A method for separating overlapping signals of satellite-borne ADS-B

Abstract

The present invention belongs to the field of satellite communication technology, and discloses a method for separating overlapping signals of satellite-borne ADS-B, comprising the following steps: first, the signal receiving end pre-processes the signal; then, a composite neural network architecture combining a self-attention mechanism, a long short-term memory network and a fully connected layer is used to optimize the information flow and feature extraction process to perform deep feature learning, thereby realizing the recognition and extraction of overlapping signals. The present invention significantly improves the receiving capability of satellite-borne asynchronous burst signals, and is suitable for the application of future space-based passive monitoring systems.

Description

Technical Field

[0001] The invention belongs to the technical field of satellite communications, and in particular relates to a method for separating overlapping satellite-borne ADS-B signals. Background Art

[0002] Automatic Dependent Surveillance-Broadcast (ADS-B) is the main technology for air traffic control, but since the ground communication network cannot cover areas such as oceans, deserts, and rainforests, these areas cannot meet the requirements of the surveillance system. Satellite communication systems have the ability to provide long-distance, wide coverage, and global all-day services, and are an effective way to solve the above problems. Satellite-borne ADS-B technology came into being.

[0003] The wide coverage of satellite communication systems has greatly enhanced our ability to communicate in remote areas and even around the world, but this wide coverage has inevitably brought about the problem of signal collision. This is because, with the limited number of satellite communication channels and the increasing demand for users and services, multiple signals may overlap on the same communication channel, resulting in signal interference.

[0004] Traditional separation methods are not applicable on satellites, and an algorithm must be designed that can effectively solve the signal collision problem in underdetermined situations. Summary of the invention

[0005] In order to solve the defects of the prior art, the present invention proposes a method for separating overlapping satellite-borne ADS-B signals. The method is based on a separation algorithm of a self-attention neural network. The method captures long-distance dependencies through the self-attention neural network, which is particularly effective for processing and analyzing time series data. It can deeply analyze the intrinsic characteristics of overlapping signals and achieve more accurate signal separation by learning the mutual influence between different signals.

[0006] In order to achieve the above object, the present invention is achieved through the following technical solutions:

[0007] The present invention is a method for separating overlapping satellite-borne ADS-B signals, which specifically comprises the following steps:

[0008] Step 1: Normalization: Satellite received data Normalize the data Each signal in Normalize to the range of [0, 1], adjust the amplitude of all signals to the range of [0, 1] to obtain normalized data ;

[0009] Step 2: Vectorization: Normalize the data by multiplying it by a constant C Perform scale up and convert the scaled data into a feature vector set ;

[0010] Step 3: Positional Encoding: For the feature vector set obtained in step 2 Each vector in After complex position encoding, the encoded feature vector is obtained , through the encoded feature vector To encode the position information, providing the attention mechanism model with information about the position of each element in the satellite reception data input sequence;

[0011] Step 4: Self-attention neural network processing: After the complex position encoding in step 3 is completed, the encoded feature vector Input into the self-attention neural network to obtain the feature set processed by the self-attention neural network , the feature set after processing by the self-attention neural network Pass to the next self-attention neural network;

[0012] Step 5: Long short-term memory network processing: The feature set processed by the self-attention neural network obtained in step 4 Transformed into a feature representation suitable for classification tasks;

[0013] Step 6: Fully connected neural network processing: The fully connected layer is responsible for the final linear transformation of the features transferred from the long short-term memory network;

[0014] Step 7, signal separation: decode the high and low levels obtained after passing through the self-attention neural network, long short-term memory network and fully connected neural network to obtain the source signal.

[0015] A further improvement of the present invention is that: in step 3, the complex position code is used to provide time information and position information, and the feature vector set Each vector in Its position in the time dimension Add complex position encoding, the formula is as follows:

[0016]

[0017] in, Indicates the position of an element in a sequence. Represents the dimension, represents the total dimension of the position encoding vector, represents pi, Is an imaginary unit.

