Method for identifying low-altitude sea-skimming target by multi-channel detector based on deep learning

Abstract

The invention belongs to the field of multi-channel detector detection simulation design, and discloses a method for identifying a low-altitude sea-skimming target by a multi-channel detector based on deep learning. Establishing a simulation environment, and determining rendezvous conditions; the optimal starting position and the distance measurement information of the multi-channel detector are obtained through detection parameter calculation; and when the detection information is identified to be a target or a sea surface, a deep learning neural network model is introduced for judgment. According to the method, the characteristic that the deep learning neural network model can learn, process and analyze big data is utilized, and through analysis of the training set, the detector can have better accuracy in target identification, so that the identification performance of the detector is greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-channel detector detection simulation design, and in particular relates to a method for multi-channel detector to recognize low-altitude sea-skimming targets based on deep learning. Background Art

[0002] Detectors play an important role in our daily lives. With the development of human technology, the detection environment faced by detectors has become increasingly complex, and detectors require more precise detection systems to ensure accurate detection of targets.

[0003] When the detector detects a target, it calculates an optimal detection point to ensure complete detection of the target. Detection parameter estimation refers to the process of predicting the optimal point position, height and other parameters in advance during the flight of the detector. For low-altitude sea-skimming targets, due to the interference of sea surface information, the detector cannot determine whether it is a target or a sea surface when detecting an object, so ranging information is needed to assist in judgment. Each beam of the detection field of a multi-channel detector has a ranging function, which can detect the distance between the detection point and the emission source in the corresponding direction. Multiple beams are detected simultaneously to obtain distance information in each branch direction. Based on this information, the target and the sea surface can be distinguished, and finally it is determined whether the target is detected. Since the optimal detection point is greatly affected by distance and angle, and the recognition of the target will affect the calculation of the optimal detection point position, the accuracy of target recognition is very important.

[0004] In recent years, artificial intelligence has gradually emerged, and deep learning models have developed from simple neural networks to more complex structures. At the same time, they have also been combined with more fields, making outstanding contributions to the development of various fields. Introducing deep learning in the detector detection system and utilizing its ability to learn and analyze large amounts of data can make the detector detection more intelligent, more adaptable to complex environments, and effectively improve the detection capability of targets. The present invention introduces a deep learning neural network model in the detection point parameter estimation, and improves the accuracy of target recognition by learning the patterns and rules of the input data.

[0005] At present, there are a lot of research results on target recognition methods based on traditional radar echo characteristics, and various characteristics of the target have been analyzed, such as orbital characteristics, attitude characteristics, polarization characteristics, etc., but its classification and recognition performance still needs to be improved, and there has been no research on the recognition of low-altitude sea-skimming targets through ranging information characteristics. The present invention proposes a method for recognizing low-altitude sea-skimming targets using a multi-channel detector based on deep learning, and the use of a deep learning neural network model can effectively improve the accuracy of recognizing low-altitude sea-skimming targets. Summary of the invention

[0006] The present invention aims to solve the problems raised in the background technology and proposes a method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning.

[0007] In order to achieve the purpose of the present invention, the present invention discloses a method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning, comprising the following steps:

[0008] Step 1: Determine the rendezvous conditions, including the relative position, attitude angle, and aerodynamic angle of the target and the detector, and build the sea surface environment;

[0009] Step 2: After multi-channel detector detection and calculation, the best detection point, target orientation and output data are obtained, which includes the ranging information of each channel beam of the detector;

[0010] Step 3: Use the ranging information and the corresponding classification labels as a training set to train the deep learning neural network model;

[0011] Step 4: The trained neural network model can determine whether the target is detected.

[0012] Furthermore, in step 2, the method for calculating the multi-channel detector ranging information is as follows:

[0013] There are multiple detection devices installed around the detector. After setting the parameters according to the status of different detection devices, such as beam inclination angle, beam angle, etc., a line segment is drawn with the launch point as the starting point, and the sea surface element and the target surface element are traversed along the direction of the line segment to determine whether the line segment intersects with the surface element.

