Intelligent separation method for multiple collision ADS-B signals

Abstract

The application belongs to the technical field of satellite communication, and discloses an intelligent separation method of multiple collision ADS-B signals, which comprises the following steps: performing normalization processing on multiple collision data received by a satellite, performing waveform feature mapping, outputting a feature vector, inputting the feature vector into a residual down-sampling processing module to obtain deep features, performing up-sampling operation to obtain up-sampled features, fusing the up-sampled features with skip features to obtain fused features, performing residual convolution refinement on the fused features to output a new refined feature vector, converting the new refined feature vector into a one-dimensional vector, performing multilayer perceptron global mapping and signal separation, and restoring the signal into a final numerical signal, i.e., an ADS-B digital signal sequence. The application significantly enhances the feature expression capability and time domain modeling depth of the model, and overcomes the limitation that single-layer convolution has limited modeling capability when processing sequences with long-distance dependent features.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication technology, specifically relating to an intelligent separation method for multiple collision ADS-B signals. Background Technology

[0002] Automatic Dependent Surveillance-Broadcast (ADS-B) is one of the key technologies in modern air traffic control. However, due to insufficient coverage of terrestrial communication networks in remote areas such as oceans and deserts, long-distance surveillance must rely on satellite communication systems. Satellite communication offers wide coverage and strong global service capabilities, but channel resources are limited, and the problem of overlapping and colliding multiple signals on satellite channels is particularly prominent. In existing technologies, separating double-overlapping signals is already very difficult, and for three ADS-B signals simultaneously transmitted by three aircraft—i.e., triple-collision signals—the power ratio and timing relationships are even more complex and variable. Traditional time-domain decision or phase demodulation methods are almost incapable of directly and effectively processing this type of data.

[0003] Extensive research has been conducted by scholars both domestically and internationally on the ADS-B signal collision problem, resulting in various separation algorithms. Generally, existing methods can be categorized into two main types: those based on multi-antenna arrays and those based on single-antenna receivers. For the former, antenna arrays with spatial dimensions can separate overlapping signals based on differences in direction of arrival (DOA). For example, existing technologies have proposed Projection Algorithm (PA) and Extended Projection Algorithm (EPA), which utilize orthogonal projection separation of overlapping Mode S signals received by multiple antennas to achieve de-aliasing of secondary radar response signals. For higher-order overlap scenarios, existing technologies have proposed Iterative Projection Algorithm (IPA) to achieve blind signal separation. Existing technologies also treat the ADS-B signal separation problem as an optimization problem, using alternating direction multipliers to address the non-convex blind adaptive beamforming problem. However, in many practical scenarios, such as spaceborne ADS-B, limitations in satellite platform size and cost necessitate the use of a single-antenna receiver, making it impossible to deploy large antenna arrays. Therefore, a new separation algorithm is urgently needed to solve the problem of accurate separation of three overlapping ADS-B signals in a spaceborne system. Summary of the Invention

[0004] To address the aforementioned technical issues, this application proposes an intelligent separation method for multi-collision ADS-B signals. This method extracts local features of the input signal through multiple downsampling convolutional modules, gradually restores the temporal resolution through corresponding upsampling modules, and fuses detailed features from the encoding stage during the decoding process. The output end employs a multilayer fully connected network (MLP) structure to map high-dimensional features to the existence probabilities of the three signals. Finally, threshold decision or binarization methods are used to convert the probability sequence into three independent ADS-B level signals.

[0005] To achieve the above objectives, this application employs the following technical solution:

[0006] This application discloses an intelligent separation method for multi-collision ADS-B signals, which is implemented using the SplitNet-3 model. The intelligent separation method specifically includes the following steps:

[0007] Step 1, Normalization: Normalize the ADS-B multi-collision data received from satellites. Normalization processing is performed on the ADS-B multi-collision data. Each signal in Normalizing to the range [0, 1], the amplitudes of all signals are normalized to the range [0, 1] to obtain the normalized data. ;

[0008] Step 2, Waveform Feature Mapping: Map the normalized data obtained in Step 1. The normalized data is processed through a one-dimensional convolutional layer of the encoder. The transformation to a multi-channel feature space outputs the feature vectors of the multi-channel feature space.

