GNSS interference detection method based on ADS-B multi-feature transfer learning

Abstract

The invention relates to the technical field of GNSS interference detection, in particular to a GNSS interference detection method based on ADS-B multi-feature transfer learning, and the method comprises the following steps: building a first data set; processing the first data set to obtain a second data set; training an LSTM-AE model, and generating a low-dimensional potential vector; and training the DANN model. The LSTM-AE model is constructed to perform feature extraction on the ADS-B data, and the DANN model is combined to perform multi-version data transfer learning, so that the accuracy of GNSS interference detection is effectively improved, and GNSS interference phenomena in different environments can be identified more accurately. The method is compatible with data of multiple versions, and transfer learning is realized by using a DANN model, so that GNSS interference analysis can be performed on ADS-B data of different versions under a unified detection framework.

Description

Technical Field

[0001] The present invention relates to the technical field of GNSS interference detection, and in particular to a GNSS interference detection method based on ADS-B multi-feature transfer learning. Background Art

[0002] The Global Navigation Satellite System (GNSS) is the foundation of modern air navigation systems and a key technology to ensure navigation safety and service quality. However, GNSS signals are very susceptible to radio frequency interference due to their weak signal strength. As the threat of GNSS interference becomes increasingly severe, existing interference troubleshooting methods generally have problems such as low efficiency and long cycles, and they need to be improved urgently.

[0003] In recent years, the Automatic Dependent Surveillance-Broadcast (ADS-B) system has been widely used due to its advantages such as high efficiency, high precision and low cost. The ADS-B system mainly relies on GNSS to provide location information. Therefore, when the GNSS signal encounters radio frequency interference, the effectiveness and availability of airborne ADS-B will be directly affected, which may cause abnormalities in navigation integrity indicators and track information, affecting navigation safety.

[0004] With the continuous development of ADS-B technology, four main models of airborne ADS-B transmission systems have been derived: DO-260, DO-260A, DO-260B and DO-260C. Different versions of airborne ADS-B systems have different adoption of navigation integrity indicators. For details, see Figure 1 and Figure 2 Existing GNSS interference detection methods based on ADS-B data are usually designed for a certain version of navigation integrity indicators, so they are not compatible with different versions of ADS-B data, which limits the application scope and applicability of existing models. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies in the prior art and provide a GNSS interference detection method based on ADS-B multi-feature transfer learning.

[0006] The present invention is achieved through the following technical solutions:

[0007] A GNSS interference detection method based on ADS-B data, the method comprising the following steps:

[0008] Establishing a first data set, the first data set comprising initial features and navigation integrity indicators, the initial features comprising UTC time, longitude and latitude, the navigation integrity indicators comprising NIC, NACp, SIL of DO-260A version, NIC, NACp, SIL of DO-260B version, and NUCp of DO-260 version;

[0009] Performing data cleaning, coordinate conversion, feature expansion and normalization processing on the first data set to obtain a second data set;

[0010] Training an LSTM-AE model using the second data set, and performing feature extraction on the second data set using the trained LSTM-AE model to generate a low-dimensional latent vector;

[0011] A DANN model is trained based on the low-dimensional latent vector to achieve GNSS interference detection of multi-version ADS-B data.

[0012] Preferably, the data cleaning is to remove data in the data set that meets one of the following conditions:

[0013] A track feature in which 80% of the navigation integrity indicators are 0 or null values;

[0014] The altitude of the track feature is less than 100 meters;

[0015] The effective length of the track feature is less than 100 point tracks;

[0016] Track points with the same time value;

[0017] Track points with null latitude and longitude.

[0018] Preferably, the coordinate conversion comprises the following steps:

[0019] Convert the WGS84 coordinates in the dataset to two-dimensional rectangular coordinates.

[0020] Preferably, the feature extension comprises the following steps:

[0021] Calculating the track point intervals in the data set;

[0022] Perform logarithmic transformation on the track point intervals;

[0023] The track point intervals are feature expanded by a sliding window method.

[0024] Preferably, "feature expansion of track point intervals by a sliding window method" comprises the following steps:

[0025] Assuming the length of the first data set is N, the number of integrity indicator features is m, and the number of track features is n, the time series of the first data set is:

[0026] X= <X (1) ,X (2) ,…,X (t) ,…,X (N) >

[0027] Among them, the data set at time t is:

[0028]

[0029] Assuming the sliding window length is w and the sliding step size is s, the sequence of the second data set after being processed by the sliding window is:

[0030]

[0031] The maximum value, mean, sum, median and standard deviation of the integrity indicator feature of each sliding window in the sequence of the second data set are calculated, and the mean, sum and standard deviation of the position feature are calculated. At time t, each corresponding vector is:

[0032]

[0033] Preferably, the normalized calculation formula is:

[0034]

[0035] In the formula, x′ is the normalized value, x i is the value to be normalized, x max is the maximum value of the sample feature, x min is the minimum value of the sample feature.

