An ADS-B overlapping signal separation processing system based on deep learning

Abstract

This invention provides a deep learning-based ADS-B overlapping signal separation and processing system, relating to the field of aviation surveillance and signal processing technology. It includes a signal receiving module that receives mixed radio frequency signals from air traffic targets, the mixed radio frequency signals including overlapping ADS-B signals; and a signal preprocessing module that preprocesses the mixed radio frequency signals to form a time-domain sequence of overlapping signals. This deep learning-based ADS-B overlapping signal separation and processing system, by combining the signal preprocessing module and deep learning separation processing, achieves efficient decoding and analysis of ADS-B signals and provides accurate flight parameter information in signal overlap and interference environments. Furthermore, it uses a collision avoidance warning module to assess collision risks in real time and generate timely and effective warning information, thereby improving the safety and reliability of the air traffic monitoring system.

Description

Technical Field

[0001] This invention relates to the field of aviation surveillance and signal processing technology, specifically to an ADS-B overlapping signal separation and processing system based on deep learning. Background Technology

[0002] Automatic Dependent Surveillance-Broadcast (ADBBS) is a core surveillance technology in modern air traffic control. Airborne equipment periodically broadcasts aircraft identification, location, and altitude information, which is received by ground stations or other aircraft to monitor the air situation. In dense airspace such as flight training airports, multiple aircraft transmit signals simultaneously, and differences in signal transmission delays cause waveform overlap at the receiver. Traditional receivers use a first-come, first-served principle, analyzing individual signals through preamble detection and bit decision. Existing technologies focus on improving receiver sensitivity or optimizing antenna design, achieving good results in non-overlapping signal scenarios.

[0003] The core drawback of the current method is that when two or more signals overlap in the time domain, the preamble and data bits interfere with each other, and traditional receivers cannot effectively separate the mixed waveforms. The bit decision error rate increases sharply, and the receiver can only decode the stronger signal, while the remaining signals are overwhelmed by noise, leading to target loss. Co-channel interference causes incomplete monitoring data, and collision warning systems based on ADS-BIN are at risk of missing reports, making it difficult to guarantee the collision avoidance safety requirements of high-density flight training. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a deep learning-based ADS-B overlapping signal separation and processing system. The technical problem this invention aims to solve is: how to separate and process overlapping ADS-B signal time-domain sequences using a deep learning-based neural network model, thereby improving the accuracy of target surveillance and collision avoidance warning in high-density flight environments.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a deep learning-based ADS-B overlapping signal separation and processing system, comprising: a signal receiving module: used for receiving mixed radio frequency signals of air traffic targets, wherein the mixed radio frequency signals are overlapping ADS-B signals.

[0006] Signal preprocessing module: This module is used to preprocess the mixed radio frequency signals to form an overlapping signal time-domain sequence. The preprocessing includes filtering, amplitude normalization, and adding timestamp information.

[0007] Deep Neural Network Module: Used by the deep neural network module to separate the time-domain sequence of the overlapping signal to form a signal separation processing result. The separation processing adopts a neural network model, and the signal separation processing result includes the separated ADS-B signal sequence.

[0008] Signal parsing module: Used by the signal parsing module to decode the signal separation processing results to form flight parameter information.

[0009] Collision avoidance warning module: This module is used to perform real-time analysis of the flight parameter information to generate collision avoidance warning information. The real-time analysis adopts a collision detection algorithm.

[0010] Preferably, the overlap time of the overlapping ADS-B signals is 0μs-8μs, and the number of overlapping ADS-B signals is ≥2.

[0011] Preferably, the signal preprocessing module includes a bandpass filtering unit, an amplitude normalization unit, and a timestamp addition unit. The bandpass filtering unit performs the filtering process on the mixed radio frequency signal to form a filtered signal, which is a 1090MHz±2MHz frequency band signal. The amplitude normalization unit performs the amplitude normalization process on the filtered signal to form a normalized signal, which has an amplitude value range of [-1,1]. The timestamp addition unit adds timestamp information to the normalized signal to form the time-domain sequence of the overlapping signal.

[0012] Preferably, the filtering process includes dynamic noise suppression processing, which employs an adaptive notch filter. The adaptive notch filter suppresses lightning pulse interference and terrain reflection interference in the aviation environment. The filtering process also suppresses frequency components outside the 1090MHz±2MHz band in the mixed radio frequency signal.