[0014] Furthermore, in step 2, the process of determining intersection is as follows:

[0015] Assume that the endpoints of the line segment are M 1 (x 1 ,y 1 ,z 1 ) and M 2 (x 2 ,y 2 ,z 2 ), the equation of the plane where the sea surface or target element is located is:

[0016]

[0017] The equation of the line segment is:

[0018]

[0019] Where is the k parameter. When its value is in [0,1], the point represented by the equation is located on the line segment. The value of k can be obtained by combining the equations:

[0020]

[0021] If the solution is between [0,1], it means that the line segment and the plane have an intersection point, and the intersection point is:

[0022] M 1 =(x 1 +k(x 2 -x 1 ),y 1 +k(y 2 -y 1 ),z 1 +k(z 2 -z 1 ))

[0023] Finally, the distance information can be obtained according to the coordinate operation:

[0024]

[0025] Where R represents the distance between the intersection point and the detector.

[0026] Furthermore, the target orientation calculation method in step 2 is as follows:

[0027] Assume that the detector is located at the optimal position P 0 , the detection offset is D 0 The detection cone consists of 72 rays, and the angle between two adjacent rays is 5°; assuming that the target center is m 0 The target orientation is calculated mainly by traversing the triangular planes formed by two adjacent detection cone rays to determine in which triangular plane the target center is located. Assume that the target center is in the detection cone ray l k and l k+1 Between, where 0≤k≤70, L k With L k+1 Represents l k With l k+1 The vectors of these two rays are used to calculate the target orientation by the following formula:

[0028] L 1 =(P 0 +D 0 )-m 0

[0029]

[0030] Where, L 1 represents the vector from the center of the detector to the center of the target, and θ represents the target orientation.

[0031] Furthermore, in step 3, the target recognition network is built based on a convolutional neural network (CNN). The CNN network is a variant model of a multi-layer perceptron. Its design is inspired by a concept in biology, the "receptive field" cells in the visual cortex, which filter local features of the input data and can more effectively extract partial features of the input data. The target recognition network includes a convolutional layer for linear weighted filtering of the local receptive field and a pooling layer for feature screening processing.

[0032] Furthermore, the convolution calculation formula is as follows:

[0033]

[0034] In the formula, f is the input feature matrix, g is the convolution kernel, (x, y) represents the coordinates of the output feature matrix, and * represents the convolution operation. The convolution operation enables the convolutional neural network to effectively capture the local patterns in the input data. Convolutions of different channels can extract different local features in the input data. Through multiple convolution channels, the feature information of the input data can be more comprehensively described, thereby improving the network model's ability to express complex data.

[0035] Furthermore, the pooling layer uses a pooling operator with a pre-set size to perform pooling operations on the feature matrix, and traverses the entire feature matrix in a sliding manner to screen local features. This operation helps to retain the more important local features in the feature matrix, reduce the amount of subsequent calculations, and reduce the probability of network overfitting through nonlinear calculations.

[0036] Furthermore, before the final output of CNN, a fully connected layer is used to generate a one-dimensional output result. The fully connected layer flattens the feature matrix, maps the extracted features into the sample label space and represents them with distributed features. In this way, abstract features can be represented as specific output results, making the input data easier to understand.

[0037] Compared with the prior art, the significant progress of the present invention is that the present invention utilizes the characteristics of deep learning neural network model to learn, process and analyze big data. By analyzing the training set, the detector can have better accuracy in identifying the target, thereby greatly improving the recognition performance of the detector.

[0038] In order to more clearly illustrate the functional characteristics and structural parameters of the present invention, further description is given below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0040] Figure 1 It is a flowchart of a method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning;

[0041] Figure 2 It is a schematic diagram of the detection field of a multi-channel detector;

[0042] Figure 3 is a sea surface modeling effect diagram in the embodiment;

[0043] Figure 4 is the target model used in the embodiment;

[0044] Figure 5 It is a software screenshot during the simulation process of the embodiment;

[0045] Figure 6 is a diagram of the deep learning network structure model used in the present invention;

[0046] Figure 7 It is part of the data of multi-channel detector ranging in the embodiment;

