A method for separating intermediate frequency signals based on machine learning

Abstract

The present invention discloses a method for separating intermediate frequency signals based on machine learning. The method includes: first, preprocessing the input data and transforming the data form, then inputting it into a feature extractor based on a convolutional neural network model, and then using the extracted features for signal detection; then using the detection result as a mask template to extract the features of the detected signals; finally, inputting the features into a recurrent neural network for signal separation. The method of the present invention directly uses intermediate frequency signals, effectively reducing the information loss caused by manual feature extraction, enhancing the adaptability of the model to the environment using machine learning methods, and at the same time providing an efficient algorithm for simultaneous signal separation and recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of information processing, and particularly to a method and system for separating intermediate frequency signals based on machine learning. Background Art

[0002] In modern warfare, electronic information operation systems and their weapon systems play an important role. Information operation is completed through the collection, transmission, and storage of information, aiming to understand the enemy's intentions and then make correct decisions. Such a process can be represented by the so-called "observe - locate - decide - act" loop model. The warning system is an important part of observation, and this system needs to detect, classify, identify, and locate radiation sources in the radio frequency band of the environment. Among them, target classification is the basis for subsequent target recognition, location, status recognition, risk level assessment, and further situation generation.

[0003] Signal separation is the process of separately extracting the emission signals of different radiation sources in the electromagnetic space to obtain the signal of a single radiation source. Currently, most methods use the received intermediate frequency signal to extract PDW features to form a pulse stream, and then separate the pulse trains of each radiation source of interest from the randomly interleaved pulse stream. Therefore, signal separation is also known as signal deinterleaving. Commonly used PDW is five-dimensional parameters, including Time of Arrival (TOA), Direction of Arrival, Radio Frequency (RF), Pulse Amplitude (PA), and Pulse Width. In the field of signal separation, multiple methods are generally used in engineering practice. Most of these methods use expert knowledge to construct a discrimination system, which can be mainly divided into methods based on inter-pulse features and methods based on intra-pulse features. In recent years, many machine learning-based methods have also emerged, comprehensively using intra-pulse and inter-pulse features.

[0004] However, due to only having five-dimensional parameters, the PDW features cannot well characterize the characteristics of modern radars. With the continuous development of radars, more and more complex modulation methods have emerged. Especially for RF parameters, modulation methods such as linear frequency modulation and non-linear frequency modulation have appeared, resulting in the fact that for a single pulse, RF is not a fixed value. For PA parameters, due to the scanning characteristics of the radar, they usually change as the scan goes from sidelobe to main lobe and then back to sidelobe. And these intra-pulse change characteristics can help the separation method based on intra-pulse features to separate radar pulse signals. In addition, the incorrect measurement caused by noise has always been an important problem that plagues the application of the PDW method in practice. The measurement of PDW is very prone to the problem of leading and trailing edge failures, which makes the radar pulses and noise become similar, and it is very easy to have difficulty balancing false alarms and missed alarms during parameter measurement. This situation leads to a large number of "cracked", "extended" and even "merged" pulse parameter measurement phenomena, resulting in final separation errors. With the widespread use of electronic devices, the scenarios have become more and more complex. The method based on expert knowledge needs to perform real-time parameter adjustment for the environment, which requires a lot of manpower. In summary, there is an urgent need for a new architecture for radar emitter signal separation and a new signal characterization method at present. Summary of the Invention

[0005] In view of the above defects, the present invention designs a signal separation method based on machine learning. By designing a separation algorithm that directly uses intermediate frequency sampling signals, the information loss caused by extracting PDW is avoided, and useful information is retained as much as possible. At the same time, a deep learning-based method is used to solve the problem of non-linear mapping modeling, and the useful information retained is used to extract features that can help separation, improve the separability of data, and avoid error accumulation in the intermediate process. The new framework will generate the machine learning feature expression of the separated single emitter, which can be used for subsequent tasks such as state recognition, model recognition, and parameter recognition, greatly improving the effect of electronic warfare reconnaissance and helping to design better interference methods in the future.