[0009] Step 3, Residual downsampling block processing: Input the feature vector obtained in step 2 into the residual downsampling processing module of the encoder, and obtain deep features with global perception capability after multiple residual downsamplings;

[0010] Step 4: Initialize upsampling: Use the deep features with global perception capabilities as the initial input features of the decoder, and perform upsampling operation starting from the bottleneck layer of the decoder to obtain upsampled features;

[0011] Step 5, Skip Feature Fusion: The upsampled features obtained in Step 4 are fused with the corresponding layer skip features stored in the encoder in the channel dimension to obtain the fused features;

[0012] Step 6, Residual Convolution Refinement: The features obtained after fusion in Step 5 are processed by convolution using the same residual block structure as the encoder, and a new refined feature vector is output.

[0013] Step 7: Decode the feature vector: Perform column-major vectorization on the new refined feature vector output in Step 6 and convert it into a one-dimensional vector;

[0014] Step 8, Multilayer Perceptron Global Mapping: Apply two layers of linear transformation with PReLU activation function to the one-dimensional vector obtained in Step 7, and then apply another layer of linear mapping to obtain the output vector;

[0015] Step 9, Signal Separation: Transform the output vector obtained in Step 8 into a tensor, apply the sigmoid function to each element of the tensor to obtain the existence probability of the three separated signals at each sampling point, decode the sequence of the existence probability of the three separated signals at each sampling point, and recover the final numerical signal, i.e., the ADS-B digital signal sequence.

[0016] A further improvement of this application is that, in step 2, the encoder extracts the normalized data through layer-by-layer convolution operations. The deep features are extracted, and the temporal resolution is gradually reduced, assuming the normalized data... The sampling length is , denoted as: The initial input is a single channel, that is Normalized data First, a one-dimensional convolutional layer is used to increase the number of channels from 1 to... Achieve the transformation from normalized data Mapping to a multi-channel feature space, outputting the feature vectors of the multi-channel feature space. Represented as:

[0017]

[0018] in, The kernel size is [size]. For stride size, Number of output channels Input the number of channels. express Activation function, if input It is a positive number. It's the input itself; if the input... It is a negative number. It is 0.

[0019] A further improvement of this application is that step 3 specifically includes the following steps:

[0020] Step 3.1, Residual Block Convolution: First, perform two one-dimensional convolutions on the encoder's residual downsampling processing module to obtain the main branch output. :

[0021]

[0022] in, This is a characteristic of the current stage. This represents the number of signal sampling points.

[0023] Step 3.2, Batch Normalization: Output the main branch Batch normalization is performed using the following formula:

[0024]

[0025] in, This is the normalized output of the main branch;

[0026] Step 3.3, Transform and Summation: When the number of input channels... Not equal to output channel At that time, a 1×1 convolution is used to process the features of the current stage. Project. For projection features, when the number of input channels Equal to output channel hour, The normalized main branch output With the transformed Add element by element, and through The output of the activated residual downsampling processing module is a deep feature with global perception capability. .

[0027] A further improvement of this application is that, in step 4, the decoder includes a total of The nth upsampling step, each upsampling step is denoted as the nth upsampling step. The layers proceed sequentially from the bottleneck layer to the decoder's output layer, as follows:

[0028]

[0029] in, For upsampling features, , For the first The current stage characteristics of the layer.

[0030] A further improvement of this application is that, in step 5, the upsampled features are compared with the corresponding layer skip features stored in the encoder. By fusing along the channel dimension, we obtain the fused features:

[0031]

[0032] in, The characteristics after fusion This is an upsampling feature.

[0033] A further improvement of this application is that, in step 6, the fused features obtained in step 5... By employing two convolutional layers and using a ResidualBlock1d structure with the encoder, a new refined feature vector is obtained. .

[0034] A further improvement of this application is that, in step 7, the decoder's output is set to... ,in Indicates the number of channels. For the sampling length, the new refined feature vector Column-major vectorization, the formula is as follows:

[0035]

[0036] in, This is the result of column-major vectorization.

[0037] A further improvement of this application is that, in step 8, two linear transformations with PReLU activation functions are sequentially applied to the one-dimensional vector obtained in step 7, followed by a linear mapping to obtain the output vector. The formula is as follows:

[0038] , ,

[0039] , ,

[0040] , ,

[0041] in, This represents a linear transformation with the PReLU activation function. , and The weights correspond to the first layer, the second layer, and the output layer, respectively. , and These correspond to the biases of the first layer, the second layer, and the output layer, respectively. , These represent the intermediate layer outputs after being transformed by the activation function.