[0018] A further improvement of the present invention is that the step 4 specifically comprises the following steps:

[0019] Step 4.1, linear transformation of input: the encoded feature vector of the input Perform linear transformation to generate three new feature sets: query ,key Sum , the formula is as follows:

[0020]

[0021] in, is a trainable weight matrix;

[0022] Step 4.2: Calculate attention score: Using query With all the keys The compatibility score is calculated by the dot product between them, and then normalized by a normalization function (softmax function) to calculate the final attention score. , helps the self-attention neural network determine the relative importance of each element in the input data set, the attention score The calculation formula is as follows:

[0023]

[0024] in, is the dimension of the key vector, Represents key value The transpose of

[0025] Step 4.3, weighted sum: Add the attention score calculated in step 4.2 With value Perform weighted summation to obtain the final feature set processed by the self-attention neural network , that is, to integrate different input features according to the importance of the features, the formula is:

[0026]

[0027] Among them, the feature set after the self-attention neural network The different aspects and relationships of the input features are fused and passed to the next self-attention mechanism network.

[0028] A further improvement of the present invention is that in step 5, the long short-term memory network includes three modules: an input gate, a forget gate and an output gate. The feature set is transformed into Transformed into a feature representation suitable for classification tasks, specifically including the following steps:

[0029] Step 5.1. Forget Gate: Feature set processed by the self-attention neural network As the input sequence of the LSTM network, the input of each time step is represented as ,in is the time step. The LSTM network first uses the forget gate to determine the current moment How much information from the previous time step should be forgotten? Sigmoid function Compress the result of the linear combination to between 0 and 1 and output a forgetting factor :

[0030]

[0031] in, is the weight matrix of the forget gate, is the bias term, is the hidden state at the previous time step, is the feature vector after attention mechanism processing in The value of the time step;

[0032] Step 5.2, Input Gate: The input gate determines how much new information is taken from the current input is added to the current cell state. The output of the forget gate Control the old state The retention amount, the activation value of the input gate Controls the addition of update information generated by the current input and previous hidden state through the candidate value layer, as follows:

[0033]

[0034] in, represents the activation value of the input gate, represents the candidate value generated by the tanh function, Represents the new cell state, which is the old state and the new candidate value Combined to get, , Represents the weight matrix of the input gate, which is used to control the addition of new information. represents the bias term of the input gate, represents the weight matrix used to generate candidate states that may be added to the current cell state, Represents the bias term when generating candidate states; , : The outputs of the forget gate and the input gate control the forgetting of old state information and the addition of new information respectively;

[0035] Step 5.3, Output Gate: The output gate determines what information will be output from the cell state. Output to hidden state , the tanh function performs a nonlinear transformation on the updated unit state, and then passes through the output gate The adjustment of the output hidden state , the output hidden state of the LSTM network is the feature set after processing by the self-attention neural network The result of time series modeling, hidden state Used in subsequent neural network layers, such as fully connected layers or other decision layers. The formula is as follows:

[0036]

[0037] in, represents the bias term of the output gate, Indicates the new cell state.

[0038] A further improvement of the present invention is that the step 6 specifically comprises the following steps:

[0039] Step 6.1: Connect the fully connected layer to perform the final linear transformation on the features passed from the LSTM network to adapt to the dimension of the target output. The output dimension of the fully connected layer is 1 single neuron.

[0040] Step 6.2: At the end of the model, hide the output state of the LSTM network The linear transformation through the fully connected layer and Sigmoid activation is converted to the predicted probability, the formula is as follows:

[0041]

[0042] in, Represents an activation function. If the input is a positive number, the output is the input itself. If the input is a negative number, the output is 0. is the hidden state of the LSTM network at the last time step, and They are the weights and biases of the fully connected layer respectively;

[0043] Step 6.3: The output is processed by the Sigmoid activation function, and the formula is as follows:

[0044]

[0045] in, It is the value passed to the Sigmoid function and is the output of the ReLU function.

