A radar echo signal enhancement detection method based on weak feature targets

Abstract

The application discloses a radar echo signal enhancement detection method based on a weak feature target, and comprises the following steps: acquiring an original echo signal, performing signal enhancement on the original echo signal to obtain a first echo signal; loading a multi-dimensional feature fusion detection network to detect the first echo signal and acquire a preliminary detection result; the preliminary detection result comprises state information at a k moment and is expressed as; inputting the preliminary detection result into a tracking filter to predict a predicted state at a next moment, and judging whether to add a track according to the predicted state; the predicted state is expressed as (); judging whether to trigger feedback reexamination; if the feedback reexamination is triggered, a local reexamination is activated at a predicted position, a missing track is generated, and a track is maintained; and track information is output. According to the above technical scheme, the detection probability of a weak and low observable target and the track continuity rate under a high clutter background can be significantly improved.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing and target detection technology, and more specifically, to a radar echo signal enhancement and detection method based on weak feature targets. Background Technology

[0002] The core task of radar is to detect, identify, and track targets in complex jamming environments (such as strong sea clutter, ground clutter, and electronic interference). With the rapid development of stealth technology and small unmanned aerial vehicles (UAVs) and other "weak-feature targets," the reflected signals of these targets are extremely weak, with very low signal-to-noise ratios (SNR) and inconspicuous features, posing a severe challenge to detection technology.

[0003] Traditional radar target detection primarily relies on constant false alarm rate (CFAR) processing and its improved algorithms. These methods set adaptive detection thresholds based on estimates of the statistical characteristics of background clutter, performing well in uniform clutter environments. However, in non-uniform and non-stationary real-world clutter environments, CFAR performance degrades significantly, easily leading to missed detections of weak targets or excessively high false alarm rates. In recent years, deep learning-based detection methods have been introduced, which can learn complex features from data through learning models. However, they have significant limitations in engineering applications: First, as purely data-driven models, their decision-making process lacks physical interpretability, and their generalization ability is questionable in extreme scenarios where training data is not covered. Second, existing methods mostly adopt a serial open-loop architecture of "detect first, track later," isolating the detection and tracking modules. Once the front-end detection fails, subsequent tracking cannot recover, and the continuous spatiotemporal context information provided by tracking cannot be used to assist and improve detection performance.

[0004] Therefore, a technical solution is needed that can deeply integrate physical priors and data intelligence to achieve deep interaction and closed-loop optimization of detection and tracking, thereby fundamentally improving the radar system's ability to stably detect targets with weak features. Summary of the Invention

[0005] To achieve the above objectives, this application provides a radar echo signal enhancement and detection method based on weak feature targets, comprising the following steps: The original echo signal is acquired, and the original echo signal is enhanced to obtain the first echo signal; the information of the first echo signal includes: measurement point, timing, distance, and velocity; A multi-dimensional feature fusion detection network is loaded to detect the first echo signal and obtain preliminary detection results. These preliminary results include temporal sequence, target presence markers, and state information. The state information includes distance information and Doppler information. The state information of a measurement point at time k is represented as follows: ; The preliminary detection results are input into the tracking filter to predict the predicted state at the next moment. Based on the predicted state, it is determined whether to add the track. The predicted state is represented as ( ); Determine whether a feedback re-inspection has been triggered; if a feedback re-inspection is triggered, activate a local re-inspection at the predicted location, generate a defect-filled track, and maintain the track. Output trajectory information, which includes the target's distance, speed, and bearing.

[0006] The original echo signal is a two-dimensional data matrix of distance-slow time; Signal enhancement of the original echo signal includes: Loading a signal enhancement network; the signal enhancement network uses an adversarial network as its backbone. Based on prior information about the radar's operating environment, it models the statistical characteristics of clutter and specific interference to form physical constraint terms, which are then generated through training. Each distance cell or each distance gate of the distance-slow time two-dimensional data matrix is ​​used as a measurement point. The slow time series of each measurement point is input into the signal enhancement network, and the enhanced two-dimensional data matrix is ​​output to form the first echo signal.

[0007] Among them, the total loss function of the signal enhancement network Defined as: ,in, , , Weighting coefficients respectively For signal reconstruction loss, To combat losses, Loss due to physical constraints; Signal reconstruction loss is used to ensure that the output signal approximates the optimal augmented signal; Adversarial loss is used to improve the overall realism of the generated signal; Physical constraint loss makes the statistical properties conform to the prior model of the target signal or suppresses the energy of a specific frequency band to combat interference.

