Navigation interference signal identification method based on federal learning

Abstract

The invention discloses a navigation interference signal identification method based on federated learning. The method comprises the following steps: S1, acquiring a satellite intermediate frequency signal and a spectrogram corresponding to the signal; s2, constructing a distributed interference identification model, and establishing a plurality of clients; s3, each client carries out network training and updating to obtain respective network weight; s4, judging whether the current updating frequency reaches a preset period or not, and if yes, entering a summarization and optimization process of model parameters; if not, continuing local training updating; s5, summarizing and optimizing the model parameters of all the clients, and returning the updated model parameters to the clients; s6, calculating values of all loss functions, and adjusting parameters of an output layer of the model; s7, updating all network parameters of all clients; s8, repeatedly executing local updating and global parameter summarization until a preset training frequency is reached, and finally obtaining global network parameters; and S9, outputting the trained model. According to the identification method, interference can be rapidly identified, and the identification accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of satellite navigation communication technology, and in particular to a navigation interference signal recognition method based on federated learning. Background Art

[0002] As the core of the positioning, navigation and timing system, the global satellite navigation system has played an important role in the military and civilian fields and has developed into an indispensable infrastructure construction for a country. Its technology is everywhere, from mobile phones, cars to communication systems and logistics industries, all of which rely on its positioning system or precise time synchronization. Its existence has affected the application and development of all aspects of modern life. However, since the power of the satellite signal received by the navigation receiver is very weak, this is the inherent vulnerability of the navigation signal and is easily affected by interference. Therefore, how to improve the interference detection performance of satellite navigation signals has always been the focus of people's research. And with the continuous development of modern science and technology, navigation electronic warfare has become increasingly complex. Navigation confrontation capability is an important manifestation of a country's scientific and military strength. Whether in peacetime or wartime, it is crucial to ensure the normal operation of the satellite navigation system. The core issue of interference identification is how to accurately and quickly identify the type of interference. Summary of the invention

[0003] In view of the above-mentioned technical problems to be solved, the present invention provides a navigation interference signal identification method based on federated learning. By performing network training through federated learning, interference can be quickly identified and the accuracy of identification can be improved.

[0004] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0005] A method for identifying navigation interference signals based on federated learning comprises the following steps:

[0006] Step S1, obtaining the I and Q paths of the satellite intermediate frequency signal and the navigation interference signal, performing FFT on the I / Q data, and obtaining a spectrum diagram corresponding to the signal;

[0007] Step S2, constructing a distributed interference identification model and establishing multiple clients;

[0008] Step S3, input the obtained spectrum data into the recognition model for training, and for each client, perform network training and update to obtain respective network weights;

[0009] Step S4, determine whether the current update number is an integer multiple of the local update number, if yes, continue to execute steps S5 and S6, otherwise go to step S7;

[0010] Step S5, sending the model parameters on all clients to a central server for overall aggregation and optimization, and then sending the updated model parameters back to each client;

[0011] Step S6, calculating the values ​​of all loss functions and adjusting the output layer parameters of the model to minimize the prediction error;

[0012] Step S7, updating all network parameters of all clients;

[0013] Step S8, repeating steps S3 to S7 until the training is completed to obtain the final global network parameters;

[0014] Step S9: output the trained model.

[0015] As a further improvement of the above technical solution:

[0016] Preferably, the step S1 includes the following contents:

[0017] S1-1, receiving satellite interference signals through a navigation receiver;

[0018] S1-2, performs down-conversion, filtering and ADC conversion on the satellite signal and then outputs an intermediate frequency signal;

[0019] S1-3, processing the intermediate frequency signal in a tracking loop to obtain I / Q data;

[0020] S1-4, perform FFT on the I / Q data to obtain a spectrum diagram.

[0021] Preferably, the interference signal includes at least amplitude modulation interference, chirp interference, frequency modulation interference, pulse interference, narrowband interference, and deceptive interference.

[0022] Preferably, in step S3, the local model parameter of each client is set to w i (t), where t = 0, 1, 2, ..., N represents the number of training iterations;

[0023] When t=0, all client parameters are initialized to the same parameter w i (0); when t>0, w is calculated based on the parameters of the previous iteration and the local loss function i (t); Perform global aggregation after a round of local updates:

[0024]

[0025] Where D represents the total number of sample sets, D i represents the number of samples of each client, i represents the client number, and M represents the total number of clients;

[0026] In each training iteration, the client will perform a local update followed by a global aggregation step.

