play any role in the identification of the mode, so only the entire synchronization bit and part of the data bit are taken as the training set sample when the mode is taken as the training set sample. Figure 2 (d) is also the waveform chart of the first half of the mode 5, and the reason is the same as described in the mode S.

[0016] Figure 3 The result chart recognized by the method proposed in the application.

[0017] Figure 4 The flow chart of the staged capsule neural network. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative effort fall within the protection scope of the application.

[0019] It should be noted that all directionality indications (such as up, down, left, right, front, back, etc.) in the embodiments of the application are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directionality indications also change accordingly.

[0020] In the application, unless otherwise explicitly specified and limited, the terms "connection", "fixation" and the like should be understood in a broad sense, for example, "fixation" can be fixed connection, or detachable connection, or integral; "connection" can be mechanical connection, or electrical connection. For a person of ordinary skill in the art, the specific meanings of the above terms in the application can be understood according to the specific circumstances.

[0021] In addition, the technical solutions of the various embodiments of the application can be combined with each other, but it must be based on the fact that a person of ordinary skill in the art can realize it, and when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, and is not within the protection scope required by the application.

[0022] The specific embodiments, technical difficulties and points of the application will be further introduced below in combination with the design examples.

[0023] In combination Figure 1 The IFF signal intelligent classification and identification method described in the application has the following steps:

[0024] Step 1, select the working mode as the response signal, and generate a training sample set.

[0025] Because the operating frequency band of the IFF system is fixed, the simulation assumes that it passes through an additive Gaussian noise channel.

[0026] Step 1-1: Based on the prior information of the response signal, create a signal generation module that can generate random codes of different modes, as well as an additive Gaussian channel simulation module, to prepare for the simulated generation of the signal.

[0027] Step 1-2: The signals randomly generated in different modes are processed by the noise channel simulation module to generate signals with a signal-to-noise ratio of 0 to 10 dB, and these signals are used as training samples.

[0028] Steps 1-3: Similar to Steps 1-2, randomly generate signals for each mode, shift the signals to the left or right by 15-50% of the length, and then convert them into signals with a signal-to-noise ratio of 0-10dB. Use these signals as training samples.

[0029] Steps 1-4: The two types of training samples mentioned above together constitute the training sample set x. i ∈R(i=0,…,n-1).

[0030] Step 2, refer to Figure 4 The hierarchical capsule neural network is trained to generate learning parameters.

[0031] Step 2-1: Input training sample x i For each element in R (i = 0, ..., n-1), normalization and reconstruction are performed. The data processing equations are as follows:

[0032]

[0033]

[0034]

[0035] y i =γx i +β(i=0,…,n-1) (4)

[0036] Where, μ x Representing data x i The mean, Representing data x i The variance, x i Representing data x i Normalized to ε data x i Normalized variance offset, y i Representing data x i The reconstructed value.

[0037] Step 2-2: Reconstruct data yi (i = 0, …, n-1) are convolved and pooled, and the convolution processing equation is as follows:

[0038]

[0039] wherein Z represents a convolution output matrix, W represents a convolution kernel matrix, m represents the number of convolution kernels in the convolution kernel matrix, k represents the length of the convolution kernel vector in the convolution kernel matrix, and l represents the size of the convolution feature map. When a space is inserted between the convolution kernel elements, the length of the convolution kernel vector is extended, and the calculation formula of the output feature map size l1 is as follows:

[0040]

[0041] wherein s represents the convolution step, and d represents the number of spaces inserted between the convolution kernel elements. When the feature map size l2 is calculated, the calculation formula of the ordinary convolution is as follows:

[0042]

[0043] The pooling processing can be regarded as a special convolution. The pooling layer used in this paper is an average pooling layer, which can be regarded as a convolution calculation with a step s = 2 and a convolution kernel column vector The calculation process is consistent with formula (5). After the data is reconstructed through the hollow convolution, the ordinary convolution, and the pooling processing, the data is Z' ∈ R l ' ×m wherein l' represents the number of rows of the data after the convolution and the pooling processing.

[0044] Step 2-3, the data matrix Z' is transformed into a matrix with a specific dimension, and the number of elements in the matrix is unchanged. After the reconstruction processing, the data matrix is represented as Z" ∈ R a×b wherein a x b = l' x m.

