Integrated signal processing method and device based on u-net and medium

Abstract

The application discloses an integrated signal processing method and device based on U-Net, and a medium, and belongs to the cross field of signal processing and artificial intelligence, and comprises the following steps: in data preprocessing, a digital signal segment is processed into a signal feature map; a U-Net segmentation network is used to segment the signal feature map to give the probability that each signal frame feature point in the signal feature map belongs to different targets, unknown targets or noise; and a vector containing the target category number of all signal frame feature points in the signal feature map is outputted, so that each signal frame feature point belongs to a certain known target, an unknown target or noise. The application can realize integrated processing of signal detection, sorting and identification, simplifies the signal processing process and provides strong signal resolution capability.

Description

Technical Field

[0001] This invention relates to the intersection of signal processing and artificial intelligence, and more specifically, to an integrated signal processing method, device, and medium based on U-Net. Background Technology

[0002] Classical electronic warfare digital signal processing generally consists of several stages, including signal detection, pulse parameter measurement (often called preprocessing or parameter estimation), signal sorting, and target identification. Figure 1 As shown.

[0003] Signal detection refers to the process of extracting useful signals from the background (usually noise). James Tsui, in his classic work "Wideband Digital Receivers" [M]. Beijing: Electronic Industry Press, introduces two of the most common and classic signal detection methods in digital reconnaissance receivers: time-domain and frequency-domain detection. Many other signal detection algorithms exist, such as using correlation detection to improve detection sensitivity and using dynamic thresholds to adapt to constantly changing noise floor conditions.

[0004] The traditional concept of signal sorting refers to separating PDW sequences belonging to different radiation sources that are interleaved in the time domain, based on the differences in signals between targets from different radiation sources, to obtain different PDW subsequences. There are many signal sorting methods, among which the well-known Cumulative Difference Histogram (CDIF) algorithm based on PRI information proposed by Mardia et al. in the paper "New techniques for the deinterleaving of repetitive sequences[J].IEE Proceedings,1989,136(4):149-154." is very representative. In addition to sorting methods based on PRI information, multi-parameter clustering, plane transformation, and other algorithms are also common sorting methods.

[0005] Target identification, as the name suggests, is the process of identifying radiation source targets. Depending on the specific application requirements, the identification results can include various factors such as model, attributes, individual characteristics, and operating mode. In 2003, Talbot et al. proposed a typical individual identification system architecture: first, the signal is received through a radio frequency receiving subsystem; then, after signal processing, the received signal time series undergoes various preprocessing steps such as filtering and denoising, pulse detection, and signal demodulation according to actual needs; next, individual feature extraction is performed to obtain fine-grained features containing individual radiation source information; finally, comparison with a database is conducted, and a classification algorithm is used to determine the specific radiation source emitting the signal, thus achieving individual radiation source identification. This framework is also applicable to general target identification.

[0006] In classic electronic warfare signal processing architectures, detection, sorting, and identification algorithms are highly diverse, with different combinations resulting in varying performance and adapting to different system requirements. However, regardless of these variations, detection, sorting, and identification always function as relatively independent, sequential processing steps within the overall process flow—a fundamental characteristic of classic signal processing architectures. Nevertheless, the signal processing techniques in existing solutions require further improvement. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide an integrated signal processing method, device and medium based on U-Net, which can realize integrated processing of signal detection, sorting and recognition, simplify the signal processing process and provide powerful signal resolution capabilities.

[0008] The objective of this invention is achieved through the following solution:

[0009] An integrated signal processing method based on U-Net includes the following steps:

[0010] In data preprocessing, digital signal segments are... Processed into signal feature maps

[0011] The signal feature map is segmented using a U-Net segmentation network. Give the feature points of each signal frame. The probability of belonging to different targets, unknown targets, or noise.

[0012] The output includes a signal feature map. A vector of target category numbers for all signal frame feature points. Each signal frame feature point is determined to belong to either a known target, an unknown target, or noise.

[0013] Furthermore, the signal feature map It is a supervector with length N and number of channels 3, which are frequency characteristic channel FFC, amplitude characteristic channel AFC and phase characteristic channel PFC.

[0014] Furthermore, data preprocessing also includes the step of generating labels. This is used to construct a dataset and train the U-Net segmentation network using a supervised learning method.

[0015] Furthermore, the U-Net segmentation network is used to segment the signal feature map. Give the feature points of each signal frame. The probability of belonging to different targets, unknown targets, or noise. Specifically, it includes sub-steps:

[0016] Process the signal feature map output after preprocessing using the U-Net segmentation network after training convergence Output the segmentation result thereof, and the segmentation result is composed of N prediction vectors 0 ≤ i < N, the supervector formed:

[0017] Where

[0018]

[0019] And it is a K + 2-dimensional vector, where y i (0) ～ y i (K-1) 、y i (K) 、y i (K) The meanings of are respectively The probabilities of belonging to the 0# to (K - 1)# known targets, unknown targets, and noise; the segmentation result Actually equivalently indicates the probability that each signal frame in the digital signal segment belongs to which target, or unknown target, or noise.

[0020] Furthermore, the output includes the vector of the target category numbers of all signal frame feature points in the signal feature map Specifically, it includes sub-steps: Specifically, it includes sub-steps:

[0021] Convert the probabilities of the target categories to which each signal frame feature point output by the U-Net segmentation network belongs Into the target category to which it belongs through the following method Where Y i is The target number to which the i-th signal frame feature point in belongs:

[0022] Find out the index number of the maximum value in y i (0) ～ y i (K+1) in turn:

[0023]

[0024] where, 0 ≤ i < N.