[0046] A further improvement of the present invention is that in step 7, the model composed of the self-attention neural network, the long short-term memory network and the fully connected neural network generates high and low levels of different signals by learning the characteristics of the signals, and converts the high and low levels back to their original ADS-B message format, thereby realizing accurate identification and separation of each signal source.

[0047] The beneficial effects achieved by the present invention are:

[0048] 1. The self-attention neural network used in the present invention can effectively separate the overlapping signals of satellite-borne S-mode ADS-B, improving the accuracy and efficiency of signal processing. Through the self-attention mechanism, the network model can automatically learn the dependencies between signals, thereby achieving more accurate signal recognition and separation in a complex signal environment;

[0049] 2. By adding a self-attention neural network in the feature extraction stage, the present invention can effectively extract the spatiotemporal characteristics of the signal and significantly improve the overall performance and stability of the system.

[0050] 3. Through complex position encoding, the position information can be naturally embedded in the imaginary part of the data, while the original data content is retained in the real part. Such separation allows the model to more clearly distinguish between the data content and its position information during processing, which helps to improve the model's understanding of the data structure.

[0051] 4. The present invention significantly improves the receiving capability of satellite-borne asynchronous burst signals and is suitable for the application of future space-based passive monitoring systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of the satellite-borne ADS-B overlapping signal separation method of the present invention.

[0053] Figure 2 These are the training and testing curves of random overlapping signals at a 20dB signal-to-noise ratio.

[0054] Figure 3 This is a graph of the bit error rate of random overlapping signals under a signal-to-noise ratio of 5 to 20dB.

[0055] Figure 4 It is a diagram of overlapping signals received by satellite.

[0056] Figure 5 It is the high and low levels of the source signal with earlier arrival time obtained after the overlapping signal is processed by the model.

[0057] Figure 6 It is the high and low levels of the source signal with later arrival time obtained after the overlapping signal is processed by the model. DETAILED DESCRIPTION

[0058] The following will disclose the embodiments of the present invention with drawings. For the purpose of clear description, many practical details will be described together in the following description. However, it should be understood that these practical details should not be used to limit the present invention. That is, in some embodiments of the present invention, these practical details are not necessary. In addition, for the purpose of simplifying the drawings, some conventional structures and components will be depicted in a simple schematic manner in the drawings.

[0059] like Figure 1 As shown, the present invention is a method for separating overlapping satellite-borne ADS-B signals, and the separation method specifically comprises the following steps:

[0060] Step 1: Normalization: Satellite received data Normalize the data Each signal in Normalize to the range of [0, 1], adjust the amplitude of all signals to the range of [0, 1] to obtain normalized data , the formula is as follows:

[0061]

[0062] in is the original signal, is the maximum amplitude in the signal, is the minimum amplitude in the signal.

[0063] Step 2: Vectorization: First, normalize the data Multiply by a constant C to enlarge the scale, the formula is as follows:

[0064]

[0065] Then, the rescaled data is taken Converted to feature vector, the formula is as follows:

[0066]

[0067] in, Represents the adjusted scale data points, is the scaled data, and the scaled data is converted into a feature vector set It is the input for the next step of processing.

[0068] Step 3: Positional Encoding: For the feature vector set obtained in step 2 Each vector in After complex position encoding, the encoded feature vector is obtained , through the encoded feature vector To encode the position information, providing the attention mechanism model with information about the position of each element in the satellite reception data input sequence;

[0069] In step 3, the complex position encoding is used to provide time information and position information, which is the feature vector set Each vector in Its position in the time dimension Add complex position encoding, the formula is as follows:

[0070]

[0071] in, Indicates the position of an element in a sequence. Represents the dimension, represents the total dimension of the position encoding vector, represents pi, is the imaginary unit, and the encoded eigenvector , which contains the position information and will be used in the next step of the self-attention mechanism processing.