[0036] Preferably, “training the LSTM-AE model through the second data set, and extracting features from the second data set through the trained LSTM-AE model to generate a low-dimensional latent vector” comprises the following steps:

[0037] Constructing a first LSTM-AE model and a second LSTM-AE model, wherein the LSTM-AE model includes an encoder and a decoder, wherein the encoder is used to receive original data and compress it into a low-dimensional latent space representation, and the decoder is used to reconstruct the original input data from the low-dimensional latent space representation;

[0038] The first LSTM-AE model is trained with the DO-260 version of ADS-B data, and the second LSTM-AE model is trained with the DO-260A version of ADS-B data and the DO-260B version of ADS-B data, wherein the training objective is to minimize the reconstruction error;

[0039] After training, the encoder is used to extract features from the ADS-B data to obtain a low-dimensional latent vector. Multiple versions of the low-dimensional latent vector have the same dimension.

[0040] Preferably, it is characterized in that "training the DANN model based on the low-dimensional potential vector" comprises the following steps:

[0041] The low-dimensional latent vectors of the DO-260A version and the DO-260B version are used as source domain datasets, and the low-dimensional latent vectors of the DO-260 version are used as target domain datasets;

[0042] Training the encoder and the category classifier of the DANN model with the source domain dataset, optimizing the supervision task loss, until the DANN model converges;

[0043] The encoder, category classifier and domain discriminator of the DANN model are trained with the source domain dataset and the target domain dataset, the supervision task loss is optimized, adversarial loss is introduced, and the encoder is adversarially updated through a gradient reversal layer;

[0044] Based on the trained DANN model, GNSS interference detection is performed on the target domain dataset to achieve GNSS interference detection for multiple versions of ADS-B data.

[0045] The beneficial effects of the present invention are:

[0046] The present invention constructs an LSTM-AE model to extract features from ADS-B data, and combines the DANN model to perform multi-version data migration learning, which effectively improves the accuracy of GNSS interference detection and enables more accurate identification of GNSS interference phenomena in different environments. The method is compatible with DO-260, DO-260A and DO-260B versions of data, and uses the DANN model to implement migration learning, so that different versions of ADS-B data can be used for GNSS interference analysis under a unified detection framework, thereby enhancing the applicability and generalization ability of the method. Compared with traditional GNSS interference detection methods that usually rely on a large amount of labeled training data, the present invention uses the migration learning ability of the DANN model to enable different versions of ADS-B data to learn from each other, greatly reducing the demand for labeled data and reducing data annotation costs.

[0047] Furthermore, the time series features are extracted through the sliding window method, and the statistical features (such as maximum value, mean, median, standard deviation, etc.) are calculated, which improves the stability of GNSS interference detection and enables the system to work effectively in complex environments.

[0048] Furthermore, the LSTM-AE model is used to extract features from ADS-B data and generate low-dimensional latent vectors, which not only improves the feature expression ability of the model, but also effectively reduces the computational complexity, making the detection system more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the navigation integrity indicator of different ADS-B versions.

[0050] Figure 2 It is a description of the ADS-B navigation integrity indicators.

[0051] Figure 3 It is a framework diagram of the interference detection system of the present invention.

[0052] Figure 4 It is a schematic diagram of the autoencoder structure.

[0053] Figure 5 This is a schematic diagram of the DANN structure. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and the best embodiment. Based on the embodiments in the invention, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the invention.

[0055] Reference Figure 3 The present invention provides a GNSS interference detection method based on ADS-B data, the method comprising the following steps:

[0056] S1, establishing a first data set, wherein the first data set includes initial features and navigation integrity indicators, wherein the initial features include UTC time, longitude and latitude, and the navigation integrity indicators include NIC, NACp, SIL of DO-260A version, NIC, NACp, SIL of DO-260B version, and NUCp of DO-260 version;

[0057] S2, performing data cleaning, coordinate conversion, feature expansion and normalization processing on the first data set to obtain a second data set;

[0058] S3, training an LSTM-AE model using the second data set, and performing feature extraction on the second data set using the trained LSTM-AE model to generate a low-dimensional latent vector;

[0059] S4, training a DANN model based on the low-dimensional latent vector to implement GNSS interference detection of multi-version ADS-B data.