[0006] To achieve the above object, a method for separating intermediate frequency signals based on machine learning adopted by the present invention includes the following steps:

[0007] S1: Intermediate frequency signal preprocessing: First, perform Hilbert transform on the intermediate frequency sampling signal, and then perform short-time discrete-time Fourier transform to generate a time-frequency diagram x(d, t, f);

[0008] S2: Intermediate frequency feature extraction: Use a feature extractor based on a deep convolutional neural network to extract features from the time-frequency diagram x(d, t, f) to obtain compressed features y(c, t, f):

[0009] y(c, t, f) = f(x(d, t, f))

[0010] S3: Signal Detection: Based on the compressed feature y(c, t, f), use a neural network-based classifier det(c, t, f) to classify the compressed feature, determine whether each time-frequency point contains a radar signal, and obtain the judgment result cls(t, f):

[0011] cls(t, f) = det(y(c, t, f))

[0012] S4: Feature Rearrangement: Use the time-frequency point coordinates with signals in the result of cls(t, f) to extract y(c, t, f) and rearrange it according to the time sequence to form a sequence z(c, l) expressed by pulses;

[0013] S5: Sorting and Recognition: Classify the rearranged feature z(c, l) using a sequence neural network model to obtain the category of each feature, and simultaneously complete separation and recognition.

[0014] Optionally, step S1 specifically includes: Sampling the original continuous intermediate-frequency signal g(x), segmenting the sampled signal, where the length of each segment is determined by the subsequent model input size, denoted as L, and the segmented signal is denoted as s(n), n = 1...L, and performing Hilbert transform and short-time discrete-time Fourier transform on the signal.

[0015] Optionally, in step S1:

[0016] The expression of the Hilbert transform is specifically:

[0017]

[0018] The expression of the short-time discrete-time Fourier transform is specifically:

[0019]

[0020] Where, |x| represents the modulo operation, and arg(x) represents the argument operation.

[0021] Optionally, step S2 specifically includes: Performing a normalization operation on x(0, t, f) and x(1, t, f), and using a U-shaped convolutional neural network model to extract features from a time-frequency diagram with a fixed input size, where the time-frequency diagram needs to satisfy the input size T*F, where T = L / G and F = H.

[0022] Optionally, in step S4:

[0023] The expression of the normalization operation is specifically:

[0024]

[0025] The U-shaped convolutional neural network model includes a left tower and a right tower. The left tower performs downsampling operations using convolutions with strides:

[0026]

[0027] The right tower consists of a group of transposed convolutions with strides, which can perform upsampling on the fused features;

[0028] The size of the finally output feature map is the same as that of the original time-frequency map:

[0029]

[0030] Optionally, the U-shaped convolutional neural network model has the same number of upsampling and downsampling convolutions, and the parameters of the convolutions are as follows: Usually, the convolution kernel size K = 3; for the first layer of convolution, the input channel D = 2, and for non-first layer convolutions, D is the output channel C of the previous layer; for padding P = 3; the stride S of the convolution is selected according to specific circumstances to satisfy the gradual reduction of the (t, f) size.

[0031] Optionally, step S3 specifically includes: using a neural network to classify the compressed feature y(c, t, f), generating the probability of the presence or absence of a signal at each time-frequency point using softmax, selecting the category with a higher probability to determine the presence or absence of a signal at that time-frequency point, and then generating a template cls(t, f) of the presence or absence of a signal.

[0032] Optionally, step S4 specifically includes: using the obtained detection result cls(t, f) and the compressed feature y(c, t, f), rearranging these two variables in the memory storage order to form cls(t*f) = cls(l) and y(c, t*f) = y(c, l), and then according to the result of cls(l), extracting y(c , l) along the coordinate axis l and rearranging it to form a sequence z(c, l) of pulse expressions.

