Global Navigation Satellite Systems (GNSS) provide users with positioning, velocity, and timing (PVT) services. It is the only positioning and navigation system that offers high precision, global coverage, and all-weather availability. After decades of development, GNSS has been widely and deeply applied across various sectors.

Because GNSS plays a fundamental and key role in military, commercial and civil fields, its vulnerability has also received more and more attention. As we all know, the power of the signal emitted by the navigation satellite is very weak when it reaches the ground, and GNSS users are extremely vulnerable to radio frequency interference and spoofing attacks [1, 2]. Studies have shown that [3] an interference source with a radiation power of only 1 W can render GNSS receivers within a range of approximately 15 km unable to function properly.

Radio frequency interference can be categorized into unintentional and intentional interference. Unintentional interference is generally caused by spectrum leakage or improper operation of radar or communication systems adjacent to the navigation frequency band. Intentional interference is typically used to attack an opponent’s navigation equipment or to protect one’s privacy [4]. Typical GNSS interference includes continuous wave interference, narrowband interference, pulse interference, sweep chirp interference and broadband interference [5–7]. The impact of different types of interference on GNSS receivers can be quantitatively characterized by the equivalent C/N₀ [8].

To address the threat of radio frequency interference, researchers have proposed a series of interference suppression methods. The basic principle is similar: utilizing the sparsity of interference in a certain dimension to detect and eliminate it. For example, due to the sparsity of continuous wave interference and narrowband interference in the frequency domain, Frequency Domain Pulse Blanking (FDPB) and Adaptive Filtering (AF) [9–12] have been proposed for suppression. In light of the sparsity of pulse interference in the time domain, Time Domain Pulse Blanking (TDPB) [13, 14] has been proposed for suppression. For chirp interference, it has been found that [15, 16] it also exhibits time-domain sparsity within a specific sweep frequency range, allowing it to be suppressed by setting the time-domain pulse to zero. Broadband interference is unique in that it does not exhibit sparsity in either the time or frequency domains; it must be received by an array antenna or polarized antenna to demonstrate sparsity in the spatial or polarization domain for suppression. This paper does not address broadband interference.

The interference mitigation method of radio frequency has been a topic not only in electronics but also in space and astrophysics. Among these, deep learning methods were already employed. References [17–19] investigate how to detect and identify radio-frequency interference using deep learning methods, which have achieved good results in the field of radio interferometry. However, these methods do not provide further approaches for interference cancellation or signal recovery. References [20, 21] study the use of Bayesian methods to eliminate interference and restore signals. This method performs well under the condition that the prior distribution of the data is known. However, the interference faced by GNSS receiver is unknown, dynamic, and uncertain. Moreover, under different types of interference, the data exhibit different probability distributions. Therefore, it is difficult to directly apply this method in GNSS interference mitigation. For a given type of interference, different mitigation methods yield varying suppression effects. This paper aims to design an interference mitigation method that not only identifies interference, but also automatically selects the optimal approach to suppress interference based on its characteristics, thereby enhancing the performance and robustness of the receiver. This method is referred to as the intelligent interference mitigation (IIM) method.

The innovative contributions of this paper can be summarized as follows: (a) We propose a GNSS IIM method that utilizes spectrum sensing to convert input data into spectrograms, followed by the application of a deep learning network to output the optimal interference mitigation method. (b) We present an implementation framework for the IIM method, which employs Short Time Fourier Transform (STFT) for time-frequency two-dimensional spectral sensing, with GoogLeNet chosen as the pre-trained model to build our deep learning network. (c) Finally, we conduct experiments using a data collector and a software-defined receiver (SDR) for interference mitigation and signal processing to evaluate the performance of the proposed method under various interference conditions. The results demonstrate the effectiveness of the proposed method and its ability to mitigate different types of interfering signals.

The content of this article is organized as follows: Section 2 establishes the signal model for the GNSS receiver; Section 3 introduces three typical GNSS interference mitigation methods—FDPB, AF, and TDPB—and analyzes their interference suppression effects. Section 4 proposes the basic principles and implementation framework of the GNSS IIM method. Sections 5, 6 design simulations and open-sky experiments to verify the performance of the proposed method. Finally, the main conclusions drawn from this study are presented.