[0060] Furthermore, under the influence of GNSS interference, the ADS-B track and navigation integrity indicators will produce abnormal situations, so it is necessary to take into account both the track characteristics and the integrity characteristics when selecting features. The initial features selected by the first data set of this embodiment include three track characteristics: UTC time, longitude, and latitude. The airborne ADS-B carried by flights in my country's airspace currently includes three versions: DO-260, DO-260A, and DO-260B. This article selects the navigation integrity features of the DO-260A / B version: NIC, NACp, SIL and the navigation integrity features of the DO-260 version: NUCp.

[0061] Furthermore, the conditions for data cleaning in step S2 include:

[0062] Condition 1: The track characteristics in which 80% of the navigation integrity indicators NIC or NUCp are 0 or null values.

[0063] Condition 2: The height of the track feature is less than 100 meters. Such flights are close to the airport surface, so the ADS-B signal is easily affected by the shielding of surface buildings, resulting in poor track quality.

[0064] Condition 3: The effective length of the track feature is less than 100 points.

[0065] Condition 4: Track points with the same time value.

[0066] Condition 5: Track point whose latitude and longitude are null values.

[0067] The specific method of data cleaning is: clearing the data in the first data set that meets one of the above conditions.

[0068] Furthermore, the ADS-B original position data of the first data set adopts WGS84 coordinates, and the coordinates need to be projected before training to ensure their equidistance. Therefore, in step S2 of this embodiment, the conformal conic projection method commonly used in the civil aviation industry is used to perform coordinate conversion, and the WGS84 coordinates expressed in longitude and latitude are converted into two-dimensional rectangular coordinates.

[0069] Furthermore, the feature expansion in step S2 includes the following steps:

[0070] S201, calculate the track point interval in the data set, the purpose is to enable the neural network to learn the change pattern of flight position information over time, rather than just the numerical value of the data itself, to enhance the applicability of the model. The calculation formula of the track point interval is:

[0071]

[0072] In the formula, (x t ,y t) is the flight coordinate at time t, (x t―1 ,y t―1 ) are the flight coordinates at time t-1.

[0073] S202, performing logarithmic transformation on the track point intervals to prevent extreme values ​​from affecting the model.

[0074] S203, feature expansion is performed on the track point intervals by a sliding window method. Due to various non-GNSS interference factors such as flight maneuvers, ADS-B signal shielding, weather and equipment failure, the navigation integrity indicators and tracks in the ADS-B data may also produce abnormal fluctuations. In order to suppress the impact of such random abnormal fluctuations on the interference detection results and reduce the false detection rate of the model, this embodiment calculates different statistical values ​​of the specified features in each sliding window based on the original features of the ADS-B data and based on the sliding window technology, and constructs a data set as a new feature, and uses the new extended features to reduce the interference of noise in the original data on the model.

[0075] The specific method of step S203 is:

[0076] Assuming the length of the first data set is N, the number of integrity indicator features is m, and the number of track features is n, the time series of the first data set is:

[0077] X= <X (1) ,X (2) ,…,X (t) ,…,X (N) >

[0078] Among them, the data set at time t is:

[0079]

[0080] Assuming the sliding window length is w and the sliding step size is s, the sequence of the second data set after being processed by the sliding window is:

[0081]

[0082] The maximum value, mean, sum, median and standard deviation of the integrity indicator feature of each sliding window in the sequence of the second data set are calculated, and the mean, sum and standard deviation of the position feature are calculated. At time t, each corresponding vector is:

[0083]

[0084] Furthermore, the normalized calculation formula in step S2 is:

[0085]

[0086] In the formula, x′ is the normalized value, x i is the value to be normalized, x max is the maximum value of the sample feature, x min is the minimum value of the sample feature.

[0087] Further, refer to Figure 3 ,Autoencoder (AE) is an unsupervised neural network model architecture, which usually consists of two parts: encoder and decoder. Its structural characteristics are that the input layer and output layer have the same dimension, and the network is designed to minimize the difference between input and output.

[0088] Assume that the observed data at time t has a total of m features, and the input vector is in represents the observed value of the mth feature at time t. The encoder can transform the high-dimensional input X (t) The input vector is converted to a low-dimensional representation vector H containing c hidden features ( t ) , thereby forcing the neural network to learn the most informative features and remove redundant information, while the decoder is used to transform the latent variables H ( t ) Restore to the same dimensions as the initial The Euclidean distance between the input vector and the reconstructed vector is recorded as the reconstruction error.