[0033] Optionally, step S5 specifically includes: generating a learnable parameter matrix w(c, t, f), using the detection result cls(t, f), rearranging the features of w(c, t, f) as y(c, t, f) to generate v(c, l), and concatenating v(c, l) with z(c, l) to generate a new feature z'(2c, l). Using a sequence neural network to sort the sequence feature z'(2c, l) of the features, the model maps z(c, l) to j(n, l), generates the category probability using the Softmax function, and the finally sorted category i(l) is the category with the highest probability.

[0034] To achieve the above object, the present application also provides a mid-frequency signal separation system based on machine learning. The system includes:

[0035] An intermediate frequency signal preprocessing module, which is used to first perform Hilbert transform on the intermediate frequency sampled signal, and then perform short-time discrete-time Fourier transform to generate a time-frequency diagram x(d, t, f);

[0036] An intermediate frequency feature extraction module, which is used to use a feature extractor based on a deep convolutional neural network to extract features from the time-frequency diagram x(d, t, f) to obtain compressed features y(c, t, f):

[0037] y(c, t, f) = f(x(d, t, f))

[0038] A signal detection module, which is used to, based on the compressed features y(c, t, f), use a neural network-based classifier det(c, t, f) to classify the compressed features, determine whether each time-frequency point contains a radar signal, and obtain a judgment result cls(t, f):

[0039] cls(t, f) = det(y(c, t, f))

[0040] A feature rearrangement module, which is used to use the time-frequency point coordinates with signals in the result of cls(t, f) to extract y(c, t, f) and rearrange them in sequence to form a sequence z(c, l) expressed as pulses;

[0041] A sorting and recognition module, which is used to classify the rearranged features z(c, l) using a sequence neural network model to obtain the category of each feature, and at the same time complete separation and recognition.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides a method for separating intermediate frequency signals based on a neural network model. This method is a new signal separation framework that directly uses intermediate frequency signals for sorting, greatly reducing the information loss caused by the downsampling behavior in the process of generating PDW in traditional methods and further reducing the accuracy. At the same time, by widely using neural network models, the present invention can enhance the adaptability of the model to the environment through the online learning method of the neural network. The present invention can output the radiation signal feature expression of a single specific radiation source after separation, providing more information than the traditional PDW characterization, and can provide a good basis for subsequent state recognition, model recognition, and modulation parameter recognition, increasing the recognition accuracy, enhancing the effectiveness of electronic warfare reconnaissance, and at the same time providing better assistance for electronic warfare interference. Description of the Drawings

[0043] Figure 1 It is a flow framework diagram of the method for separating intermediate frequency signals based on a neural network model of the present invention;

[0044] Figure 2It is a schematic diagram of the rearrangement operation of the present invention.

[0045] The realization of the object, functional features and advantages of the present invention will be further described with reference to the embodiments and the accompanying drawings. Specific embodiments

[0046] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0047] In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0048] As Figure 1 shown, this embodiment discloses a signal separation algorithm based on machine learning. First, the original intermediate frequency signal is preprocessed to generate a time-frequency diagram, and then the processed time-frequency diagram is input into a convolutional neural network for inference to obtain a feature map of the same size. Subsequently, a classifier is used for signal detection to obtain the coordinates where the radar signal is located, the features of the radar signal are obtained using the coordinates, and then a recurrent neural network is used for signal separation.

[0049] This algorithm consists of two major parts, inference and training. Among them, inference is an operation performed after deployment and is the specific process of signal separation. These processes require the use of a large number of parameters to adapt to complex environments, while training is only before deployment and is the process of obtaining algorithm parameters. Specifically, the inference includes the following steps:

[0050] Step 1, Intermediate frequency signal preprocessing: First, perform Hilbert transform on the intermediate frequency sampled signal (which can also be directly performed by dedicated hardware), and then perform short-time discrete-time Fourier transform to generate a time-frequency diagram x(d, t, f).