[0013] Preferably, the deep neural network model is built using an encoder-decoder structure. The deep neural network model uses overlapping signal time-domain sequences and corresponding separated signal sequences as training data. The deep neural network model includes an input layer, an encoder layer, a latent representation layer, a decoder layer, and an output layer. The input layer receives the overlapping signal time-domain sequences. The encoder layer extracts features from the overlapping signal time-domain sequences in chronological order to generate feature sequences. The feature sequences are temporal features reflecting signal amplitude changes and temporal relationships. The latent representation layer performs weighted aggregation on the feature sequences to form a latent representation vector. The latent representation vector represents the overlap relationship. The decoder layer generates an output signal time-domain sequence based on the latent representation vector and the feature sequences. The number of output signal time-domain sequences is ≥2. The output layer outputs the output signal time-domain sequences as the signal separation processing result.

[0014] Preferably, the deep neural network module is trained in an end-to-end manner. The construction of the training set in this end-to-end manner includes superimposing known clean ADS-B signal sequences in the time domain at a preset overlap time to synthesize a simulated co-channel interference time-domain sequence of the overlapping signal as training input, and the clean ADS-B signal sequence as the target output corresponding to the training input. The number of clean ADS-B signal sequences is ≥2. The end-to-end method includes joint optimization of the overlapping signal time-domain sequence and the corresponding target clean ADS-B signal sequence. The loss function used in the joint optimization is: Where L is the loss function value, dimensionless; K is the number of ADS-B signal channels participating in the separation, dimensionless; and T is the number of sampling points for each signal channel, dimensionless. The sampled voltage value of the k-th target pure ADS-B baseband signal at the t-th sampling point is given, in volts. The voltage value of the predicted signal at the t-th sampling point on the k-th output channel of the deep neural network module is expressed in volts. This is the normalized reference voltage, in volts.

[0015] Preferably, the feature extraction employs a bidirectional recurrent neural network to extract forward and reverse sampling points from the overlapping signal time-domain sequence. The sampling points contain contextual information. The weighted aggregation uses an attention mechanism to calculate the attention weight of the feature vector at each time step in the feature sequence. The decoder layer uses an autoregressive approach to progressively generate multiple separate output signal time-domain sequences. The output layer includes format conversion and amplitude adjustment of the output signal time-domain sequences. The training set includes signal overlap time, signal-to-noise ratio, signal power ratio, and total number of training samples. The signal overlap time ranges from 0 μs to 8 μs, the signal-to-noise ratio ranges from 5 dB to 30 dB, the signal power ratio ranges from 0.1 to 10, and the total number of training samples is ≥100,000.

[0016] Preferably, the signal parsing module includes a cyclic redundancy check (CR) unit, a data extraction unit, and a data verification unit. The CR unit demodulates and performs bit decision on the signal separation processing result to obtain an ADS-B data packet. The CR unit performs CR on the ADS-B data packet to form a verification packet. The data extraction unit parses the verification packet to extract flight parameters. The data verification unit performs a rationality check on the flight parameters to form flight parameter information, which includes flight parameter information that passes the rationality check and flight parameter information that fails the rationality check.

[0017] Preferably, the parsing process includes extracting the aircraft identification code, latitude and longitude coordinates, barometric altitude and time information, and the rationality check includes verifying the location coordinate range, verifying the continuity of altitude data and verifying the timestamp order.

[0018] Preferably, the collision avoidance warning module calculates a collision risk score based on the conflict detection algorithm, and performs graded alarm judgment processing on the collision risk score to form the collision avoidance alarm information. The collision avoidance alarm information includes no alarm, warning level alarm, and emergency collision avoidance alarm. The graded alarm judgment processing adopts threshold interval division, which includes a distance alarm threshold and a height alarm threshold. The distance alarm threshold is 5000 meters, and the height alarm threshold is 300 meters.

[0019] Preferably, the collision avoidance warning module includes an alarm output unit. The alarm output unit includes, when the prompt-level alarm is triggered, displaying a bright yellow potential risk area marker on the display and playing an intermittent alarm sound. The intermittent alarm sound has a playback frequency of 1000Hz and a period of 1 second. When the emergency collision avoidance alarm is triggered, displaying a flashing red collision area marker on the display and playing a continuous rapid alarm sound. The continuous rapid alarm sound has a playback frequency of 2000Hz and a period of 0.5 seconds.