[0047] Figure 8 is the target recognition accuracy result in the embodiment. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0049] like Figure 1 As shown, a method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning, the specific implementation method includes the following steps:

[0050] (1) Determine the rendezvous conditions, including the relative position, attitude angle, and aerodynamic angle of the probe and target, and build the sea surface environment;

[0051] (2) After multi-channel detector detection and calculation, the optimal detection point, target orientation and output data are obtained, which includes the ranging information of each beam of the detector;

[0052] (3) Using the ranging information and the corresponding classification labels as training sets to train the deep learning neural network model;

[0053] (4) The trained neural network model can determine whether the target is detected.

[0054] In step (1), the specific sea surface modeling method is the Philips spectrum model method, which is based on the wave spectrum model. It has slightly poor flexibility but good authenticity and is suitable for simulating the characteristics of wind and waves in the open ocean. It regards the movement of the sea surface as a stable random process. The wave spectrum can not only reflect the important part of the wave, but also reflect the peripheral environment of the sea surface.

[0055] The wave height field is made into many sine waves and its mathematical expression is as follows:

[0056]

[0057] The wave is perpendicular to the xoz plane, so in the above formula represents the sea surface time domain coordinate, Represents the sea surface height field, wave spectrum function In represents the frequency domain coordinate, That is, it is synthesized by the frequency components on the two coordinate axes. Assuming that the side length of the sea surface to be simulated is R, then the sampling interval in the frequency domain is 2π / L, that is:

[0058]

[0059] The expression of the Philips spectral function is as follows:

[0060]

[0061] Among them, A controls the amplitude. Represents wind direction, It is the sea surface wave orthogonal to the wind direction. Adding the Fourier expression of the sea wave to the time variable, we get the dynamic ocean spectrum function:

[0062]

[0063] It is a representation of the dispersion relation, with the relation where g is the acceleration due to gravity, express The conjugate of . The height of each grid point is calculated through 2D-IFFT transformation to build a real-time scene.

[0064] First, draw a sea level grid, and then calculate the height of each grid point. This step is obtained by the inverse Fourier transform of the spectral function, and the degree of squeezing between grid points is calculated according to the settings. After the height calculation is completed, the software can simulate nearly real waves, and then use the texture mapping function to attach a map to the grid to complete the sea surface simulation.

[0065] In step (2), the method for calculating the best detection point is as follows:

[0066] During the rendezvous process, it is assumed that the detector moves relative to the trajectory L, that is, the relative motion vector of the detector is Assume that the position of the detector is P 0 , the detection offset is M 0 , the target center position is P 1 , a straight line L can be drawn through the center of the target 1 for:

[0067]

[0068] L vs L 1 The detection ray cone is established with the center of the detector as the vertex. Assume that the straight line L 1 Intersects the cone at point P 2 , then P 2 The expression is:

[0069]

[0070] The detection ray cone is composed of multiple rays during the simulation process. Therefore, the intersection point can be solved by using the line-plane intersection method based on the plane where two adjacent rays are coplanar. When the distance from the intersection point to the center of the detector is greater than the detection radius of the detector, it is discarded; if it is less than the detection radius, this point is recorded.

[0071] Assume that the straight line L 1 With ray l 1 With l 2 The coplanar planes intersect, and the coordinates of the detector center are P 3 (P 0x +M 0x ,P 0y +M 0y ,P 0z +M 0z ), we get the following equation:

[0072]

[0073] By l 1 and l 2 The direction vector of the two rays is the normal vector of the coplanar surface:

[0074] l 1 × 2 =(b 1 c 2 -b 2 c 1 ,a 2 c 1 -a 1 c 2 ,a 1 b2 -a 2 b 1 )

[0075] The plane equation is obtained through the normal vector:

[0076] (b 1 c 2 -b 2 c 1 )(xx 0 )+(a 2 c 1 -a 1 c 2 )(yy 0 )+(a 1 b 2 -a 2 b 1 )(zz 0 )=0

[0077] Relative plane equation to L 1 Solving the equations of the straight line, we get:

