An Adaptive Anti-interference Algorithm Based on Autoencoder and Meta-learning

Abstract

An adaptive anti-interference algorithm based on autoencoder and meta-learning is disclosed. This invention addresses the problem that traditional anti-interference algorithms cannot automatically adjust parameters to adapt to different types of interference, whether known or unknown. The steps of this invention are: Step 1, offline meta-learning; Step 2, online learning and fine-tuning. This invention belongs to the field of communication technology.

Description

Technical Field

[0001] This invention relates to an adaptive anti-interference algorithm, belonging to the field of communication technology. Background Technology

[0002] In modern information-based collaborative environments, such as unmanned aerial vehicle (UAV) networks, multiple nodes are often discrete, and communication and collaboration between different nodes rely on data links. Therefore, maintaining reliable data link transmission is a key research direction. For wireless data link channels, the communication quality is mainly affected by the quality of the channel itself and communication interference from adversaries. However, the open and exposed nature of wireless channels makes them highly susceptible to interference. Therefore, during communication, it is crucial to estimate the current channel and the interference received, and adaptively adjust the communication equipment parameters to adapt to different channel scenarios and different types of interference.

[0003] Traditional anti-jamming algorithms typically require adjusting their modulation and coding schemes and the magnitude of their frequency based on the type, frequency, and power of the interference. When channel conditions are good and enemy interference power is low, some transmission efficiency can be sacrificed, such as using lower-order modulation and coding schemes to improve the system's bit error rate performance. However, when channel conditions are poor, enemy interference power is high, and the interference type is complex, the algorithm often chooses to evade the current interference frequency by using frequency hopping to continuously change the current communication frequency, thereby reducing the impact of interference on normal communication and improving system reliability. While this method of anti-jamming by changing the parameters of a specific module in a traditional communication system is simple and quick, it often has the following drawbacks.

[0004] First, there is the issue of optimality. The modulation, coding methods, and frequency points preset by traditional anti-interference algorithms are often preset and fixed. For example, the modulation is often a fixed modulation constellation diagram of different orders, and the coding is a convolutional code or LDPC code with different code rates and code types. It is certainly feasible to combine them in different ways to adapt to different types and intensities of interference, but it is often not optimal.

[0005] Secondly, there are limitations. Traditional anti-jamming methods are often based on the premise that the interference can be identified. However, in modern communication warfare, enemy interference is often a superposition of multiple types of interference or intelligent agent interference. It is difficult to derive the best coding method to counter the current interference through traditional mathematical methods. Moreover, for broadband interference, the frequency points of traditional frequency hopping technology are often completely submerged. At this time, how to counter the interference becomes a problem. Finally, there is the issue of robustness. When enemy interference changes in real time and rapidly, traditional anti-jamming methods often suffer significant performance loss due to delays because they need to identify and adjust first. Furthermore, the anti-jamming benefits brought by a fixed modulation method are often unstable for different types of interference. Summary of the Invention

[0006] To address the problem that traditional anti-interference algorithms cannot automatically adjust parameters to adapt to different types of interference, whether known or unknown, this invention proposes an adaptive anti-interference algorithm based on autoencoders and meta-learning.

[0007] The technical solution adopted by the present invention to solve the above problems is as follows: The specific steps of the present invention are as follows:

[0008] Step 1: Offline Meta-Learning;

[0009] Step 2: Online learning and fine-tuning.

[0010] Furthermore, the offline meta-learning steps in step one are as follows:

[0011] Step A: Establish the offline training system model: During offline training, the neural network parameters of the encoder and decoder at the transmitter and receiver are randomly initialized. For each iteration cycle, the transmitter randomly generates a binary bit sequence, which, after passing through the channel via the encoder, is affected by channel contamination and interference. The type and power of interference added during training vary depending on the task. After receiving the signal, the receiver sends it to the decoder to obtain the prediction result. Within one iteration cycle, the meta-learner performs gradient descent based on the support set of each task to obtain the temporarily stored receiver parameters θ. R Then, the actual direction in which the model parameters should be updated is obtained through the query set in the task; after completing the above inner loop for all tasks in one iteration, the updated parameters φ of the decoder model are obtained by combining the results. R Since the training is offline, the receiver can pass the trained parameters to the transmitter's encoder to jointly update the transmitter's encoder parameters φ. T Repeat the above iterative process until the receiver loss gradually converges, completing the offline training process of this system.

[0012] Step B: Generate interference set;

[0013] Step C: Construct the meta-training task set.