[0020] This invention provides an ADS-B overlapping signal separation and processing system based on deep learning. It has the following advantages:

[0021] This invention presents a deep learning-based ADS-B overlapping signal separation and processing system, addressing the problem that traditional receivers cannot accurately separate multiple ADS-B signals when they overlap in dense air traffic areas or flight training airspace. By employing a deep neural network model, the ADS-B overlapping signal separation and processing system can accurately separate overlapping signals in complex airspace environments, improving signal decoding accuracy and ensuring real-time monitoring of flight targets and data integrity.

[0022] This deep learning-based ADS-B overlapping signal separation and processing system achieves efficient decoding and analysis of ADS-B signals by combining a signal preprocessing module and deep learning-based separation processing. In environments with signal overlap and interference, the system provides accurate flight parameter information and, through a collision avoidance warning module, assesses collision risks in real time, generating timely and effective warning information, thus improving the safety and reliability of the air traffic monitoring system. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the ADS-B overlapping signal separation and processing system. Figure 2 This is a detailed diagram of the signal preprocessing module; Figure 3 This is a schematic diagram of a deep neural network module structure; Figure 4 This is the flowchart of the signal analysis module; Figure 5 This is the flowchart for the collision avoidance warning module. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1 like Figure 1-5 As shown, this embodiment of the invention provides a deep learning-based ADS-B overlapping signal separation and processing system, including a signal receiving module: the signal receiving module receives mixed radio frequency signals of air traffic targets, the mixed radio frequency signals being overlapping ADS-B signals. The overlap time of the overlapping ADS-B signals is 0μs-8μs, and the number of overlapping ADS-B signals is ≥2.

[0026] Signal preprocessing module: The signal preprocessing module preprocesses the mixed radio frequency signals to form an overlapping signal time-domain sequence. Preprocessing includes filtering, amplitude normalization, and adding timestamp information. The signal preprocessing module includes a bandpass filtering unit, an amplitude normalization unit, and a timestamp adding unit. The bandpass filtering unit filters the mixed radio frequency signals to form a filtered signal (1090MHz±2MHz frequency band). The amplitude normalization unit normalizes the filtered signal to form a normalized signal with an amplitude range of [-1,1]. The timestamp adding unit adds timestamp information to the normalized signal to form the overlapping signal time-domain sequence.

[0027] The filtering process includes dynamic noise suppression, which employs an adaptive notch filter. The adaptive notch filter suppresses lightning pulse interference and terrain reflection interference in the aviation environment. The filtering process also suppresses frequency components outside the 1090MHz±2MHz band in the mixed radio frequency signal.

[0028] Amplitude normalization unit: The mixed radio frequency signal is received by the signal receiving module and then subjected to bandpass filtering to obtain the filtered signal. The signal obtained after filtering... The amplitude range is [ , ], which refers to the minimum and maximum amplitude values ​​of the signal.

[0029] In one example, the minimum amplitude of the filtered signal is =0.2V, the maximum amplitude of the filtered signal is =5.0V.

[0030] 2. Normalization processing The amplitude normalization unit performs a normalization operation on the filtered signal. Normalization is performed. The formula for normalization is:

[0031] The formula maps the signal amplitude to the range [-1, 1], at a certain moment... =2.5V, and =5.0V, =0.2V, then: Signal =2.5V, after normalization is =−0.042. After normalization, all input signals are within the same amplitude range, avoiding instability in deep learning model training caused by differences in signal amplitude.

[0032] In one example, the signal after amplitude normalization is denoted as: Each of them The amplitude values ​​have been mapped to the interval [−1, 1]. A certain segment of the normalized signal is: the first sampling point: =0.35, the second sampling point: =−0.10, the 3rd sampling point: =0.78.

[0033] The sampling points are a sequence of amplitude values ​​arranged in the sampling order.

[0034] Sampling period and start time settings: Before adding a timestamp to the normalized signal, the timestamp addition unit defines two known parameters: sampling frequency : The system samples signals in the 1090MHz band at a frequency of [frequency value missing]. =2 MHz.

[0035] The sampling period is: Start time : The timestamp addition unit obtains the current absolute time from the receiver's internal clock or an external time synchronization system as the start time of this signal segment.

[0036] In one example, the start time is =2025-11-19 14:05:10.000000.

