Airport runway foreign object detection (FOD) radar echo noise reduction and classification method

Abstract

The invention discloses an airport runway foreign object detection (FOD) radar echo noise reduction and classification method, and the method comprises the steps: constructing a fusion network which comprises a conditional generative adversarial network, a diffusion network and an entropy evaluation mechanism for promoting the game symmetry of the generative adversarial network, and introducing block discrimination and overall constraint into the generative adversarial network, a diffusion network is constructed at the rear end of the generative adversarial network to further denoise the output data and improve the signal-to-noise ratio, an information feedback path from a discriminator to a generator is also established, and the real sense of the detail texture and the overall structure of the generated data is improved. The generator can concentrate the training center of gravity in an uncertain area of the discriminator according to the feedback of the discriminator, so that the learning ability of the generator in an easily confused area is enhanced, and meanwhile, the output of the discriminator is pushed to be close to the maximum entropy state to realize game symmetry; according to the technical scheme, noise can be removed, the target feature information in the radar echo data can be accurately extracted, and the classification accuracy is improved.

Description

Technical Field

[0001] The present application relates to the technical field of computer systems and radars, and in particular to a method for reducing and classifying radar echoes for foreign object detection (FOD) on airport runways. Background Art

[0002] With breakthroughs in deep learning using models such as Transformer, deep learning-based methods have been widely used in various data processing tasks. These methods can achieve more efficient, accurate, and adaptive data analysis and processing by automatically learning the features and distribution patterns in the data. In practical applications, airport runway foreign object detection (FOD) is a typical scenario that requires extremely high accuracy and real-time processing of radar data. FOD detection aims to identify foreign objects (such as metal fragments, tool parts, bird debris, etc.) that may exist on the runway surface. These foreign objects can seriously threaten flight safety.

[0003] In the field of radar data processing, traditional image processing methods often have difficulty in effectively separating FOD targets from runway textures under complex backgrounds and high noise conditions, especially for the foreground and background separation and denoising of complex target echoes. Traditional methods have gradually exposed many shortcomings. Specifically: 1. Traditional manual processing methods rely on experience and manual rules, which are inefficient and highly subjective, and it is difficult to ensure accuracy when processing high-density, high-noise data; 2. Classical methods based on edge detection and other algorithms are sensitive to parameters and difficult to adapt to different types of radar images, which can easily cause target edge information to be lost and affect structural continuity; 3. Point noise (such as speckle noise) problems common in microscopic imaging devices also appear in radar imaging, resulting in reduced image quality and signal-to-noise ratio, further increasing the difficulty of target detection and feature extraction; 4. Traditional denoising algorithms usually rely on linear filtering, empirical parameters or rule matching, and have the following typical defects: 41) Lack of structural modeling capabilities, unable to fully understand the spatial or temporal characteristics of the target; 42) Difficulty in parameter selection, and it is difficult to maintain stable performance under different noise intensities and distribution forms; 43) Poor performance for complex noise such as non-Gaussian or mixed noise; 44) High computational cost and slow processing speed when processing large-scale radar data, making it difficult to meet real-time requirements.

[0004] Therefore, in order to overcome the above problems, it is necessary to design a method that can effectively solve the problems of traditional methods such as insufficient structure, lack of details, incomplete noise reduction, and difficulty in training convergence, so as to improve the noise reduction quality, target retention and model training efficiency of radar images in complex backgrounds. Summary of the Invention

[0005] The purpose of this application is to propose a foreground and background separation and denoising method using a deep learning method to solve problems in radar data processing more accurately and efficiently.

[0006] The technical solution of the present application is to provide a method for denoising and classifying radar echoes for foreign object detection (FOD) on airport runways, the method comprising:

[0007] Step 1: collect radar echo data of the target object, preprocess the radar echo data, and obtain the original RD matrix;

[0008] Step 2: Setting a first training sample set based on the original RD matrix. Specifically, filtering and target labeling are performed on the original RD matrix, and the labeled RD matrix and the original RD matrix are used as training samples to generate the first training sample set.

[0009] Step 3: Construct a fusion network. The fusion network includes a conditional generative adversarial network module, a diffusion network module, and an entropy evaluation mechanism for promoting symmetry in the generative adversarial network game. Specifically, it includes:

[0010] Step 31: Construct a conditional generative adversarial network module including a generator and a discriminator. The discriminator adopts a patchGAN network. The generator learns the process of mapping the original RD matrix and the noise vector to the target RD matrix. The discriminator is used to perform block discrimination on the target RD matrix to determine whether each component block of the target RD matrix is ​​true, so as to guide the generator to generate data with real local details. At the same time, the L1 constraint function is introduced in the conditional generative adversarial network module. The L1 constraint function is used to constrain the overall difference between the target RD matrix output by the generator and the real RD matrix, so that the generator generates data with real global structure.

[0011] Step 32: construct a diffusion network module at the back end of the conditional generative adversarial network module, and use the diffusion network module to perform secondary denoising on the target RD matrix output by the conditional generative adversarial network module;

[0012] Step 33: Using the discriminator's judgment results on each component block of the target RD matrix as local entropy, a learnable entropy mapping function is constructed. The mapping result of the entropy mapping function is multiplied element-by-element by the original RD matrix and then added to the input of the generator as a perturbation term, forming an information feedback mechanism from the discriminator to the generator. This allows the generator to dynamically perceive the changing trend of the discriminator's judgment results on each component block, thereby improving the generator's response ability to areas with low confidence in the discriminator's output, and promoting symmetry in the generation-discrimination game process.

[0013] Step 34, extracting training samples in batches from the first training sample set to train the fusion network and optimize model parameters;

[0014] Step 4: Use the trained fusion network to perform denoising on all original RD matrices in the first training sample set, and use the processed RD matrices as training samples to generate a second training sample set;

[0015] Step 5: constructing a radar echo classification network model based on a neural network, and training the radar echo classification network model using the second training sample set;

[0016] In step 6, the trained generator, diffusion network module, and classification network model are retained to perform noise reduction and classification on the actually collected radar echoes.

[0017] Furthermore, step 31 specifically includes:

[0018] The original RD matrix is ​​recorded as x, the noise vector is recorded as z, and the target RD matrix output by the generator G is recorded as The true RD matrix with labeled information is recorded as y, and the output of the patchGAN network is a tensor with a single element D i Represents the discriminator's discrimination result on a single component block in the target RD matrix, i = 1, 2, ..., N, and sets the target loss function of the generator G and the discriminator D to:

[0019]

[0020] Where, E x,y is the expectation of mapping the original RD matrix to the target RD matrix, E x,z is the expectation of mapping the noise vector to the target RD matrix, D i (x,y) is the discriminator’s judgment result on the i-th block in the real RD matrix, G(x,z) is the output of the generator G, and D i (x, G(x, z)) is the discriminator’s judgment result on the i-th block in the target RD matrix;

[0021] The loss of the L1 constraint function is set to:

[0022] L L1 (G)=E x,y,z [||G(x,z)-y||1]

[0023] Where, E x,y,z is the expectation of the original RD matrix x, the true RD matrix and the noise vector z under the joint distribution; the loss of the L1 constraint function is superimposed on the target loss function to obtain the final target loss function of the conditional generative adversarial network module:

[0024]

[0025] Where λ is the gain parameter.