[0042] A further improvement of this application is that step 9 specifically includes the following steps:

[0043] Step 9.1: Output the vector from Step 8 Transform into tensors :

[0044]

[0045] in, The sampling length;

[0046] Step 9.2, for tensors By applying the sigmoid function element by element, the existence probability of the three separated signals at each sampling point is obtained. :

[0047]

[0048] in, The tensor of the three separated signals at each sampling point; The number of separate signal paths, , The sampling length;

[0049] Step 9.3: Based on the empirical threshold, a hard decision will be made to determine the probability of existence. The signal is converted to a 0 / 1 level sequence, and three independent ADS-B digital signal sequences are inferred and reconstructed based on the SplitNet-3 model.

[0050] A further improvement of this application is that the SplitNet-3 model includes:

[0051] Encoder: Implements sequence compression and feature extraction. The encoder extracts the temporal features of the signal step by step through multiple convolutional layers. The temporal features include convolutional layers, pooling layers, and residual blocks. Each residual block is followed by a max pooling operation to achieve downsampling.

[0052] Decoder: Implements feature restoration and signal reconstruction, including upsampling and residual blocks;

[0053] A cross-layer skip connection is set up between the encoder and the decoder to splice and fuse the decoder upsampling results with the features of the corresponding layer of the encoder in the channel dimension;

[0054] The output is a set of three fully connected multilayer perceptron (MLP) output heads.

[0055] The beneficial effects of this application are:

[0056] 1. The introduction of residual connections in this application alleviates the gradient vanishing problem commonly encountered in deep networks during training, making the model more likely to converge. At the same time, it retains more temporal features related to the original signal in the output layer, improving the accuracy and fidelity of the reconstructed waveform.

[0057] 2. Unlike general signal separation networks, the SplitNet-3 model proposed in this application is customized for the highly regular structure of ADS-B signals. Specifically, the SplitNet-3 model combines the upsampling and downsampling paths of U-Net and a multi-level residual unit structure in the one-dimensional sequence input, simultaneously modeling the local pulse shape and global time series dependencies at different scales, significantly enhancing the model's feature representation ability and temporal modeling depth.

[0058] 3. This application introduces a multi-scale downsampling mechanism and cascades multiple U-Net substructures, enabling the network to have the ability to go from global modeling to local reconstruction, thereby achieving complete modeling and effective reconstruction of ADS-B messages up to 112 bits, overcoming the limitation of single-layer convolution in terms of modeling ability when processing sequences with long-distance dependent features. Attached Figure Description

[0059] Figure 1 This is a flowchart of the intelligent separation method for multiple collision ADS-B signals in this application.

[0060] Figure 2 It is a time-domain plot of the triple collision signal.

[0061] Figure 3 This is a schematic diagram of the signal 1 separation result.

[0062] Figure 4 This is a schematic diagram of the signal 2 separation result.

[0063] Figure 5 This is a schematic diagram of the signal separation results.

[0064] Figure 6 This is a graph showing the change in bit error rate as a function of signal-to-noise ratio at different reception times.

[0065] Figure 7 It is a curve showing the average bit error rate as a function of the signal-to-noise ratio.

[0066] Figure 8 This is a schematic diagram of the SplitNet-3 model in this application. Detailed Implementation

[0067] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the present invention. That is, in some embodiments of the present invention, these practical details are not essential. In addition, for the sake of simplicity, some conventional structures and components will be shown in the drawings in a simple schematic manner.

[0068] This application presents an intelligent separation method for multi-collision ADS-B signals, implemented using the SplitNet-3 model. Figure 8 As shown, the SplitNet-3 model includes an encoder for sequence compression and feature extraction. The encoder progressively extracts temporal features of the signal through multiple convolutional layers, including convolutional layers, pooling layers, and residual blocks. Each residual block is followed by a max pooling operation with a stride of 2 to achieve downsampling. This results in the output sequence after the "residual block + pooling" module having its time length halved while the number of channels doubles, for a total of 9 groups. The decoder performs feature restoration and signal reconstruction. Its structure is symmetrical to the encoder, including upsampling and residual blocks, restoring the sequence length to twice its original size and progressively reducing the number of channels, for a total of 9 groups. This enables reconstructive prediction of signal location. A cross-layer skip connection is established between the encoder and decoder, concatenating and fusing the decoder's upsampling results with the features of the corresponding layer in the encoder along the channel dimension. This allows the decoder to simultaneously utilize the global semantic information and high-resolution details captured by the encoder, compensating for information loss during downsampling. The output is a multilayer perceptron (MLP) output head composed of three fully connected layers. This structure can explicitly capture long-range dependencies across time steps while maintaining controllable model parameters.