For the sake of simplicity, we consider the reception of a single GNSS satellite signal due to the very low cross-correlation of the pseudorandom spreading codes. The signal received by a GNSS receiver, after being amplified, down-converted, filtered, and sampled, can be expressed in complex baseband form as: x n = 2 P s d n − τ 0 p n − τ 0 e j 2 π f 0 n + φ 0 + η n + i n (1)

Where P s is the power of the GNSS signal, d n is the navigation information modulated in the signal, p n is the pseudo-random code sequence, f 0 is the Doppler frequency shift, τ 0 and φ 0 are the pseudo-code phase and the carrier phase respectively, n is the time subscript. i n is a generalized interference signal, and different interference types have different expressions [4]. η n is the zero mean additive white Gaussian noise (AWGN) in the complex form. The real part and the imaginary part are independent and identically distributed. The variance is σ 2 , which can be described by Equation 2 as follows: σ 2 = N 0 B f e (2)

Where N 0 is the noise spectral density, B f e is the receiver front-end bandwidth.

When the receiver adopts active interference mitigation processing, the generalized interference mitigation process can be modeled by Equation 3 as follows: y n = Η x n (3)

In Equation 4, y n is the interference mitigation output, Η ⋅ represents a certain mathematical transformation, different interference mitigation methods have different mathematical transformation forms, which will be introduced in detail in Section 3.

After that, the receiver further performs correlation despreading on the anti-jamming output. The goal is to estimate the Doppler frequency shift and pseudo-code phase of the GNSS signal. This process can be achieved by maximizing the ambiguity function [8]: f ^ , τ ^ = arg max f , τ C f , τ = arg max f , τ ∑ n = 1 M y n p n − τ e − j 2 π f n (4)

In Equation 4, C f , τ is the ambiguity function, the symbol “argmax” is a mathematical notation used to describe the argument that maximizes a given function or expression. In the above equation, arg max f , τ C f , τ are the value of f and τ that make C f , τ as large as possible. f ^ and τ ^ are the estimated values of the doppler frequency shift and pseudo-code phase, respectively. M is the length of the sampled data used for the correlation despreading processing, and the corresponding signal duration is T = M / f s , where f s is the sampling rate. This formula describes the relationship between the ambiguity function and the doppler frequency shift as well as the pseudo-code phase. The value of the ambiguity function is maximized when the estimated values of the doppler frequency shift and pseudo-code phase are equal to that of the input GNSS signal.

The purpose of anti-jamming is to suppress the interfering signals and retain the useful signal to the greatest extent. The effect can be evaluated by the C/N₀ after correlation dispreading, which can be described by Equation 5 as follows: C / N 0 = 1 T ⋅ E C f ^ , τ ^ 2 V a r C f ^ , τ ^ (5) where E ⋅ represents the mean value and V a r ⋅ represents the variance.

Without considering special forms of receiving antennas such as array antennas or polarized antennas, conventional GNSS receivers employ three representative interference mitigation methods: FDPB, AF, or TDPB. The following sections introduce each method separately.

Considering that the interference environment faced by the receiver is unknown, dynamic, and uncertain, this paper proposes an IIM method, which automatically selects the optimal method to suppress interference based on the characteristics of interference, thereby improving the interference mitigation performance and robustness of the receiver.

Firstly, based on the simulation platform in Section 3, the Monte Carlo simulation method is used to generate sample data. The specific generation process is as follows: the signal received by the GNSS receiver is generated according to the signal model in Formula 1, and the global parameter settings are shown in Table 1. The parameters of the three types of interference are consistent with Section 3.4. 1,000 samples are generated under each interference, and a sample set containing 3,000 samples is generated in total. For each sample, by traversing three interference mitigation methods, one can obtain their respective output C/N₀. The interference mitigation method with the highest output C/N₀ is selected as the label for the sample.

The entire sample set is divided into three parts: 2,400 samples are used as the training set (800 samples for each disturbance), 300 samples are used as the verification set (100 samples for each disturbance), and 300 samples are used as the test set (100 samples for each disturbance), so 80% of the whole sample is used for training, and 20% is used for verification and testing.

The experiment utilizes the PyTorch [24] deep learning framework to complete the model construction, training and testing. The hyperparameters of the neural network are set as follows: the optimizer selects Adam; the learning rate is set to 0.001; the batch size is set to 20; the maximum number of training epoch is set to 10; and the loss function is cross entropy loss.

Load the trained network model into the IIM method proposed in Section 4, validate it on the test set, and compare it with three traditional interference mitigation methods. The comparison results of interference mitigation performance are shown in Figure 7.