[0026] Furthermore, in step 32, the diffusion network module includes a forward diffusion part and a reverse denoising part. During training, the forward diffusion part is used to add noise to the target RD matrix output by the generator, and the reverse denoising part is used to learn the process of mapping the noisy target RD matrix into a noise-free matrix; during inference, only the reverse denoising part is used to directly denoise the target RD matrix output by the generator to obtain a real and clean target RD matrix.

[0027] Furthermore, the forward diffusion part is used to add sequence noise to the target RD matrix output by the conditional generation adversarial network module according to the training cycle order. For an initially noise-free target RD matrix In each training cycle t∈{1,2,...,M}, noise is added as follows:

[0028]

[0029] Where, α t and σ t is a hyperparameter adjustable according to time domain information, M is the length of the training cycle, is the target RD matrix with sequence noise added at period t, and ε is the noise term of standard normal distribution;

[0030] The reverse denoising part is used to reversely restore the original target RD matrix and learn the recovery process. Its mathematical model is set as The loss function of the reverse denoising part is set as:

[0031]

[0032] Where, E ε,t is the joint sampling expectation of noise ε and period t.

[0033] Furthermore, step 33 specifically includes:

[0034] The output tensor of the discriminator at training cycle t is D t , tensor D t The local entropy contained is D 1,t ,D 2,t ,...,D N,t , using each local entropy as input to construct a learnable entropy mapping function:

[0035] w t =W(D t-1 )

[0036] Where w t is the function value corresponding to each local entropy in the tth training cycle, D t-1is the local entropy value in the t-1th training cycle, W(·) represents the entropy mapping function, where D t-1 When it approaches the maximum entropy state, W(D t-1 ) tends to a preset constant tensor

[0037] The mapping result w of the entropy mapping function t After being multiplied element-by-element with the original RD matrix, it is added to the input of the generator G as a perturbation term, so that the input of the generator G is:

[0038] x t =x0+x0W(D t-1 )

[0039] Where x0 is the original RD matrix before the superposition of the disturbance term, x t is the original RD matrix after superimposing the perturbation term, x0W(D t-1 ) is a disturbance term; at the same time, a disturbance dynamic control mechanism is set in the process of superimposing the disturbance term, and the disturbance dynamic control mechanism is used to calculate the superimposed disturbance expectation based on the value of the entropy mapping function in the current training cycle, and whether the current training cycle needs to superimpose the disturbance term is determined according to the size of the superimposed disturbance expectation. When the calculated superimposed disturbance expectation is greater than or equal to the preset expected value, the disturbance term is superimposed. When the calculated superimposed disturbance expectation is less than the preset expected value, the disturbance term is not superimposed. The calculation formula of the superimposed disturbance expectation is:

[0040]

[0041] Where, is the position corresponding to the current training cycle, and E[·] is the expectation.

[0042] Furthermore, step 34 specifically includes:

[0043] Set the corresponding condition information c according to the target object to be detected, extract training sample data in batches from the first training set, and perform multiple rounds of training on the fusion network according to the batches. Each round of training specifically includes:

[0044] The disturbance dynamic control mechanism is used to determine whether to inject disturbance. If not, the original RD matrix x, noise vector z and corresponding condition information c are used as the input of the generator. If so, the entropy mapping function is used to calculate the disturbance term based on the local entropy output by the discriminator, and the disturbance term is superimposed on the original RD matrix. Together with the original RD matrix x, noise vector z and corresponding condition information c, the generator is asked to output the target RD matrix y and learn the process of mapping the original RD matrix x and noise vector z to the target RD matrix. When the disturbance term is first injected, the discriminator has no output, D t-1 From the mean The RD matrix and the target RD matrix are sampled element by element from a normal distribution with a standard deviation of 0.001; the labeled RD matrix and the target RD matrix are used as the input of the discriminator, which is used to perform block discrimination and output the local entropy corresponding to each component block of the target RD matrix. The target loss function of the conditional generative adversarial network module is simultaneously calculated. By maximizing the loss of the discriminator, its ability to accurately distinguish between the true RD matrix and the target RD matrix is ​​improved. By minimizing the loss of the generator, the target RD matrix output by it is gradually close to the true one, and the network parameters of the generator and discriminator are updated by gradient descent.

[0045] Each time the conditional generative adversarial network module is trained, the diffusion network module is trained once using the target RD matrix output by the generator as a sample. Specifically, the forward diffusion part is first used to add sequence noise to the target RD matrix in the order of the training cycle, and then the reverse denoising part is used to denoise the target RD matrix after adding sequence noise. The loss function of the reverse denoising part is calculated, and the output data is gradually approached to the original target RD matrix by minimizing the loss of the reverse denoising part. The network parameters of the reverse denoising part are updated by gradient descent.

[0046] Furthermore, step 2 specifically includes:

[0047] Step 21, performing time domain clutter interference suppression on the original RD matrix, and marking the location and type of the target object in the RD matrix after clutter suppression;

[0048] Step 22: Normalize the labeled RD matrix and the original RD matrix to the range of [0, 1] respectively, use the two types of normalized matrices as training samples to generate a first training sample set, and divide the first training sample set into a first training set, a first validation set, and a first test set; wherein the first training sample set contains multiple groups of training samples, each group of training samples includes the corresponding original RD matrix and the real RD matrix with labeled information.

[0049] Furthermore, step 5 specifically includes:

[0050] Step 51, shuffling the data in the second training set and dividing the shuffled data into batches;

[0051] Step 52: Perform multiple rounds of training on the classification network model using different batches of data. After each batch of data training is completed, a new batch of data is extracted from the second training set and trained again until the training meets the preset stopping condition.

[0052] Step 53: After each training of a predetermined number of batches, the classification network model is evaluated and a version is saved using the second validation set. After the training is completed, the version with the best performance on the validation set is used as the optimal model.

[0053] Furthermore, in step 52, the process of training the classification network model using a single batch of data specifically includes:

[0054] Step 521: Use the second training set as input to the classification network model, and use the input layer of the classification network model to perform feature extraction on a single RD matrix in the second training set to obtain a set of token sequences.

[0055] Step 522: Add position codes to the single token sequence by element-by-element addition.

[0056] In step 523, the single token sequence after adding the position encoding is input into the multi-head self-attention mechanism, and the multi-head self-attention mechanism is used to perform further feature extraction to obtain the weighted features of the single token sequence. Specifically:

[0057] First, use each self-attention head to extract features to obtain the query vector Q, key vector K, and value vector V of a single token sequence. Then, based on Q, K, and V of a single token sequence, the attention matrix of the single token sequence is calculated, which can be expressed as:

[0058]

[0059] In the formula, the subscript T is the matrix transpose, and the softmax function is used to normalize the calculation results; finally, the multi-head self-attention mechanism outputs the splicing features of multiple heads:

[0060] Multi Head(Q,K,V)=Concat(head1,head2,...,head h )W o

[0061]

[0062] Where h is the number of self-attention heads, j is the index of each self-attention head, ω O is the output weight matrix, is the weight matrix corresponding to the query vector in the j-th self-attention head, is the weight matrix corresponding to the key vector in the j-th self-attention head, is the weight matrix corresponding to the median vector of the j-th self-attention head;

[0063] In step 524, the weighted features of the single token sequence are processed using a feedforward neural network (FFN) to ultimately obtain a classification result of the radar echo.