[0037] The process of adding timestamps: The timestamp addition unit calculates the corresponding timestamp for each normalized sampling point in the order of sampling.

[0038] The timestamp of the nth sampling point is calculated as follows: In the example above: Sampling period =0.5 μs, start time =14:05:10.000000.

[0039] First sampling point: The corresponding data pairs are: ( , (1))=(14:05:10.000000, 0.35).

[0040] Second sampling point: The corresponding data pairs are: ( , (2))=(14:05:10.0000005, −0.10).

[0041] 3rd sampling point: The corresponding data pairs are: ( , (3))=(14:05:10.000001, 0.78).

[0042] Until the Nth sampling point, the timestamp adding unit binds all sampling points with their corresponding timestamps, generating a sequence: {( , (1)),( , (2)),…,( , (N))}.

[0043] Forming overlapping signal time-domain sequences: The above-mentioned sampling point sequence with bound timestamps is the time-domain sequence of the overlapping signal, including: Amplitude information: reflects the total voltage magnitude of the overlapping ADS-B signals at the same sampling time; Time information: reflects the position of the sampling point on the absolute time axis.

[0044] The deep neural network module separates overlapping signal time-domain sequences to form a signal separation result. This separation process uses a neural network model and includes the separated ADS-B signal sequence. The deep neural network model employs an encoder-decoder structure, using the overlapping signal time-domain sequences and their corresponding separated signal sequences as training data. The model includes an input layer, encoder layer, latent representation layer, decoder layer, and output layer. The input layer receives the overlapping signal time-domain sequences. The encoder layer extracts features from the overlapping signal time-domain sequences in chronological order to generate feature sequences. These feature sequences reflect temporal variations and temporal relationships. The latent representation layer weights and aggregates the feature sequences to form a latent representation vector, which represents the overlap relationship. The decoder layer generates an output signal time-domain sequence based on the latent representation vector and the feature sequences. The number of output signal time-domain sequences is ≥2. The output layer outputs the output signal time-domain sequences as the signal separation result.

[0045] The deep neural network module is trained end-to-end. The construction of the training set in the end-to-end approach involves superimposing known clean ADS-B signal sequences in the time domain at preset overlap times to synthesize an overlapping signal time-domain sequence simulating co-channel interference as the training input. The clean ADS-B signal sequence serves as the target output corresponding to the training input. The number of clean ADS-B signal sequences is ≥2. The end-to-end approach includes joint optimization of the overlapping signal time-domain sequence and the corresponding target clean ADS-B signal sequence. The loss function used for joint optimization is: Where L is the loss function value, dimensionless; K is the number of ADS-B signal channels participating in the separation, dimensionless; and T is the number of sampling points for each signal channel, dimensionless. The sampled voltage value of the k-th target pure ADS-B baseband signal at the t-th sampling point is given, in volts. The voltage value of the predicted signal at the t-th sampling point on the k-th output channel of the deep neural network module is expressed in volts. This is the normalized reference voltage, in volts.

[0046] Feature extraction employs a bidirectional recurrent neural network to extract forward and reverse sampling points from the overlapping signal time-domain sequence. The sampling points contain contextual information. Weighted aggregation uses an attention mechanism to calculate the attention weight of the feature vector at each time step in the feature sequence. The decoder layer uses an autoregressive approach to progressively generate multiple separate output signal time-domain sequences. The output layer includes format conversion and amplitude adjustment of the output signal time-domain sequence. The training set includes signal overlap time, signal-to-noise ratio, signal power ratio, and the total number of training samples. The signal overlap time ranges from 0μs to 8μs, the signal-to-noise ratio ranges from 5dB to 30dB, the signal power ratio ranges from 0.1 to 10, and the total number of training samples is ≥100,000.

[0047] Signal parsing module: The signal parsing module decodes the signal separation processing results to form flight parameter information. The signal parsing module includes a cyclic redundancy check (CR) unit, a data extraction unit, and a data verification unit. The CR unit demodulates and performs bit decision on the signal separation processing results to obtain ADS-B data packets. The CR unit performs CR on the ADS-B data packets to form verification packets. The data extraction unit parses the verification packets to extract flight parameters. The data verification unit performs a rationality check on the flight parameters to form flight parameter information, which includes flight parameters that pass the rationality check and those that fail. The parsing process includes extracting the aircraft identification code, latitude and longitude coordinates, barometric altitude, and time information. The rationality check includes verifying the position coordinate range, altitude data continuity, and timestamp sequence.