[0014] Furthermore, the steps of the offline training algorithm in step A are as follows:

[0015] Step 1: Randomly initialize encoder and decoder parameters φ T φ R α represents the inner loop step size, and β represents the outer loop step size;

[0016] Step 2, repeat steps 3 through 8;

[0017] Step 3: The current interference task set is sampled as T. i ~p(T); p(T) represents the distribution of the interference task set;

[0018] Step 4: Repeat steps 5 and 6 for all tasks;

[0019] Step 5: Calculate the current task T i The loss function on the decoder corresponds to the gradient at θ. R value at

[0020] Step 6: Calculate the corresponding gradient at encoder θ based on the feedback loop. T value at

[0021] Step 7: Update the temporary model parameters in the inner loop.

[0022] Step 8, utilize The encoder and decoder parameters are updated together until the iteration converges and the loop ends.

[0023] Furthermore, in step B, the Bi-Pilot MetaNet system selects six common types of interference: single-tone interference (RS), multi-tone interference (RM), linear sweep interference (RL), tracking interference (RF), band-stop interference (RB), and impulse interference (RP). During the simulation, the corresponding interference waveforms are generated in the numerical simulation software based on the mathematical model of the interference, and the interference dataset in the baseband is obtained by sampling.

[0024] Furthermore, in step C, the meta-training, validation, and testing consist of different task sets, and each task comprises a support set and a query set. Therefore, in terms of scope, the dataset of traditional machine learning is equivalent to the support set or query set of a single task in the meta-learning task set. The difference between tasks mainly lies in the different interference data. This system uses the six different interference types mentioned in step B, along with different interference powers, to obtain different meta-learning task set samples.

[0025] Furthermore, the online learning fine-tuning steps in step two are as follows:

[0026] Step a: The receiver encodes a random pilot sequence P1 and transmits it to the transmitter. It can be assumed that the pilot sequence is designed to contain all the combinations of numbers to avoid the generalization problem of neural network training, and that both the receiver and the transmitter know it. The channel state information and the interference received are transmitted to the transmitter through the encoded pilot.

[0027] Step b: After receiving the pilot sequence affected by the channel, the transmitter first trains the channel to reproduce the current channel information, then fixes the channel layer and fine-tunes a set of autoencoders to adapt to the current channel. The encoder and decoder at the transmitter only adjust the weights of the last few layers of neurons. This fine-tuning only requires a small batch of samples, and a pilot sequence is sufficient to meet the requirements. After the transmitter fine-tunes to adapt to the current channel state information, the transmitter also transmits a pilot sequence P2 to the receiver through the new encoder.

[0028] Step c: After receiving the pilot signal from the transmitter, the receiver updates the decoder parameters only according to the pilot signal information to match the encoder of the transmitter.

[0029] Step d: Normal communication. The transmitter sends data, and the receiver performs corresponding demodulation.

[0030] Furthermore, the general algorithm for online learning fine-tuning in step two is as follows:

[0031] Step (1): Initialize encoder and decoder parameters based on MAML offline training results. This represents the parameters of the transmitter encoder neural network. This represents the parameters of the transmitter decoder neural network. This represents the parameters of the receiver encoder neural network. E represents the neural network parameters of the receiver decoder. T D represents the transmitter encoder. T E represents the transmitter decoder. R D represents the receiver encoder. R This represents the receiver decoder, initializing learning rates α1, α2, and α3, and randomly initializing the interference simulation layer φ. I ;

[0032] Step (2), repeat steps (3) to (7);

[0033] Step (3): The receiver transmits pilot P1, and the transmitter receives R1 = h[E]. R (P1)]+inter;

[0034] Step (4): Fixed encoder E at the transmitter end T Training the interference simulation layer:

[0035] Step (5): Fix the interference simulation layer I at the transmitter and fine-tune the training encoder and decoder:

[0036] Let the random training bit sequence be m, then the decoder output is:

[0037] Encoder update:

[0038] Decoder update:

[0039] Step (6): The transmitter transmits pilot P2 through the updated encoder, and the receiver receives it as R2 = h[E]. T (P2)]+inter;

[0040] Step (7): Fine-tune the training decoder at the receiving end. The loop ends and normal communication resumes when the receiving end's loss converges or exceeds the maximum number of iterations.