[0027] Preferably, the loss function is:

[0028]

[0029] Among them, loss(w) represents the loss of the global parameters of the i-th client using the local sample set, D i represents the size of the local sample set, and M represents the total number of clients.

[0030] Preferably, in step S3, represents the parameters of the client at node i after possible global aggregation; in the tth iteration, if there is no global aggregation, then If global aggregation is performed,

[0031] For the i-th client, the local update rule is:

[0032]

[0033] Here, η represents the learning rate.

[0034] Compared with the prior art, the navigation interference signal identification method based on federated learning provided by the present invention has the following advantages:

[0035] The navigation interference signal identification method based on federated learning of the present invention can be applied in many complex scenarios, including but not limited to the protection of important infrastructure in cities. Global optimization of multiple clients is achieved through federated learning. Interference signal classifiers are trained locally on each device using federated learning, and model updates are aggregated and averaged on a central server. Multiple convolutional layers are used on the client to enhance the generalization ability of the model, and the spectrum diagram of GNSS interference signals is classified using an improved FedAvg framework. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the distributed network structure of the present invention.

[0037] Figure 2 It is a schematic diagram of the process of the present invention.

[0038] Figure 3 This is a visualization result diagram of the accuracy comparison of the algorithm output after 400 rounds of iteration during the experimental verification of the present invention.

[0039] Figure 4 This is a confusion matrix result diagram after the algorithm is trained during the experimental verification of the present invention. DETAILED DESCRIPTION

[0040] The specific embodiments of the present invention are described in detail below. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0041] like Figure 1 and Figure 2 As shown, the navigation interference signal identification method based on federated learning of the present invention comprises the following steps:

[0042] Step S1, obtain the I and Q paths of the satellite intermediate frequency signal and the navigation interference signal, perform FFT on the I / Q data, and obtain the spectrum diagram corresponding to the signal. The interference signal at least includes amplitude modulation interference, chirp interference, frequency modulation interference, pulse interference, narrowband interference, and deception interference.

[0043] Includes the following:

[0044] S1-1, receiving satellite interference signals through a navigation receiver, the interference signals at least including amplitude modulation interference, chirp interference, frequency modulation interference, pulse interference, narrowband interference, and deception interference;

[0045] S1-2, performs down-conversion, filtering and ADC conversion on the satellite signal and then outputs an intermediate frequency signal;

[0046] S1-3, processing the intermediate frequency signal in a tracking loop to obtain I / Q data;

[0047] S1-4, perform FFT on the I / Q data to obtain a spectrum diagram.

[0048] Step S2, constructing a distributed interference identification model, establishing multiple clients, and initializing parameters of each client.

[0049] The signal received by the receiver may consist of real signals, interference signals, deceptive signals and noise signals (no signal).

[0050] r(t)=g(t)+j(t)+s(t)+n(t)

[0051] Where: g(t) is the real signal, j(t) is the suppressed signal, s(t) is the deception signal, and n(t) is the Gaussian white noise signal with zero mean.

[0052] The model divides signals into seven categories, namely, true signal, AM interference, chirp interference, FM interference, pulse interference, narrowband interference, and deceptive interference. H0 represents the true signal.

[0053] H1 represents amplitude modulation interference,

[0054]

[0055] This type of interference is the simplest type of interference and is also the most studied interference. They are also called continuous wave (CW) jammers. AM noise interference is different from the AM / CW simple interference described here; the AM interference in this embodiment refers to single-tone or multi-tone sinusoidal signals weighted with various amplitudes. AM interference can be single-tone interference or multi-tone interference. AM interference consists of k=1,...,..., k single tones (k=1 for single-tone case, k>1 for multi-tone case). It is mainly characterized by three parameters, namely and They represent the power at the antenna and the corresponding carrier frequency and phase of the kth interference component, respectively.

[0056] H2 represents chirp interference,

[0057]

[0058] The interference signals have their frequency modulated linearly over time. They are generated by linearly sweeping their frequency over a certain frequency range and a certain time period, after which the process starts again with the initial frequency. J , the starting frequency of the scan f J (At time T swp = 0), minimum and maximum frequency sweep f min and f max , and the scanning period T swp , which is the jammer from f min Scan to f max The variable b = ±1 is a flag to determine whether there is an upward linear frequency modulation (b = 1) or a downward linear frequency modulation (b = -1); θ J represents the initial phase of the interference signal;

[0059] is the instantaneous frequency of the interference signal.