[0025] Furthermore, the generated label is used to construct a data set and adopt the supervised learning method U-Net segmentation network, including sub-steps:

[0026] The dataset consists of data pairs. For a label corresponding to a digital signal segment of length N, it is a hypervector of length N:

[0027]

[0028] where the label vector is the label corresponding to the i-th signal frame feature point in and indicates the i-th signal frame in the digital signal segment belongs to the target category:

[0029]

[0030] is a one-hot vector of length K + 2 and forms a one-to-one correspondence with the target number ranging from 0 to K + 1;

[0031] When all elements except in the are 0, it means the corresponding target number is j; when and 0 ≤ j < K, it means belongs to the j-th # known target; when and j = K, it means belongs to an unknown target; when and j = K + 1, it means does not belong to any target, or belongs to noise;

[0032] If two or more signals overlap in time domain, the value of the label is uncertain; in this case, there are two processing methods as follows: When calculating the normalized frequency value f of the signal frame feature point in i it may correspond to any overlapping target, then the label vector and the i target number corresponding to f are the same; or, directly take the target number corresponding to the signal with the largest amplitude among the multiple overlapping signals.

[0033] Furthermore, the construction of the dataset includes sub-steps:

[0034] Construct a dataset consisting of data pairs generated by supplementing the actually detected data with data augmentation, where is and Corresponding tags;

[0035] The data pair generated by supplementing the actually detected data with data augmentation Specifically includes the pulse sequence template combined with the known signal And the signal environment parameters of the unknown signal to generate a PDW sequence, and then generate Data pair.

[0036] Further, the pulse sequence template combining the known signal And the signal environment parameters of the unknown signal to generate a PDW sequence, and then generate Data pair, including sub-steps:

[0037] a) Using the intercepted digital signal, if the data coverage of all single known target signals in the target space meets the requirements, but the coverage of the interleaved signal does not meet the requirements, then new data is generated by superimposing the existing data;

[0038] b) Using the pulse sequence template of the target, if the data coverage of some or all single known target signals does not meet the requirements, then the overlapping data of multiple targets is generated using the PDW sequence templates of each known target in the target space, and then The PDW sequence template can completely represent the PDW sequence of the target characteristics, denoted as The supervector composed of M pulse descriptor words 0 ≤ i < M, denoted as:

[0039]

[0040] The PDW sequence in needs to cover all working modes of the target, and the number of pulses M is indefinite, and according to the defined target space definition The pulse descriptor word in

[0041] An integrated signal processing device based on U-Net, including:

[0042] A data preprocessing module for processing digital signal segments Into a signal feature map

[0043] A deep learning module for using the U-Net segmentation network to segment the signal feature map Giving the probability that each signal frame feature point Belongs to different targets, or unknown targets, or noise

[0044] The post-processing module outputs a signal feature map. A vector of target category numbers for all signal frame feature points. Each signal frame feature point is determined to belong to either a known target, an unknown target, or noise.

[0045] A readable storage medium storing a computer program, the computer program being loaded by a processor and executed as described in any of the preceding methods.

[0046] The beneficial effects of this invention include:

[0047] The integrated signal processing model proposed in this embodiment of the invention, through the powerful segmentation capability of U-Net, can achieve integrated processing of signal detection, sorting, and recognition, and can also simultaneously reject unknown signals. Because the joint features of the signal in the time domain, frequency domain, amplitude domain, and phase domain are preserved as much as possible during the preprocessing stage, the method of this embodiment can effectively distinguish multiple interleaved target signals as long as there is a difference in any dimension of the time domain, frequency domain, or modulation domain.

[0048] The technical solution of this invention uses a deep learning algorithm, requiring a dataset to train the U-Net segmentation model. When generating the dataset, if multiple signals overlap in the temporal domain, the training data label is taken from the target number corresponding to the signal with the highest power among the multiple signals, inevitably discarding information from other signals. Therefore, if the signal environment has severe signal overlap, the performance of the method in this invention's technical solution will be affected. Thus, a boundary condition for the method described in this invention's technical solution is that the signal duty cycle cannot be too high. For reference, in the embodiments of this invention, the average signal duty cycle is around 5%, and the model segmentation accuracy reaches 99.85%.

[0049] The technical solution of this invention simplifies the signal processing flow. Instead of performing detection, sorting and recognition processes serially, it uses a U-Net network to synchronously complete signal detection, sorting, recognition and rejection of unknown signals.

[0050] The technical solution of this invention provides a powerful signal resolution capability. Intertwined signals can be separated as long as there is a difference in any one dimension of their frequency, pulse width, repetition rate, or intra-pulse modulation parameters. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a classic signal processing flow;

[0053] Figure 2 This is a block diagram of an integrated signal processing model based on U-Net, according to an embodiment of the present invention.

[0054] Figure 3 The tag value of the time-domain overlapping signal in this embodiment of the invention;

[0055] Figure 4 The training curves of the U-Net segmentation network in this embodiment of the invention are shown.

[0056] Figure 5 This is the processing result of target mode 2 (only the intrapulse modulation is different) in embodiment 2 of the present invention;

[0057] Figure 6 This is the processing result for target 0#+1# (only the pulse width is different) in an embodiment of the present invention;

[0058] Figure 7 This is the processing result of target mode 1 (only the frequency is different) 1#+3# in the embodiment of the present invention;

[0059] Figure 8 This is the processing result of target mode 1 (0#2#) and target mode 1 (4#) frequency point 2 (only PRI is different) in this embodiment of the invention;

[0060] Figure 9 The processing result of target 0#+2#+5# (with unknown signal added) in this embodiment of the invention;

[0061] Figure 10 The segmentation error in this embodiment of the invention is due to a poor signal-to-noise ratio.

