Deep learning false alarm suppression method based on multi-dimensional features

Abstract

The invention belongs to the technical field of signal processing target detection, and discloses a deep learning false alarm suppression method based on multi-dimensional features. The method comprises the following steps: preprocessing input echo data; performing low-threshold detection on the pre-processed accumulated synthesis data in an energy domain, and determining a suspected false alarm target; carrying out feature modeling by combining the multi-dimensional features of the trace points, and constructing a feature map of the trace points; constructing an inter-frame multi-dimensional feature space; and judging the suspected false alarm target in the inter-frame multi-dimensional feature space by using a deep learning model, and outputting a trace point result. According to the method, radar echo targeted features are modeled, the convolutional neural network is constructed for learning separability features of a real target trace point and a false alarm point, and effective distinguishing of the false alarm point and the real target trace point is realized, so that the false alarm point can be inhibited more autonomously and efficiently, the detection capability and the detection speed of a radar on a target are improved, and the detection efficiency of the radar on the target is improved. The radar detection performance is improved.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of signal processing target detection, and in particular to a deep learning false alarm suppression method based on multi-dimensional features. Background Art

[0002] As the main means of target detection and surveillance, radar is widely used in public and national defense security fields such as air and sea target surveillance and early warning detection, playing an irreplaceable role. However, radar detection environments are complex, such as dense ground clutter, electromagnetic interference, strong sea clutter, meteorological clutter and other complex scenes; the characteristics of the detection objects are diverse, such as low, small and slow moving targets, stealth aircraft, drones and other targets with low flying altitude, slow movement speed, small RCS, high maneuverability and low observability. The detection performance of radar is difficult to meet actual needs, and faces the problem of high false alarm rate and low detection rate in complex environments.

[0003] Traditional false alarm suppression methods mostly use conventional CFAR and clutter map methods. Such methods cannot completely eliminate clutter, and a large amount of residual is still retained in the detection results. For example, due to the narrow notch of the MTI filter, it is difficult to filter out clutter above the city and coastal areas, and the residual sea clutter after cancellation is serious. In addition, due to the strong amplitude of the clutter in the near detection area, the residual clutter is serious after cancellation, resulting in a low signal-to-clutter ratio of targets such as drones, making false alarms difficult to suppress and detection more difficult. The clutter map technology used to suppress clutter is very easy to cause target loss in the face of weak target scenes, such as targets with slow speed and small RCS such as drones, resulting in serious target loss. The problem of suppressing a large number of false alarms in a complex environment with strong clutter is even more challenging. Since the time domain amplitude of the strong clutter echo signal is strong and the target intensity is weak, the clutter is easily regarded as a target, causing a large number of false alarms. At the same time, due to the spectrum expansion caused by the movement of sea clutter, the target spectrum is easily submerged in the widened clutter spectrum to form missed alarms. In addition, the strong clutter signal has the characteristics of nonlinearity and non-stationarity, which further increases the difficulty of accurate target detection in a complex clutter environment.

[0004] Recently proposed detectors based on new methods such as sparse representation theory, fractal theory, and chaos theory are mainly used to suppress false alarms in non-stationary and chaotic clutter. However, due to the limitation of search resources, the dwell time at a certain wave position is limited, and the time and frequency resolution cannot be optimized at the same time. At the same time, the performance of detectors based on chaos analysis is seriously dependent on the accuracy of initial state estimation, and the performance is poor under low signal-to-clutter ratio conditions. The methods using sparse transform domain, sparse dictionary, and wavelet transform factors face problems such as difficulty in control, complex dictionary construction and training, and poor adaptability, and are not general. Summary of the invention

[0005] The purpose of the present invention is to provide a deep learning false alarm suppression method based on multi-dimensional features, which adopts a deep learning algorithm to construct a data-driven multi-dimensional feature false alarm suppression model, distinguish false alarms from real targets from the feature dimension, so as to further reduce the influence of clutter and even unintentional interference and noise, and improve the radar's detection capability for targets.

[0006] To achieve the above object, the present invention provides a deep learning false alarm suppression method based on multi-dimensional features, comprising the following steps: Step 1: Input echo data and pre-process the input echo data; Step 2: Perform low threshold detection in the energy domain on the accumulated synthetic data after preprocessing to determine suspected false alarm targets; Step 3: Combine the multi-dimensional features of the traces to perform feature modeling and construct a feature map of the traces; Step 4: Construct a multi-dimensional feature space between frames; Step 5: Use the deep learning model to judge the suspected false alarm targets in the multi-dimensional feature space between frames. The real point trace targets are judged as 1, and the false point trace targets are judged as 0. The false point trace targets judged as 0 are filtered out, and the real point trace targets judged as 1 are retained as the final point trace result output.

