CN-122017741-A - Radar behavior recognition method integrating parameter sensing mechanism and multi-scale causal convolution

CN122017741ACN 122017741 ACN122017741 ACN 122017741ACN-122017741-A

Abstract

The invention discloses a radar behavior recognition method integrating a parameter sensing mechanism and multi-scale causal convolution. The method comprises the steps of firstly constructing radar pulse parameter time sequence data in a simulation mode, sending the radar pulse parameter time sequence data into a parameter perception embedding module, modeling the importance of different parameter dimensions to obtain parameter time sequence characteristics with enhanced parameter perception, then carrying out characteristic extraction on various behavior scales of radar parameters, carrying out dynamic weighted fusion on the extracted time sequence characteristics, then adopting a radar working mode switching behavior constraint mechanism to carry out consistency constraint and state transition modeling on identification results of adjacent time windows, carrying out radar working mode discrimination on a constrained characteristic fusion matrix, and finally carrying out light weight processing on a model, and deploying the processed model on an FPGA or an embedded processing platform. The invention realizes the real-time, high-precision and low-calculation intelligent recognition of the radar working mode, reduces the deployment cost and improves the generalization capability and the environment adaptability.

Inventors

TAO SHIFEI
LUO WEI
SONG HAIWEI
YAN MU
WU YOU
DING HAO
WANG JINYANG
NIE XINGYANG

Assignees

南京理工大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (10)