[0062] Figure 11 The segmentation error in this embodiment of the invention is that an unknown target pulse with parameters close to those of a known target was randomly generated.

[0063] Figure 12 The segmentation error in this embodiment of the invention is signal time-domain overlap. Detailed Implementation

[0064] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0065] The technical solution of this invention involves using a U-Net network to achieve signal detection, sorting, and identification. Given the current state of the technology, this invention proposes an end-to-end integrated signal processing solution based on a U-Net network. Its purpose is to fully utilize the powerful feature representation and processing capabilities of deep learning technology to provide a new approach and method for electronic warfare signal processing. Specific improvements include: a) simplifying the signal processing flow, eliminating the sequential execution of detection, sorting, and identification processes, and instead using the U-Net network to synchronously complete signal detection, sorting, identification, and rejection of unknown signals; b) providing powerful signal resolution capabilities, allowing signals to be separated even when intertwined, provided there is a difference in any dimension of their frequency, pulse width, repetition rate, or intra-pulse modulation parameters.

[0066] In further inventive conception, the technical solutions of the embodiments of the present invention involve the following concept definitions:

[0067] Target space: A complete set of radiation sources used to describe the problem scenario. In the technical solution of this invention, the radiation sources in the target space include noise, known targets, and unknown targets. Noise is self-explanatory; known targets refer to targets whose signal characteristics have been fully understood, which can be achieved through various means. For example, the digital signal of the target has been fully intercepted, or the parameters of the target signal have been mastered, or the PDW sequence of the target signal has been mastered. It must be emphasized that the mastery of the target signal parameters or PDW sequence must be sufficient to reconstruct the target signal waveform, otherwise it will affect the training effect; all other targets besides known targets are classified as unknown targets. In the technical solution of this invention, unknown targets are simulated by randomly generating pulse parameters (including frequency, pulse width, modulation, etc.) within a preset range.

[0068] The technical solution of the present invention uses tape The character '+' represents a hypervector, while the character '→' represents a regular vector. The elements that make up a hypervector are vectors, while the elements that make up a regular vector are scalars.

[0069] In this invention, given a target space consisting of K known targets, unknown targets, and noise, the integrated signal processing model based on U-Net (hereinafter referred to as the processing model) is described below. Figure 2 As shown. The input to the processing model is a digital signal segment with a time domain length of sigLen. It contains N signal frames of length sigFrameLen. in Clearly, sigLen = N × sigFrameLen. (Model processing) Output Each signal frame has its own target number. The possible values ​​of the target number are 0 to K+1: if the target number is 0 to K-1, it means that the corresponding signal frame belongs to the known target number 0# to (K-1)#. The value K indicates that the signal frame belongs to an unknown target. The value K+1 indicates that the signal frame is a piece of noise.

[0070] If the processing model can achieve the above-mentioned expected goals, it can effectively distinguish whether each signal frame in a signal segment is noise or a target signal, indicating that it has signal detection capability; it can effectively distinguish whether each signal frame is a known target or an unknown target, indicating that it has the ability to reject unknown targets; and it can effectively distinguish which known target each signal frame is, indicating that it has sorting and recognition capabilities. This shows that the processing model is a signal processing model with signal detection + sorting + recognition + unknown target rejection capabilities, therefore, this invention names it the integrated signal processing model.

[0071] like Figure 2 As shown, the processing model includes three steps: data preprocessing, U-Net segmentation model, and post-processing. The first step is preprocessing, which uses the standard FAP representation method to segment the signal fragments. Processed into standard signal feature maps (StandardSignal Feature Map), The first step characterizes the overall frequency, amplitude, phase, and time-domain features of the signal segment; the second step uses the U-Net segmentation network for segmentation. Give the feature points of each signal frame. The probability that a (signal frame feature point) belongs to different targets, unknown targets, or noise. This is a crucial step in simultaneously detecting, sorting, and recognizing signals; the third step involves simple post-processing to output the target category vector. It is a collection The vector of target category numbers for all signal frame feature points in the signal segment. Each signal frame feature point belongs to a known target, an unknown target, or noise—which is equivalent to classifying all signal frames in the entire signal segment.

[0072] To obtain a high-performance U-Net segmentation network, it is necessary to construct a network consisting of... A dataset composed of data pairs is used to train the U-Net segmentation network with supervised learning methods, where is the corresponding label. There are many methods for generating data pairs, which can be generated based on the actually received data supplemented with data augmentation. The data augmentation method is shown in the embodiments of the present invention: combining the pulse sequence template of the known signal and the signal environment parameters of the unknown signal to generate the PDW sequence, and then generating data pairs based on the PDW sequence using the simulation method.

[0073] In a further embodiment, in the data preprocessing link, the main function of the data preprocessing module is to process the signal segment into a standard signal feature map The method for generating the standard signal feature map is to sequentially measure and calculate the normalized frequency f (0 ≤ i < N) of the signal frame i , the normalized amplitude A i and the normalized phase or phase difference (referring to the phase difference between two adjacent signal frames) p i , and use the signal frame feature points (or called signal frame feature vectors) to characterize the features of the signal frame , thereby obtaining the feature characterization of the signal segment From a formal perspective it is a two-dimensional array, but it is more appropriate to regard it as a supervector with a length of N and a channel number CN (Channel Number) = 3. Its three channels are the frequency feature channel FFC (Frequency Feature Channel), the amplitude feature channel AFC (Amplitude Feature Channel), and the phase (or phase difference) feature channel PFC (Phase Feature Channel). If it is used for dataset construction, the data preprocessing module also needs to complete the function of generating the labels

[0074] used for training the U-Net segmentation network.