[0048] Demodulation and bit decision examples: The signal separation and processing result of a certain path output by the deep neural network module is a normalized ADS-B baseband signal with a sampling rate of 4MHz, each bit corresponding to 4 sampling points, and an amplitude range of [−1,1]. Within a small segment corresponding to a frame of a message, the sampled data is, for example:

[0049] Table 1: ADS-B baseband signal after sampling point normalization.

[0050] The cyclic redundancy check unit performs energy determination using every 4 sampling points as one bit, with the determination threshold set to 0.3. The first bit corresponds to sampling points 1-4, with an average amplitude of approximately (0.82+0.79+0.20+0.18) / 4≈0.497, which is greater than 0.3, so it is determined to be bit 1.

[0051] The second bit corresponds to sampling points 5-8, with an average amplitude of approximately (−0.05+0.02−0.10+0.00) / 4≈−0.033, which is less than 0.3, so it is determined to be bit 0.

[0052] The entire separated signal is processed to obtain a 112-bit ADS-B data message, where the first 88 bits are the data field and the last 24 bits are the CRC field carried in the message, represented in hexadecimal as follows: The hexadecimal string corresponding to the original bit sequence of the message is: 8D40621D58B402F38135C7A5, where the first 11 bytes are the data field and the last 3 bytes are the CRC field.

[0053] Example of CRC check: The Cyclic Redundancy Check (CRC) unit performs a CRC calculation on the first 11 bytes of the above message using the generator polynomial specified by ADS-B. The internal 24-bit CRC register is initially set to 0. 88 data bits are input sequentially, and after shifting and bitwise XOR operations, the calculated CRC value is obtained:

[0054] The calculated CRC value is 0x35C7A5, and the CRC field received at the end of the message is also 0x35C7A5.

[0055] Since the two are consistent, the frame message is determined to have passed the CRC check. The cyclic redundancy check unit marks the message as valid and outputs it as a check message. If the two are inconsistent, the message is determined to be an erroneous message and will not enter the subsequent data extraction process.

[0056] Collision Avoidance Warning Module: This module analyzes flight parameter information in real time to generate collision avoidance warning information. The real-time analysis employs a conflict detection algorithm. Based on this algorithm, the module calculates a collision risk score and then performs tiered warning processing to generate collision avoidance warning information. This includes no warning, a warning level, and an emergency collision avoidance warning. The tiered warning processing uses threshold range division, which includes a distance warning threshold of 5000 meters and an altitude warning threshold of 300 meters.

[0057] The collision avoidance warning module includes an alarm output unit. When a warning-level alarm is triggered, the alarm output unit displays a bright yellow potential risk area marker on the display and plays an intermittent alarm sound. The intermittent alarm sound plays at a frequency of 1000Hz and a period of 1 second. When an emergency collision avoidance alarm is triggered, a flashing red collision area marker is superimposed on the display and a continuous rapid alarm sound plays at a frequency of 2000Hz and a period of 0.5 seconds.

[0058] Example 2 This embodiment uses an end-to-end trained deep neural network module to separate overlapping ADS-B signal time-domain sequences, achieving efficient and accurate signal separation and providing reliable data support for flight parameter extraction and collision avoidance warning.

[0059] 1. Preparation of the training set The training set for the deep neural network module includes multiple overlapping signal time-domain sequences and corresponding target clean signal sequences. Input data is obtained through a signal receiving module and a preprocessing module, and each overlapping signal is synchronized by a timestamp-adding unit to form a timestamped overlapping signal time-domain sequence.

[0060] In one example, there are two signal sources, aircraft A and aircraft B, which generate the following clean signal sequences: Aircraft A's signal : ={1.2,1.1,1.3,1.5,1.4}V.

[0061] Aircraft B's signal : ={0.5,0.4,0.6,0.7,0.5}V.

[0062] By superimposing the signals and adding noise, we obtain the overlapping signal sequence x(t): The target signal is separated from the overlapping signal x(t) using a neural network. and .

[0063] 2. Structure of Deep Neural Networks Deep neural networks employ an encoder-decoder structure, comprising the following components: Input layer: Receives the preprocessed overlapping signal time-domain sequence x(t).

[0064] Encoder layer: A bidirectional gated recurrent unit is used to extract temporal features from the time-domain sequence of the input overlapping signal, generating a temporal feature vector. .