1. A radar behavior recognition method integrating a parameter sensing mechanism and multi-scale causal convolution is characterized by comprising the following steps: Step 1, generating a time domain radar pulse description word PDW sequence of a radar in different behavioral modes in a simulation mode, introducing missing pulses, false pulses and measurement errors into the PDW sequence to simulate a real electromagnetic environment, thereby constructing radar pulse parameter time sequence data, and performing Z-Score standardization processing on the radar pulse parameter time sequence data to eliminate the influence of different dimension differences on training; Step 2, sending the radar pulse parameter time sequence data into a parameter sensing embedding module, and modeling the importance of different parameter dimensions to obtain parameter time sequence characteristics with enhanced parameter sensing; Step 3, adopting four layers of one-dimensional causal convolution time sequence networks as feature extractors, and realizing feature extraction of multiple behavior scales of radar parameters by setting causal convolution structures of different time receptive fields; Step 4, based on radar parameter variation characteristics reflected in parameter time sequence behaviors, dynamically weighting and fusing extracted time sequence characteristics of different time scales, so that a network can adaptively select time scale characteristics with more discrimination for identification according to importance differences of parameter variation under different radar working modes; step 5, adopting a radar working mode switching behavior constraint mechanism to carry out consistency constraint and state transition modeling on the identification results of the adjacent time windows; step 6, judging the radar working mode of the constrained feature fusion matrix by adopting a classification structure based on convolution mapping, and outputting the current working mode category and the corresponding identification result of the radar; and 7, carrying out light weight processing on the full convolution causal time sequence network on the premise of ensuring the identification performance, and deploying the model after light weight processing on an FPGA or an embedded stream processing platform to realize the online and real-time identification of the radar working mode.
2. The method for identifying radar behavior by fusing a parameter sensing mechanism and multi-scale causal convolution according to claim 1, wherein in step 1, the radar pulse parameter time sequence data is subjected to Z-Score standardization processing, and the influence of different dimension differences on training is eliminated, and the specific formula is as follows: Wherein, the Representing the radar pulse signal, And Respectively representing the mean value and standard deviation of corresponding radar behavior parameters; Meanwhile, sliding window processing is carried out on the longer or shorter pulse signals, so that the radar pulse parameter time sequence data are uniformly fixed to be the length T=256.
3. The radar behavior recognition method of fusion of a parameter sensing mechanism and multi-scale causal convolution according to claim 1, wherein step 2 is characterized in that the radar pulse parameter time sequence data is sent to a parameter sensing embedding module, and the importance of different parameter dimensions is modeled to obtain parameter time sequence characteristics with enhanced parameter sensing, and the method specifically comprises the following steps: the parameter sensing module inputs the parameter time sequence matrix Carrying out global statistics in the parameter dimension direction, calculating the time aggregation characteristics of each parameter, and generating a parameter weight vector through one-dimensional convolution and Sigmoid activation function, wherein the specific formula is as follows: Wherein, the Is the first The parameter perceived weights after the calculation of the individual channels, The representation employs a Sigmoid activation function, And representing the radar behavior parameter time sequence matrix after global average pooling.
4. The radar behavior recognition method integrating a parameter sensing mechanism and multi-scale causal convolution according to claim 1, wherein the one-dimensional causal convolution time sequence network in step 3 adopts causal convolution structures with different expansion rates, and the calculation mode is as follows: Wherein, the Is the first Time window number The convolved output of the layer(s), Is the first The weight of the layer is determined by the weight of the layer, Is the first Time sequence of radar parameter weighted by time window, convolution kernel Expansion ratio of Is set to 1,2, 4 and 8 which are increased along with the number of convolution layers and are different in expansion rate And radar pulse parameter variation characteristics corresponding to different time scales.
5. The method for identifying radar behavior by fusion of a parameter sensing mechanism and multi-scale causal convolution according to claim 1, wherein in step 4, the extracted time sequence features of different time scales are dynamically weighted and fused based on radar parameter variation characteristics reflected in parameter time sequence behaviors, and the features are fused The method meets the following conditions: Wherein the weight coefficient Adaptively generating according to global statistical information of each scale feature, Is the first The characteristic output of the layer is that, Is the number of radar behavior characteristics.
6. The radar behavior recognition method of fusion parameter sensing mechanism and multi-scale causal convolution according to claim 1, wherein the radar working mode switching behavior constraint mechanism is adopted in step 5 to perform consistency constraint and state transition modeling on recognition results of adjacent time windows, specifically as follows: Step 5.1, dividing a radar pulse description PDW parameter time sequence into a plurality of continuous time windows according to time sequence, wherein each time window comprises T pulse parameter sampling points; step 5.2, radar behavior feature fusion matrix for the t-th time window Carrying out one-dimensional convolution mapping to obtain a radar working mode prediction score vector corresponding to the current time window: Wherein, the As the number of classes of radar operation modes, Is the first Discrimination scores of radar-like working modes; Step 5.3, normalizing the radar working mode predictive score vector to obtain instantaneous recognition probability ; Step 5.4, constructing a switching behavior constraint model between radar working modes according to the working flow and the mode switching rule of the radar system, wherein the constraint model uses a state transition matrix The form is expressed as: Wherein, the And the behavior constraint weight representing the radar switching from the ith operating mode to the jth operating mode.
7. The radar behavior recognition method of fusion parameter sensing mechanism and multi-scale causal convolution according to claim 1, wherein the feature fusion matrix after constraint is subjected to radar working mode discrimination by adopting a classification structure based on convolution mapping in step 6, and the current working mode category and the corresponding recognition result of the radar are output, specifically as follows: Step 6.1, probability of current instantaneous identification Recognition probability after constraint with previous time window Performing joint constraint calculation to obtain a confidence vector after introducing the constraint of the switching behavior: Wherein, the For the behavioral constraint weights of the radar switching from the ith mode of operation to the jth mode of operation, Is the first The current instantaneous probability of recognition of the radar-like operating mode, Constraint of the first time window The probability of recognition of the mode of operation of the radar-like, The total category number is the total category number of the radar working mode; step 6.2, probability vector based on constraint Determining the radar working mode category corresponding to the t time window by adopting a maximum confidence judgment criterion: Wherein, the Representing the result of the operation mode identification of the radar in the current time window, Representing constrained first Confidence of radar-like operation mode.
8. The radar behavior recognition method integrating a parameter sensing mechanism and multi-scale causal convolution according to claim 1, wherein in step 7, on the premise of ensuring recognition performance, a full convolution causal time sequence network is subjected to light weight processing, and a model after light weight processing is deployed on an FPGA or an embedded stream processing platform, so as to realize online and real-time recognition of a radar working mode, and specifically comprises the following steps: step 7.1, evaluating the importance of the convolution kernels by counting the gradient amplitude of the loss function relative to the weight in the training process of each convolution kernel, and removing the convolution kernels with lower contribution according to the importance sorting result, thereby reducing the scale of model parameters, wherein the formula is as follows: Wherein, the Is the first The convolution kernel importance parameter of the layer, In order to train the data set, As a function of the loss, Is the first Weights of the layers; step 7.2, deleting the convolution kernels with importance lower than a threshold value through importance evaluation of the convolution kernels, reconnecting the model after pruning and training; 7.3, after the convolution kernel pruning is completed, carrying out numerical quantization processing on the residual network weights, and mapping the floating point number weights into low-order integer quantization representation; And 7.4, deploying the model subjected to light weight treatment on an FPGA or an embedded stream processing platform, and realizing online and real-time identification of the radar working mode.
9. The computer device is characterized by comprising a memory and a processor, wherein the memory and the processor are in communication connection, the memory stores computer instructions, and the processor executes the computer instructions so as to execute the radar behavior recognition method of fusion parameter perception mechanism and multi-scale causal convolution according to any one of claims 1-8.
10. A computer readable storage medium having stored thereon a computer program, wherein the program when executed by a processor implements the steps of a method of radar behavior identification incorporating a parameter sensing mechanism as claimed in any one of claims 1 to 8 with a multi-scale causal convolution.