[0075] In a further embodiment, in the U-Net segmentation network, the role of the U-Net segmentation network in the technical solution of the embodiments of the present invention is to process the [[ID=ST44]] output of the preprocessing and output its segmentation result. The segmentation result is a supervector composed of N prediction vectors (0 ≤ i < N):

[0076] Where

[0077]

[0078] And it is a K+2 dimensional vector, where y i (0) ～ y i (K-1) y i (K) y i (K) The meanings are respectively The probability of belonging to known targets (0#~(K-1)#), unknown targets, and noise. The segmentation result is clearly... In effect, it indicates a segment of digital signal. The probability that each signal frame belongs to a target, an unknown target, or noise.

[0079] Key design considerations for the U-Net segmentation network in this embodiment of the invention include:

[0080] 1) The input layer shape is (N, 3);

[0081] 2) All convolutional layers use 1D convolution operations;

[0082] 3) Number of channels in the output layer = K + 2;

[0083] 4) The activation function of the output layer is usually the Softmax function.

[0084] Other design parameters of the U-Net segmentation network, such as the upsampling / downsampling ratio, the number of symmetrical convolutional layers, and the number of filters, are not strictly required and need to be optimized according to the specific needs of the problem.

[0085] The U-Net segmentation network needs to converge after training to function effectively. The goal of training is to minimize the difference between the labels and the output of the U-Net segmentation network. The related principles and implementations are common in the field of deep learning research and are not within the scope of this invention. Existing common deep learning development frameworks (such as Python + PyTorch, Python + TensorFlow, Python + Keras—the frameworks used in implementation examples, Caffee, Paddle, etc.) can all provide U-Net network training support, which will not be elaborated further.

[0086] In a further implementation, during the post-processing stage, the U-Net segmentation network outputs the probability of the target category to which each signal frame feature point belongs. Post-processing is required to convert this probability into the corresponding target category. Where Y i yes The target number to which the i-th signal frame feature point in belongs.

[0087] The output of the U-Net segmentation network Calculate the target class vector The method is to sequentially find the y i (0) ～ y i (K+1) (0 ≤ i < N), the index number of the maximum value:

[0088]

[0089] According to the requirements in the specific scenario, more complex post-processing strategies can also be adopted. For example, the target class of sporadic signal frame feature points that suddenly appear can be adjusted according to the classes of nearby signal frame feature points. These works are not the focus of this invention and will not be elaborated.

[0090] In a further implementation scheme, on the dataset, to obtain a high-performance U-Net segmentation network, it is necessary to construct a dataset to offline train the U-Net segmentation network. The dataset consists of a large number of Data pairs.

[0091] The method of generating with signal segments has been described in the above technical solution. For the Corresponding to it is a hypervector of length N

[0092]

[0093] Among them, the label vector is the same as The i-th signal frame feature point in The corresponding label indicates the target class to which it belongs. Obviously, this is also equivalent to indicating the i-th signal frame in the digital signal segment Belonging to the target class:

[0094]

[0095] is a one-hot vector of length K + 2, forming a one-to-one correspondence with the target number (taking values 0 to K + 1). When in the elements composing All other elements are 0, indicating The corresponding target number is j: When And 0 ≤ j < K, it indicates (Or the i-th signal frame) belongs to the j-th known target; when ​​And when j = K, it means (or the i-th signal frame) belongs to an unknown target; when And when j = K+1, it means (or the i-th signal frame) does not belong to any target, or in other words, it belongs to a small segment of noise.

[0096] If two or more signals overlap in the time domain, the value of the tag becomes uncertain. In this case, the calculation... The normalized frequency value f of the signal frame feature points. i It may correspond to any of the overlapping targets, in principle, the label vector. Should and f i The corresponding target number is consistent (and naturally also consistent with A). i and P i (Corresponding). As a concise method, the technical solution of this invention directly takes the target number corresponding to the one with the largest amplitude among multiple overlapping signals, which can guarantee the correctness of the label in most cases. For example Figure 3 As shown, a certain signal is composed of the superposition of target signal #0 and target signal #1. The pulse amplitude of target signal #0 is larger, so the tag values ​​corresponding to its 8 signal frames should be 0,0,0,0,0,0,0,1 respectively.

[0097] In a further implementation scheme, regarding data augmentation, the dataset consists of a large number of... The data pairs are constructed as described above, and are generated accordingly. The data pair requires the use of a large number of digital signal segments. It can be generated based on the results of detection of radiation source targets. (Based on the generation...) Digital signal segments used at the time Number of targets It comes in several types. If A signal originating from a single target is called For a single objective Data pairs; if It is composed of the superposition of two different target signals, hence it is called a 2-target signal. Data pairs; 3 objectives Data pairs or more targets The same applies to the data pairs.

[0098] If the data coverage of the intercepted signal is sufficient, digital signal segments It can be directly and randomly extracted from the already captured digital signal data. If the data coverage of the captured signal is insufficient, data augmentation methods are needed to generate new digital signal segments, and then the data can be calculated. Data pairs are used to supplement the data set.

[0099] a) Using the intercepted digital signals

[0100] If the data coverage of all individual known target signals in the target space is sufficient, but the coverage of the interleaved signals is insufficient, new data can be generated by superimposing the existing data.

[0101] b) Using the pulse sequence template of the target [[ID=??]] [[ID=??]]

[0102] If the data coverage of some or all individual known target signals is insufficient, the PDW sequence templates of each known target in the target space can be used to generate overlapping data of multiple targets, and then The so-called PDW sequence template refers to the PDW sequence that can completely represent the target characteristics, denoted as It is a supervector composed of several (the number is denoted as M) pulse description words (0 ≤ i < M), denoted as:

[0103]

[0104] The PDW sequences in need to cover all working modes of the target. The number of pulses M is indefinite. Generally speaking, the more complex the target is, the more pulses are required. Obtaining the PDW sequence template usually requires long-term interception and analysis processing of the target. Specifically, any effective classical or intelligent signal processing method can be used. The relevant content is not within the scope of this invention and will not be elaborated.