[0064] Furthermore, the feedforward neural network FFN is expressed as:

[0065] FFN(e)=Linear(ReLU(Linear(e)))+e

[0066] Where e is the weighted feature of a single token sequence, Linear(e) is the first-layer linear transformation, and Linear(ReLU(Linear(x))) is the second-layer linear transformation. The weighted features of a single token sequence are processed using a feedforward neural network (FFN), including:

[0067] After the weighted features of the token sequence enter the feedforward neural network FFN, the feedforward neural network FFN performs the first linear transformation on the weighted features through matrix multiplication, uses ReLU to activate the linear transformation output of the previous layer, and then performs a second linear transformation on the result after the activation function, and then maps the features to the target space.

[0068] The beneficial effects of this application are:

[0069] First, the technical solution of this application adopts the PatchGAN network to set the discriminator, and uses the PatchGAN network to perform block discrimination on the output of the generator, paying more attention to local area details, and can train the generator G to learn high-frequency texture features. At the same time, the L1 constraint function is introduced to constrain the global structure of the generator's output, reduce the deviation, and ensure that the generated image is consistent with the real image in terms of overall structure. The patchGAN network and the L1 constraint function cooperate with each other, and can simultaneously improve the realism of the detailed texture and overall structure of the generated data during training, effectively imitating the real data distribution.

[0070] Second, the technical solution of the present application constructs a diffusion network module at the back end of the conditional generative adversarial network module, and uses the diffusion network module to perform more detailed noise removal based on the generator output results, thereby improving the signal-to-noise ratio of the final output data, eliminating blurring artifacts and residual interference, and obtaining cleaner radar data; the diffusion network module is located at the back end of the conditional generative adversarial network module, and the generator can directly use the data output by the generator as a sample during training, and perform synchronous training with the conditional generative adversarial network module without the need to prepare additional training samples, thereby avoiding the waste of sample data and the incompatibility of the two models.

[0071] Third, the technical solution of the present application sets up an entropy evaluation mechanism in the fusion network to promote the symmetry of the generative adversarial network game. The entropy evaluation mechanism can use the entropy mapping function to feed back the output tensor of the discriminator to the generator during the training process, and establish an information feedback path from the discriminator to the generator, so that the generator can focus the training on the uncertain area of ​​the discriminator, enhance the learning ability of the generator in the easily confused area, guide the generator to output more realistic data, and at the same time push the output of the discriminator to the maximum entropy state, so as to achieve the symmetry of the game process, further improve the overall model's restoration effect on difficult-to-identify targets and targets with blurred edges, and enhance robustness; the entropy evaluation mechanism can also dynamically control the addition of disturbances, and dynamically adjust the input of the disturbance according to the state of the discriminator output entropy value, so as to avoid the disturbance term tending to the constant tensor when the discriminator output is in the maximum entropy state, causing adverse interference to the training of the generator.

[0072] Fourth, the technical solution in this application can remove noise through a fusion network, accurately extract target information in the foreground part of the radar echo, and has high-fidelity reconstruction capabilities in strong noise environments. It can significantly improve the accuracy and quality of subsequent classification processes, and provides a practical, efficient, and deployable end-to-end solution for radar image denoising. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The advantages of the above and / or additional aspects of the present application will become apparent and readily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0074] Figure 1 is a schematic flow chart of a method for denoising and classifying radar echoes for airport runway foreign object detection (FOD) according to one embodiment of the present application;

[0075] Figure 2 Schematic diagram of the radar noise reduction process obtained by using the method of the present application in Example 1;

[0076] Figure 3 It is the image of a foreign object entering the airport runway during a certain period of time in Example 2;

[0077] Figure 4 This is the radar data positioning map reported by the host computer after noise reduction and classification recognition using the method of this application in Example 2. DETAILED DESCRIPTION

[0078] In order to more clearly understand the above-mentioned objectives, features and advantages of the present application, the present application is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features therein can be combined with each other in the absence of conflict.

[0079] In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present application is not limited to the specific embodiments disclosed below.

[0080] like Figure 1 As shown, this embodiment provides a radar echo denoising and classification method based on a neural network, including:

[0081] Step 1: Collect radar echo data of the target object and preprocess the radar echo data to obtain the original RD matrix (Range-Doppler Matrix, which is a two-dimensional matrix used to represent the distance between the target and the radar).

[0082] Foreign objects on the airport runway are taken as targets to be detected. The target objects are scanned using radar equipment to obtain radar echo data. The radar echo data is then subjected to carrier frequency removal (orthogonal demodulation), pulse compression, and two-way FFT (i.e., Fourier transform in the range and Doppler directions) processing to obtain the original RD matrix. The RD matrix is ​​a two-dimensional graph of distance × velocity, showing the energy distribution of the target in the distance and velocity dimensions. Each element in the graph represents the reflection intensity.

[0083] In this embodiment, radar equipment with strong noise resistance, such as pulse radar, synthetic aperture radar (SAR), etc., can be used to collect low-noise radar data as much as possible when collecting echo data of the target object, so as to reduce the errors generated in the subsequent training of the neural network model, accelerate the convergence speed of the model, improve the robustness and accuracy of the model, and also enhance the generalization ability of the model in actual application scenarios.

[0084] Step 2: Set a first training sample set based on the original RD matrix. Specifically, filter and target label the original RD matrix, and use the labeled RD matrix and the original RD matrix as training samples to generate the first training sample set.

[0085] Step 21: Suppress the time domain clutter interference on the original RD matrix, and mark the location and type of the target object in the RD matrix after clutter suppression.

[0086] In this embodiment, time-domain clutter suppression refers to the process of suppressing and removing clutter interference in the radar signal's time domain (i.e., on the time axis) through specific signal processing techniques (such as time-domain filtering, adaptive filtering, and time-domain waveform analysis). Clutter refers to interference signals in radar echoes generated by non-target signals (such as environmental noise and weather factors).

[0087] In step 22, the labeled RD matrix and the original RD matrix are normalized to the range of [0, 1] respectively, the two types of normalized matrices are used as training samples to generate a first training sample set, and the first training sample set is divided into a first training set, a first validation set, and a first test set.

[0088] Among them, the first training sample set contains multiple groups of training samples, each group of training samples includes the corresponding original RD matrix and the real RD matrix with labeled information; the first training set is used to train the model, and the first validation set is used to evaluate the model during the training process so as to adjust the hyperparameters (such as learning rate, network structure, etc.) in time to prevent overfitting; the first test set is used to evaluate the final performance of the model after the training is completed to test the generalization ability of the model.

[0089] In this embodiment, since the training process of the neural network relies on the backpropagation algorithm, the efficiency and stability of the algorithm usually depend on the range of the input data. If the numerical distribution of the input data is very uneven (individual values ​​are too large or too small), gradient explosion or gradient vanishing objects will occur during training, resulting in the inability to effectively update the weights, thereby affecting the accuracy of the training results; normalizing the data to the range of [0,1] can ensure that the values ​​of the input features are within a relatively small range, making the gradient calculation more stable, helping to accelerate the convergence of the model, and facilitating the training of the neural network.

[0090] Step 3: Construct a fusion network. The fusion network includes a conditional generative adversarial network module, a diffusion network module, and an entropy evaluation mechanism for promoting symmetry in the generative adversarial network game. Specifically, the following steps are involved:

[0091] It should be noted that this application proposes a novel fusion model that leverages the sampling advantages of generative adversarial networks (GANs) and the core denoising concepts of diffusion models to create a highly efficient single model for radar range profile denoising. It also proposes a method for verifying the convergence of the GAN model. A radar range profile is a two-dimensional radar image, specifically represented by an RD matrix. This is a two-dimensional image generated by the radar system by collecting and processing return echo signals, and is generated by a back-end computer.