[0051] Step 2, Intermediate frequency feature extraction: Use a feature extractor based on a deep convolutional neural network to extract features from the time-frequency diagram x(d, t, f) to obtain compressed features y(c, t, f):

[0052] y(c, t, f) = f(x(d, t, f))

[0053] Step 3, Signal detection: Based on the compressed features y(c, t, f), use a neural network-based classifier det(c, t, f) to classify the compressed features, that is, judge whether each time-frequency point contains a radar signal, and obtain a judgment result cls(t, f):

[0054] cls(t, f) = det(y(c, t, f))

[0055] Step 4, Feature Rearrangement: Using the time-frequency point coordinates with signals in the result of cls(t,f), extract y(c,t,f) and rearrange it in time sequence to form a sequence z(c,l) expressed as pulses.

[0056] Step 5, Sorting and Recognition: Classify the rearranged features z(c,l) using a sequence neural network model to obtain the category of each feature, and simultaneously complete separation and recognition. At the same time, in this step, a clustering method can also be used to only perform sorting without recognition.

[0057] Further, the specific method of Step 1 is as follows:

[0058] Step 1.1, Intermediate Frequency Signal Sampling: To enable fast processing, sample the original continuous intermediate frequency signal g(x). A sampling rate of 100MHz is commonly used, but this sampling rate is not limited. The precision of the sampled signal can use 8-bit, 16-bit, and 32-bit integer or floating-point types, which can be adjusted according to requirements. This step needs to be processed through a dedicated ADC device, and subsequent digital processing will be carried out.

[0059] Step 1.2, Sampled Signal Segmentation: Segment the sampled signal. The length of each segment is determined by the input size of the subsequent model, denoted as L. The method for obtaining the value of L will be given in the subsequent steps. The segmented signal is denoted as s(n), where n = 1...L.

[0060] Step 1.3, Hilbert Transform: To facilitate subsequent Fourier transform processing, perform a Hilbert transform on the signal:

[0061]

[0062] Step 1.4, Fourier Transform: Perform a short-time discrete-time Fourier transform on the signal,

[0063]

[0064]

[0065] where |x| represents the modulus operation, and arg(x) represents the argument operation. The obtained result is also called a time-frequency diagram x(d,t,f), where d = 0, 1. In this step, G represents the time interval size of the short-time discrete-time Fourier transform, and H represents the size of the time window. At the beginning and end, zero-padding operations need to be performed on the signal segments, and the number of zeros added is G / 2 to meet the requirements of data operation boundaries in the calculation. After performing the Fourier transform, for the further data, the specific method of Step 2 is as follows:

[0066] Step 2.1, Data Preprocessing: Perform a normalization operation on x(0,t,f) and x(1,t,f):

[0067]

[0068] where % is the modulo operation.

[0069] Step 2.2, Feature Extraction Process: Use a U-shaped convolutional neural network model to extract features from a time-frequency graph with a fixed input size. The time-frequency graph needs to meet the input size T*F, where T = L / G and F = H. Generally, in this method, a fixed T is used to determine L and G, but as long as the above conditions are met, the algorithm can run normally. The U-shaped convolutional neural network consists of two parts, called the left tower and the right tower. Among them, the left tower uses convolution with a stride for downsampling:

[0070]

[0071] The right tower consists of a group of transposed convolutions with a stride, and these transposed convolutions can upsample the fused features. These fused features are formed by concatenating the features output by the left tower of the corresponding layer and the upsampled features generated by the previous transposed convolution. Finally, the size of the output feature map is the same as that of the original time-frequency graph:

[0072]

[0073]

[0074] It should be noted that during the convolution process, due to the length and width of the time-frequency graph, that is, T and F, they may not satisfy (T - 1) % S t == 0, (F - 1) % S f == 0. In this case, additional zero-padding is required during the upsampling process of the transposed convolution. The number of zero-paddings is (T - 1) % S t and (F - 1) % S f , and this zero-padding is reflected in the value ranges of t and f. The position of zero-padding is at the lower right edge. Simply put, after zero-padding by this method, the feature map should have the same size as the time-frequency graph or feature map of the corresponding left tower.