[0008] Furthermore, the detection of the first echo signal includes: The first echo signal is preprocessed using a three-channel input to output a one-dimensional sequence, a time-frequency image, and a range-Doppler spectrum slice. The feature extraction network is loaded, and features are extracted from one-dimensional sequences, time-frequency images, and distance-Doppler spectral slices, respectively, and preliminary detection results are output.

[0009] The three-channel input preprocessing includes: A fixed-length slow time sequence centered on the measurement point is extracted from the first echo signal as a one-dimensional input vector. The one-dimensional input vector is a one-dimensional sequence containing the original phase information. A short-time Fourier transform is performed on a one-dimensional input vector to generate a time-frequency image that shows the distribution of energy over time and frequency. Coherent accumulation is performed on the slow-time matrix containing a distance window of the current measurement point to form a 64x8 range-Doppler spectrum slice.

[0010] Among them, the tracking filter adopts an interactive multi-model-probabilistic data association filter, which analyzes the target's motion state through a uniform motion model and a uniform turning model, and generates the target's predicted state at the next moment.

[0011] Furthermore, based on the predicted status, determine whether to add the track indicator: In the predicted state ( Searching within the associated gate near the measurement point, if the state information of the measurement point at time k+1 is... Within the associated gate, successful tracking is determined, and the status information is displayed. Add the flight path; otherwise, it will be considered a missed detection in the initial test.

[0012] Further, determine whether a feedback review is triggered: If the initial detection is deemed a missed detection, and the number of consecutive successfully tracked frames exceeds a specified threshold, a feedback re-detection is triggered, generating a feedback instruction. The feedback instruction includes the prediction state at time k+1. ).

[0013] Furthermore, predicted location refers to the predicted state ( Using as the center, define a rectangular region of a specified size on the distance-Doppler two-dimensional plane; Activating local re-detection refers to capturing the state of the missing trace within the local re-detection gate. ); For the missing trace state ( ) and predicted state ( Perform motion consistency verification. If the speed change is within a reasonable range, the motion consistency verification is passed. If the motion consistency verification is passed, maintain the track and add the defective track to the track.

[0014] This method significantly improves the detection probability and track continuity of weak and low-observable targets in high clutter backgrounds by constructing a knowledge-data jointly driven signal enhancement network, a multi-dimensional feature fusion detection network, and a detection-tracking adaptive closed-loop verification mechanism. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the steps of a radar echo signal enhancement and detection method based on weak feature targets provided in an embodiment of the present invention. Detailed Implementation

[0016] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] Figure 1 The specific steps of the radar echo signal enhancement and detection method based on weak feature targets are provided, including: Step S100: Acquire the original echo signal, enhance the original echo signal to obtain the first echo signal; In this step, the radar system first receives the original intermediate frequency echo signal, and after conventional processing such as down-conversion and pulse compression, the original echo signal is generated. The original echo signal is a range-slow time two-dimensional data matrix, where the slow time dimension includes time k and k+1.

[0018] Then, the original echo signal is enhanced, specifically including the following steps: First, load the signal enhancement network; Specifically, the signal enhancement network uses a cGAN as its backbone. Based on prior information about the radar operating environment (such as sea state and terrain), it models the statistical characteristics of clutter (such as K-distribution and Weibull distribution) and specific interference (such as narrowband interference) to form physical constraint terms, which are then generated through training.

[0019] The generator of the signal augmentation network adopts a U-Net structure with residual connections. The input data is a segment of the original echo signal containing clutter and weak targets (e.g., a one-dimensional sequence of length 256), and the output is an augmented signal segment of the same length. The discriminator of the signal augmentation network adopts a classic CNN (Convolutional Neural Network) classifier structure to distinguish the generator-augmented signal from the ideal clean signal.

[0020] Total loss function of signal enhancement network Defined as: ,in, , , Weighting coefficients respectively For signal reconstruction loss, To combat losses, This represents the loss due to physical constraints.

[0021] 1) Signal reconstruction loss is used to ensure that the output signal approximates the optimal enhanced signal; Specifically, the signal reconstruction loss uses smoothing L1 loss to measure the amplitude difference between the generated signal and the target signal, expressed as: ,in, The number of signal sampling points. The enhanced signal output by the generator is at the first... The amplitude value of each sampling point For the corresponding ideal clean target signal in the first... The amplitude values ​​of each sampling point (the amplitude values ​​are provided by simulation or high signal-to-noise ratio measured data during training); in, The function is defined as: ,in, is a coefficient, with a value of 1.0.