[0092] Step 31, construct a conditional generative adversarial network module including a generator G and a discriminator D. The discriminator D adopts a patchGAN (Patch Generative Adversarial Network Discriminator) network. The generator is used to learn the process of mapping the original RD matrix and the noise vector to the target RD matrix. The discriminator is used to perform block discrimination on the target RD matrix to determine whether each component block of the target RD matrix is ​​true, so as to guide the generator to generate data with real local details. At the same time, the L1 constraint function is introduced in the conditional generative adversarial network module. The L1 constraint function is used to constrain the overall difference between the target RD matrix output by the generator and the true RD matrix, so that the generator generates data with real global structure.

[0093] The original RD matrix is ​​recorded as x, the noise vector is recorded as z, and the target RD matrix output by the generator G is recorded as The true RD matrix with labeled information is recorded as y, and the output of the patchGAN network is a tensor with a single element D i Represents the discriminator's judgment result on a single component block in the target RD matrix. The number of elements in this tensor is equal to the number of blocks that the patchGAN network divides the target RD matrix into, both of which are N. Then i = 1, 2, ..., N. The target loss function of the generator G and the discriminator D is set to:

[0094]

[0095] Where, For a single element D i The corresponding target loss function, E x,y is the expectation of mapping the original RD matrix to the target RD matrix, E x,z is the expectation of mapping the noise vector to the target RD matrix, D i (x, y) is the discriminator's judgment result on the i-th block in the real RD matrix (the judgment result is the probability value / confidence, representing the possibility of being true), G(x, z) is the output of the generator G, D i (x, G(x, z)) is the discriminator’s judgment result on the i-th block in the target RD matrix;

[0096] The loss of the L1 constraint function is set to:

[0097] L L1 (G)=E x,y,z [||G(x,z)-y||1]

[0098] Where, E x,y,zis the expectation of the original RD matrix x, the true RD matrix and the noise vector z under the joint distribution; the loss of the L1 constraint function is superimposed on the target loss function to obtain the final target loss function of the conditional generative adversarial network module:

[0099]

[0100] Where λ is the gain parameter, and the value of λ is 100.

[0101] In this embodiment, a discriminator D is constructed based on the patchGAN network, and an L1 constraint function is introduced into the conditional generative adversarial network. The patchGAN network focuses on local details (i.e., high-frequency textures) and can perform block discrimination on the output of the generator G, which is conducive to training the generator G to generate high-frequency details and prevent blurring. The L1 constraint function focuses on the overall structure (i.e., low-frequency structure) and can constrain and guide the generator G to maintain global consistency, ensuring that the output structure is consistent, the position is accurate, and the background is correct. The patchGAN network and the L1 constraint function cooperate with each other to enhance the target edge and texture details while ensuring the authenticity of the overall structure of the output data, so that the target RD matrix output by the generator G is closer to the real RD matrix in terms of global structure and local details. The core training mechanism of the conditional generative adversarial network module is the minimum maximum game. The generator G is trained to minimize the loss as the goal, so as to generate realistic data as much as possible. The discriminator D is trained to maximize the loss as the goal, so as to distinguish between true and false as accurately as possible. The two play against each other, thereby promoting the learning of the entire model.

[0102] Step 32: construct a diffusion network module at the back end of the conditional generative adversarial network module, and use the diffusion network module to perform secondary denoising on the target RD matrix output by the conditional generative adversarial network module.

[0103] The diffusion network module is divided into a forward diffusion part and a reverse denoising part. During training, the forward diffusion part adds noise to the target RD matrix output by the generator, and the reverse denoising part learns the process of mapping the noisy target RD matrix into a noise-free matrix to fit the actual filtered data distribution; during inference (i.e., actual application), only the reverse denoising part is used to directly denoise the target RD matrix output by the generator to obtain a more realistic and clean target RD matrix.

[0104] The forward diffusion part is used to add sequence noise to the target RD matrix output by the conditional generation adversarial network module according to the training cycle order. For an initially noise-free target RD matrix In each training cycle t∈{1,2,...,M}, noise is added as follows:

[0105]

[0106] Where, α t and σ t are hyperparameters that can be adjusted according to time domain information, respectively expressed as and Signal-to-noise ratio at t = M Small enough, M is the length of the training cycle, is the target RD matrix output by the conditional generative adversarial network module, The target RD matrix with sequence noise added at t period (i.e. On the basis of , a target RD matrix of the noise amount corresponding to the period t in the sequence noise is added), and ε is the noise term of the standard normal distribution.

[0107] The reverse denoising part is used to reversely recover the original target RD matrix and learn the recovery process. Its mathematical model is set as The loss function of the reverse denoising part is set as:

[0108]

[0109] Where, E ε,t is the joint sampling expectation of noise ε and period t.

[0110] It should be noted that although the conditional generative adversarial network module can generate the target RD matrix, the output may still contain artifacts, blur, or non-real distribution characteristics under complex noise backgrounds or blurred targets. The diffusion network module can further perform fine-grained modeling and secondary denoising on the output of the conditional generative adversarial network module, remove residual interference, and improve the output signal-to-noise ratio, which helps to obtain an RD matrix that is closer to the real target characteristics and enhance the denoising and fitting capabilities of the entire fusion network. Moreover, the diffusion network module is set at the back end of the conditional generative adversarial network module. During training, the data output by the conditional generative adversarial network module can be directly used as samples to achieve synchronous training with the conditional generative adversarial network module. There is no need to prepare additional training samples, which avoids the waste of sample data and the incompatibility of the two models.

[0111] In step 33, the discriminator's discrimination results for each component block of the target RD matrix are used as local entropy to construct a learnable entropy mapping function. The mapping result of the entropy mapping function is multiplied element-by-element with the original RD matrix and then added to the input of the generator as a disturbance term, forming an information feedback mechanism from the discriminator to the generator, so that the generator can dynamically perceive the changing trend of the discriminator's discrimination results for each component block, so as to improve the generator's response ability to areas with low confidence in the discriminator output (and areas that the discriminator is difficult to discriminate), and ultimately promote the generation-discrimination game process to reach a more balanced symmetric state.

[0112] In this embodiment, confidence refers to the degree of certainty of the discriminator in its output judgment result, which is usually expressed as an output probability value. For example, a judgment result of 0.9 indicates that the discriminator is very confident that the block is real, and the confidence is high. A judgment result of 0.5 indicates that the discriminator is not sure whether the block is true or false, and the confidence is low. A judgment result of 0.1 indicates that the discriminator is very confident that the block is false, and the confidence is high.

[0113] It should be noted that the discriminator D is used to evaluate the output results of the generator G. Since the discriminator D uses the patchGAN network, it is equivalent to introducing block discrimination in the evaluation of the output results of the generator G. In this evaluation mode, the discriminator can understand where the generator performs poorly, while the generator has no direct spatial clues to understand where the discriminator performs poorly, which makes the minimax game favorable to the discriminator. Without further intervention, this will hinder the balance of the game and lead to failure to converge. Therefore, it is necessary to find the information that the discriminator should pass to the generator to restore the symmetry of the game.

