CN-122017743-A - Double-channel deep learning sorting method for intra-pulse and inter-pulse feature fusion

CN122017743ACN 122017743 ACN122017743 ACN 122017743ACN-122017743-A

Abstract

The invention discloses a dual-channel deep learning sorting method for intra-pulse and inter-pulse feature fusion, which changes the traditional serial architecture, constructs an end-to-end parallel processing frame based on feature fusion, creatively introduces a cross-modal attention mechanism, and realizes deep interaction and collaborative reasoning of intra-pulse modulation features and inter-pulse statistical features. The method effectively overcomes the limitation of single characteristic dimension, and improves the sorting accuracy and robustness under the parameter fuzzy scene.

Inventors

FAN YIFEI
SUN BINGBING
GUO ZIXUN
WANG LING
XIE JIAN
WANG HAITAO

Assignees

西北工业大学

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (9)

1. The double-channel deep learning sorting method for the intra-pulse and inter-pulse feature fusion is characterized by comprising the following steps of: Step 1, extracting and preprocessing the isomerism characteristic map in parallel; for a received radar pulse sequence, two heterogeneous feature maps in and between pulses are extracted in parallel and converted into a two-dimensional image format suitable for convolutional neural network processing; step 1-1, extracting an intra-pulse feature map, carrying out continuous wavelet transformation on a baseband waveform signal of each pulse, and generating a two-dimensional time-frequency map capable of representing intra-pulse modulation information of the baseband waveform signal; step 1-2, extracting a pulse sequence TOA, and converting a one-dimensional TOA sequence into a two-dimensional time-PRI diagram capable of representing a PRI modulation rule by using a PRI conversion algorithm; the method comprises the steps of 1-3, preprocessing a feature map, namely respectively carrying out targeted preprocessing on the generated intra-pulse feature map and inter-pulse feature map, wherein the preprocessing comprises logarithmic transformation, edge clipping, filtering denoising, size normalization and gray enhancement operation so as to obtain network input with uniform format; step 2, double-channel deep feature learning; inputting the preprocessed intra-pulse feature map and inter-pulse feature map into a parallel double-channel deep convolutional neural network architecture, and independently extracting deep abstract features; step 2-1, constructing a parallel backbone network, adopting two parallel deep convolutional neural networks as feature extractors, and respectively receiving an intra-pulse feature map and an inter-pulse feature map as inputs; Step 2-2, generating depth feature vectors, namely deeply learning modes in an input image by multi-layer convolution and pooling operation of each backbone network, and finally respectively outputting high-dimensional intra-pulse modulation depth feature vectors and inter-pulse time sequence regular depth feature vectors as the basis of subsequent cross-mode fusion; step 3, cross-modal attention feature fusion and enhancement; The inter-modal attention fusion module is utilized to perform asymmetric bidirectional interaction on the intra-pulse depth features and inter-pulse depth features extracted by the two channels, so that the depth complementation and mutual enhancement of information are realized; Step 3-1, constructing a bidirectional cross attention mechanism; The bidirectional cross attention mechanism realizes feature enhancement through asymmetric interaction, namely firstly, the inter-pulse feature actively queries information to inter-pulse feature, and calculates inter-pulse feature representation effectively refined and enhanced by inter-pulse information according to the association degree information of the inter-pulse feature; Step 3-2, fusing the enhancement features; two mutually enhanced feature vectors, namely enhanced intra-pulse features and enhanced inter-pulse features, are finally obtained through bidirectional interaction, and are fused to obtain a final feature representation containing intra-pulse and inter-pulse comprehensive information; step 4, the multi-task parallel decoding and end-to-end output of the fusion characteristics; the fused feature vectors are sent to two independent classification heads, are decoded in parallel in a multi-task learning mode, and meanwhile, the sorting and recognition results of the pulses are output; Step 4-1, designing parallel classifiers; constructing two independent classification modules which are respectively used as an intra-pulse modulation recognition head and an inter-pulse modulation recognition head; step 4-2, outputting a sorting result; finally, for each input pulse, the model can simultaneously give the pulse modulation type and PRI modulation type, thereby completing the signal sorting task; Step 4-3, joint optimization training; the entire network is trained under a unified end-to-end framework by jointly optimizing the loss functions of the two classification tasks.