[0045] Step 2-4, Z" is input to the low-level capsule , i = 1, 2, …, h, wherein h represents the number of capsules, and b represents the number of neurons (vector length) in each capsule. A conversion matrix W ij ∈ R p×b is applied, p represents the number of neurons of the output capsule, the input is converted into a prediction vector u j|i ∈ R p×1 , and the calculation formula is as follows:

[0046]

[0047] wherein the weight matrix W ij is learned through back propagation. All obtained prediction vectors are weighted and summed, and the summation formula is as follows:

[0048]

[0049] wherein the input is called the high-level capsule, is the weight coefficient, also called the coupling coefficient, which is learned by the dynamic routing algorithm, and is ensured by the nonlinear activation function squashing that short vectors can be compressed to near 0 and long vectors to near 1 in length, and the direction of the vector remains unchanged, and the activation function is in the form of:

[0050]

[0051] For the above learning parameters is updated by the iterative dynamic routing algorithm, the number of iterations is r, and the iteration parameter vector is defined as (initial value is 0), and the iteration specific calculation process is as follows:

[0052] ①Calculate the value of the vector , that is, all routing weights of the capsule i, since the first iteration is initialized to 0, so are all equal in the first round of iteration, that is, 1 / p, p refers to the number of higher-level capsules;

[0053]

[0054] ② The prediction vector is weighted and summed;

[0055] ③The last step vector is passed through the nonlinear function squash;

[0056]

[0057] ④Update the weight . The weight is updated by the dot product of the output of the capsule j j|i and the prediction vector u + the original weight is the new weight value, after updating the weight, the next round of iteration is performed;

[0058]

[0059] ⑤After r iterations, the trained is generated. Thus the parameter training of the neural network is completed.

[0060] Step 3, using the trained neural network to test the measured data, and generating the neural network output vector v j, and a data attention distribution probability is obtained by using an attention mechanism algorithm.

[0061] Step 4, multi-layer perception and data fitting of the data mode are performed by using a full connection layer processing, and finally a mode recognition result is output.

[0062] Embodiment 1

[0063] 1. Simulation conditions:

[0064] The simulation experiment of the present application is carried out on an AMD3700X CPU 3.8GHz, GTX2080Ti, Windows1064bit system, a Tensorflow1.15.0 running platform, and a pyqt5 encapsulated detection system, and the above completes the construction of the inquiry and response simulation signals for training and the simulation experiment of the intelligent classification and identification of the continuous friend-or-foe identification signals.

[0065] 2. Simulation experiment content

[0066] The waveform diagram of the 240MHz sampling rate response friend-or-foe identification signal used in the simulation experiment of the present application is as shown in Figure 2 , wherein Figure 2 (a) is a MarkX series response signal waveform diagram, Figure 2 (b) is a mode 4 signal waveform diagram, Figure 2 (c) is a waveform diagram of the first half of the mode S signal, because the data link of the mode S signal is too long and does not play any role in the identification of the mode, so when the mode is taken as a training set sample, only the entire synchronization bit and part of the data bit are taken. Figure 2 (d) is also a waveform diagram of the first half of the mode 5 signal, and the reason is the same as described in mode S. The specific parameter settings are as follows:

[0067] 1) The first 2000 sample points of the input data are cut and normalized 2000*1;

[0068] 2) 64*12 hole convolution is performed, the expansion rate is 5, the output is (2000-(12-1)*5)*64: 1945*64, and the feature map is extracted;

[0069] 3) 64*12 normal convolution is performed, the step is 4, the output is (1945-(12-1)) / 4*64: 484*64;

[0070] 4) Pooling is performed, the pooling layer Max pooling2 (step)*2 (pooling size), the output is 448 / 4*64: 121*64;

[0071] 5) 256*8 normal convolution is performed, the step is 2, the output is (121-(8-1)) / 2*256: 57*256;

[0072] 6) Perform data reshaping operation, (57*32)*8:1824*8, use the result of this layer as the initial capsule layer, and compress multiple sets of local feature maps into low-level capsules;

[0073] 7) Perform iterative routing algorithm calculations and output the data: 64*10;

[0074] 8) Perform attention mechanism processing to output 64*1;

[0075] 9) Perform fully connected processing to output signal pattern classification and recognition feature results 10*1.

[0076] 3. Analysis of simulation experiment results:

[0077] The simulation results of this invention are as follows: Figure 3 As shown. Figure 3 The right-hand side of the diagram intelligently classifies and identifies continuous IFF (Identification Friend or Foe) signals, detecting pattern signals starting from a specific location. The middle right section sums all detected signals and shows a total of 97 identified pattern signals. Figure 3 The left side shows the waveform of the third identified signal. It can be seen that the waveform is basically the same as the simulated mode S signal, indicating excellent recognition performance. The final recognition accuracy for continuous signals at a 13dB signal-to-noise ratio is 98.67%, and the false alarm rate for the entire test sample set is 1.235%.

[0078] The simulation experiments above demonstrate that the identification method of this invention can complete cognitive identification tasks of different modes for intelligent classification and recognition of friend-or-foe identification signals, and the method is feasible.