[0114] The output tensor of the discriminator D at training cycle t is D t , tensor D t The local entropy contained (local entropy is the value corresponding to each element, that is, the probability value of discrimination) is D 1,t ,D 2,t ,...,D N,t , using each local entropy as input to construct a learnable entropy mapping function, which is expressed as:

[0115] w t =W(D t-1 )

[0116] Where w t is the function value corresponding to each local entropy in the tth training cycle, D t-1 is the local entropy value in the t-1th training cycle, W(·) represents the entropy mapping function, whose value range is in the matrix space of the generator D input, and W(·) satisfies the following rules: D t-1 When it approaches the maximum entropy state, W(D t-1 ) tends to a preset constant tensor. Specifically, for the tensor D t-1 Any local entropy D in i,t-1 , when D i,t-1 →1 o'clock, A preset constant.

[0117] The mapping result w of the entropy mapping function t After being multiplied element-by-element with the original RD matrix, it is added to the input of the generator G as a perturbation term, so that the input of the generator G is:

[0118] x t =x0+x0W(D t-1 )

[0119] Where x0 is the original RD matrix before the superposition of the disturbance term, x t is the original RD matrix after superimposing the perturbation term, x0W(D t-1 ) is the disturbance term.

[0120] At the same time, a disturbance dynamic control mechanism is set in the process of superimposing disturbance terms. The disturbance dynamic control mechanism is used to calculate the superimposed disturbance expectation based on the value of the entropy mapping function in the current training cycle, and whether the current training cycle needs to superimpose disturbance terms is determined according to the size of the superimposed disturbance expectation. When the calculated superimposed disturbance expectation is greater than or equal to the preset expected value, the disturbance term is superimposed. When the calculated superimposed disturbance expectation is less than the preset expected value, the disturbance term is not superimposed. The calculation formula for the superimposed disturbance expectation is:

[0121]

[0122] Where, is the position corresponding to the current training cycle. E[·] is the expectation (the arithmetic mean of each local entropy can be taken); the dynamic control mechanism of perturbations during training can also be used to generate whether the adversarial network converges under inflammatory conditions. When the expectation E[W(D t-1 )] gradually stabilizes and approaches the preset range, which proves that the output tensor of the discriminator tends to be stable, and each local entropy is close to 1, and at the same time, it shows that the data output by the generator is gradually realistic.

[0123] Among them, during the reasoning process (i.e., formal data processing after training), the discriminator D no longer works, D t-1 The mean value is The tensor is sampled element by element from a normal distribution with a standard deviation of 0.001, that is, it simulates the output of the discriminator D to maintain the perturbation structure during training and avoid performance degradation during generator reasoning.

[0124] It should be noted that the purpose of setting the perturbation dynamic control mechanism is to dynamically adjust the perturbation input according to the state of the discrimination, so as to avoid the perturbation term tending to the constant tensor causing adverse interference to the training of the generator when the output of the discriminator is in the maximum entropy state; specifically, the perturbation is added to find the areas that are difficult for the discriminator to judge, and to feed back the positions of these areas to the generator, guiding it to strengthen the generation quality of these areas. When the discriminator is confident in judging all areas and there is no ambiguity, its local entropy tends to 1 (the output is close to 1, which means that the discriminator believes that the generator output data is true, and the output is close to 0, which means that the discriminator believes that the generator output data is false). t-1)All elements are close to constants, which means that the discriminator will not tell the generator any useful information. There is almost no information in the perturbation term. The perturbation term itself is a kind of interference. Without transmitting any information, it will not only not play a guiding role, but also disturb the generator input. Therefore, there is no need to add the perturbation term at this time. It is necessary to set up a dynamic control mechanism for perturbations to skip the periods that do not require perturbations.

[0125] In this embodiment, for the entropy mapping function, W(D t-1 ) and W(D t ) are not correlated in time, and these function values ​​are approximately Gaussian distributed as a whole. They are random disturbances with no time trend, and will not affect the next step due to a small / large change in the weight of the previous step, and will not affect the accuracy of the training results. For the discriminator D, the disturbance term x0W(D t-1 ) initially follows Brownian motion and diffuses through various periods during training (because the discriminator has not yet learned where the generator is not doing well, the perturbation diffuses randomly in each area like Brownian motion). As training progresses, the discriminator becomes more and more able to identify areas where the generator is not doing well, and the perturbation gradually diffuses to these key areas, becoming a directed perturbation, helping the generator D to better learn to eliminate noise. This mechanism is an indirect communication mechanism between the generator and the discriminator, and it can also transmit signals to the discriminator, allowing it to pass through W(D t ) seeks the maximum entropy state, that is, the better the data output by the generator, the more difficult it is for the discriminator to judge whether it is true or false, W(D t ) gradually tends to be smooth, which makes the output of the discriminator move closer to the maximum entropy state and realizes the symmetry of the game process.

[0126] Step 34, extracting training samples in batches from the first training sample set to train the fusion network and optimize model parameters, specifically includes the following steps:

[0127] Set the corresponding condition information c according to the target object to be detected (for example, if the target object is a foreign vehicle, the condition information c can be set to "a vehicle"), initialize the parameters of the fusion network, extract training sample data in batches from the first training set, and perform multiple rounds of training on the fusion network according to the batches. A single set of training samples includes the labeled RD matrix and the original RD matrix. Each round of training specifically includes:

[0128] The conditional generative adversarial network module is trained by alternating training. Specifically, the perturbation dynamic control mechanism is used to determine whether to inject perturbations. If not, the original RD matrix x and the noise vector z (a batch of noise is sampled from the Gaussian distribution, usually a random noise vector z of a fixed dimension is selected, such as z∈R 100) and the corresponding conditional information c as the input of the generator G. If so, the entropy mapping function is used to calculate the perturbation term based on the local entropy output by the discriminator D, and the perturbation term is superimposed on the original RD matrix. Together with the original RD matrix x, the noise vector z, and the corresponding conditional information c, the generator G is made to output the target RD matrix y and learn the process of mapping the original RD matrix x and the noise vector z to the target RD matrix. When the perturbation term is first injected, the discriminator has no output, and D t-1 From the mean The target RD matrix is ​​obtained by element-by-element sampling from a normal distribution with a standard deviation of 0.001; the labeled RD matrix (i.e., the true RD matrix) and the target RD matrix output by the generator G are used as the input of the discriminator D, and the discriminator D is instructed to discriminate the target RD matrix in blocks, and output the local entropy corresponding to each component block of the target RD matrix. The target loss function of the conditional generative adversarial network module is simultaneously calculated, and its ability to accurately distinguish between the true RD matrix and the target RD matrix is ​​improved by maximizing the loss of the discriminator. The target RD matrix output by the generator is gradually close to the true one by minimizing the loss of the generator, and the network parameters of the generator and discriminator are updated by gradient descent.

[0129] Each time the alternating training of the conditional generative adversarial network module is completed, the diffusion network module is trained once with the target RD matrix output by the generator G as a sample. Specifically, the forward diffusion part is first used to add sequence noise to the target RD matrix in the order of the training cycle, and then the reverse denoising part is used to denoise the target RD matrix after adding sequence noise. The loss function of the reverse denoising part is calculated, and the output data is gradually approached to the original target RD matrix by minimizing the loss of the reverse denoising part. The network parameters of the reverse denoising part are updated by gradient descent.

[0130] When both the conditional generative adversarial network module and the diffusion network module meet the iteration conditions, the training is terminated, and the generator D of the conditional generative adversarial network module and the reverse denoising part of the diffusion network module are retained; the iteration conditions of the conditional generative adversarial network module are: the target loss is stable in the first preset range and the expected E[W(D t-1 )] is stable in the second preset range; the iteration condition of the diffusion network module is: the loss function L DM Stable in the third preset range.