[0069] like Figure 1 As shown, the intelligent separation method for multiple collision ADS-B signals specifically includes the following steps:

[0070] Step 1, Normalization: Normalize the ADS-B multi-collision data received from satellites. Normalization processing is performed on the ADS-B multi-collision data. Each signal in Normalizing to the range [0, 1], the amplitudes of all signals are normalized to the range [0, 1] to obtain the normalized data. The formula is as follows:

[0071]

[0072] in It is the original signal. It is the maximum amplitude in the signal. It is the minimum amplitude in the signal.

[0073] Step 2, Waveform Feature Mapping: Map the normalized data obtained in Step 1. The normalized data is processed through a one-dimensional convolutional layer of the encoder. The transformation to a multi-channel feature space outputs the feature vectors of the multi-channel feature space.

[0074] In step 2, the encoder extracts the normalized data through layer-by-layer convolution operations. The deep features are extracted, and the temporal resolution is gradually reduced, assuming the normalized data... The sampling length is , denoted as: The initial input is a single channel, that is Normalized data First, a one-dimensional convolutional layer is used to increase the number of channels from 1 to... Achieve the transformation from normalized data Mapping to a multi-channel feature space, outputting the feature vectors of the multi-channel feature space. Represented as:

[0075]

[0076] in, The kernel size is 3. For a stride size of 1, Number of output channels Input the number of channels. express Activation function, if input It is a positive number. It's the input itself; if the input... It is a negative number. It is 0.

[0077] Step 3, Residual Downsampling Block Processing: The feature vector obtained in Step 2 is input into the encoder's residual downsampling processing module. After multiple residual downsampling operations, deep features with global perception capabilities are obtained. This specifically includes the following steps:

[0078] Step 3.1, Residual Block Convolution: First, perform two one-dimensional convolutions on the encoder's residual downsampling processing module to obtain the main branch output. :

[0079]

[0080] in, This is a characteristic of the current stage. This represents the number of signal sampling points.

[0081] Step 3.2, Batch Normalization: Output the main branch Batch normalization is performed using the following formula:

[0082]

[0083] in, This is the normalized output of the main branch;

[0084] Step 3.3, Transform and Summation: When the number of input channels... Not equal to output channel At that time, a 1×1 convolution is used to process the features of the current stage. Project. For projection features, when the number of input channels Equal to output channel hour, The normalized main branch output With the transformed Add element by element, and through The output of the activated residual downsampling processing module is a deep feature with global perception capability. .

[0085] Step 4, Initial Upsampling: Using deep features with global perception capabilities as the initial input features of the decoder, upsampling is performed starting from the bottleneck layer of the decoder to obtain upsampled features, preparing for the detailed features in the subsequent fusion encoding stage. In this step, the decoder contains a total of... The nth upsampling step, each upsampling step is denoted as the nth upsampling step. The layers proceed sequentially from the bottleneck layer to the decoder's output layer, as follows:

[0086]

[0087] in, For upsampling features, , For the first The current stage characteristics of the layer.

[0088] Step 5, Skip Feature Fusion: Skip feature fusion combines shallow and deep features through specific connection methods, enabling the model to utilize both detailed and semantic information. This involves combining the upsampled features with the corresponding layer skip features stored in the encoder. By fusing along the channel dimension, we obtain the fused features:

[0089]

[0090] in, The characteristics after fusion This is an upsampling feature.

[0091] Step 6, Residual Convolution Thinning: The fused features obtained in Step 5 High-resolution detail information is preserved, and the fused features are... By employing two convolutional layers and using a ResidualBlock1d structure with the encoder, a new refined feature vector is obtained. .

[0092] Step 7: Decode the feature vector: Perform column-major vectorization on the new refined feature vector output from Step 6 and convert it into a one-dimensional vector; in this step, let the decoder's final output be... ,in Indicates the number of channels. For the sampling length, the new refined feature vector Column-major vectorization, the formula is as follows:

[0093]

[0094] in, This is the result of column-major vectorization.