2. The method for two-channel deep learning and sorting by fusion of intra-pulse and inter-pulse features according to claim 1, wherein the step 1 specifically comprises: received radar pulse sequence Generating a time-frequency diagram using continuous wavelet transform The concrete representation is as follows: Wherein, the Is a wavelet function Complex conjugate function of (2), wavelet function From mother wavelets Through the dimensions And translation parameters The transformation is obtained by the following steps: The generated CWT amplitude matrix is the intra-pulse feature map ; Time-series of pulse arrival Pulse sequence signal modeled as one dimension : On a fine time grid Sampling to obtain discrete time sequence : Wherein the discrete sampling time points Time of arrival sequence The relation of (2) is: Wherein, the Is the time resolution; For a pair of It is subjected to short-time Fourier transform to obtain a time-frequency energy distribution matrix : Wherein, the Is the frequency at which the frequency is to be determined, Is a function of the window and, Indicating the number of samples per window sliding forward, Is the length of the window function; Frequency axis Converting into PRI axis to obtain time-PRI two-dimensional image, i.e. inter-pulse feature map ; The two feature images are subjected to feature preprocessing, logarithmic transformation and time axis two-end cutting aiming at the intra-pulse feature image, smoothing processing is carried out on the inter-pulse feature image by adopting a median filter, micro break points caused by pulse loss in spectral lines are filled by assisting in morphological closing operation, structural enhancement is carried out, and finally, the size of the two feature images is normalized to 224 by adopting LANCZOS interpolation algorithm 224 Pixels, into an 8-bit gray scale map, and histogram equalization is applied to maximize image contrast.
3. The method for two-channel deep learning and sorting by fusion of intra-pulse and inter-pulse features according to claim 2, wherein the step 2 specifically comprises: the backbone network adopts ResNet-18 with FC layer removed and inputs And Through respective ResNet-18 channels, respectively, and after self-adaptive global average pooling, 512-dimension depth feature vectors are output And 。
4. The method for two-channel deep learning and sorting by fusion of features between pulses according to claim 3, wherein in said step 3, a cross-attention mechanism based on scaled dot product attention is used, and Enhancement The calculation process of (2) is as follows: To be input feature vector And Respectively mapped to query vectors Key vector Sum vector Three semantic spaces: Wherein, the 、 And Is a learnable weight matrix; calculate the attention weight and normalize And Is transposed of (a) To obtain the similarity between the query and the keys, i.e. the attention score, each part quantized by the score matrix is compared with A degree of association between each of the parts of (a); dividing the attention score by a scaling factor To prevent the gradient from being too small, and then normalized to a attention weight matrix by a Softmax function : Wherein, the Is the dimension of the key vector and, =256; Using the weight matrix Vector of values Weighted summation is carried out to obtain final enhanced feature output : By symmetrical calculation, i.e. As a means of , As a means of And Obtaining inter-pulse enhancement features Final fusion of And Obtaining 。
5. The method for two-channel deep learning and sorting by fusion of intra-pulse and inter-pulse features according to claim 4, wherein the step 4 is specifically: The inter-pulse PRI classifying head adopts a three-layer MLP structure, and integrates batch normalization and Dropout layers; By jointly optimizing the loss functions of the two tasks, the end-to-end training is realized, and the loss functions are defined as: Wherein, the And Prediction logits of real labels and network outputs respectively representing the intra-pulse modulation tasks; And Real labels and predictions logits corresponding to inter-pulse PRI tasks; The weight coefficient for balancing the two tasks is set to be 1.0; representing a standard cross entropy loss function, and the calculation formula is as follows: Wherein, the In order to train the number of samples, Is a model to sample Is used for the prediction result of (a), Is a sample Is a real tag of (1); For each input pulse, the model can output the pulse modulation type and PRI modulation type at the same time, so as to finish the radar signal sorting task.
6. An electronic device comprising a processor and a memory, the memory for storing a computer program, the processor for executing the computer program stored by the memory to cause the electronic device to perform the method of any one of claims 1 to 5.
7. A computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the method according to any one of claims 1 to 5.
8. A chip comprising a processor for calling and running a computer program from a memory, causing a device on which the chip is mounted to perform the method of any one of claims 1 to 5.
9. A computer program product comprising a computer storage medium storing a computer program comprising instructions executable by at least one processor, the instructions when executed by the at least one processor implementing the method of any one of claims 1 to 5.