[0007] Furthermore, in step 1, the input echo data is preprocessed including angle measurement and distance measurement.

[0008] Furthermore, in step 2, low threshold detection is performed on the pre-processed echo data in the energy domain, and characteristic information of the suspected false alarm target including the real distance, speed and angle can be obtained.

[0009] Furthermore, in step 3, the feature map of the dot trace is constructed. To visualize the feature map of each layer of the convolutional neural network through feature reconstruction, the following steps are included: S1: The feature information extracted by the shallow layer of the convolutional neural network includes amplitude, texture, contour, color, and size; S2: The feature information extracted by the middle layer of the convolutional neural network includes edge contours and spatial relationships; S3: The feature information extracted by the deep layer of convolutional neural network includes intrinsic correlation features.

[0010] Furthermore, in step 4, constructing the inter-frame multi-dimensional feature space includes the following steps: S4: firstly, obtaining feature information of the suspected false alarm target obtained through preprocessing and low threshold detection; S5: Taking the range gate and Doppler gate currently measured by the suspected false alarm target as the center, intercept the echo intensity data information of the adjacent range gate and the adjacent Doppler gate on the accumulated synthetic data after preprocessing to form the inter-frame multi-dimensional feature data; S6: Correspond the multi-dimensional feature data between frames to the dimensions of distance-Doppler-frame respectively.

[0011] Furthermore, the characteristic information of the suspected false alarm target in step S4 includes the real distance, speed and angle of the suspected false alarm target.

[0012] Furthermore, in step 5, the deep learning model used is a CNN model, and the steps of using the CNN model to determine the suspected false alarm target in the multi-dimensional feature space between frames are as follows: S7: Input the data processed in step 4, a convolution layer followed by a nonlinear activation function, and then enter the pooling layer; S8: Add several combinations of convolutional layers-nonlinear activation functions-convolutional layers into the stacked layers through addition operations; S9: The nonlinear activation function entering the stacked layer passes through the fully connected layer, giving different weights to the high-dimensional attribute features of the suspected false alarm targets extracted by the convolutional neural network; S9: Obtain a value similar to the confidence level, and divide the input echo data into real point targets and false point targets according to the obtained value.

[0013] Beneficial effects: The present invention provides a method for suppressing false alarms based on deep learning of multidimensional features. Different from the traditional method of filtering false alarm points by energy, the present invention proposes a method for suppressing false alarms based on deep learning of multidimensional features. After preprocessing, the method constructs a point feature map based on low threshold detection point traces and multidimensional feature information such as distance and orientation, and removes false points in the feature domain using the designed deep learning network. The present invention can achieve good suppression of a large number of false points, making the target point traces clearly visible and improving the target detection rate. The present invention has the following advantages: (1) good false alarm suppression effect, which can effectively reduce the number of false alarm points and effectively highlight the real target points without obvious target loss; (2) strong autonomy, which can autonomously learn data characteristics and extract separable features for classification without manual intervention; (3) strong environmental adaptability, which has good false alarm suppression ability in the face of different weather and different scene targets; (4) strong portability, and the algorithm can be applied to different radar platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a framework diagram of a deep learning false alarm suppression method based on multi-dimensional features involved in an embodiment of the present invention; Figure 2 is an example diagram of an inter-frame multi-dimensional feature of a false alarm involved in an embodiment of the present invention; Figure 3 is an example diagram of inter-frame multi-dimensional features of a real target involved in an embodiment of the present invention; Figure 4 is a target feature map extracted by a shallow layer of a neural network involved in an embodiment of the present invention; Figure 5 is a target feature map extracted by the middle layer of the neural network involved in the embodiment of the present invention; Figure 6 It is the target feature extraction of the deep layer of the neural network involved in the embodiment of the present invention; Figure 7 It is a structural diagram of a CNN model used in multi-dimensional feature false alarm suppression according to an embodiment of the present invention; Figure 8 It is a distance-azimuth diagram and a distance-time diagram of the original point trace involved in the embodiment of the present invention; Fig. 9 It is a distance-azimuth diagram and a distance-time diagram of a point trace after false alarm suppression according to an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The preferred mechanism and implementation method of the present invention are further described below in conjunction with the accompanying drawings and specific implementation methods.