[0105] The definition of the pulse description word [[ID=??]] is related to whether it can fully represent the target signal characteristics and the specific signal pattern of the target. For example, for the pulse description word defined by the classic five parameters (frequency, pulse width, arrival time, amplitude, and azimuth), if the target signal only contains conventional signals, the PDW sequence template can restore the target signal waveform. In this case, the PDW sequence template can represent the target signal characteristics; if the target signal contains in-pulse modulation signals, the PDW sequence cannot restore the target signal waveform. In this case, the PDW sequence template cannot represent the target signal characteristics, and more complex pulse description words need to be defined.

[0106] Since the target spaces corresponding to specific problems are various and difficult to uniformly define, the specific signal patterns involved are also difficult to uniformly define. Therefore in is also difficult to uniformly define. In short, the target space should be defined according to specific problems, and then Related examples are shown in the embodiments of this invention. Note: There are some question marks in the translation where the original text seems to have some incomplete or unclear parts in the provided tags. You may need to double-check the original text for those specific tags and correct the translation accordingly if necessary.

[0107] Assume that PDW sequence templates of K known targets in the target space have been generated using the target signal parameters, and they are respectively Next, take the case of two targets as an example to illustrate the data enhancement method, and the method of generating more target data pairs can be deduced by analogy. There are two possibilities for the case of noiseless + targets. Either both targets are known targets, or one known target plus one unknown target. We will illustrate with the latter case.

[0108] 1) Select the i-th (0 ≤ i < K) known target, and assume that the range of the arrival time in i is [TOAStart <00 / / 00041>, TOAEnd i -TOAStart i ] > sigLen;

[0109] 2) Randomly intercept a segment with a duration of sigLen from , and let its starting time be T0. Intercept the pulse sequence segment contained in this time period from , and record the number of pulses in this segment as M i . Subtract T0 from the TOA of all pulses therein, and then randomly select the signal power within the preset power range to calculate the signal amplitude, and set the pulse amplitude in the pulse sequence segment to this signal amplitude value to obtain

[0110] [[ID=[]34]]3) Generate the pulse sequence segment of the unknown target M k is the number of pulses randomly set (but not greater than a certain preset maximum value), and the parameters of each pulse are randomly set within the preset range respectively. The pulse parameter range may include (but is not limited to, depending on the specific problem) the arrival time range, the frequency range, the pulse width range, the intra-pulse modulation type and its modulation parameter range, etc.;

[0111] 4) Use [[ID=[]43]]the parameters of each pulse in to generate the signal segment of the known target is the number of pulses randomly set (but not greater than a certain preset maximum value), and the parameters of each pulse are randomly set within the preset range respectively. The pulse parameter range may include (but is not limited to, depending on the specific problem) the arrival time range, the frequency range, the pulse width range, the intra-pulse modulation type and its modulation parameter range, etc.;

[0111] 4) Use the parameters of each pulse in to generate the signal segment of the known target Generate the signal segment of the unknown target

[0112] 5) Superimpose and Then superimpose the baseband noise, and then limit the amplitude according to the preset maximum amplitude, and add quantization noise to obtain the signal segment

[0113] 6) Use the method described in the above technical solution. generate

[0114] 7) Put and The signals are mixed together, sorted by amplitude from smallest to largest, and traversed from front to back, with each pulse's corresponding signal frame feature point labeled to generate the signal. When labeling, if the time domains of the following pulse and the preceding pulse overlap, the label of the following pulse will cover the label of the preceding pulse. This is done to ensure that, when signals overlap, the label corresponding to the feature point of the signal frame matches the target number with the largest amplitude.

[0115] 8) Repeat steps 1) through 6) above to generate more. Continue until the dataset meets the training requirements.

[0116] The integrated signal processing model proposed in this invention, through the powerful segmentation capabilities of U-Net, can achieve integrated processing of signal detection, sorting, and recognition, and can also simultaneously reject unknown signals. Because the preprocessing stage preserves the joint features of the signal in the time, frequency, amplitude, and phase domains as much as possible, the method of this invention can effectively distinguish multiple interleaved target signals as long as there is a difference in any dimension of the time, frequency, or modulation domain.

[0117] This invention employs a deep learning algorithm, requiring a dataset to train the U-Net segmentation model. When generating the dataset, if multiple signals overlap in the temporal domain, the training data labels are taken from the target number corresponding to the signal with the highest power among the multiple signals, inevitably discarding information from other signals. Therefore, if the signal environment exhibits severe signal overlap, the performance of the method in this invention will be affected. Thus, a boundary condition for the method described in this invention is that the signal duty cycle cannot be too high. For reference, in the embodiments of this invention, the average signal duty cycle is around 5%, and the model segmentation accuracy reaches 99.85%.

[0118] The present invention uses a method of randomly generating pulse parameters to simulate unknown target signals and classifies all unknown target signals into one category. Therefore, the method of the present invention can only generally achieve the rejection of unknown targets and cannot distinguish further information or patterns of unknown targets.

[0119] The embodiments of the present invention are not limited to the examples described in this section. The signal parameter settings and system parameter settings in the example scenarios are shown in Table 1, and the parameter settings of the known target signal are shown in Table 2.

[0120] Table 1. Signal and system parameter settings for the implementation example.