[0075] Specifically, the model has the same number of upsampling and downsampling convolutions, and the parameters of the convolution are as follows: Usually, the kernel size K = 3; for the first layer of convolution, the input channel D = 2, and for non-first layer convolutions, D is the output channel C of the previous layer. For padding P = 3; the stride S of the convolution can be determined according to the specific situation as long as it can meet the requirement of gradually reducing the size of (t, f).

[0076] Further, the specific method of step 3 is as follows:

[0077] Step 3.1, Neural network classification: Use a neural network to classify the compressed feature y(c, t, f). The neural network is a single-layer convolutional neural network with a stride of 1 and a convolutional kernel size of 3. The number of input channels of the network is the number of channels C of the feature, and the number of output channels is 2, meaning 0 / 1, that is, there are two numerical values to judge the probabilities h(o, t, f) of no signal or signal respectively, where o = 0, 1.

[0078] Step 3.2, Probability generation and signal detection: Use softmax to generate the probabilities of signal presence or absence at each time-frequency point:

[0079]

[0080] Then select the category with a higher probability to determine the signal presence or absence at this time-frequency point, that is, 0: no signal; 1: signal, and then generate a template cls(t, f) for signal presence or absence:

[0081]

[0082] Further, the specific method of step 4 is as follows:

[0083] Step 4.1, Feature rearrangement: Use the obtained detection result cls(t, f) and the compressed feature y(c, t, f), rearrange these two variables in the memory storage order, form cls(t*f) = cls(l) and y(c, t*f) = y(c, l), and then according to the result of cls(l), extract y(c, l) with signal (that is, cls(l) = 1) along the coordinate axis l, and perform rearrangement to form a sequence z(c, l) of pulse expressions, as Figure 1 shown.

[0084] Further, the specific method of step 5 is as follows:

[0085] Step 5.1, Position information encoding: Generate a learnable parameter matrix w(c, t, f), use the detection result cls(t, f), and according to the method described in step 4.1, regard w(c, t, f) as y(c, t, f) for feature rearrangement to generate v(c, l), and splice v(c, l) with z(c, l) to generate a new feature z'(2c, l). This step can ensure that both the feature and its position information on the time-frequency diagram are included in the new feature.

[0086] Step 5.2. Neural network separation and recognition integration: Use a sequence neural network to sort the sequence feature z'(2c, l) of feature pairs. The available neural network models include various models for processing sequence problems such as convolutional neural network, RNN, LSTM, and attention model (Attention). Since these sequence models have been widely used, they will not be elaborated here. The model maps z(c, l) to j(n, l), where n is the total number of radiation source categories. Similarly, the Softmax function is also used here to generate category probabilities, and the finally sorted category i(l) is the category with the highest probability. Through the category i(l), the pulse sequence feature z of a single radiation source is output i (c, l).

[0087] The training operation is a method for obtaining the model parameters of neural networks, convolutional neural networks, and sequence neural networks, and is a minimum loss function optimization method based on statistical learning. Among the above steps, steps 2.2, 3.1, and 5.1 require neural network training. The neural network model in step 2.2 is a feature extraction model and needs to provide features in steps 3.1 and 5.1. Therefore, it participates in the training as a pre-model in the latter training. In the training data, there are labels for each time-frequency point on the time-frequency diagram that can be used as supervision information. Therefore, in the model training of step 3.1, a simple cross-entropy loss function is used, and the momentum stochastic optimization method (Adam) is used to optimize it; for the sequence model in step 5.1, the label can be regarded as a feature and the same rearrangement operation as in step 4.1 is performed, and used as the category supervision information of the sequence:

[0088]

[0089] where p(c, f, t) and q(c, f, t) are the calculated probability and the true probability respectively:

[0090]

[0091] g(f, t) is the data category label.