[0089]

[0090] When training an autoencoder, as the codec continuously receives and processes a large amount of training data, the data goes through the encoding and decoding process, and the weights in the network gradually reach a stable state under the goal of minimizing the reconstruction error. During the encoding process, the network selectively retains the information that best represents the essence of the data, while discarding redundant or unimportant details to adapt to the spatial limitations of lower dimensions, thereby achieving effective compression of high-dimensional features and extraction of key features.

[0091] Long Short-Term Memory Autoencoder LSTM-AE is a deep learning model that combines LSTM and AE. This allows LSTM-AE to not only rely on information at the current time point, but to effectively remember and transmit long-term information, while learning the temporal dependencies and dynamic characteristics of time series data. In addition, LSTM-AE, with its unique design, can flexibly process time series data of various lengths, such as ADS-B data with different time series lengths. This makes LSTM-AE more advantageous when processing complex time series tasks such as flight data.

[0092] Step S3, based on the feature extraction module of the LSTM-AE model, maps the data sets of multiple versions of ADS-B with different numbers of features into a common feature space, making the number of features or feature dimensions the same. And through the learning ability of the autoencoder of the LSTM-AE model, more meaningful features are captured, thereby improving the effect of the classification task.

[0093] The specific method of step S3 is:

[0094] S301, constructing a first LSTM-AE model and a second LSTM-AE model, wherein the LSTM-AE model includes an encoder and a decoder, wherein the encoder is used to receive the original data and compress it to a low-dimensional latent space representation, and the decoder is used to reconstruct the original input data from the low-dimensional latent space representation. In order to achieve consistency in feature dimensions, the latent vectors output by the encoders of the two LSTM-AE models are set to have the same dimension.

[0095] S302, training the first LSTM-AE model with the DO-260 version of ADS-B data, and training the second LSTM-AE model with the DO-260A version of ADS-B data and the DO-260B version of ADS-B data, wherein the training objective is to minimize the reconstruction error. During the training process, the latent vector learned by the encoder not only retains the main features of the original data, but also maintains consistent feature representation across data sets.

[0096] S303, after completing the training, the encoder is used to extract features from the ADS-B data to obtain low-dimensional latent vectors, and the dimensions of multiple versions of the low-dimensional latent vectors are the same. These low-dimensional latent vectors not only have a uniform dimension, but also help capture the essential characteristics of the data, which is beneficial to subsequent task processing.

[0097] Furthermore, since airborne ADS-B devices are dispersed and the navigation integrity indicators of different versions of ADS-B airborne devices are also different, a single model cannot detect the GNSS interference of all flights in the airspace. To address the above problems, this embodiment uses a domain adversarial neural network (DANN) for interference detection.

[0098] The DANN model structure is as follows Figure 4 As shown in Figure 1, it consists of three components: encoder, category classifier, and domain classifier. The encoder is responsible for extracting feature representations from the input data; the category classifier predicts the category label based on these features; and the domain classifier aims to determine whether the sample is from the source domain or the target domain based on the feature representation.

[0099] DANN uses an adversarial training mechanism, under which the encoder and the category classifier are collaboratively optimized to minimize the supervised task loss on the source domain, that is, the classification error. At the same time, the encoder also needs to optimize its parameters so that it is difficult for the domain classifier to distinguish whether the feature is from the source domain or the target domain. Specifically, this process introduces a gradient reversal layer (GRL) so that during the back propagation process, the update direction for the encoder is opposite to the loss gradient direction of the domain classifier. The goal of the domain classifier is to minimize its own binary classification error, thereby accurately discerning the domain source of the feature.

[0100] In this embodiment, the ADS-B version that accounts for the largest proportion in the airspace is used as the source domain data, and the ADS-B data of other versions is used as the target domain data. By utilizing the characteristics of DANN, the source data and the target data are projected onto the same coding space, so that the encoder learns feature representations that are representative of both the source domain and the target domain and are independent of the specific domain, thereby improving the generalization performance of the model in the unlabeled target domain, thereby realizing GNSS interference detection for different versions of ADS-B devices.

[0101] The specific method of step S4 is:

[0102] S401, using the low-dimensional latent vectors of the DO-260A version and the low-dimensional latent vectors of the DO-260B version as source domain data sets, and using the low-dimensional latent vectors of the DO-260 version as target domain data sets;

[0103] S402, initialization stage: training the encoder and category classifier of the DANN model with the source domain dataset, optimizing the supervision task loss, until the DANN model converges;

[0104] S403, adversarial training stage: train the encoder, category classifier and domain discriminator of the DANN model with the source domain data set and the target domain data set, optimize the supervision task loss, introduce adversarial loss, and adversarially update the encoder through the gradient reversal layer, so that the features generated by it are both helpful for category classification and difficult for the domain classifier to identify the source domain. Finally, GNSS interference detection of multiple versions of ADS-B can be achieved.