[0060] H3 indicates FM interference.

[0061]

[0062] FM interference can also be single tone or multi-tone. The carrier frequency is time-dependent. k Incorporated into the signal model, this parameter is the modulation index of the kth tone. For the single tone case, k=1.

[0063] H4 indicates pulse interference,

[0064]

[0065] The pulse interference signal is only effective during the repetitive time period, and the effective period of the pulse is called the "duty cycle". τ (t) is a rectangular pulse with a width of τ (duty cycle), f r is the pulse repetition frequency, δ(t) is the Dirac pulse, is the convolution operator.

[0066] H5 represents narrowband interference,

[0067]

[0068] Narrowband noise jammers are jammers that span a narrow band of the signal spectrum, for example by sending a higher power signal in that region of the spectrum. Where β is the modulation index, n(ζ) represents a signal with zero mean and Variance of a stationary random process.

[0069] H6 represents deceptive interference.

[0070]

[0071] Where N is the number of signals captured by the receiver. A represents the signal amplitude. i and C i Represents data code and C / A code respectively. kT s Represents the sampling time. τ i Indicates the carrier phase offset (rad). IF is the local intermediate frequency. The carrier Doppler frequency shift (Hz) and code phase delay are expressed as fd i and φ i .

[0072] Each type of signal has different spectral features. The network model is used to extract features and classify the signal. The model is trained using the improved FedAvg framework, taking into account the CNN architecture. The CNN architecture consists of two convolutional layers, two pooling layers, and a fully connected layer with a ReLU activation function. The learning rate is 0.02, and the SGD optimizer is used. The last layer is a softmax layer, which produces classification results and divides the signal into 7 categories. The cross entropy is used for the loss function.

[0073] like Figure 1 As shown in the figure, multiple clients are established and parameters of each client are initialized. Federated learning requires training local clients. All clients use local data sets to perform one or more rounds of updated network parameters, which are sent to the server through the communication network. The server aggregates the received client parameters by using the proportion of each client sample as an aggregation rule to obtain global parameters, and then sends the global parameters to each client for training and updating.

[0074] In step S3, the spectrum graph data is input into the recognition model for training, and for each client, network training is performed separately to update the respective network weights.

[0075] Set the local model parameters of each client to w i (t), where t=0,1,2,...,N represents the number of training iterations. When t=0, all client parameters are initialized to the same parameter w i (0). When t>0, w is calculated based on the parameters of the previous iteration and the local loss function i (t). In this way, the gradient descent process of the local dataset and parameters is updated locally. The server performs global aggregation after a round of local updates:

[0076]

[0077] Where D represents the total number of sample sets, D i represents the number of samples for each client, i represents the client number, and M represents the total number of clients.

[0078] In each training iteration, the client will perform a local update followed by a global aggregation step.

[0079] Step S4, determine whether the current partial update number is an integer multiple of the local update number. If yes, execute steps S5 and S6; otherwise, go to step S7.

[0080] Step S5, the model parameters on all clients are sent to a central server, which performs overall aggregation and optimization, and then sends the updated model parameters back to each client.

[0081] Step S6, calculate the values ​​of all loss functions, and adjust the output layer parameters of the model according to the values ​​of the loss functions to minimize the prediction error. The loss function is:

[0082]

[0083] Among them, loss(w) represents the loss of the global parameters of the i-th client using the local sample set, D i represents the size of the local sample set, and M represents the total number of clients. The global loss function may not be directly available on the server, and the loss of each client needs to be sent to the server through the transmission network.

[0084] represents the parameters of the client at node i after possible global aggregation. In the tth iteration, if there is no global aggregation, then If aggregation is done globally, then

[0085] For the i-th client, the local update rule is:

[0086]

[0087] Among them, η represents the learning rate.

[0088] Step S7, all network parameters of all clients are updated, and each client uses the new network parameters for training. The server sends the aggregated network parameters to each client for updating, and the network parameters include the weights and biases of the neural network model.

[0089] Step S8, repeat steps S3 to S7, until the number of training rounds reaches 400 or the model loss value reaches the convergence condition, end the training, obtain the final global network parameters, and apply the obtained global network parameters.