[0022] 2) Adversarial loss is used to improve the overall realism of the generated signal; In practice, the loss function of Wasserstein GAN is used to improve the overall realism of the generated signal, expressed as: ,in, The evaluation value output by the discriminator. This represents the enhanced signal output by the generator.

[0023] 3) Physical constraint loss makes the statistical characteristics conform to the prior model of the target signal or suppresses the energy of a specific frequency band to combat interference; In practical implementation, sea clutter characteristic constraints are introduced: the amplitude statistics of sea clutter are described based on the K-distribution model, and the shape parameter α and scale parameter b of the K-distribution model are related to the sea state. A physical constraint loss term is designed to force the amplitude distribution of the enhanced signal in the clutter-dominant region to deviate from the K-distribution with given parameters, while maintaining the peak characteristics of the target region.

[0024] The loss function for physical constraint loss is expressed as: ,in, Let KL divergence represent the amplitude distribution of the enhanced signal in the clutter-dominated region. Compared with the typical K-distribution model The degree of deviation is used to penalize the similarity analysis between the enhanced clutter region distribution and the typical sea clutter distribution; For kurtosis calculation, used to measure the impulse characteristics of the signal in the target region, it encourages the enhancement signal to maintain the target peak.

[0025] Before training the signal enhancement network, training data needs to be prepared: using electromagnetic simulation software, radar echo data containing small targets of different sizes and motion states under different sea states (levels 1-5) and different radar parameters are generated as the original feature data; at the same time, some field-measured sea clutter data are collected to supplement the original feature data; pure target signals are generated through simulation, or ideal target signals in the corresponding scenarios are extracted from high signal-to-noise ratio measured data as labels; the original feature data and the corresponding labels constitute the training data.

[0026] Based on the training data, a two-stage training method is used to train the signal enhancement network: the first stage uses only the reconstruction loss to pre-train the generator; the second stage adds adversarial loss and physical constraint loss for joint fine-tuning, where the weight coefficients are set to... Finally, the construction of the signal enhancement network was completed.

[0027] Then, each distance cell or each distance gate of the distance-slow time two-dimensional data matrix is ​​used as a measurement point. The slow time series of each measurement point is input into the signal enhancement network, and the enhanced two-dimensional data matrix is ​​output to form the first echo signal. Therefore, the information of the first echo signal includes: measurement point, timing, distance and velocity.

[0028] Step S110: Load the multi-dimensional feature fusion detection network to detect the first echo signal and obtain preliminary detection results; the preliminary detection results are composed of time sequence, target existence probability, distance information and Doppler information.

[0029] The detection of the first echo signal includes the following steps: 1) Perform three-channel input preprocessing on the first echo signal, including: A slow time series of a fixed length (e.g., 128 points) centered on the measurement point is extracted from the first echo signal as a one-dimensional input vector; the one-dimensional input vector is a one-dimensional sequence containing the original phase information; A short-time Fourier transform is performed on a one-dimensional input vector to generate a time-frequency image that shows the distribution of energy over time and frequency; the time-frequency image is 64x64 pixels and can highlight the micro-motion features of the target. Coherent accumulation (e.g., 64 pulses) is performed on a slow-time matrix containing a range window (e.g., 8 adjacent cells) of the current measurement point to form a 64x8 range-Doppler spectral slice; the range-Doppler spectral slice can provide information on the coupling between range and Doppler.

[0030] 2) Load the feature extraction network, extract features from the one-dimensional sequence, time-frequency image, and range-Doppler spectral slice respectively, and output the preliminary detection results; The feature extraction network includes multiple sub-networks that support feature extraction of different dimensions of information, including: A one-dimensional CNN is implemented by four layers of one-dimensional convolutions (kernel size [3,5,7,9]), with each layer followed by batch normalization and LeakyReLU activation, used to extract features from one-dimensional sequences; The image channels are implemented through two independent branches of a lightweight 2D CNN backbone (such as a simplified version of MobileNetV2), with the two branches sharing some parameters for extracting features from time-frequency images and range-Doppler spectral slices.