[0092] In some specific examples, the neural network model structure and the size of intermediate data used are as follows in the specific implementation steps:

[0093] Step 1.1. For the input intermediate frequency data, use ADC for sampling, with a sampling rate of 100 MHz and a sampling accuracy of 32-bit floating point;

[0094] Step 1.2. Collect a segment of sampled data with a length of 2000000 (floating point);

[0095] Step 1.3. Perform Hilbert transform with the length unchanged;

[0096] Step 1.4: Perform short-time Fourier transform with an interval of 80 (points) and a time window size of 50 (points) to obtain the time-frequency Figure 2 *50*25000.

[0097] Step 2.1: Data preprocessing, normalize both the amplitude and phase angle of the time-frequency diagram to [0, 1];

[0098] Step 2.2: Feature extraction, use the following convolutional neural network for feature extraction:

[0099] Perform initial convolution on the time-frequency diagram to generate features:

[0100]

[0101] Input the features generated by the initial convolution into the left tower for downsampling:

[0102]

[0103] Input the downsampling result of the left tower into the right tower for upsampling:

[0104]

[0105] Generate features with a dimension of 16*50*25000.

[0106] Step 3.1: Use a classification neural network for signal detection:

[0107]

[0108] Step 3.1: Generate probabilities and then obtain the category, with a size of 1*50*25000.

[0109] Step 4.1: Feature rearrangement, assuming there are 4000 time-frequency points with signals here, then the obtained sequence feature size is 16*4000.

[0110] Step 5.1: Position encoding and rearrangement: Perform the rearrangement operation in Step 4.1 on the learnable parameter matrix (with a size of 16*50*25000), and then concatenate it with the sequence features to obtain a feature size of 32*4000;

[0111] Step 5.2: Signal separation and recognition integration, use the following neural network model to extract features and generate categories. There are two categories in total:

[0112]

[0113] Take the category with the highest probability as the classification result, generate a category of 1*4000, and obtain the corresponding single radiation source features through the category.

[0114] It should be noted that in this text, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or system comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or system. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or system comprising such element.

[0115] The serial numbers of the above embodiments of the present invention are only for description and do not represent the superiority or inferiority of the embodiments.

[0116] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product is stored in a storage medium as described above (such as ROM / RAM, magnetic disk, optical disc) and includes several instructions to enable a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in various embodiments of the present invention.

[0117] The above are only the preferred embodiments of the present invention and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for separating intermediate frequency signals based on machine learning, characterized in that, The method includes the following steps: S1: Intermediate frequency signal preprocessing: First, perform Hilbert transform on the intermediate frequency sampled signal, and then perform short-time discrete-time Fourier transform to generate a time-frequency map x(d, t, f); S2: Intermediate frequency feature extraction: Use a feature extractor based on a deep convolutional neural network to extract features from the time-frequency map x(d, t, f) to obtain compressed features y(c, t, f): y(c, t, f) = f(x(d, t, f)) S3: Signal detection: Based on the compressed feature y(c , t,f), use the neural network-based classifier det(c,t,f) to classify the compressed feature, and determine whether each time-frequency point contains a radar signal to obtain the judgment result cls(t,f): cls(t, f) = det(y(c, t, f)) S4: Feature rearrangement: Use the time-frequency point coordinates with signals in the result of cls(t, f) to extract y(c, t, f) and rearrange it in chronological order to form a sequence z(c, l) of pulse expressions; S5: Sorting and recognition: Classify the rearranged feature z(c , l) using a sequence neural network model to obtain the category of each feature, and at the same time complete separation and recognition.

2. The method for separating intermediate frequency signals based on machine learning according to claim 1, wherein In step S1, specifically: Sample the original continuous intermediate frequency signal g(x), segment the sampled signal, and the length of each segment is determined by the subsequent model input size, denoted as L. The segmented signal is denoted as s(n), n = 1...L. Perform Hilbert transform and short-time discrete-time Fourier transform on the signal.

3. The method for separating intermediate frequency signals based on machine learning according to claim 2, wherein In step S1: The expression of the Hilbert transform is specifically: The expression of the short-time discrete-time Fourier transform is specifically: Where, |x| represents the modulus operation, and arg(x) represents the argument operation.