[0065] Latent representation layer: The attention mechanism is used to aggregate the temporal features output by the encoder to form a latent representation vector z.

[0066] Decoder layer: Based on the latent representation vector z and the temporal features output by the encoder. Multiple separate signal time-domain sequences are generated step by step. (n), n=1,2,…,N.

[0067] Output layer: Adjusts the amplitude and converts the format of the time-domain sequence of the signal output by the decoder, and outputs the separated signal sequence.

[0068] 3. Joint Optimization and Loss Function The neural network is trained using an end-to-end joint optimization approach. The loss function is calculated based on the difference between the predicted signal and the target signal output by the network.

[0069] The loss function formula is as follows: Where L is the loss function value, dimensionless; K is the number of ADS-B signal channels participating in the separation, dimensionless; and T is the number of sampling points for each signal channel, dimensionless. The sampled voltage value of the k-th target pure ADS-B baseband signal at the t-th sampling point is given, in volts. The voltage value of the predicted signal at the t-th sampling point on the k-th output channel of the deep neural network module is expressed in volts. This is the normalized reference voltage, in volts.

[0070] By minimizing the loss function L, the network optimizes its internal parameters to improve the predicted separation signal. Approaching target signal The goal is to separate independent signals from overlapping signals.

[0071] 4. Network Training and Separation Process One overlapping signal sample in the training set is x={1.7,1.5,1.9,2.2,1.9}, and the target signal is... ={1.2,1.1,1.3,1.5,1.4} and ={0.5,0.4,0.6,0.7,0.5}.

[0072] Network output: Loss calculation: By minimizing the loss function, the output separated signal gradually approaches the target signal.

[0073] Example 3 This embodiment calculates a collision risk score based on a conflict detection algorithm, performs graded alarm determination, and generates different levels of collision avoidance alarm information to improve flight safety.

[0074] 1. Collision Detection Algorithm Collision detection algorithms are used to assess the relative position and velocity of two aircraft and calculate whether there is a potential risk of collision between them. Collision detection algorithms are based on the following factors:

[0075] Relative distance of target aircraft : The straight-line distance between the two aircraft; the relative velocity of the target aircraft v(t): the relative velocity component between the two.

[0076] Time to minimum distance : The time from the current moment until the minimum distance of the aircraft occurs.

[0077] The current position of aircraft A is ( , ), speed is ( , The current position of aircraft B is ( ). , ), speed is ( , ).

[0078] Relative distance of target aircraft Calculated using the following formula: Where t is the predicted time from the current moment to the nearest contact moment.

[0079] 2. Collision Risk Score Calculation Based on relative distance and relative velocity Calculate the collision risk score R. The score is used to assess the risk of a collision between two aircraft and is calculated based on the following formula:

[0080] in, =5000m is the preset safe distance threshold. =250m / s is the preset maximum relative velocity threshold, and α and β are the coefficients for adjusting the weights, which are set to α=0.6 and β=0.4 based on historical data.

[0081] At a certain moment, the relative distance between target aircraft A and aircraft B is... =4500m, relative speed =180m / s. Substitute the value into the formula to calculate:

[0082] The resulting collision risk score is 0.828, indicating a moderate risk between aircraft A and aircraft B.

[0083] 3. Classified alarm determination Based on the collision risk score R, the collision avoidance warning module performs graded alarm judgment and processing. The grading criteria are as follows:

[0084] No alarm: R<0.5, alert level alarm: 0.5≤R<0.8, emergency collision avoidance alarm: R≥0.8.

[0085] According to the classification criteria, when the calculated R=0.828, the system will trigger an emergency collision avoidance alarm.

[0086] 4. Collision avoidance alarm information output The collision avoidance warning module generates collision avoidance warning information based on the collision risk score, and outputs the warning level and detailed data to the user or other relevant systems. The system calculates the collision risk score between aircraft A and aircraft B to be 0.828, and the collision avoidance warning information includes:

[0087] Alarm type: Emergency collision avoidance alarm.

[0088] Target aircraft information: Aircraft A: ABC123, current position: 40.7128°N, 74.0060°W. Aircraft B: XYZ789, current position: 40.7100°N, 74.0050°W. Distance: 4500 meters, relative speed: 180 m / s, alarm time: 2024-10-17 14:05:10 UTC.