[0105] S404, based on the trained DANN model, GNSS interference detection is performed on the target domain data set to achieve GNSS interference detection on multiple versions of ADS-B data.

[0106] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A GNSS interference detection method based on ADS-B data, characterized in that: The method comprises the following steps: Establishing a first data set, the first data set comprising initial features and navigation integrity indicators, the initial features comprising UTC time, longitude and latitude, the navigation integrity indicators comprising NIC, NACp, SIL of DO-260A version, NIC, NACp, SIL of DO-260B version, and NUCp of DO-260 version; Performing data cleaning, coordinate conversion, feature expansion and normalization processing on the first data set to obtain a second data set; Training an LSTM-AE model using the second data set, and performing feature extraction on the second data set using the trained LSTM-AE model to generate a low-dimensional latent vector; A DANN model is trained based on the low-dimensional latent vector to achieve GNSS interference detection of multi-version ADS-B data.

2. A GNSS interference detection method based on ADS-B data according to claim 1, characterized in that: The data cleaning is to remove the data in the data set that meets one of the following conditions: A track feature in which 80% of the navigation integrity indicators are 0 or null values; The altitude of the track feature is less than 100 meters; The effective length of the track feature is less than 100 point tracks; Track points with the same time value; Track points with null latitude and longitude.

3. The GNSS interference detection method based on ADS-B data according to claim 1, characterized in that: The coordinate conversion comprises the following steps: Convert the WGS84 coordinates in the dataset to two-dimensional rectangular coordinates.

4. The GNSS interference detection method based on ADS-B data according to claim 1, characterized in that: The feature extension includes the following steps: Calculating the track point intervals in the data set; Perform logarithmic transformation on the track point intervals; The track point intervals are feature expanded by a sliding window method.

5. The GNSS interference detection method based on ADS-B data according to claim 1, characterized in that: "Feature expansion of track point intervals by sliding window method" includes the following steps: Assuming the length of the first data set is N, the number of integrity indicator features is m, and the number of track features is n, the time series of the first data set is: X=<X (1) ,X (2) ,…,X (t) ,…,X (N) > Among them, the data set at time t is: Assuming the sliding window length is w and the sliding step size is s, the sequence of the second data set after being processed by the sliding window is: The maximum value, mean, sum, median and standard deviation of the integrity indicator feature of each sliding window in the sequence of the second data set are calculated, and the mean, sum and standard deviation of the position feature are calculated. At time t, each corresponding vector is:

6. The GNSS interference detection method based on ADS-B multi-feature transfer learning according to claim 1 is characterized in that: The normalized calculation formula is: In the formula, x′ is the normalized value, x i is the value to be normalized, x max is the maximum value of the sample feature, x min is the minimum value of the sample feature.

7. The GNSS interference detection method based on ADS-B multi-feature transfer learning according to claim 1 is characterized in that: “Training the LSTM-AE model through the second data set, and extracting features from the second data set through the trained LSTM-AE model to generate a low-dimensional latent vector” includes the following steps: Constructing a first LSTM-AE model and a second LSTM-AE model, wherein the LSTM-AE model includes an encoder and a decoder, wherein the encoder is used to receive original data and compress it into a low-dimensional latent space representation, and the decoder is used to reconstruct the original input data from the low-dimensional latent space representation; The first LSTM-AE model is trained with the DO-260 version of ADS-B data, and the second LSTM-AE model is trained with the DO-260A version of ADS-B data and the DO-260B version of ADS-B data, wherein the training objective is to minimize the reconstruction error; After training, the encoder is used to extract features from the ADS-B data to obtain a low-dimensional latent vector. Multiple versions of the low-dimensional latent vector have the same dimension.

8. The GNSS interference detection method based on ADS-B multi-feature transfer learning according to claim 1, characterized in that: “Training a DANN model based on the low-dimensional latent vector” includes the following steps: The low-dimensional latent vectors of the DO-260A version and the DO-260B version are used as source domain datasets, and the low-dimensional latent vectors of the DO-260 version are used as target domain datasets; Training the encoder and the category classifier of the DANN model with the source domain dataset, optimizing the supervision task loss, until the DANN model converges; The encoder, category classifier and domain discriminator of the DANN model are trained with the source domain dataset and the target domain dataset, the supervision task loss is optimized, adversarial loss is introduced, and the encoder is adversarially updated through a gradient reversal layer; Based on the trained DANN model, GNSS interference detection is performed on the target domain dataset to achieve GNSS interference detection for multiple versions of ADS-B data.

Images