[0072] Step 4: Self-attention neural network processing: After the complex position encoding in step 3 is completed, the encoded feature vector Input into the self-attention neural network to obtain the feature set processed by the self-attention neural network , the feature set after processing by the self-attention neural network Passed to the next self-attention neural network to enhance the model's understanding of the relationship between different positions. The processing steps are as follows:

[0073] Step 4.1, linear transformation of input: the encoded feature vector of the input Perform linear transformation to generate three new feature sets: query ,key Sum , the formula is as follows:

[0074]

[0075] in, is a trainable weight matrix;

[0076] Step 4.2: Calculate attention score: Using query With all the keys The compatibility score is calculated by the dot product between them, and then normalized by a normalization function (softmax function) to calculate the final attention score. , helps the self-attention neural network determine the relative importance of each element in the input data set, the attention score The calculation formula is as follows:

[0077]

[0078] in, is the dimension of the key vector, Represents key value The transpose of

[0079] Step 4.3, weighted sum: Add the attention score calculated in step 4.2 With value Perform weighted summation to obtain the final feature set processed by the self-attention neural network , that is, to integrate different input features according to the importance of the features, the formula is:

[0080]

[0081] Among them, the feature set after the self-attention neural network The different aspects and relationships of the input features are integrated and passed to the next self-attention mechanism network. It is the new feature matrix output after network layer processing, which integrates different aspects and relationships of the input features.

[0082] Step 5: Long short-term memory network processing: The feature set processed by the self-attention neural network obtained in step 4 Transformed into a feature representation suitable for classification tasks, the specific steps are as follows:

[0083] Step 5.1. Forget Gate: Feature set processed by the self-attention neural network As the input sequence of the LSTM network, the input of each time step is represented as ,in is the time step. The LSTM network first uses the forget gate to determine the current moment How much information from the previous time step should be forgotten? Sigmoid function Compress the result of the linear combination to between 0 and 1 and output a forgetting factor :

[0084]

[0085] in, is the weight matrix of the forget gate, is the bias term, is the hidden state at the previous time step, is the feature vector after attention mechanism processing in The value of the time step;

[0086] Step 5.2, Input Gate: The input gate determines how much new information is taken from the current input is added to the current cell state. The output of the forget gate Control the old state The retention amount, the activation value of the input gate Controls the addition of update information generated by the current input and previous hidden state through the candidate value layer, as follows:

[0087]

[0088] in, represents the activation value of the input gate, represents the candidate value generated by the tanh function, Represents the new cell state, which is the old state and the new candidate value Combined to get, , Represents the weight matrix of the input gate, which is used to control the addition of new information. represents the bias term of the input gate, represents the weight matrix used to generate candidate states that may be added to the current cell state, Represents the bias term when generating candidate states; , : The outputs of the forget gate and the input gate control the forgetting of old state information and the addition of new information respectively;

[0089] Step 5.3, Output Gate: The output gate determines what information will be output from the cell state. Output to hidden state , the tanh function performs a nonlinear transformation on the updated unit state, and then passes through the output gate The adjustment of the output hidden state , the output hidden state of the LSTM network is the feature set after processing by the self-attention neural network The result of time series modeling, hidden state Used in subsequent neural network layers, such as fully connected layers or other decision layers. The formula is as follows:

[0090]

[0091] in, represents the bias term of the output gate, Indicates the new cell state.

[0092] Step 6: Fully connected neural network processing: The fully connected layer is responsible for the final linear transformation of the features transferred from the long short-term memory network.

[0093] Fully connected The fully connected layer is responsible for the final linear transformation of the features passed from the long short-term memory network to adapt to the dimension of the target output. The output dimension of the fully connected layer is 1 single neuron;

[0094] At the end of the model, the output hidden state of the LSTM network is The linear transformation through the fully connected layer and Sigmoid activation is converted to the predicted probability, the formula is as follows:

[0095]

[0096] in, Represents an activation function. If the input is a positive number, the output is the input itself. If the input is a negative number, the output is 0. is the hidden state of the LSTM network at the last time step, and They are the weights and biases of the fully connected layer respectively;

[0097] Next, the output is processed through the Sigmoid activation function, the formula is as follows:

[0098]

[0099] in, It is the value passed to the Sigmoid function and is the output of the ReLU function.