[0121]

[0122]

[0123] Table 2 Known target signal parameters

[0124]

[0125] In Table 2, targets #0 and #1 have identical frequencies and repetition rates, with pulse widths differing by only 0.5 μs. The operating modes 1 of targets #0 and #2, and the second group frequency conversion point of the first operating mode of target #4, have identical frequencies and pulse widths. The repetition rates and pulse widths of the first operating mode of targets #1 and #3 are identical. These settings are used to verify the integrated signal processing model's ability to distinguish signals with differences in frequency or time domain. The second operating mode of targets #2 and #4 are set to have identical center frequencies, pulse widths, and repetition rates, with intra-pulse modulation of linear frequency modulation and 13-bit Barker code modulation, respectively. This setting aims to verify the integrated signal processing model's ability to distinguish signals with differences only in the modulation domain.

[0126] Based on the U-Net segmentation network configuration in the above embodiments, the U-Net segmentation network is designed, and its specific structure is shown in Table 3.

[0127] Table 3 U-Net Segmentation Network Structure

[0128]

[0129]

[0130] After constructing the U-Net segmentation network, the next step is to generate the dataset according to the method described in the above embodiment. To save storage space, Keras's built-in method is used to generate training and validation datasets for each training batch.

[0131] The main training parameter settings for training the U-Net segmentation network are as follows:

[0132] 1) A maximum of 300 epochs can be trained;

[0133] 2) Train 20 batches per epoch;

[0134] 3) Each batch of training data contains 56 samples. Among them, single target There are 6 data pairs: one for each of the 5 known targets (0# to 5#) and one for each of the unknown targets; 30 data pairs with 2 targets: two data pairs are generated for every combination of 2 targets; and 20 data pairs with 3 targets: one data pair is generated for every combination of 3 targets.

[0135] 4) Use the Adam optimization algorithm;

[0136] 5) The loss function is the multi-class cross-entropy loss function;

[0137] 6) The initial learning rate is set to 0.001;

[0138] 7) Validation loss (val_loss): If there is no improvement after 16 consecutive epochs, the learning rate is halved; if there is no improvement after 40 consecutive epochs, training is terminated.

[0139] Final training results are as follows Figure 4 As shown, after 300 epochs of training, the peak segmentation accuracy on the validation set reached 99.85%, and the corresponding U-Net network parameters were saved as the final U-Net segmentation network parameters. It should be noted that the 99.85% represents the accuracy of the U-Net segmentation network in correctly segmenting known targets, unknown targets, and noise. A large portion of this accuracy is contributed by correctly segmenting the signal frame feature points of noise and signals; the accuracy in segmenting known target signals is not this high.

[0140] Some test results can be found in Figures 5-9 As shown, each figure, from top to bottom, displays the segmentation results of the signal feature map, the label value (which is the theoretical target category of the signal frame feature points), the frequency feature channel, the amplitude feature channel, and the phase difference feature channel. The symbol "★" indicates the label value of an incorrect signal processing result, which is the theoretical target number.

[0141] observe Figures 5-9 :

[0142] a) Figures 5-9 This demonstrates that the U-Net integrated signal processing model can effectively distinguish signal frame feature points of noise or targets, regardless of whether the target is known or unknown, indicating that it has signal detection capabilities.

[0143] b) Figure 9 This demonstrates that the U-Net integrated signal processing model can distinguish between known and unknown targets, indicating its ability to reject unknown targets.

[0144] c) Figures 5-8This indicates that, regardless of frequency, pulse width, repetition rate, or intra-pulse modulation, as long as multiple targets are separable in any dimension, the U-Net integrated signal processing model can distinguish them well, demonstrating its ability to sort and identify known target signals.

[0145] According to the test results, the U-Net segmentation network missegmented approximately 0.18% of the signal frame feature points, mainly due to several reasons. Similar reasons could also lead to errors or deviations in the processing results of classic signal processing methods.

[0146] a) Error segmentation points may occur at the pulse edge.

[0147] Figure 5 and Figure 9 This phenomenon can be observed in various signals. This is because at the pulse edge, the signal in the signal frame may only account for a small proportion, which may cause the measurement of the corresponding signal frame feature points to deviate from the true value, further leading to deviations in the output of the integrated signal processing model.

[0148] b) Poor signal-to-noise ratio

[0149] Figure 10 In the test signal generation, there are three pulses, originating from targets 0#, 1#, and 3# respectively. The signal frame feature points corresponding to the third pulse (within the dashed rectangle) are all incorrectly segmented as noise. As can be seen from the amplitude feature channel (AFC), the amplitude of this pulse is completely indistinguishable from the background noise. The frequency feature channel (FFC) also shows that its frequency characteristics are chaotic and cannot match the signal parameters of target 3#.

[0150] c) Random signal parameters are close to the known target

[0151] Figure 11 In the image, the area within the dashed box represents an unknown target. As can be seen from the frequency characteristic channel FFC, its frequency is very close to that of targets #0 and #2 (251MHz). Therefore, some signal frame feature points are misidentified as target #0, and some as target #2. If the signal parameters of known targets are avoided when generating signals for unknown targets, the probability of this problem occurring is expected to be very low; however, for simplicity, this implementation example does not employ such a method.

[0152] d) Pulses overlap in the time domain

[0153] Figure 12In the image, two pulses from targets #2 and #0 overlap, causing most of the signal frame feature points within the dashed box to be misclassified as target #2. Note that targets #0 and #2 have identical pulse widths and frequencies; therefore, classifying the overlapping portion as either target would be correct. Because the integrated signal processing model architecture cannot simultaneously contain information about two targets in the signal feature map, nor does it support output signal frame feature points belonging to both target A and target B, this misclassification, while inconsistent with the label values, is an inevitable consequence of the algorithm architecture and does not incur any additional loss.