[0089] Depending on the collision warning level, the system will take the following measures: display a flashing red collision zone indicator on the monitor and play a continuous, rapid alarm sound.

[0090] Example 4 This embodiment collects environmental parameters under extreme weather conditions and calculates environmental impact coefficients to dynamically adjust the reasonableness judgment threshold of flight parameters and collision risk score, thereby improving the accuracy and reliability of flight parameter verification and collision avoidance warning under complex airspace conditions.

[0091] 1. Parsing and processing At a specific moment, the verified ADS-B message content is 8D40621D58B402F38135C7A5.

[0092] The following is an example of the analyzed flight parameters: Table 2: Flight Parameter Analysis Data Table.

[0093] The data extraction unit organizes the data into flight parameter records and transmits them to the data verification unit.

[0094] 2. Standard rationality check of the data validation unit 2.1 Routine Validation Location coordinate range verification: The longitude range should be between -180° and +180°, and the latitude range should be between -90° and +90°.

[0095] High continuity verification: The previous valid message was at an altitude of 34,000 ft, and the time interval was approximately 5 seconds. The altitude change was as follows: The standard allows for a height variation range of ±2000ft, and the height continuity verification has been passed.

[0096] Timestamp order verification: The previous time was 14:05:05.20, and the current time is 14:05:10.200, satisfying the sequential consistency requirement. The flight parameters are valid records under the standard rules.

[0097] 3. Calculation of Environmental Acquisition and Impact Coefficients In airspace where extreme weather is possible, the system needs to collect environmental parameters, including: Table 3: Real-time environmental parameter collection data.

[0098] The system calculates the ratio for each quantity: Wind speed ratio: 38 / 20=1.90, temperature difference ratio: 2.2 / 1.0=2.20, humidity ratio: 1.5 / 1.0=1.50, interference ratio: 0.68 / 1.00=0.68.

[0099] Weights are assigned based on the impact of each environmental factor: Each environmental factor is assigned an impact level score, and based on historical data and engineering experience, an impact level of 1-5 is given for each factor.

[0100] Table 4: Distribution of Environmental Impact Levels

[0101] The weights are obtained by normalizing the influence levels: The sum of all impact levels is 5+4+3+4=16.

[0102] The weights of each factor are then calculated as follows: Wind speed: 5 / 16 = 0.3125, temperature gradient = 0.25, humidity change = 0.1875, disturbance index = 0.25.

[0103] The weights were optimized as follows: wind speed weight: 0.30, temperature difference weight: 0.25, humidity weight: 0.20, and interference weight: 0.25.

[0104] The environmental impact factor is calculated as follows: The environmental impact coefficient is obtained as follows: =1.59 indicates that the current airspace environment has deteriorated.

[0105] 4. Dynamically adjust the reasonableness check threshold. 4.1 Dynamically adjust the position coordinate range Original latitude deviation tolerance: ±0.050°.

[0106] Rule revision: That is, the dynamic position deviation tolerance is ±0.0795°, and the system uses the tolerance to judge whether the position drift is reasonable.

[0107] 4.2 Dynamically adjust the range of height continuity Original height change threshold: 2000ft.

[0108] Rule revision: Altitude changes exceeding 3180 ft are considered abnormal. The current altitude change is 1000 ft; 1000 < 3180, so altitude continuity verification passes.

[0109] It becomes more adaptable after environmental modification.

[0110] 5. Environmental correction for collision risk values The collision avoidance warning module calculates the initial collision risk value based on the ADS-B trajectory, for example: =0.52.

[0111] The system corrects for risk based on the environmental impact factor: The corrected risk score is 0.827, which is higher than the emergency alarm threshold of 0.80, so the system directly triggers an emergency collision avoidance alarm.

[0112] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based ADS-B overlapping signal separation and processing system, characterized in that, include: Signal receiving module: used to receive mixed radio frequency signals from air traffic targets, wherein the mixed radio frequency signals are overlapping ADS-B signals; Signal preprocessing module: used to preprocess the mixed radio frequency signals to form an overlapping signal time-domain sequence. The preprocessing includes filtering, amplitude normalization, and adding timestamp information. Deep neural network module: used to separate overlapping signal time-domain sequences to form signal separation results. The separation process adopts a neural network model, and the signal separation results include separated ADS-B signal sequences. Signal parsing module: used to decode the signal separation and processing results to form flight parameter information; Collision avoidance warning module: used to analyze flight parameter information in real time and generate collision avoidance warning information. The real-time analysis adopts a collision detection algorithm.