[0031] The features extracted by the feature extraction subnetwork are compressed into 128-dimensional feature vectors and then concatenated to obtain a 384-dimensional fused feature vector. The fused feature vectors are passed through a two-layer fully connected network and then fed into two parallel output heads: one of which is a classification head that outputs the probability of the target's existence. The other is a regression head, which outputs the fine offset of the target within the range-Doppler spectral slice. .

[0032] Set global detection threshold (e.g., 0.7), if Exceeding the global detection threshold The system determines the measurement point corresponding to the feature as the target, and calculates the precise distance and Doppler value based on its range cell and regression offset, outputting the preliminary detection result. The preliminary detection result includes time series, target presence markers, and state information, where the state information includes distance and Doppler information; the state information of a measurement point at time k is represented as... .

[0033] Step S120: Input the preliminary detection results into the tracking filter to predict the predicted state at the next moment, and determine whether to add the track based on the predicted state; The tracking filter employs an interactive multi-model-probabilistic data association filter, which analyzes the target's motion state using a uniform motion model and a uniform turning model to generate the target's predicted state at the next moment. The predicted state can be compared with the actual state at the next moment to determine whether to add it to the track.

[0034] In this step, the preliminary detection result at time k is input into the tracking filter, and the tracking filter adjusts its state information based on the result at time k. Predict the target state at time k+1, generate the predicted state, and represent it as ( ),in This provides the predicted distance information at time k+1. This is the predicted Doppler information at time k+1.

[0035] In the predicted state ( Searching within the associated gate near the measurement point, if the state information of the measurement point at time k+1 is... Within the associated gate, successful tracking is determined, and the status information is displayed. Add the flight path; otherwise, it will be considered a missed detection in the initial test.

[0036] Step S130: Determine whether feedback re-inspection is triggered; If the initial detection is deemed a missed detection, and the number of consecutive successfully tracked frames exceeds a specified threshold, a feedback re-detection is triggered, generating a feedback instruction. The feedback instruction includes the prediction state at time k+1. ).

[0037] Step S131: If a feedback re-inspection is triggered, activate a local re-inspection at the predicted location, generate a missing track, and maintain the track; Among them, the predicted position refers to the predicted state ( Centered on a range-Doppler two-dimensional plane, a rectangular region of a specified size is delineated, which serves as a local complex detection gate; wherein, the specified size is only 1 / 5 of that of a conventional search gate (e.g., ±2 units in the range direction and ±3 units in the Doppler direction).

[0038] Activating local re-detection refers to capturing the state of the missing trace within the local re-detection gate. ), capturing and filling in missing traces can be achieved in two ways: One approach is to directly use a lower detection threshold to compensate for defects; this method sacrifices the local false alarm rate in exchange for a higher detection probability; for example, by lowering the global detection threshold. Adjusted to = 0.5× Extract more The peak points are used as the filler traces; Another approach is to activate a pre-trained lightweight detection model (e.g., a simplified model obtained by knowledge distillation or transfer learning of the main detection network) specifically designed to handle extremely low signal-to-noise ratio signals within this small window to obtain the missing traces.

[0039] For the missing trace state ( ) and predicted state ( Perform motion consistency verification. If the speed change is within a reasonable range, the motion consistency verification is passed. If the motion consistency verification is passed, maintain the track: add the defective track to the track.

[0040] Step S140: Output track information, which includes the target's distance, speed, and bearing.

[0041] This invention utilizes an enhanced network driven by knowledge and data to more accurately suppress complex clutter and enhance the essential features of the target, providing high-quality input for subsequent detection. Through a multi-dimensional feature fusion detection network, it comprehensively utilizes complementary information of the target in the signal, time-frequency, and image domains, improving the ability to identify weak and mutated features. Furthermore, this invention employs a closed-loop feedback mechanism for detection and tracking, breaking the limitations of traditional serial processing. It intelligently guides front-end detection using tracking prediction information, enabling proactive searching and re-detection of missed targets. This significantly enhances the system's continuous and stable tracking capability in harsh environments and effectively reduces the risk of track breakage due to single-frame detection failures.

[0042] While improving performance, this invention avoids the huge computational load caused by global processing through a local re-examination mechanism, which meets the stringent requirements of radar systems for algorithm real-time performance and computing resources.