4. The method for separating intermediate frequency signals based on machine learning according to claim 1, characterized in that Step S2 specifically includes: Perform a normalization operation on x(0, t, f) and x(1, t, f), and use a U-shaped convolutional neural network model to extract features from a time-frequency map with a fixed input size. The time-frequency map needs to satisfy the input size T*F, where T = L / G and F = H.

5. The method for separating intermediate frequency signals based on machine learning according to claim 4, wherein In step S4: The expression of the normalization operation is specifically: The U-shaped convolutional neural network model includes a left tower and a right tower. The left tower performs downsampling operations using convolutions with strides: The right tower consists of a group of transposed convolutions with strides and can perform upsampling on the fused features; The size of the finally output feature map is the same as that of the original time-frequency map:

6. The method for separating intermediate frequency signals based on machine learning according to claim 5, characterized in that The U-shaped convolutional neural network model has the same number of upsampling and downsampling convolutions, and the parameters of the convolution are as follows: Generally, the convolution kernel size K = 3; for the first layer of convolution, the input channel D = 2, and for non-first layer convolutions, D is the output channel C of the previous layer; for padding P = 3; the convolution stride S is selected according to specific circumstances to satisfy gradually reducing the (t , f) size.

7. The method for separating intermediate frequency signals based on machine learning according to claim 1, wherein Step S3 specifically includes: classifying the compressed feature y(c , t,f) using a neural network, generating the probability of the presence or absence of a signal at each time-frequency point using softmax, selecting the class with a higher probability to determine the presence or absence of a signal at that time-frequency point, and then generating a template cls(t , f) for the presence or absence of a signal.

8. The method for separating intermediate frequency signals based on machine learning according to claim 1, wherein, Step S4 specifically includes: Using the obtained detection result cls(t , f) and the compressed feature y(c , t,f), these two variables are rearranged in the memory storage order to form cls(t*f) = cls(l) and y(c , t*f) = y(c , l). Then, according to the result of cls(l), the y(c , l) with cls(l) = 1 is extracted along the axis l and rearranged to form a sequence z(c , l) expressed in pulses.

9. The method for separating intermediate frequency signals based on machine learning according to claim 1, characterized in that Step S5 specifically includes: generating a learnable parameter matrix w(c , t,f), using the detection result cls(t , f), taking w(c , t,f) as y(c , t,f) for feature rearrangement to generate v(c , l), and concatenating v(c , l) with z(c , l) to generate a new feature z'(2c , l), sorting the sequence feature z'(2c , l) of the feature pair using a sequence neural network, mapping z(c , l) to j(n,l) by the model, generating class probabilities using the Softmax function, and finally the sorted class i(l) is the class with the highest probability.

10. A mid-frequency signal separation system based on machine learning, characterized in that, The system includes: The intermediate-frequency signal preprocessing module is used to first perform Hilbert transform on the intermediate-frequency sampled signal, and then perform short-time discrete-time Fourier transform to generate a time-frequency diagram x(d , t,f); Intermediate frequency feature extraction module, which is used to utilize a feature extractor based on a deep convolutional neural network to perform feature extraction on the time-frequency diagram x(d , t,f) to obtain compressed features y(c , t,f): y(c, t, f) = f(x(d, t, f)) A signal detection module, which is used to classify the compressed features by using a neural network-based classifier det(c , t,f) based on the compressed features y(c , t,f), and determine whether each time-frequency point contains a radar signal to obtain a judgment result cls(t , f): cls(t, f) = det(y(c, t, f)) Feature rearrangement module, which is used to utilize the time-frequency point coordinates with signals in the result of cls(t , f), extract y(c , t, f) and rearrange them in time sequence to form a sequence z(c , l); Sorting and recognition module, which is used to classify the rearranged feature z(c , l) using a sequence neural network model to obtain the category of each feature, and at the same time complete separation and recognition.

Images