[0131] During training, the generator and discriminator are evaluated using the first validation set after a predetermined number of training rounds to prevent overfitting. The generator's loss is used to determine whether it is generating samples that are closer to real data. If the generator's loss gradually decreases, it indicates that the generator is generating increasingly realistic data. The discriminator's loss is used to determine whether the discriminator can effectively distinguish between generated and real data. If the discriminator's loss gradually increases, it indicates that the discriminator's ability to accurately distinguish between real and generated data has improved. After training, the generator and the reverse denoising part of the diffusion network are tested using the first test set. If the output data denoising rate reaches above 96%, the training is successful.

[0132] Step 4: Use the trained fusion network to perform denoising on all original RD matrices in the first training sample set, and use the processed RD matrices as training samples to generate a second training sample set.

[0133] All original RD matrices are denoised in turn using the generator D in the fusion network and the reverse denoising part in the diffusion network module. Data enhancement is performed on the RD matrix after clutter suppression, and the position of the target object is marked. The marked RD matrix is ​​then divided into the second training set and the second validation set.

[0134] In this embodiment, the RD matrix data after clutter suppression can be enhanced by methods such as rotation, translation, and scaling to increase the diversity and quantity of the data.

[0135] Step 5: constructing a radar echo classification network model based on a neural network, and training the radar echo classification network model using the second training sample set, specifically comprising the following steps:

[0136] Step 51, shuffling the data in the second training set and dividing the shuffled data into batches;

[0137] In this example, to ensure the effectiveness and versatility of model training, the dataset was shuffled to prevent the sequential nature of the data from affecting the model's learning. Subsequently, the dataset was divided into batches to accommodate the neural network's batch training mechanism, improving training efficiency and facilitating stable model convergence.

[0138] Step 52: Use different batches of data to perform multiple rounds of training on the classification network model. After each batch of data training is completed, a new batch of data is extracted from the second training set and trained again until the training meets the preset stopping condition. The process of training the classification network model using a single batch of data specifically includes:

[0139] Step 521: Use the second training set as input to the classification network model, and use the input layer of the classification network model to perform feature extraction on a single RD matrix in the second training set to obtain a set of token sequences.

[0140] In this embodiment, the input layer in the classification network model can perform feature extraction on the RD matrix through a linear mapping or a convolutional layer, and convert the RD matrix into several token sequences (taking several token sequences corresponding to one RD matrix as a group), and each token sequence is represented as a vector.

[0141] Step 522: Add position codes to the single token sequence by element-by-element addition.

[0142] The positional encoding is added to the token sequence by using the input sequence plus positional encoding method (i.e., input + positional encoding, each token sequence in a single group is added to the positional encoding element by element).

[0143] In this embodiment, the position encoding may adopt a sine / cosine function or a learnable embedding vector; the position encoding is added to preserve the sequential information of the input data.

[0144] Step 523: Input the single token sequence after adding the position encoding into the multi-head self-attention mechanism, and use the multi-head self-attention mechanism to perform further feature extraction to obtain the weighted features of the single token sequence;

[0145] Multiple self-attention heads are used to process a single group of token sequences after adding position encoding. After each self-attention head performs feature extraction, the query vector Q, key vector K and value vector V of the single group of token sequences are obtained, which can be expressed as:

[0146] Q=Fω Q ,K=Fω K ,V=Fω V

[0147] Where F is the input token sequence, ω Q 、ω K 、ω V The weight matrix learned by the classification network model is used to generate query vectors, key vectors, and value vectors respectively;

[0148] Then, the attention matrix of a single token sequence is calculated based on the query vector Q, key vector K and value vector V of the single token sequence, which is expressed as:

[0149]

[0150] Where, the subscript T is the matrix transpose, QK T is the dot product of the query vector and the key vector, d k is the dimension of the key vector, and the softmax function is used to normalize the calculation results;

[0151] Finally, the multi-head self-attention mechanism outputs the concatenated features of multiple heads:

[0152] Multi Head(Q,K,V)=Concat(head1,head2,...,head h )ω o

[0153]

[0154] Where h is the number of self-attention heads, j is the index of each self-attention head, ω O is the output weight matrix, is the weight matrix corresponding to the query vector in the j-th self-attention head, is the weight matrix corresponding to the key vector in the j-th self-attention head, is the weight matrix corresponding to the median vector of the j-th self-attention head.

[0155] Step 524: Process the weighted features of the single token sequence using a feed-forward neural network (FFN) to obtain a radar echo classification result.

[0156] The feed-forward neural network (FFN) consists of two linear layers (i.e., fully connected layers). The ReLU activation function is used between the two linear layers. Residual connections are added before and after the feed-forward neural network (i.e., the input weighted features are directly added to the output of the second linear layer) to enhance the training stability of the model. The feed-forward neural network FFN is expressed as:

[0157] FFN(e)=Linear(ReLU(Linear(e)))+e

[0158] Where e is the weighted feature of a single token sequence, Linear(e) is the first-layer linear transformation, and Linear(ReLU(Linear(x))) is the second-layer linear transformation.

[0159] The weighted features of the token sequence are processed using the feedforward neural network FFN, specifically including:

[0160] After the weighted features of the token sequence enter the feedforward neural network FFN, the feedforward neural network FFN performs the first linear transformation on the weighted features through matrix multiplication, and uses ReLU to activate the linear transformation output of the previous layer. The role of ReLU is to speed up training and improve the expressiveness of the model. The result after the activation function is then subjected to a second linear transformation to map the features to the target space.

[0161] Step 53: After each training of a predetermined number of batches, the classification network model is evaluated and a version is saved using the second validation set. After the training is completed, the version with the best performance on the validation set is used as the optimal model.

[0162] In this embodiment, gradient descent is used to update model parameters during the training of a radar echo classification network model to reduce the value of the loss function and optimize model performance. Specifically, after data is input into the neural network model, the model performs calculations layer by layer through a forward propagation process, ultimately outputting a prediction result. At this point, the system compares the model's predicted output with the true label of each sample in the dataset, using a preset loss function to quantify the difference between the predicted result and the true value, namely the loss value. This loss value reflects the model's current predictive performance. Using the loss value backpropagation technique, the impact of each model parameter on the loss value, namely the gradient, is calculated layer by layer, from the output layer to the input layer. This process helps determine how each parameter should be adjusted to reduce the loss value and improve the model's predictive accuracy.

[0163] In step 6, the trained generator, the reverse denoising part of the diffusion network module, and the classification network model are retained, and the generator, the reverse denoising part of the diffusion network module, and the classification network model are deployed to the carrier device to denoise and classify the actual collected radar echoes.

[0164] In this embodiment, the carrier device may include an edge device (such as an embedded GPU, FPGA, ARM chip, etc.), an edge server, a cloud server, etc., and the generator and classifier can be deployed on the edge side to achieve fast reasoning.

[0165] Example 1:

[0166] According to the method of steps 1 to 2 in this application, the radar echo data of the target object is collected and set as a training sample. Some data is extracted from the training sample as a test sample. According to the method of step 3, a fusion network is established, and the fusion network is trained using the training sample to obtain a converged fusion network. The trained fusion network is used as a test object, and the process of its processing of radar echo data is visualized.