[0016] The core idea of ​​the present invention is to pre-process the radar echo data to obtain a multi-dimensional echo signal, which is directly input into a deep learning network for feature extraction and target classification. The network's powerful feature learning and mapping capabilities are used to achieve effective classification of targets, and the network model is optimized by iteratively evaluating the classification results, thereby simplifying the processing flow and improving the radar's detection performance. The key to this technology is the pre-processing of raw data, the generation of labels, and the construction of a classification model. In complex environment detection and processing, it is extremely critical to pre-process the echo data and generate input that meets the requirements of a deep learning network.

[0017] like Figures 1 to 9 As shown, an embodiment of the present invention discloses a technical solution for a deep learning false alarm suppression method based on multi-dimensional features. Figure 1 This is a framework diagram of a deep learning false alarm suppression method based on multi-dimensional features involved in an embodiment of the present invention. Example 1

[0018] A deep learning false alarm suppression method based on multi-dimensional features comprises the following steps: Step 1: Input echo data and pre-process the input echo data, including angle measurement, distance measurement, etc.

[0019] Step 2: Perform low threshold detection in the energy domain on the accumulated synthetic data after preprocessing to determine suspected false alarm targets.

[0020] Step 3: Combine the multi-dimensional features of the traces, including time, distance, azimuth, elevation and other information, to perform feature modeling and construct a feature map of the traces; To ensure the stability and interpretability of the false alarm suppression model, the feature maps of each layer of the convolutional neural network are visualized through feature reconstruction. Figures 4 to 6 It can be seen that the shallow layer of the convolutional neural network extracts information such as amplitude, texture, contour, color, size, etc., the middle layer extracts features such as edge contours and spatial relationships, and the deep layer extracts indescribable high-level semantic information, that is, intrinsic correlation features.

[0021] Step 4: Construct a multi-dimensional feature space between frames; The construction of the inter-frame multi-dimensional feature space is firstly carried out by low threshold detection and preprocessing to obtain the real distance, speed, angle and other characteristic information of the suspected target. Then, with the range gate and Doppler gate currently measured by the suspected target as the center, the echo intensity data of the adjacent range gate and the adjacent Doppler gate are intercepted on the accumulated synthetic data after signal preprocessing to form the inter-frame multi-dimensional feature data, which correspond to the distance-Doppler-frame dimensions respectively. The constructed inter-frame multi-dimensional feature diagram is shown in Figure 2 and Figure 3 . Figures 2 to 3 This is an example diagram of the visualization of multi-dimensional features between frames, where each sub-image is a frame, and the horizontal and vertical axes are Doppler and distance dimensions, respectively.

[0022] Step 5: Use the deep learning model to judge the suspected false alarm targets in the multi-dimensional feature space between frames. The real point trace targets are judged as 1, and the false point trace targets are judged as 0. The false point trace targets judged as 0 are filtered out, and the real point trace targets judged as 1 are retained as the final point trace result output; The present invention adopts CNN model to carry out refined discrimination of real targets and false alarms in the above-mentioned inter-frame multidimensional feature space. CNN is a hierarchical model. Through a series of convolution, pooling and non-linear activation function mapping operations, the low-level to high-level feature information contained in the original data can be extracted layer by layer and abstracted layer by layer. Each convolution layer is followed by a non-linear activation function to further enhance the feature expression ability of the network. Finally, through the fully connected layer, different weights are given to each deep feature, and finally a value similar to the confidence is obtained and the input data is classified according to it. In step 5, a deep learning model is used to judge the suspected false alarm target in the inter-frame multidimensional feature space, and the specific operation steps are: Input the data processed in step 4, a convolution layer followed by a nonlinear activation function, and then enter the pooling layer; The combination of several convolutional layers-nonlinear activation functions-convolutional layers is added into the stacked layers; The nonlinear activation function entering the stacked layers passes through the fully connected layer, giving different weights to the high-dimensional attribute features of the suspected false alarm targets extracted by the convolutional neural network; A numerical value similar to the confidence level is obtained, and the input echo data is divided into real point track targets and false point track targets according to the obtained numerical value. Example 2