[0154] It should be noted that, within the scope of protection defined in the claims of this invention, the following embodiments can be combined and / or extended or replaced in any logical manner from the above specific embodiments, such as the disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0155] Example 1

[0156] An integrated signal processing method based on U-Net includes the following steps:

[0157] In data preprocessing, digital signal segments are... Processed into signal feature maps

[0158] The signal feature map is segmented using a U-Net segmentation network. Give the feature points of each signal frame. The probability of belonging to different targets, unknown targets, or noise.

[0159] The output includes a signal feature map. A vector of target category numbers for all signal frame feature points. Each signal frame feature point is determined to belong to either a known target, an unknown target, or noise.

[0160] Example 2

[0161] Based on Example 1, the signal feature map It is a supervector with length N and number of channels 3, whose three channels are frequency characteristic channel FFC, amplitude characteristic channel AFC and phase characteristic channel PFC.

[0162] Example 3

[0163] Based on Example 1, the data preprocessing also includes the step of generating labels. This is used to construct a dataset and train the U-Net segmentation network using a supervised learning method.

[0164] Example 4

[0165] Based on Example 1, use the U-Net segmentation network to segment the signal feature map Give the feature points of each signal frame The probabilities of belonging to different targets, or unknown targets, or noise Specifically, it includes sub-steps:

[0166] Use the U-Net segmentation network after training convergence to process the signal feature map output after preprocessing Output the segmentation result, and the segmentation result is composed of N prediction vectors 0 ≤ i < N, the supervector formed:

[0167] Where

[0168]

[0169] is a K + 2-dimensional vector, where y i (0) ～ y i (K-1) y i (K) y i (K) The meanings of are The probabilities of belonging to known targets from 0# to (K - 1)#, unknown targets, and noise; the segmentation result Actually equivalently indicates the probabilities of which target, or unknown target, or noise each signal frame in the digital signal segment belongs to respectively.

[0170] Example 5

[0171] Based on Example 1, the output contains the vector of the target category numbers of all signal frame feature points in the signal feature map Specifically, it includes sub-steps: Specifically, it includes sub-steps:

[0172] Convert the probabilities of the target categories to which each signal frame feature point output by the U-Net segmentation network belongs into the target category to which it belongs through the following method Where Y i is the target number to which the i-th signal frame feature point in belongs:

[0173] Find out y i (0) ～ y i(K+1) Index number of the maximum value in:

[0174]

[0175] where 0 ≤ i < N.

[0176] Example 6

[0177] Based on Example 3, the generated label is used to construct a data set and a supervised learning method U-Net segmentation network, including sub-steps:

[0178] The data set consists of data pairs. For a signal of length N [[ID=​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ and f i corresponds to the same target number; or, directly take the target number corresponding to the one with the largest amplitude among multiple overlapping signals.

[0185] Example 7

[0186] Based on Example 3, the construction of the data set includes sub-steps:

[0187] Construct a data set composed of data pairs generated by supplementing the actually detected data with data augmentation, where is the label corresponding to and ;

[0188] In the data pairs generated by supplementing the actually detected data with data augmentation The data pairs specifically include combining the pulse sequence template of the known signal and the signal environment parameters of the unknown signal to generate a PDW sequence, and then generating from the PDW sequence by simulation methods data pairs.

[0189] Example 8

[0190] Based on Example 7, the combination of the pulse sequence template of the known signal and the signal environment parameters of the unknown signal to generate a PDW sequence, and then generating from the PDW sequence by simulation methods data pairs includes sub-steps:

[0191] a) Using the intercepted digital signal, if the data coverage of all single known target signals in the target space meets the requirements, but the coverage of the interleaved signals does not meet the requirements, then use the existing data to superimpose and generate new data;

[0192] b) Using the pulse sequence template of the target, if the data coverage of some or all single known target signals does not meet the requirements, then generate multi-target overlapping data using the PDW sequence templates of each known target in the target space, and then generate The PDW sequence template that can completely represent the target characteristics, denoted as a supervector composed of M pulse descriptor words , 0 ≤ i < M, denoted as:

[0193]

[0194] The PDW sequence in needs to cover all working modes of the target, where the number of pulses M is indefinite and is defined according to the defined target space the pulse descriptor word in

[0195] Example 9

[0196] An integrated signal processing device based on U-Net, comprising:

[0197] The data preprocessing module is used to process digital signal segments during data preprocessing. Processed into signal feature maps

[0198] A deep learning module is used to segment the signal feature map using a U-Net segmentation network. Give the feature points of each signal frame. The probability of belonging to different targets, unknown targets, or noise.

[0199] The post-processing module outputs a signal feature map. A vector of target category numbers for all signal frame feature points. Each signal frame feature point is determined to belong to either a known target, an unknown target, or noise.

[0200] Example 10

[0201] A readable storage medium storing a computer program, the computer program being loaded by a processor and executed as described in any one of Examples 1 to 8.

[0202] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0203] According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0204] In another aspect, embodiments of the present invention also provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

[0205] All parts not covered in this invention are the same as or can be implemented using existing technologies.

[0206] The above technical solution is only one embodiment of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the above specific embodiments of the present invention. Therefore, the methods described above are only preferred and are not restrictive.