[0167] The trained fusion network is tested using the test samples, and the results are as follows: Figure 2 The radar noise reduction process shown, Figure 2 (a) is the image corresponding to the original RD matrix. From this figure, we can see that there is a lot of clutter interference in the original RD matrix. Figure 2 (b) is the image corresponding to the target RD matrix generated after processing by the conditional generative adversarial network module. From this figure, it can be seen that most of the clutter interference in the original RD matrix is ​​filtered out, leaving relatively clean data. Figure 2 (c) in the figure is the target RD matrix after adding sequence noise to the forward diffusion part of the diffusion network module. Figure 2 (d) is the target RD matrix after reverse recovery by the reverse denoising part in the diffusion network module. From this figure, it can be seen that the reverse denoising part can perform secondary denoising on the noisy target RD matrix to obtain relatively clean data. Figure 2 (e) is the reverse denoising part of the diffusion network module directly Figure 2 (b) in the figure is the image after secondary denoising (equivalent to the image obtained by processing the original RD matrix when only the generator and the reverse denoising part are retained, that is, the actual processing process). It can be seen from the figure that after double denoising by the generator in the generative adversarial network module and the reverse denoising part in the diffusion network module, the obtained RD matrix has a clear and clean background and prominent targets, reducing most of the clutter interference, and can be accurately identified by the subsequent classification network model.

[0168] Example 2: Deploy the trained fusion network and classification network models to the test carrier equipment on the airport runway, and use the carrier equipment to detect foreign objects on the airport runway; Figure 3 The following is an image of a foreign object entering the airport runway within a certain period of time. After the carrier device detects the foreign object, it reports the radar data to the host computer, and obtains the following: Figure 4 The positioning map shown in the figure contains the timestamp (hour: minute: second. millisecond), direction angle (azimuth, pitch), longitude and latitude, as well as the amplitude value of the foreign object, the distance amplitude with the background difference, the first 7th distance amplitude (the amplitude of the 7th bin in front of the target, the radar divides the detection distance direction into small intervals, each small interval is called a bin), the last 7th distance amplitude (the amplitude of the 7th bin behind the target), and the last 8th distance amplitude.

[0169] like Figure 4As shown in the figure, at the time stamp 01:40:38.102, the direction angle of the foreign object is 7.14° in azimuth and 27.29° in elevation. The longitude and latitude are 116438551.63 and 39515108.93 respectively. The current amplitude value, the amplitude of the distance difference from the background, the amplitude of the distance before the 7th, the amplitude of the distance after the 7th, and the amplitude of the distance after the 8th are 25514841.00, 25313490.00, 85047.00, 184175.00, and 225004.00 respectively. Figure 4 It can be seen that after fusion network denoising and classification network model classification and identification, the foreign object can be accurately detected. The host computer can accurately obtain the amplitude and position of the foreign object based on the target intensity information fed back by the fusion network (RD matrix contains target intensity information, which is used to obtain the amplitude value), the target position information fed back by the classification network model, and the posture information fed back by the carrier device.

[0170] The steps in this application can be adjusted in order, combined, and deleted according to actual needs.

[0171] The units in the device of the present application can be combined, divided and deleted according to actual needs.

[0172] Although the present application is disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and are not intended to limit the application of the present application. The scope of protection of the present application is defined by the appended claims and may include various modifications, alterations and equivalents made to the invention without departing from the scope and spirit of the present application.

Claims

1. A method for denoising and classifying radar echoes for airport runway foreign object detection (FOD), characterized in that: The method includes: Step 1: collect radar echo data of the target object, preprocess the radar echo data, and obtain the original RD matrix; Step 2: Setting a first training sample set based on the original RD matrix. Specifically, filtering and target labeling are performed on the original RD matrix, and the labeled RD matrix and the original RD matrix are used as training samples to generate the first training sample set. Step 3: Construct a fusion network. The fusion network includes a conditional generative adversarial network module, a diffusion network module, and an entropy evaluation mechanism for promoting symmetry in the generative adversarial network game. Specifically, it includes: Step 31: Construct a conditional generative adversarial network module including a generator and a discriminator. The discriminator adopts a patchGAN network. The generator learns the process of mapping the original RD matrix and the noise vector to the target RD matrix. The discriminator is used to perform block discrimination on the target RD matrix to determine whether each component block of the target RD matrix is ​​true, so as to guide the generator to generate data with real local details. At the same time, the L1 constraint function is introduced in the conditional generative adversarial network module. The L1 constraint function is used to constrain the overall difference between the target RD matrix output by the generator and the real RD matrix, so that the generator generates data with real global structure. Step 32: construct a diffusion network module at the back end of the conditional generative adversarial network module, and use the diffusion network module to perform secondary denoising on the target RD matrix output by the conditional generative adversarial network module; Step 33: Using the discriminator's judgment results on each component block of the target RD matrix as local entropy, a learnable entropy mapping function is constructed. The mapping result of the entropy mapping function is multiplied element-by-element by the original RD matrix and then added to the input of the generator as a perturbation term, forming an information feedback mechanism from the discriminator to the generator. This allows the generator to dynamically perceive the changing trend of the discriminator's judgment results on each component block, thereby improving the generator's response ability to areas with low confidence in the discriminator's output, and promoting symmetry in the generation-discrimination game process. Step 34, extracting training samples in batches from the first training sample set to train the fusion network and optimize model parameters; Step 4: Use the trained fusion network to perform denoising on all original RD matrices in the first training sample set, and use the processed RD matrices as training samples to generate a second training sample set; Step 5: constructing a radar echo classification network model based on a neural network, and training the radar echo classification network model using the second training sample set; In step 6, the trained generator, diffusion network module, and classification network model are retained to perform noise reduction and classification on the actually collected radar echoes.

2. The method for denoising and classifying radar echoes for airport runway foreign object detection (FOD) according to claim 1, wherein: The step 31 specifically includes: The original RD matrix is ​​recorded as x, the noise vector is recorded as z, and the target RD matrix output by the generator G is recorded as The true RD matrix with labeled information is recorded as y, and the output of the patchGAN network is a tensor with a single element D i Represents the discriminator's discrimination result on a single component block in the target RD matrix, i = 1, 2, ..., N, and sets the target loss function of the generator G and the discriminator D to: Where, E x,y is the expectation of mapping the original RD matrix to the target RD matrix, E x,z is the expectation of mapping the noise vector to the target RD matrix, D i (x,y) is the discriminator’s judgment result on the i-th block in the real RD matrix, G(x,z) is the output of the generator G, and D i (x, G(x, z)) is the discriminator’s judgment result on the i-th block in the target RD matrix; The loss of the L1 constraint function is set to: L L1 (G)=E x,y,z [||G(x,z)-y||1] Where, E x,y,z is the expectation of the original RD matrix x, the true RD matrix and the noise vector z under the joint distribution; the loss of the L1 constraint function is superimposed on the target loss function to obtain the final target loss function of the conditional generative adversarial network module: Where λ is the gain parameter.

3. The method for denoising and classifying radar echoes for airport runway foreign object detection (FOD) according to claim 1, wherein: In step 32, the diffusion network module includes a forward diffusion part and a reverse denoising part. During training, the forward diffusion part is used to add noise to the target RD matrix output by the generator, and the reverse denoising part is used to learn the process of mapping the noisy target RD matrix into a noise-free matrix. During inference, only the reverse denoising part is used to directly denoise the target RD matrix output by the generator to obtain a real and clean target RD matrix.