[0100] Step 7, signal separation: decode the high and low levels obtained after the self-attention neural network, long short-term memory network and fully connected neural network to obtain the source signal. The received satellite-borne ADS-B overlapping signal is input into the model composed of the self-attention neural network, long short-term memory network and fully connected neural network after parameter optimization. By learning the characteristics of the signal, the high and low levels of different signals are generated, and the high and low levels are converted back to their original ADS-B message format to achieve accurate identification and separation of each signal source.

[0101] Figure 2 These are the training and testing curves of random overlapping signals at a 20dB signal-to-noise ratio.

[0102] Figure 3 This is a graph of the bit error rate of random overlapping signals under a signal-to-noise ratio of 5 to 20dB.

[0103] Figure 4It is a mixed overlapping signal received by a single channel of the satellite, which is processed in a model consisting of a self-attention neural network, a long short-term memory network, and a fully connected neural network.

[0104] Figure 5 These are the high and low levels of the source signal with an earlier arrival time obtained after the overlapping signal is processed by the model. Among them, blue is the high and low levels of the separated signal, and orange is the high and low levels of the real signal. The vertical coordinates near 1 in the picture are high levels, and those near 0 are low levels. By comparing the separated signal with the real signal, it is found that the model separates the overlapping signal more accurately.

[0105] Figure 6 The high and low levels of the late-arriving source signal are obtained after the overlapping signal is processed by a model composed of a self-attention neural network, a long short-term memory network, and a fully connected neural network. By comparing the separated signal with the real signal, it is found that the model has separated the overlapping signal more accurately. Figure 5 and Figure 6 By comparing the separated signal with the real source signal, it is found that the high and low levels of the source signal are basically separated correctly.

[0106] The present invention utilizes a long short-term memory network to process the output of the self-attention network, and converts the output of the LSTM into a predicted probability through a fully connected layer to achieve effective signal separation.

[0107] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for separating overlapping satellite-borne ADS-B signals, characterized in that: The separation method specifically comprises the following steps: Step 1: Normalization: Satellite receiving data Normalize the data Each signal in Normalize to the range of [0, 1], adjust the amplitude of all signals to the range of [0, 1] to obtain normalized data ; Step 2: Vectorization: Normalized data Perform scale up and convert the scaled data into a feature vector set ; Step 3: Complex position encoding: The feature vector set obtained in step 2 Each vector in After complex position encoding, the encoded feature vector is obtained , through the encoded feature vector To encode the position information, providing the attention mechanism model with information about the position of each element in the satellite reception data input sequence; Step 4: Self-attention neural network processing: After the complex position encoding in step 3 is completed, the encoded feature vector Input into the self-attention neural network to obtain the feature set processed by the self-attention neural network , the feature set after processing by the self-attention neural network Pass to the next self-attention neural network; Step 5: Long short-term memory network processing: The feature set processed by the self-attention neural network obtained in step 4 Transformed into a feature representation suitable for classification tasks; Step 6: Fully connected neural network processing: The fully connected layer is responsible for the final linear transformation of the features transferred from the long short-term memory network; Step 7, signal separation: decode the high and low levels obtained after passing through the self-attention neural network, long short-term memory network and fully connected neural network to obtain the source signal.

2. A method for separating overlapping satellite-borne ADS-B signals according to claim 1, characterized in that: In step 3, the complex position encoding is used to provide time information and position information, which is the feature vector set Each vector in Its position in the time dimension Add complex position encoding, the formula is as follows: ， in, Indicates the position of an element in a sequence. Represents the dimension, represents the total dimension of the position encoding vector, represents pi, Is an imaginary unit.