[0078]

[0079] Solving the equation gives the value of k, which can then be calculated through L 1 The coordinates of the intersection point are calculated from the equation of the straight line. Assume that the intersection point P 2 The distance from the detector center is less than the detection radius, and the best detection point is P opt , the current detector position is P 0 , the coordinates of the best detection point can be obtained as:

[0080]

[0081] In step (3), a 1D-CNN neural network is used, which is a variant of a convolutional neural network specifically used to process one-dimensional sequence data. 1D-CNN has certain advantages in processing sequence data and can automatically learn local features and long-term dependencies in the input data. In this invention, multi-beam ranging information at 18 consecutive time points is input into a network model based on 1D-CNN for target recognition and judgment.

[0082] The network model structure is as follows Figure 6 As shown. In order to enable the network model to better extract the implicit features of the input data, it is necessary to introduce a residual structure into the network. The main purpose of the residual structure is to allow the network to learn the identity mapping, that is, to pass the input directly to the output. This design allows the network to selectively learn the residual, thereby enhancing the network's expressive ability and making it more flexible to adapt to different input data distributions.

[0083] In order to ensure that the network can learn more general features, some ReLU activation functions are replaced with LeakeyReLU activation functions. The traditional ReLU activation function is linearly activated in the positive region, but completely suppresses the output in the negative region, which may cause neuron death, that is, the neuron stops activating. To overcome this problem, the LeakeyReLU activation function introduces a small slope in the negative region so that negative inputs can pass through the activation function. This improvement solves the neuron death problem of the ReLU activation function and improves the generalization ability of the network.

[0084] The LeakeyReLU activation function is shown below (α is a manually set parameter):

[0085]

[0086] In order to improve the network training speed and enhance its generalization ability, the network structure of the present invention also introduces a batch normalization layer. In the batch normalization layer processing, the input data is first normalized to calculate the mean and variance of the input data in the current batch, and then learnable parameters are introduced through scaling and translation operations, so that the network can adapt to different data distributions during training. This process helps to accelerate the convergence of the network, improve the stability of network training, and further enhance the generalization ability of the network.

[0087] The batch normalization layer formula is as follows:

[0088]

[0089] Among them, μ represents the mean of the input data, σ represents the data variance, m represents the number of input data, and x i represents the i-th input data, γ and β represent the learnable parameters of scaling and translation, respectively, and ∈ is a constant added to prevent the denominator from being zero.

[0090] The dataset is generally divided into three categories: detected target and sea surface, detected target but not sea surface, detected sea surface but not target. After setting the labels of training set and test set according to the corresponding categories, the neural network can be trained.

[0091] Example

[0092] The method for predicting the azimuth of low-altitude sea-skimming targets using a multi-channel detector based on deep learning provided by the present invention utilizes a multi-channel detector to obtain ranging information, and uses a deep learning neural network model to learn and analyze the obtained information to predict the target azimuth. Such prediction results have higher accuracy than traditional methods.

[0093] Figure 2 It is a schematic diagram of the detection field of the multi-channel detector involved in the present invention. The detector includes 8 detection channels, which are evenly distributed around the detector. The azimuths of the respective ranging channels are 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°, respectively. The detection azimuths of adjacent ranging channels differ by 45°.

[0094] Figure 3 It is the final effect diagram of the sea surface modeling in the embodiment, and simulates the wave effect according to the set sea level.

[0095] Figure 4 It is the target model used in the embodiment. The storage content in its file is divided into three parts: geometric shape, target structure and target key points. Each part records corresponding data and facet information for calculation in the simulation process.

[0096] Figure 5 It is a software screenshot during the simulation process of the embodiment. The intersection conditions set in this embodiment enable the detector to detect the target and the sea surface.

[0097] In this embodiment, the beam ranging information of 18 consecutive time points before the best detection point is selected as the input of the network model. In this case, relatively complete intersection information can be input into the network. Figure 7 It is the channel ranging information collected in the embodiment. Each line records eight pieces of data, which are the ranging information of the eight channels of the detector.