[0095] Step 8, Multilayer Perceptron Global Mapping: Multilayer perceptron global mapping refers to the process by which an MLP maps the global information of the input data—that is, the features of the entire input space—to the output space through multiple layers of nonlinear transformations. This mapping capability stems from the fully connected structure and nonlinear activation function of the MLP, enabling it to capture global dependencies, complex patterns, and abstract features in the input data. This is the core capability of MLP in handling nonlinear problems and implementing complex tasks.

[0096] Apply two linear transformations with PReLU activation functions to the one-dimensional vector obtained in step 7, followed by a linear mapping layer to obtain the output vector. The formula is as follows:

[0097] , ,

[0098] , ,

[0099] , ,

[0100] in, This represents a linear transformation with the PReLU activation function. , and The weights correspond to the first layer, the second layer, and the output layer, respectively. , and These correspond to the biases of the first layer, the second layer, and the output layer, respectively. , These represent the intermediate layer outputs after being transformed by the activation function.

[0101] Step 9, Signal Separation: Transform the output vector obtained in Step 8 into a tensor, apply a sigmoid function to each element of the tensor to obtain the existence probability of the three separated signals at each sampling point, decode the sequence of existence probabilities of the three separated signals at each sampling point, and recover the final numerical signal, i.e., the ADS-B digital signal sequence. This specifically includes the following steps:

[0102] Step 9.1: Output the vector from Step 8 Transform into tensors :

[0103]

[0104] in, The sampling length;

[0105] Step 9.2, for tensors By applying the sigmoid function element by element, the existence probability of the three separated signals at each sampling point is obtained. :

[0106]

[0107] in, The tensor of the three separated signals at each sampling point; The number of separate signal paths, , The sampling length;

[0108] Step 9.3: Based on the empirical threshold, a hard decision will be made to determine the probability of existence. The signal is converted to a 0 / 1 level sequence, and three independent ADS-B digital signal sequences are inferred and reconstructed based on the SplitNet-3 model.

[0109] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A smart separation method for multi-collision ADS-B signals, characterized in that: The separation method is implemented using the SplitNet-3 model, and the intelligent separation method specifically includes the following steps: Step 1, Normalization: Normalize the ADS-B multi-collision data received from satellites. Normalization processing is performed on the ADS-B multi-collision data. Each signal in Normalization to the range [0, 1]: Normalize the amplitude of all signals to the range [0, 1] to obtain the normalized data. ; Step 2, Waveform Feature Mapping: Map the normalized data obtained in Step 1. The normalized data is processed through a one-dimensional convolutional layer of the encoder. The transformation to a multi-channel feature space outputs the feature vectors of the multi-channel feature space. Step 3, Residual downsampling block processing: Input the feature vector obtained in step 2 into the residual downsampling processing module of the encoder, and obtain deep features with global perception capability after multiple residual downsamplings; Step 4: Initialize upsampling: Use the deep features with global perception capabilities as the initial input features of the decoder, and perform upsampling operation starting from the bottleneck layer of the decoder to obtain upsampled features; Step 5, Skip Feature Fusion: The upsampled features obtained in Step 4 are fused with the corresponding layer skip features stored in the encoder in the channel dimension to obtain the fused features; Step 6, Residual Convolution Refinement: The features obtained after fusion in Step 5 are processed by convolution using the same residual block structure as the encoder, and a new refined feature vector is output. Step 7: Decode the feature vector: Perform column-major vectorization on the new refined feature vector output in Step 6 and convert it into a one-dimensional vector; Step 8, Multilayer Perceptron Global Mapping: Apply two layers of linear transformation with PReLU activation function to the one-dimensional vector obtained in Step 7, and then apply another layer of linear mapping to obtain the output vector; Step 9, Signal Separation: Transform the output vector obtained in Step 8 into a tensor, apply the sigmoid function to each element of the tensor to obtain the existence probability of the three separated signals at each sampling point, decode the sequence of the existence probability of the three separated signals at each sampling point, and recover the final numerical signal, i.e., the ADS-B digital signal sequence.