Figure 7A shows the C/N₀ results obtained from the first 30 samples in the test set (including 10 narrowband interference, 10 pulse interference, and 10 chirp interference samples each). It can be seen that traditional interference mitigation methods have good suppression effects on one or two types of interference, but it is difficult to achieve good suppression effects on all three types of interference at the same time. However, the IIM method proposed in this paper can achieve this. Although the output C/N₀ obtained by the IIM method is not the highest on individual samples (because there is also a probability of errors in the prediction results of the network model), overall, the output C/N₀ obtained by it is the highest in most cases. Figure 7B shows the statistical results on the entire test set. It can be seen from the figure that the IIM method is the optimal interference mitigation method in a statistical sense, and the proportion of the C/N₀ above a given threshold (In engineering applications, it is generally agreed that the initial C/N₀ reduced by 3 dB is used as the threshold.) is more than 12% higher than traditional methods.

It should be particularly noted that, compared to traditional methods, IIM incorporates a spectrum sensing module and a deep learning network. The introduction of the deep learning network significantly increases the computational complexity by thousands of times. As a result, a dedicated GPU chip must be added specifically to accelerate the inference process of the deep learning model. Additionally, the inference process of the deep learning model introduces an extra processing latency of several milliseconds. The increase in computational complexity and processing latency limits the deployment of IIM in low-cost and real-time demanding GNSS receivers.

To further verify the proposed method, an open sky experiment is carried out. In this test, an antenna array based data collector is placed at a fixed position in the open wild for GNSS signal collection, while interfering signals are simulated and added in our SDR. This method is often used in GNSS jamming experiments [25, 26] to avoid impacting civil GNSS users. The open sky experiment configuration is shown in Figure 8.

The data collector includes 7 antennas, a front-end, an A/D converter, and a lighting port output. It should be mentioned that only data from antenna 1 (i.e., the center antenna) is used in our experiment. The output baseband data (sampling frequency is 62/3 MHz and duration is 2,300 ms) of the data collector are stored on a personal computer (PC) and processed by the SDR.

For the SDR, firstly, interference signals are generated and added to the baseband data at 1,000 ms, the parameters of the three types of interference signals are consistent with Section 3.4. For each type of interference, 30 random samples are generated, totaling 90 samples. For each sample, the SDR is used for processing, and the performance of different interference mitigation methods is compared.

Figure 9 shows the processing results of the SDR for the B3I signal (PRN 1) under sweep chirp interference (ISR is 64.53 dB, sweep bandwidth is 16.27 MHz, and sweep period is 0.935 ms). It can be observed that before the interference is turned on (0–1000 ms), with direct pass-through processing, the output C/N₀ is around 43 dB Hz. After applying the IIM algorithm, the output C/N₀ is slightly lower than that of the direct pass-through processing, due to the insertion loss introduced by the interference mitigation processing. After the interference is turned on (1000 ms–2300 ms), the output C/N₀ of the direct pass-through processing drops rapidly, and the receiver loses track of the signal, while the output C/N₀ of the IIM algorithm remains essentially unchanged, allowing the receiver to maintain normal tracking of the signal, as seen in Figure 10.

Figure 11 compares the time-frequency diagrams of the output signals obtained from direct pass-through and IIM. It can be seen that after applying IIM processing, the power of the chirp interference is significantly reduced, with a power level similar to that of the surrounding thermal noise, indicating that the IIM algorithm effectively suppresses the interference.

Finally, Figure 12 presents the statistical results across all 90 sample sets. For each sample, we take the average of the output C/N₀ results after the interference is turned on (with a duration of 1.3 s) as the final output C/N₀ result. It can be observed from the figure that the IIM method is statistically the optimal method, with its proportion above the threshold (initial C/N₀ minus 3 dB, i.e., 40 dB Hz) being more than 16% higher than that of traditional methods. This result is basically consistent with the outcomes derived from the simulation, with only minor differences in specific numerical values.

This article examines the interference challenges encountered by GNSS receivers and introduces a deep learning-based approach for GNSS interference mitigation. The method begins by analyzing the input signal in a two-dimensional spectrogram and then employs a deep learning network to determine the optimal strategy among three interference mitigation algorithms. The selected optimal interference mitigation strategy is subsequently conveyed to the signal capture tracking and PVT calculation processes via a switching mechanism. Compared to traditional interference mitigation techniques, the method outlined in this paper is capable of adapting to various types of interference and different interference parameters, demonstrating effective suppression across a range of interference scenarios. Simulation and open-sky experimental results have confirmed the efficacy of this approach.

In conclusion, the proposed technique is a potent interference mitigation tool that can substantially broaden the operational range of GNSS receivers under interference conditions. Future enhancements could focus on reducing computational complexity and enhancing performance in scenarios with multiple concurrent interference signals.