[0090] Step S9: output the trained model for navigation interference signal recognition and classify various interference signals.

[0091] Experimental verification

[0092] In this experimental verification, model training was performed for 6 typical interference types, namely, amplitude modulation interference, chirp interference, frequency modulation interference, pulse interference, narrowband interference and deceptive interference. Each interference type provides 10,000 sample data for model training. In addition, the present invention is also applicable to the identification and classification of other interference scenarios, such as deceptive interference, complex environments with mixed multiple interference signals, and dynamic interference scenarios. In these cases, the federated learning architecture can effectively cope with interference signals distributed on different clients, ensuring the robustness and generalization ability of the model in different application scenarios. These scenarios include but are not limited to real-time navigation in autonomous driving, precise positioning in drone flight, safety monitoring of intelligent transportation systems, etc., and the safety and stability of the overall system are improved by quickly identifying and classifying different interference signals.

[0093] Figure 3 The interference classification results are compared under 400 rounds of communication when the number of clients of the proposed federated learning training model is set to M=20, 30, and 40, and the number of clients of the FedAvg algorithm is set to 20. The interference classification algorithm proposed in this embodiment is superior to the traditional CNN and FedAvg algorithms in terms of classification accuracy.

[0094] Figure 4 The confusion matrix when the number of clients is set to 20. It can be seen that the recognition accuracy of each interference signal class is relatively high. Among them, the accuracy of DME interference signal, deception interference signal and real signal is the highest, all higher than 99%.

[0095] The above implementation cases are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, they are not intended to limit the present invention. Therefore, any simple modification, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for identifying navigation interference signals based on federated learning, characterized in that: The following steps are involved: Step S1, obtaining the I and Q paths of the satellite intermediate frequency signal and the navigation interference signal, performing FFT on the I / Q data, and obtaining a spectrum diagram corresponding to the signal; Step S2, constructing a distributed interference identification model and establishing multiple clients; Step S3, input the obtained spectrum data into the recognition model for training, and for each client, perform network training and update to obtain respective network weights; Step S4, determine whether the current update number is an integer multiple of the local update number, if yes, continue to execute steps S5 and S6, otherwise go to step S7; Step S5, sending the model parameters on all clients to a central server for overall aggregation and optimization, and then sending the updated model parameters back to each client; Step S6, calculating the values ​​of all loss functions and adjusting the output layer parameters of the model to minimize the prediction error; Step S7, updating all network parameters of all clients; Step S8, repeating steps S3 to S7 until the training is completed to obtain the final global network parameters; Step S9: output the trained model.

2. The method for identifying navigation interference signals based on federated learning according to claim 1, characterized in that: In the step S1, Includes the following: S1-1, receiving satellite interference signals through a navigation receiver; S1-2, performs down-conversion, filtering and ADC conversion on the satellite signal and then outputs an intermediate frequency signal; S1-3, processing the intermediate frequency signal in a tracking loop to obtain I / Q data; S1-4, perform FFT on the I / Q data to obtain a spectrum diagram.

3. The method for identifying navigation interference signals based on federated learning according to claim 2, characterized in that: The interference signal at least includes amplitude modulation interference, chirp interference, frequency modulation interference, pulse interference, narrowband interference, and deceptive interference.

4. The method for identifying navigation interference signals based on federated learning according to claim 1, characterized in that: In step S3, the local model parameters of each client are set to w i (t), where t = 0, 1, 2, ..., N represents the number of training iterations; When t=0, all client parameters are initialized to the same parameter w i (0); when t>0, w is calculated based on the parameters of the previous iteration and the local loss function i (t); Perform global aggregation after a round of local updates: Where D represents the total number of sample sets, D i represents the number of samples of each client, i represents the client number, and M represents the total number of clients; In each training iteration, the client will perform a local update followed by a global aggregation step.

5. The method for identifying navigation interference signals based on federated learning according to claim 4, characterized in that: The loss function is: Among them, loss(w) represents the loss of the global parameters of the i-th client using the local sample set, D i represents the size of the local sample set, and M represents the total number of clients.

6. The method for identifying navigation interference signals based on federated learning according to claim 4, characterized in that: In step S3, it is assumed that represents the parameters of the client at node i after possible global aggregation; in the tth iteration, if there is no global aggregation, then If global aggregation is performed, For the i-th client, the local update rule is: Among them, η represents the learning rate.

Images