[0023] A deep learning false alarm suppression method based on multi-dimensional features in this embodiment is basically the same as that in Embodiment 1, except that: 1. Modeling: (1) In step 4, the multi-dimensional feature space between frames is constructed as follows: Figure 2 and Figure 3 An example of a false alarm inter-frame multi-dimensional feature and a real target inter-frame multi-dimensional feature is given. Each sub-image in each column of the figure is a frame, the horizontal axis is the Doppler gate, and the vertical axis is the range gate. The red dots in the figure represent the corresponding range gate and Doppler gate positions of the suspected target in these frames, that is, and In this example, a total of 5 frames are selected for joint processing, of which the third frame is the current processing frame, that is, The value is 2. This example also shows that 2 frames of data need to be cached during the processing, that is, the current processing frame is the first two frames of the actual received frame, but because the corresponding dwell time of each frame is very short, it has little impact on the instantaneity of radar target detection. Each frame is intercepted with 8 range gates and 24 Doppler gates around the suspected target. That is, the extracted features of each frame are composed of 17 range gates. It consists of 49 Doppler gates.

[0024] like Figure 2 As shown in the figure, although strong target points can be detected in the 3rd and 4th frames, they are not reflected in other adjacent frames. Figure 3 The target shown is detected in the 3rd and 4th frames, and relatively strong echoes can also be observed at the corresponding positions (near the red dot) of the suspected target in the 1st and 5th frames. Although the signal-to-noise ratio is low due to the beam deviating far from the target and the detection threshold cannot be too low, it also indirectly confirms the existence of the target, and equivalently achieves the effects of inter-frame similarity and multi-frame accumulation.

[0025] (2) The false alarm suppression model based on the CNN model in step 5 is constructed as follows: Based on the typical ResNet18 structure, the present invention constructs a residual convolutional neural network RDF-ResNet for inter-frame distance-Doppler multi-dimensional features. Figure 7 As shown in the figure, Conv is a convolutional layer, Relu is a nonlinear activation layer using the relu function, MaxPool is a maximum pooling layer, Add is an addition operation, Concat is a stacking layer, and Gemm is a linear connection layer based on general matrix multiplication. The proposed RDF-ResNet has a total of 4 residual modules. The input of the network is , compared with the typical ResNet18 The input is different. If the ResNet18 structure is directly used, many detailed features will disappear due to the overly large convolution kernel and the overly deep convolution pooling layer. Therefore, the RDF-ResNet proposed in the present invention draws on the idea of ​​the residual module and redesigns and optimizes the convolution kernel to match the dimensional size and distribution characteristics of the multi-dimensional echo features between frames.

[0026] 2. Model training process: (1) Sample annotation: The suspected points after low threshold detection are labeled, and the labels of positive and negative samples need to be labeled to support the classification training of the network. In the labeling process, the present invention not only directly labels based on the point information, but also assists in labeling based on the extracted features, with the help of expert knowledge and manual experience, observes a large number of feature samples, and visualizes the sample features to determine whether the sample features at the fuzzy boundary are closer to positive samples or negative samples, thereby manually labeling the fuzzy samples and obtaining more accurate labels for the fuzzy samples, providing more samples and labels for subsequent network training and optimization, and improving the applicability and generalization ability of the network.

[0027] (2) Deep learning model settings: First, set the optimizer. The gradient descent algorithm in the deep learning method, that is, the choice of the optimizer, is extremely important for model training. Common deep learning optimizers include stochastic gradient descent (SGD), Adagrad, Adadelta, mean square estimation RMSprop, and adaptive moment estimation Adam. The Adam algorithm combines the advantages of the Adadelta and RMSprop algorithms. When calculating the learning rate of each parameter, it not only uses the exponential decay average of the square of the gradient (second-order moment), but also considers the exponential decay average of the gradient (first-order moment). During iteration, the learning rate is within a certain range, and the parameters decrease steadily. Therefore, the present invention adopts the Adam algorithm as a gradient descent algorithm, which has better adaptability.

[0028] Secondly, the loss function expresses the preference of the network. It is usually combined with the optimization problem as a learning criterion, and the model is solved and evaluated by minimizing the loss function. The cross entropy loss function is used here, which is essentially the application of cross entropy in information theory to classification problems. Minimizing cross entropy is equivalent to minimizing the relative entropy between the observed value and the estimated value, that is, the Kullback-Leibler divergence of the probability distribution of the two, which is an unbiased estimate. The cross entropy is defined as follows: ; The above cross entropy represents the KL divergence of the probability distribution between the observed values ​​and the estimated values ​​of the target and clutter samples. Minimizing the cross entropy makes the observed values ​​closest to the estimated values, that is, the training model has the best effect.

[0029] (3) Hardware platform: Based on the labeled training samples, the present invention was trained on Nvidia V100 for 2 days to obtain the optimal clutter suppression model.