[0098] In this embodiment, the trajectory is divided into three categories: detecting the target and the sea surface, detecting the target but not the sea surface, and detecting the sea surface but not the target. The trained target recognition network is used to predict 600 test sets, and the target recognition accuracy is 94.6%. Figure 8 It is the curve of the accuracy of the training set and the test set changing with the number of training rounds.

[0099] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0100] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning, characterized in that: The following steps are involved: Step 1: Determine the rendezvous conditions, including the relative position, attitude angle, and aerodynamic angle of the target and the detector, and build the sea surface environment; Step 2: After multi-channel detector detection and calculation, the best detection point, target orientation and output data are obtained, which includes the ranging information of each channel beam of the detector; Step 3: Use the ranging information and the corresponding classification labels as a training set to train the deep learning neural network model; Step 4: The trained neural network model can determine whether the target is detected.

2. According to the method of claim 1, the method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning is characterized in that: In step 2, the multi-channel detector ranging information calculation method is as follows: There are multiple detection devices installed around the detector. After setting the parameters according to the status of different detection devices, a line segment is drawn with the launch point as the starting point, and the sea surface element and the target surface element are traversed along the direction of the line segment to determine whether the line segment intersects with the surface element.

3. The method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning according to claim 2, characterized in that: In step 2, the process of determining intersection is as follows: Assume that the endpoints of the line segment are M1(x1, y1, z1) and M2(x2, y2, z2), and the equation of the plane where the sea surface or target element is located is: The equation of the line segment is: Where is the k parameter. When its value is in [0,1], the point represented by the equation is located on the line segment. The value of k can be obtained by combining the equations: If the solution is between [0,1], it means that the line segment and the plane have an intersection point, and the intersection point is: M1=(x1+k(x2-x1),y1+k(y2-y1),z1+k(z2-z1)) Finally, the distance information is obtained according to the coordinate calculation: Where D is the distance between the intersection point and the detector.

4. The method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning according to claim 3, characterized in that: The target orientation calculation method in step 2 is as follows: Assume that the detector is located at the optimal position P0, the detection offset is D0, the detection cone consists of 72 rays, and the angle between two adjacent rays is 5°; Assume that the target center is m0, the target orientation is calculated mainly by traversing the triangular plane formed by the two adjacent detection cone rays to determine in which triangular plane the target center is located. Assume that the target center is in the detection cone ray l k and l k+1 Between, where 0≤k≤70, L k With L k+1 Represents l k With l k+1 The vectors of these two rays are used to calculate the target orientation by the following formula: L1=(P0+D0)-m0 Where L1 represents the vector from the center of the detector to the center of the target, and θ represents the target orientation.

5. The method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning according to claim 1, characterized in that: In step 3, the target recognition network is built based on a convolutional neural network (CNN), which includes a convolutional layer for linear weighted filtering of the local receptive field and a pooling layer. Used for feature screening processing.

6. The method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning according to claim 5, characterized in that: The convolution calculation formula is as follows: In the formula, f is the input feature matrix, g is the convolution kernel, (x, y) represents the coordinates of the output feature matrix, and * represents the convolution operation. The convolution operation enables the convolutional neural network to effectively capture the local patterns in the input data. The convolution of different channels extracts different local features in the input data, and describes the feature information of the input data more comprehensively through multiple convolution channels, thereby improving the network model's ability to express complex data.

7. The method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning according to claim 6, characterized in that: In the pooling layer, a pooling operator with a pre-set size is used to perform pooling operations on the feature matrix. The entire feature matrix is ​​traversed by sliding to screen local features. This operation helps to retain the more important local features in the feature matrix, reduce the amount of subsequent calculations, and reduce the probability of network overfitting through nonlinear calculations.

8. The method for identifying low-altitude sea-skimming targets using a multi-channel detector based on deep learning according to claim 7, characterized in that: Before the final output of CNN, a fully connected layer is used to generate a one-dimensional output result. The fully connected layer flattens the feature matrix, maps the extracted features into the sample label space and represents them with distributed features. In this way, the abstract features are represented as specific output results, making the input data easier to understand.

Images