Claims

1. A method for intelligent classification and recognition of IFF signals, characterized in that: Based on the principle of fully exploiting the limited temporal features of signals using a segmented capsule neural network, this method first simulates training sets of samples with different signal-to-noise ratios using signals of known modes. Then, the segmented capsule network is used to train the simulated training sets. Finally, a deep learning detection model is built using the trained network, thereby achieving continuous detection and recognition of real signals. The steps are as follows: Step 1: Select the working mode as response signal and generate the training sample set x. i ∈R, i=0,…,n-1; Step 2: Train the hierarchical capsule neural network using the training sample set to generate learning parameters, as follows: Step 2-1: Input the training sample set x i For each element in R, normalization and reconstruction are performed. The data processing equations are as follows: Where, μ x Representing data x i The mean, Representing data x i variance Representing data x i Normalized to ε data x i The normalized variance offset, y i Representing data x i The reconstructed value; Step 2-2: Reconstruct data y i Perform convolution and pooling processing, where i = 0, ..., n-1. The convolution processing equation is as follows: Where Z represents the convolution output matrix, W represents the convolution kernel matrix, m represents the number of convolution kernels in the kernel matrix, k represents the length of the convolution kernel vector in the kernel matrix, and l represents the size of the convolution feature map. During dilated convolution, spaces are inserted between kernel elements to extend the kernel vector length. The formula for calculating the output feature map size l1 is as follows: Where s represents the convolution stride, and d represents the number of spaces inserted between kernel elements; the formula for calculating the feature map size l2 during ordinary convolution is: All pooling layers used are average pooling layers with a stride of s = 2, and the convolution kernel column vectors are... The convolution calculation, j = 1, ..., m-1, is consistent with equation (5); the reconstructed data is processed by dilated convolution, ordinary convolution, and pooling, and the data is Z'∈R. l'×m Where l' represents the number of rows of data after convolution and pooling; Steps 2-3: Reshape the data matrix Z' into a matrix of a specific dimension, while keeping the number of elements in the matrix unchanged. After reconstruction, the data matrix is ​​represented as Z"∈R. a×b Where a×b=l′×m; Steps 2-4, Z" as a low-level capsule The input is i = 1, 2, ..., h, where h represents the number of capsules and b represents the number of neurons in each capsule, i.e., the vector length; a transformation matrix W is applied. ij ∈R p×b p represents the number of neurons in the output capsule, which will be the input... Convert to prediction vector The calculation formula is as follows: Wherein, for the weight matrix W ij Learning is achieved through backpropagation; all the obtained prediction vectors are then weighted and summed using the following formula: in The input known as high-level capsules, These are weighting coefficients, also known as coupling coefficients, which are learned through dynamic routing algorithms. The non-linear activation function squashing is used to ensure that short vectors can be compressed to near 0, and long vectors to near 1, while maintaining the direction of the vectors. The activation function is as follows: Regarding the learning parameters mentioned above It updates the routing algorithm through an iterative dynamic routing algorithm, with r iterations, and the iteration parameter vector is defined as follows. The initial value is 0; Step 3: Use the trained neural network to test the actual data and generate the neural network output vector v. j The attention mechanism algorithm is used to obtain the probability of attention allocation in the data; Step 4: Use fully connected layers to process data patterns for multi-layer perception and data fitting, and finally output the pattern recognition results.

2. The intelligent classification and recognition method for IFF signals according to claim 1, characterized in that, The steps are as follows: In step 1, the working mode is selected as response signal, and a training sample set is generated, as detailed below: Step 1-1: Based on the prior information of the response signal, create a signal generation module that can generate random codes of different modes, as well as an additive Gaussian channel simulation module, to prepare for the simulated generation of the signal. Steps 1-2: The signals randomly generated in different modes are processed by the noise channel simulation module to generate signals with a signal-to-noise ratio of 0 to 10 dB, and these signals are used as training samples. Steps 1-3: Similar to Steps 1-2, randomly generate signals for each mode, shift the signals to the left or right by 15-50% of the length, and then convert them into signals with a signal-to-noise ratio of 0-10dB. Use these signals as training samples. Steps 1-4: The two types of training samples mentioned above together constitute the training sample set x. i ∈R, i=0,…,n-1.

3. The intelligent classification and recognition method for IFF signals according to claim 1, characterized in that, In steps 2-4, the specific calculation process of the iteration is as follows: ① Calculate vectors The value of , i.e., all the routing weights of capsule i, due to the first iteration It was initialized to 0, therefore In the first iteration, they are all equal, i.e., 1 / p, where p refers to the number of higher-level capsules; ② The prediction vectors are then weighted and summed. ③ The vector in the final step is processed by the non-linear function squash; ④ The weights are updated using the output of capsule j. With prediction vector dot product + original weights With the new weight values, update the weights and proceed to the next iteration; ⑤ After r iterations, a trained [system / program] is generated. The parameter training of the neural network is now complete.

Images