2. The intelligent separation method for multiple collision ADS-B signals according to claim 1, characterized in that: In step 2, the encoder extracts the normalized data through layer-by-layer convolution operations. The deep features are extracted, and the temporal resolution is gradually reduced, assuming the normalized data... The sampling length is , denoted as: The initial input is a single channel, that is Normalized data First, a one-dimensional convolutional layer is used to increase the number of channels from 1 to... Achieve the transformation from normalized data Mapping to a multi-channel feature space, outputting the feature vectors of the multi-channel feature space. Represented as: ， in, The kernel size is [size]. For stride size, Number of output channels Input the number of channels. express Activation function, if input It is a positive number. It's the input itself; if the input... It is a negative number. It is 0.

3. The intelligent separation method for multiple collision ADS-B signals according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3.1, Residual Block Convolution: First, perform two one-dimensional convolutions on the encoder's residual downsampling processing module to obtain the main branch output. : ， in, This is a characteristic of the current stage. This represents the number of signal sampling points. Step 3.2, Batch Normalization: Output the main branch Batch normalization is performed using the following formula: ， in, This is the normalized output of the main branch; Step 3.3, Transform and Summation: When the number of input channels... Not equal to output channel At that time, a 1×1 convolution is used to process the features of the current stage. Project. For projection features, when the number of input channels Equal to output channel hour, The normalized main branch output With the transformed Add element by element, and through The output of the activated residual downsampling processing module is a deep feature with global perception capability. .

4. The intelligent separation method for multiple collision ADS-B signals according to claim 1, characterized in that: In step 4, the decoder contains a total of The nth upsampling step, each upsampling step is denoted as the nth upsampling step. The layers proceed sequentially from the bottleneck layer to the decoder's output layer, as follows: ， in, For upsampling features, , For the first The current stage characteristics of the layer.

5. The intelligent separation method for multiple collision ADS-B signals according to claim 4, characterized in that: In step 5, the upsampled features are compared with the corresponding layer skip features stored in the encoder. By fusing along the channel dimension, we obtain the fused features: ， in, The characteristics after fusion This is an upsampling feature.

6. The intelligent separation method for multiple collision ADS-B signals according to claim 1, characterized in that: In step 6, the fused features obtained in step 5 By employing two convolutional layers and using a ResidualBlock1d structure with the encoder, a new refined feature vector is obtained. .

7. The intelligent separation method for multiple collision ADS-B signals according to claim 6, characterized in that: In step 7, let the decoder's output be... ,in Indicates the number of channels. For the sampling length, the new refined feature vector Column-major vectorization, the formula is as follows: ， in, This is the result of column-major vectorization.

8. The intelligent separation method for multiple collision ADS-B signals according to claim 1, characterized in that: In step 8, the one-dimensional vector obtained in step 7 is subjected to two layers of linear transformation with PReLU activation function, followed by a layer of linear mapping to obtain the output vector. The formula is as follows: ， ， , ， ， ， in, This represents a linear transformation with the PReLU activation function. , and The weights correspond to the first layer, the second layer, and the output layer, respectively. , and These correspond to the biases of the first layer, the second layer, and the output layer, respectively. , These represent the intermediate layer outputs after being transformed by the activation function.

9. The intelligent separation method for multiple collision ADS-B signals according to claim 8, characterized in that: Step 9 specifically includes the following steps: Step 9.1: Output the vector from Step 8 Transform into tensors : ， in, The sampling length; Step 9.2, for tensors By applying the sigmoid function element by element, the existence probability of the three separated signals at each sampling point is obtained. : ， in, The tensor of the three separated signals at each sampling point; The number of separate signal paths, , The sampling length; Step 9.3: Based on the empirical threshold, a hard decision will be made to determine the probability of existence. The signal is converted to a 0 / 1 level sequence, and three independent ADS-B digital signal sequences are inferred and reconstructed based on the SplitNet-3 model.

10. A method for intelligent separation of multiple collision ADS-B signals according to claim 1 or 8, characterized in that: The SplitNet-3 model includes: Encoder: Implements sequence compression and feature extraction. The encoder extracts the temporal features of the signal step by step through multiple convolutional layers. The temporal features include convolutional layers, pooling layers, and residual blocks. Each residual block is followed by a max pooling operation to achieve downsampling. Decoder: Implements feature restoration and signal reconstruction, including upsampling and residual blocks; A cross-layer skip connection is set up between the encoder and the decoder to splice and fuse the decoder upsampling results with the features of the corresponding layer of the encoder in the channel dimension; The output is a set of three fully connected multilayer perceptron output heads.

Images