[0207] In addition to the examples above, other embodiments may be obtained by those skilled in the art based on the above disclosure or by making modifications using knowledge or technology in related fields. The features of each embodiment may be interchanged or replaced. Modifications and changes made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An integrated signal processing method based on U-Net, characterized in that, Includes the following steps: In data preprocessing, digital signal segments are... Processed into signal feature maps ; The signal feature map is segmented using a U-Net segmentation network. Give the feature points of each signal frame. The probability of belonging to different targets, unknown targets, or noise. ; The output includes a signal feature map. A vector of target category numbers for all signal frame feature points. Each signal frame feature point is determined to belong to either a known target, an unknown target, or noise. The signal feature map It is a supervector with length N and number of channels 3, whose three channels are frequency characteristic channel FFC, amplitude characteristic channel AFC and phase characteristic channel PFC.

2. The integrated signal processing method based on U-Net according to claim 1, characterized in that, Data preprocessing also includes the step of generating labels. This is used to construct a dataset and train the U-Net segmentation network using a supervised learning method.

3. The integrated signal processing method based on U-Net according to claim 1, characterized in that, The U-Net segmentation network is used to segment the signal feature map. Give the feature points of each signal frame. The probability of belonging to different targets, unknown targets, or noise. Specifically, it includes the following sub-steps: Process the signal feature map output after preprocessing using the U-Net segmentation network after training convergence , and output the segmentation result for it. The segmentation result consists of N prediction vectors , where 0 ≤ i < N, to form a supervector: =[ , …, ],in =[and i (0) , …,and i (K) , and i (K+1) ] T And it is a K+2 dimensional vector, where y i (0) ～ y i (K-1) y i (K) y i (K+1) The meanings are respectively The probability of belonging to known targets, unknown targets, and noise from 0 to (K-1); Segmentation results In effect, it is equivalent to indicating a segment of digital signal. The probability that each signal frame belongs to a target, an unknown target, or noise.

4. The integrated signal processing method based on U-Net according to claim 1, characterized in that, The output includes a signal feature map. A vector of target category numbers for all signal frame feature points. Specifically, it includes the following sub-steps: The probability of the target category to which each feature point in the signal frame output by the U-Net segmentation network belongs. Convert to the target category as follows: =[Y0,Y1,…,Y N-1 ], where Y i yes The target number to which the feature point of the i-th signal frame belongs: Find y in sequence i (0) ～ y i (K+1) Index number of the maximum value: Where 0 ≤ i < N.

5. The integrated signal processing method based on U-Net according to claim 2, characterized in that, The generated label The process of constructing a dataset and using the supervised learning method U-Net segmentation network includes the following sub-steps: The dataset consists of ( , The data consists of pairs of length N. The corresponding tags It is a supervector of length N: =[ , ,…, ] Among them, the label vector Is and Feature point of the i-th signal frame The corresponding labels indicate the digital signal segments. The i-th signal frame Target category: =[ , ,…, ] T ,0≤i<N, It is a uniquely hot 1-hot vector of length K+2, and forms a one-to-one correspondence with the target number whose value ranges from 0 to K+1; When the components of are such that = 1 and the other elements are 0, it represents that the corresponding target number is j; when = 1 and 0 ≤ j < K, it represents that it belongs to the j-th known target; when = 1 and j = K, it represents that it belongs to an unknown target; when = 1 and j = K + 1, it represents that it does not belong to any target, or rather belongs to noise; If two or more signals overlap in the time domain, the value of the tag becomes uncertain. In this case, two processing methods are available: calculation... The normalized frequency value f of the signal frame feature points. i If it could correspond to any of the overlapping targets, then the label vector and f i The corresponding target numbers are consistent; or, the target number corresponding to the one with the largest amplitude among multiple overlapping signals is directly taken.

6. The integrated signal processing method based on U-Net according to claim 2, characterized in that, The construction of the dataset includes the following sub-steps: The structure is generated by supplementing data with actual detection data. , The dataset consists of data pairs, where Is and Corresponding tags; The data generated is based on actual detected data supplemented with data augmentation. , The data pairs specifically include: pulse sequence templates combined with known signals. A PDW sequence is generated based on the signal environment parameters of the unknown signal, and then a simulation method is used to generate ( ). , ) data pairs.

7. The integrated signal processing method based on U-Net according to claim 6, characterized in that, The pulse sequence template that combines known signals A PDW sequence is generated based on the signal environment parameters of the unknown signal, and then a simulation method is used to generate ( ). , Data pairs, including sub-steps: a) Using the intercepted digital signals, if the data coverage of all individual known target signals in the target space meets the requirements, but the coverage of interleaved signals does not meet the requirements, then new data is generated by superimposing the existing data. b) Using the pulse sequence template of the target, if the data coverage of some or all of the single known target signals does not meet the requirements, then generate overlapping data of multiple targets using the PDW sequence templates of each known target in the target space, and then generate ( , ); the PDW sequence template is a PDW sequence that can completely represent the target characteristics, denoted as , a supervector composed of M pulse description words , 0 ≤ i < M, denoted as: =[ , ,…, ] The PDW sequence needs to cover all operating modes of the target, and the number of pulses M is variable, defined according to the target space. Medium pulse description word .

8. An integrated signal processing device based on U-Net, characterized in that, include: The data preprocessing module is used to process digital signal segments during data preprocessing. Processed into signal feature maps ; A deep learning module is used to segment the signal feature map using a U-Net segmentation network. Give the feature points of each signal frame. The probability of belonging to different targets, unknown targets, or noise. ; The post-processing module outputs a signal feature map. A vector of target category numbers for all signal frame feature points. Each signal frame feature point is determined to belong to either a known target, an unknown target, or noise. The signal feature map It is a supervector with length N and number of channels 3, whose three channels are frequency characteristic channel FFC, amplitude characteristic channel AFC and phase characteristic channel PFC.

9. A readable storage medium, characterized in that, A computer program is stored in a readable storage medium, the computer program being loaded by a processor and executing the method as described in any one of claims 1 to 7.

Images