3. The method for separating overlapping satellite-borne ADS-B signals according to claim 1, characterized in that: The step 4 specifically comprises the following steps: Step 4.1, linear transformation of input: the encoded feature vector of the input Perform linear transformation to generate three new feature sets: query ,key Sum , the formula is as follows: ， in, is a trainable weight matrix; Step 4.2: Calculate attention score: Using query With all the keys The compatibility score is calculated by the dot product between them, and then normalized by a normalization function to calculate the final attention score. , the self-attention neural network determines the relative importance of each element in the input data set, the attention score The calculation formula is as follows: ， in, is the dimension of the key vector, Representing key values The transpose of Step 4.3, weighted sum: Add the attention score calculated in step 4.2 With value Perform weighted summation to obtain the final feature set processed by the self-attention neural network , that is, to integrate different input features according to the importance of the features, the formula is: ， Among them, the feature set after the self-attention neural network Passed to the next self-attention mechanism network.

4. The method for separating overlapping satellite-borne ADS-B signals according to claim 1, characterized in that: In step 5, the long short-term memory network includes three modules: input gate, forget gate and output gate. The long short-term memory network is used to transform the feature set Transformed into a feature representation suitable for classification tasks, specifically including the following steps: Step 5.1, Forget Gate: Feature Set after Self-Attention Neural Network Processing As the input sequence of the LSTM network, the input of each time step is represented as ,in is the time step. The LSTM network first uses the forget gate to determine the current moment How much information is forgotten from the previous time step? Sigmoid function Compress the result of the linear combination to between 0 and 1 and output a forgetting factor : ， in, is the weight matrix of the forget gate, is the bias term, is the hidden state at the previous time step, is the feature vector after attention mechanism processing in The value of the time step; Step 5.2, Input Gate: Output of Forget Gate Control the old state The retention amount, the activation value of the input gate Controls the addition of update information generated by the current input and previous hidden state through the candidate value layer, as follows: ， in, represents the activation value of the input gate, represents the candidate value generated by the tanh function, Represents the new cell state, which is the old state and the new candidate value Combined to get, , Represents the weight matrix of the input gate, which is used to control the addition of new information. represents the bias term of the input gate, represents the weight matrix used to generate candidate states, Represents the bias term when generating candidate states; , : The outputs of the forget gate and the input gate control the forgetting of old state information and the addition of new information respectively; Step 5.3, output gate: through the output gate The adjustment of the output hidden state , the output hidden state of the LSTM network is the feature set after processing by the self-attention neural network The result of time series modeling is as follows: ， in, represents the bias term of the output gate, Indicates the new cell state.

5. The method for separating overlapping satellite-borne ADS-B signals according to claim 1, characterized in that: The step 6 specifically includes the following steps: Step 6.1: Connect the fully connected layer to perform the final linear transformation on the features passed from the LSTM network to adapt to the dimension of the target output. The output dimension of the fully connected layer is 1 single neuron. Step 6.2: At the end of the model, hide the output state of the LSTM network The linear transformation through the fully connected layer and Sigmoid activation is converted to the predicted probability, the formula is as follows: ， in, Represents an activation function. If the input is a positive number, the output is the input itself. If the input is a negative number, the output is 0. is the hidden state of the LSTM network at the last time step, and They are the weights and biases of the fully connected layer respectively; Step 6.3: The output is processed by the Sigmoid activation function, and the formula is as follows: ， in, It is the value passed to the Sigmoid function and is the output of the ReLU function.

6. The method for separating overlapping satellite-borne ADS-B signals according to claim 1, characterized in that: In step 7, the model composed of the self-attention neural network, the long short-term memory network and the fully connected neural network generates high and low levels of different signals by learning the characteristics of the signals, and converts the high and low levels back to their original ADS-B message formats to achieve accurate identification and separation of each signal source.

Images