Description

Double-channel deep learning sorting method for intra-pulse and inter-pulse feature fusion Technical Field The invention belongs to the technical field of radars, and particularly relates to a dual-channel deep learning sorting method for intra-pulse and inter-pulse feature fusion. Background The radar radiation source signal sorting is a key link in an electronic reconnaissance system, and is an important premise and basis for extracting radar signal characteristics and identifying targets. Along with the trend of complexity of modern electromagnetic environment, the signal presents characteristics such as high density, high agility, modulation mode are various, have presented serious challenge to signal sorting technology. The existing radar signal sorting technology routes mainly can be divided into the following three categories: The first is a conventional method based on inter-pulse parameter statistics. The method mainly relies on single dimension characteristics such as Pulse Repetition Interval (PRI) and the like to carry out histogram statistics or sequence search. However, when facing the scene that modern radar signal parameters are seriously overlapped, the limitation that the radar signal parameters are only dependent on a single characteristic dimension is prominent, the sorting precision is obviously reduced, and the problems of batch missing and batch error are easy to occur. The second category is the end-to-end sorting method based on deep learning. Such methods utilize deep neural networks to directly extract features from pulse streams and learn sorting rules, such as time-sequential modeling of pulse arrival Time (TOA) sequences by cyclic neural networks (RNNs), or clustering by segmentation networks after mapping pulse stream data into images or point cloud representations. Although the method effectively reduces the dependence on artificial feature extraction, the core mechanism is unsupervised or self-supervised clustering, and only blind separation of pulse sequences can be realized, so that semantic cognition capability of key attributes such as modulation type, PRI mode and the like is lacking. The system still needs secondary processing of the rear-end recognition module, the complexity and the time delay of a processing link are increased, and the integrated real-time processing requirement under a complex electromagnetic environment is difficult to meet. The third class is a cascading method that combines inter-pulse and intra-pulse features. The method adopts a serial processing architecture, firstly uses inter-pulse features to perform de-interlacing, and then uses intra-pulse features to perform modulation recognition. The upper limit is theoretically identified to be higher, but its cascade structure introduces the problem of "error accumulation". When the performance of the front end inter-pulse de-interlacing link is reduced due to parameter blurring, the overall accuracy is limited. In addition, there is a significant sorting performance bottleneck in complex overlapping signals due to lack of inter-pulse and intra-pulse feature synergy and information feedback. In view of the foregoing, the prior art has technical drawbacks in terms of fuzzy processing parameters, single feature dimensions, and structural error accumulation, and a new sorting method capable of cooperatively utilizing multidimensional information and avoiding error accumulation is needed. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides a dual-channel deep learning sorting method for intra-pulse and inter-pulse feature fusion, which changes the traditional serial architecture, constructs an end-to-end parallel processing framework based on feature fusion, creatively introduces a cross-modal attention mechanism, and realizes deep interaction and collaborative reasoning of intra-pulse modulation features and inter-pulse statistical features. The method effectively overcomes the limitation of single characteristic dimension, and improves the sorting accuracy and robustness under the parameter fuzzy scene. The technical scheme adopted for solving the technical problems is as follows: Step 1, extracting and preprocessing the isomerism characteristic map in parallel; for a received radar pulse sequence, two heterogeneous feature maps in and between pulses are extracted in parallel and converted into a two-dimensional image format suitable for convolutional neural network processing; step 1-1, extracting an intra-pulse feature map, carrying out continuous wavelet transformation on a baseband waveform signal of each pulse, and generating a two-dimensional time-frequency map capable of representing intra-pulse modulation information of the baseband waveform signal; step 1-2, extracting a pulse sequence TOA, and converting a one-dimensional TOA sequence into a two-dimensional time-PRI diagram capable of representing a PRI modulation rule by using a PRI conversion a