4. The method for airport runway foreign object detection (FOD) radar echo denoising and classification according to claim 3, wherein: The forward diffusion part is used to add sequence noise to the target RD matrix output by the conditional generation adversarial network module according to the training cycle order. For an initially noise-free target RD matrix In each training cycle t∈{1,2,...,M}, noise is added as follows: Where, α t and σ t is a hyperparameter adjustable according to time domain information, M is the length of the training cycle, is the target RD matrix with sequence noise added at period t, and ε is the noise term of standard normal distribution; The reverse denoising part is used to reversely restore the original target RD matrix and learn the recovery process. Its mathematical model is set as The loss function of the reverse denoising part is set as: Where, E ε,t is the joint sampling expectation of noise ε and period t.

5. The method for airport runway foreign object detection (FOD) radar echo denoising and classification according to claim 4, characterized in that: The step 33 specifically includes: The output tensor of the discriminator at training cycle t is D t , tensor D t The local entropy contained is D 1,t ,D 2,t ,...,D N,t , using each local entropy as input to construct a learnable entropy mapping function: w t =W(D t-1 ) Where w t is the function value corresponding to each local entropy in the tth training cycle, D t-1 is the local entropy value in the t-1th training cycle, W(·) represents the entropy mapping function, where D t-1 When it approaches the maximum entropy state, W(D t-1 ) tends to a preset constant tensor The mapping result w of the entropy mapping function t After being multiplied element-by-element with the original RD matrix, it is added to the input of the generator G as a perturbation term, so that the input of the generator G is: x t =x0+x0W(D t-1 ) Where x0 is the original RD matrix before the superposition of the perturbation term, x t is the original RD matrix after superimposing the perturbation term, x0W(D t-1 ) is a disturbance term; at the same time, a disturbance dynamic control mechanism is set in the process of superimposing the disturbance term, and the disturbance dynamic control mechanism is used to calculate the superimposed disturbance expectation based on the value of the entropy mapping function in the current training cycle, and whether the current training cycle needs to superimpose the disturbance term is determined according to the size of the superimposed disturbance expectation. When the calculated superimposed disturbance expectation is greater than or equal to the preset expected value, the disturbance term is superimposed. When the calculated superimposed disturbance expectation is less than the preset expected value, the disturbance term is not superimposed. The calculation formula of the superimposed disturbance expectation is: Where, is the position corresponding to the current training cycle, and E[·] is the expectation.

6. The method for denoising and classifying radar echoes for airport runway foreign object detection (FOD) according to claim 1, wherein: The step 34 specifically includes: Set the corresponding condition information c according to the target object to be detected, extract training sample data in batches from the first training set, and perform multiple rounds of training on the fusion network according to the batches. Each round of training specifically includes: The disturbance dynamic control mechanism is used to determine whether to inject disturbance. If not, the original RD matrix x, noise vector z and corresponding condition information c are used as the input of the generator. If so, the entropy mapping function is used to calculate the disturbance term based on the local entropy output by the discriminator, and the disturbance term is superimposed on the original RD matrix. Together with the original RD matrix x, noise vector z and corresponding condition information c, the generator is asked to output the target RD matrix y and learn the process of mapping the original RD matrix x and noise vector z to the target RD matrix. When the disturbance term is first injected, the discriminator has no output, D t-1 From the mean The RD matrix and the target RD matrix are sampled element by element from a normal distribution with a standard deviation of 0.001; the labeled RD matrix and the target RD matrix are used as the input of the discriminator, which is used to perform block discrimination and output the local entropy corresponding to each component block of the target RD matrix. The target loss function of the conditional generative adversarial network module is simultaneously calculated. By maximizing the loss of the discriminator, its ability to accurately distinguish between the true RD matrix and the target RD matrix is ​​improved. By minimizing the loss of the generator, the target RD matrix output by it is gradually close to the true one, and the network parameters of the generator and discriminator are updated by gradient descent. Each time the conditional generative adversarial network module is trained, the diffusion network module is trained once using the target RD matrix output by the generator as a sample. Specifically, the forward diffusion part is first used to add sequence noise to the target RD matrix in the order of the training cycle, and then the reverse denoising part is used to denoise the target RD matrix after adding sequence noise. The loss function of the reverse denoising part is calculated, and the output data is gradually approached to the original target RD matrix by minimizing the loss of the reverse denoising part. The network parameters of the reverse denoising part are updated by gradient descent.

7. The method for airport runway foreign object detection (FOD) radar echo denoising and classification according to claim 1, wherein: The step 2 specifically includes: Step 21, performing time domain clutter interference suppression on the original RD matrix, and marking the location and type of the target object in the RD matrix after clutter suppression; Step 22: Normalize the labeled RD matrix and the original RD matrix to the range of [0, 1] respectively, use the two types of normalized matrices as training samples to generate a first training sample set, and divide the first training sample set into a first training set, a first validation set, and a first test set; wherein the first training sample set contains multiple groups of training samples, each group of training samples includes the corresponding original RD matrix and the real RD matrix with labeled information.

8. The method for airport runway foreign object detection (FOD) radar echo denoising and classification according to claim 1, wherein: The step 5 specifically includes: Step 51, shuffling the data in the second training set and dividing the shuffled data into batches; Step 52: Perform multiple rounds of training on the classification network model using different batches of data. After each batch of data training is completed, a new batch of data is extracted from the second training set and trained again until the training meets the preset stopping condition. Step 53: After each training of a predetermined number of batches, the classification network model is evaluated and a version is saved using the second validation set. After the training is completed, the version with the best performance on the validation set is used as the optimal model.

9. The method for airport runway foreign object detection (FOD) radar echo denoising and classification according to claim 8, wherein: In step 52, the process of training the classification network model using a single batch of data specifically includes: Step 521: Use the second training set as input to the classification network model, and use the input layer of the classification network model to perform feature extraction on a single RD matrix in the second training set to obtain a set of token sequences. Step 522: Add position codes to the single token sequence by element-by-element addition. In step 523, the single token sequence after adding the position encoding is input into the multi-head self-attention mechanism, and the multi-head self-attention mechanism is used to perform further feature extraction to obtain the weighted features of the single token sequence. Specifically: First, use each self-attention head to extract features to obtain the query vector Q, key vector K, and value vector V of a single token sequence. Then, based on Q, K, and V of a single token sequence, the attention matrix of the single token sequence is calculated, which can be expressed as: In the formula, the subscript T is the matrix transpose, and the softmax function is used to normalize the calculation results; finally, the multi-head self-attention mechanism outputs the splicing features of multiple heads: MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W o Where h is the number of self-attention heads, j is the index of each self-attention head, ω O is the output weight matrix, is the weight matrix corresponding to the query vector in the j-th self-attention head, is the weight matrix corresponding to the key vector in the j-th self-attention head, is the weight matrix corresponding to the median vector of the j-th self-attention head; In step 524, the weighted features of the single token sequence are processed using a feedforward neural network (FFN) to ultimately obtain a classification result of the radar echo.

10. The method for airport runway foreign object detection (FOD) radar echo denoising and classification according to claim 9, characterized in that: The feedforward neural network FFN is expressed as: FFN(e)=Linear(ReLU(Linear(e)))+e Where e is the weighted feature of a single token sequence, Linear(e) is the first-layer linear transformation, and Linear(ReLU(Linear(x))) is the second-layer linear transformation. The weighted features of a single token sequence are processed using a feedforward neural network (FFN), including: After the weighted features of the token sequence enter the feedforward neural network FFN, the feedforward neural network FFN performs the first linear transformation on the weighted features through matrix multiplication, uses ReLU to activate the linear transformation output of the previous layer, and then performs a second linear transformation on the result after the activation function, and then maps the features to the target space.

Images