[0030] 3. Model usage process: First, the feature modeling method is used to perform feature modeling on the test data samples to form a sample feature map. Then, the trained clutter suppression model is used for test classification. Finally, the samples classified as clutter are filtered according to the classification results of the test samples. The remaining samples are retained as the target points obtained by the final detection. The verification effect is shown in Fig. 9 . Compared with the original trace Figure 8 By comparison, it can be found that after using the invented technology to suppress false alarms, the false alarm points in the radar dot trace diagram are greatly reduced, the target dot traces are clearly displayed, and the radar detection capability is effectively improved.

[0031] The present invention provides a deep learning false alarm suppression method based on multi-dimensional features, which achieves the purpose of effectively suppressing false alarm traces of radar detection. The present invention models the targeted features of radar echoes, constructs a convolutional neural network for learning the separable features of real target traces and false alarm points, and effectively distinguishes false alarm points from real target traces, thereby being able to suppress false alarm points more autonomously and efficiently, improve the radar's detection capability and detection speed for targets, and enhance radar detection performance.

[0032] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. However, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A deep learning false alarm suppression method based on multi-dimensional features, characterized in that: The following steps are involved: Step 1: Input echo data and pre-process the input echo data; Step 2: Perform low threshold detection in the energy domain on the accumulated synthetic data after preprocessing to determine suspected false alarm targets; Step 3: Combine the multi-dimensional features of the traces to perform feature modeling and construct a feature map of the traces; Step 4: Construct a multi-dimensional feature space between frames; Step 5: Use the deep learning model to judge the suspected false alarm targets in the multi-dimensional feature space between frames. The real point trace targets are judged as 1, and the false point trace targets are judged as 0. The false point trace targets judged as 0 are filtered out, and the real point trace targets judged as 1 are retained as the final point trace result output.

2. The deep learning false alarm suppression method based on multidimensional features according to claim 1 is characterized in that: In step 1, the input echo data is preprocessed including angle measurement and distance measurement.

3. The deep learning false alarm suppression method based on multidimensional features according to claim 1 is characterized in that: In step 2, low threshold detection is performed on the pre-processed echo data in the energy domain, and characteristic information of the suspected false alarm target including the real distance, speed and angle can be obtained.

4. The deep learning false alarm suppression method based on multi-dimensional features according to claim 1 is characterized in that: In step 3, we construct a feature map that forms a dot trace. We need to visualize the feature map of each layer of the convolutional neural network through feature reconstruction. The following steps are involved: S1: The feature information extracted by the shallow layer of the convolutional neural network includes amplitude, texture, contour, color, and size; S2: The feature information extracted by the middle layer of the convolutional neural network includes edge contours and spatial relationships; S3: The feature information extracted by the deep layer of convolutional neural network includes intrinsic correlation features.

5. The deep learning false alarm suppression method based on multi-dimensional features according to claim 1 is characterized in that: In step 4, constructing the inter-frame multi-dimensional feature space includes the following steps: S4: firstly, obtaining feature information of the suspected false alarm target obtained through preprocessing and low threshold detection; S5: Taking the range gate and Doppler gate currently measured by the suspected false alarm target as the center, intercept the echo intensity data information of the adjacent range gate and the adjacent Doppler gate on the accumulated synthetic data after preprocessing to form the inter-frame multi-dimensional feature data; S6: Correspond the multi-dimensional feature data between frames to the dimensions of distance-Doppler-frame respectively.

6. The method for suppressing false alarms based on deep learning of multi-dimensional features according to claim 5, characterized in that: The characteristic information of the suspected false alarm target in step S4 includes the real distance, speed and angle of the suspected false alarm target.

7. The deep learning false alarm suppression method based on multi-dimensional features according to claim 1 is characterized in that: In step 5, the deep learning model used is a CNN model. The steps of using the CNN model to determine the suspected false alarm target in the multi-dimensional feature space between frames are as follows: S7: Input the data processed in step 4, a convolution layer followed by a nonlinear activation function, and then enter the pooling layer; S8: Add several combinations of convolutional layers-nonlinear activation functions-convolutional layers into the stacked layers through addition operations; S9: The nonlinear activation function entering the stacked layer passes through the fully connected layer, giving different weights to the high-dimensional attribute features of the suspected false alarm targets extracted by the convolutional neural network; S9: Obtain a value similar to the confidence level, and divide the input echo data into real point targets and false point targets according to the obtained value.

Images