2. The deep learning-based ADS-B overlapping signal separation and processing system according to claim 1, characterized in that: The overlap time of the overlapping ADS-B signals is 0μs-8μs, and the number of overlapping ADS-B signals is ≥2.

3. The ADS-B overlapping signal separation and processing system based on deep learning according to claim 1, characterized in that: The signal preprocessing module includes a bandpass filtering unit, an amplitude normalization unit, and a timestamp addition unit. The bandpass filtering unit performs the filtering process on the mixed radio frequency signal to form a filtered signal, which is a 1090MHz±2MHz frequency band signal. The amplitude normalization unit performs the amplitude normalization process on the filtered signal to form a normalized signal, which has an amplitude value range of [-1,1]. The timestamp addition unit adds timestamp information to the normalized signal to form the time-domain sequence of the overlapping signal.

4. The ADS-B overlapping signal separation and processing system based on deep learning according to claim 3, characterized in that: The filtering process includes dynamic noise suppression processing, which employs an adaptive notch filter. The adaptive notch filter suppresses lightning pulse interference and terrain reflection interference in the aviation environment. The filtering process also suppresses frequency components outside the 1090MHz±2MHz band in the mixed radio frequency signal.

5. The ADS-B overlapping signal separation and processing system based on deep learning according to claim 1, characterized in that: The deep neural network model is built using an encoder-decoder structure. The deep neural network model uses the overlapping signal time-domain sequence and the corresponding separated signal sequence as training data. The deep neural network model includes an input layer, an encoder layer, a latent representation layer, a decoder layer, and an output layer. The input layer is used to receive the overlapping signal time-domain sequence, and the encoder layer extracts features from the overlapping signal time-domain sequence in chronological order to generate a feature sequence.

6. The ADS-B overlapping signal separation and processing system based on deep learning according to claim 5, characterized in that: The deep neural network model is trained using an end-to-end supervised learning approach. This approach uses the overlapping signal time-domain sequence as the network input and the separated signal sequence as the target label for the network output. The deep neural network model learns the nonlinear relationship between the overlapping signal time-domain sequence and the separated signal time-domain sequence.

7. The ADS-B overlapping signal separation and processing system based on deep learning according to claim 1, characterized in that: The signal parsing module includes a cyclic redundancy check (CR) unit, a data extraction unit, and a data verification unit. The CR unit demodulates and performs bit decision on the signal separation processing results to obtain ADS-B data packets. The CR unit performs CR on the ADS-B data packets to form verification packets. The data extraction unit parses the verification packets to extract flight parameters. The data verification unit performs a rationality check on the flight parameters to form flight parameter information, which includes flight parameter information that passes the rationality check and flight parameter information that fails the rationality check.

8. The ADS-B overlapping signal separation and processing system based on deep learning according to claim 7, characterized in that: The parsing process includes extracting the aircraft identification code, latitude and longitude coordinates, barometric altitude, and time information. The rationality check includes verifying the location coordinate range, verifying the continuity of altitude data, and verifying the timestamp order.

9. The ADS-B overlapping signal separation and processing system based on deep learning according to claim 1, characterized in that: The collision avoidance warning module calculates a collision risk score based on a conflict detection algorithm. The collision avoidance warning module performs a graded alarm judgment process on the collision risk score to form the collision avoidance alarm information. The collision avoidance alarm information includes no alarm, warning level alarm, and emergency collision avoidance alarm. The graded alarm judgment process adopts a threshold interval division. The threshold interval division includes a distance alarm threshold and a height alarm threshold. The distance alarm threshold is 5000 meters, and the height alarm threshold is 300 meters.

10. The ADS-B overlapping signal separation and processing system based on deep learning according to claim 9, characterized in that: The collision avoidance warning module includes an alarm output unit. When the prompt-level alarm is triggered, the alarm output unit displays a bright yellow potential risk area marker on the display and plays an intermittent alarm sound. The intermittent alarm sound plays at a frequency of 1000Hz and a period of 1 second. When the emergency collision avoidance alarm is triggered, a flashing red collision area marker is superimposed on the display and a continuous rapid alarm sound plays. The continuous rapid alarm sound plays at a frequency of 2000Hz and a period of 0.5 seconds.

Images