[0043] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A radar echo signal enhancement and detection method based on weak feature targets, characterized in that, Includes the following steps: The original echo signal is acquired, and the original echo signal is enhanced to obtain the first echo signal; The information in the first echo signal includes: measurement point, timing, distance, and velocity; A multi-dimensional feature fusion detection network is loaded to detect the first echo signal and obtain preliminary detection results. These preliminary results include time sequence, target presence markers, and state information. The state information includes distance information and Doppler information. The state information of a measurement point at time k is represented as follows: ; The preliminary detection results are input into the tracking filter to predict the predicted state at the next moment, and the decision to add the track is made based on the predicted state; the predicted state is represented as ( ); Determine whether a feedback re-inspection has been triggered; if a feedback re-inspection is triggered, activate a local re-inspection at the predicted location, generate a defect-filled track, and maintain the track. Output trajectory information, which includes the target's distance, speed, and bearing.

2. The radar echo signal enhancement and detection method based on weak feature targets according to claim 1, characterized in that, The original echo signal is a two-dimensional data matrix of distance-slow time; The signal enhancement of the original echo signal includes: A signal enhancement network is loaded; the signal enhancement network is based on an adversarial network, and based on prior information about the radar operating environment, it models the statistical characteristics of clutter and specific interference to form physical constraint terms, which are then generated through training. Each distance cell or each distance gate of the distance-slow time two-dimensional data matrix is ​​used as a measurement point. The slow time sequence of each measurement point is input into the signal enhancement network, and the enhanced two-dimensional data matrix is ​​output to form the first echo signal.

3. The radar echo signal enhancement and detection method based on weak feature targets according to claim 2, characterized in that, The total loss function of the signal enhancement network Defined as: ,in, , , Weighting coefficients respectively For signal reconstruction loss, To combat losses, Loss due to physical constraints; The signal reconstruction loss is used to ensure that the output signal approximates the optimal enhanced signal; The adversarial loss is used to improve the overall realism of the generated signal; Physical constraint loss makes the statistical properties conform to the prior model of the target signal or suppresses the energy of a specific frequency band to combat interference.

4. The radar echo signal enhancement and detection method based on weak feature targets according to claim 1, characterized in that, The detection of the first echo signal includes: The first echo signal is preprocessed using a three-channel input to output a one-dimensional sequence, a time-frequency image, and a range-Doppler spectrum slice. The feature extraction network is loaded, and features are extracted from one-dimensional sequences, time-frequency images, and distance-Doppler spectral slices, respectively, and preliminary detection results are output.

5. The radar echo signal enhancement and detection method based on weak feature targets according to claim 4, characterized in that, The three-channel input preprocessing includes: A fixed-length slow time sequence centered on the measurement point is extracted from the first echo signal as a one-dimensional input vector, wherein the one-dimensional input vector is a one-dimensional sequence containing the original phase information; A short-time Fourier transform is performed on the one-dimensional input vector to generate a time-frequency image that shows the distribution of energy over time and frequency. Coherent accumulation is performed on the slow-time matrix containing a distance window of the current measurement point to form a 64x8 range-Doppler spectrum slice.

6. The radar echo signal enhancement and detection method based on weak feature targets according to claim 1, characterized in that, The tracking filter employs an interactive multi-model-probabilistic data association filter, which analyzes the target's motion state using a uniform motion model and a uniform turning model to generate a predicted state of the target at the next moment.

7. The radar echo signal enhancement and detection method based on weak feature targets according to claim 1, characterized in that, The step of determining whether to add a track based on the predicted state refers to: In the predicted state ( Searching within the associated gate near the measurement point, if the state information of the measurement point at time k+1 is... Within the associated gate, successful tracking is determined, and the status information is displayed. Add the flight path; otherwise, it will be considered a missed detection in the initial test.

8. The radar echo signal enhancement and detection method based on weak feature targets according to claim 7, characterized in that, The determination of whether feedback re-check is triggered refers to: If the initial detection is deemed a missed detection, and the number of consecutive successfully tracked frames exceeds a specified threshold, a feedback re-detection is triggered, generating a feedback instruction. The feedback instruction includes the prediction state at time k+1. ).

9. The radar echo signal enhancement and detection method based on weak feature targets according to claim 8, characterized in that, The predicted position refers to the predicted state ( Using as the center, define a rectangular region of a specified size on the distance-Doppler two-dimensional plane; The activation of local re-detection refers to: capturing the state of the missing trace within the local re-detection gate ( ); For the missing trace state ( ) and predicted state ( Perform motion consistency verification. If the speed change is within a reasonable range, the motion consistency verification is passed. If the motion consistency verification is passed, maintain the track and add the defective track to the track.

Images