Description

Radar behavior recognition method integrating parameter sensing mechanism and multi-scale causal convolution Technical Field The invention relates to the technical fields of radar signal processing, electronic reconnaissance and artificial intelligence time sequence signal identification, in particular to a radar behavior identification method integrating a parameter sensing mechanism and multi-scale causal convolution. Background Radar operation modes such as scan mode, tracking mode, imaging mode, etc. are typically distinguished by parameters such as Time of Arrival (TOA), pulse repetition interval (Pulse Repetition Interval, PRI), pulse Width (PW), carrier Frequency (RF), and amplitude (Amplitude Modulation, AM). In an electronic reconnaissance system, the rapid and accurate identification of the radar working mode has important significance for target situation analysis, threat assessment and electronic countermeasure. Traditional classification methods based on machine learning, such as a support vector machine (Support Vector Machine, SVM), random forest and the like, usually take pulse parameters as static characteristics to be input into a model, and the methods cannot effectively utilize the time sequence structure of a pulse sequence and cannot describe dynamic behavior characteristics such as mode switching, repetition period, modulation mode change and the like. In recent years, deep learning is increasingly applied to the field of radar radiation source identification. The cyclic neural network (Recurrent Neural Network, RNN) and the Long-Short-Term Memory network (LSTM) can model sequences, but have the problems of complex structure, low reasoning speed, difficulty in real-time processing of scenes, difficulty in deployment on embedded platforms such as FPGA (field programmable gate array) and DSP (digital signal processor), high energy consumption, great training difficulty, remarkable gradient disappearance/explosion problems and the like, and meanwhile, part of methods do not fully consider causal characteristics of radar pulse sequences and are difficult to realize stream online identification. Therefore, it is needed to invent a radar working mode identification method with light structure, definite causal constraint, robustness to abnormal pulses and easy hardware deployment, so as to solve the problems of poor real-time performance, high deployment cost and insufficient generalization capability in the prior art. Disclosure of Invention The invention aims to provide a radar behavior recognition method integrating a parameter sensing mechanism and multi-scale causal convolution, which has the advantages of simple algorithm, high recognition precision, strong real-time performance, low deployment cost and strong generalization capability. The technical scheme for realizing the purpose of the invention is that the radar behavior identification method integrating a parameter sensing mechanism and multi-scale causal convolution comprises the following steps: Step 1, generating a time domain radar pulse description word PDW sequence of a radar in different behavioral modes in a simulation mode, introducing missing pulses, false pulses and measurement errors into the PDW sequence to simulate a real electromagnetic environment, thereby constructing radar pulse parameter time sequence data, and performing Z-Score standardization processing on the radar pulse parameter time sequence data to eliminate the influence of different dimension differences on training; Step 2, sending the radar pulse parameter time sequence data into a parameter sensing embedding module, and modeling the importance of different parameter dimensions to obtain parameter time sequence characteristics with enhanced parameter sensing; Step 3, adopting four layers of one-dimensional causal convolution time sequence networks as feature extractors, and realizing feature extraction of multiple behavior scales of radar parameters by setting causal convolution structures of different time receptive fields; Step 4, based on radar parameter variation characteristics reflected in parameter time sequence behaviors, dynamically weighting and fusing extracted time sequence characteristics of different time scales, so that a network can adaptively select time scale characteristics with more discrimination for identification according to importance differences of parameter variation under different radar working modes; step 5, adopting a radar working mode switching behavior constraint mechanism to carry out consistency constraint and state transition modeling on the identification results of the adjacent time windows; step 6, judging the radar working mode of the constrained feature fusion matrix by adopting a classification structure based on convolution mapping, and outputting the current working mode category and the corresponding identification result of the radar; and 7, carrying out light weight processing on the full convolution c