[0041] The beneficial effects of this invention are:

[0042] 1. Under AWGN channels, this system, after using the adaptive anti-interference algorithm, has better resistance to different types of interference than the traditional OFDM system;

[0043] 2. Under Rayleigh multipath channel conditions, the system, after using the adaptive anti-interference algorithm, demonstrates superior resistance to different types of interference compared to the traditional OFDM system at high signal-to-noise ratio.

[0044] 3. Compared with the unidirectional frequency fine-tuning system proposed by others, the algorithm of this invention has been shown to have faster convergence speed and better decoding performance. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the offline training of the Bi-Pilot MetaNet system;

[0046] Figure 2 This is a schematic diagram illustrating the meta-training task set configuration of the Bi-Pilot MetaNet system;

[0047] Figure 3 This is a block diagram illustrating the online fine-tuning principle of the Bi-Pilot MetaNet system;

[0048] Figure 4 This is a schematic diagram comparing the convergence speed of the Bi-Pilot MetaNet system and the unidirectional pilot system under RB band-stop interference.

[0049] Figure 5 This is a schematic diagram comparing the symbol error rate performance of an autoencoder under QPSK modulation and the traditional method in an AWGN channel.

[0050] Figure 6 This is a schematic diagram comparing the symbol error rate performance of an autoencoder under QPSK modulation and the traditional method in a Rayleigh channel.

[0051] Figure 7 This is a comparison of the symbol error rate curves before and after fine-tuning under different meta-training set ranges when RB is an unknown interference. Detailed Implementation

[0052] Specific implementation method one: Combining Figures 1 to 7 The specific steps of the adaptive anti-interference algorithm based on autoencoder and meta-learning described in this embodiment are as follows:

[0053] Step 1: Offline Meta-Learning;

[0054] Step 2: Online learning and fine-tuning.

[0055] Specific Implementation Method Two: Combining Figures 1 to 7 This embodiment describes the offline meta-learning step in step one of the adaptive anti-interference algorithm based on autoencoders and meta-learning.

[0056] Step A: Establish the offline training system model: During offline training, the neural network parameters of the encoder and decoder at the transmitter and receiver are randomly initialized. For each iteration cycle, the transmitter randomly generates a binary bit sequence, which, after passing through the channel via the encoder, is affected by channel contamination and interference. The type and power of interference added during training vary depending on the task. After receiving the signal, the receiver sends it to the decoder to obtain the prediction result. Within one iteration cycle, the meta-learner performs gradient descent based on the support set of each task to obtain the temporarily stored receiver parameters θ. R Then, the actual direction in which the model parameters should be updated is obtained through the query set in the task; after completing the above inner loop for all tasks in one iteration, the updated parameters φ of the decoder model are obtained by combining the results. R Since the training is offline, the receiver can pass the trained parameters to the transmitter's encoder to jointly update the transmitter's encoder parameters φ. T Repeat the above iterative process until the receiver loss gradually converges, completing the offline training process of this system.

[0057] Step B: Generate interference set;

[0058] Step C: Construct the meta-training task set.

[0059] like Figure 1The diagram illustrates the workflow of this system during offline training. During offline training, the encoder and decoder neural network parameters at the transmitter and receiver are randomly initialized. For each iteration cycle, the transmitter randomly generates a binary bit sequence. After passing through the channel via the encoder, this sequence is affected by channel contamination and interference. The type and power of the interference added during training vary depending on the task. After receiving the signal, the receiver sends it to the decoder to obtain the prediction result. Within one iteration cycle, the meta-learner performs gradient descent based on the support set for each task to obtain temporarily stored receiver parameters θ. R Then, the actual direction in which the model parameters should be updated is obtained through the query set in the task. After completing the above inner loop for all tasks in one iteration, the updated parameters φ of the decoder model are obtained by combining the results. R Since the training is offline, the receiver can pass the trained parameters to the transmitter's encoder, jointly updating the transmitter's encoder parameters φ. T Repeat the above iterative process until the receiver loss gradually converges, completing the offline training process of this system.

[0060] Specific implementation method three: Combining Figures 1 to 7 This embodiment describes the offline training algorithm steps in step A of the adaptive anti-interference algorithm based on autoencoder and meta-learning as follows:

[0061] Step 1: Randomly initialize encoder and decoder parameters φ T φ R α represents the inner loop step size, and β represents the outer loop step size;

[0062] Step 2, repeat steps 3 through 8;

[0063] Step 3: The current interference task set is sampled as T. i ~p(T); p(T) represents the distribution of the interference task set;

[0064] Step 4: Repeat steps 5 and 6 for all tasks;

[0065] Step 5: Calculate the current task T i The loss function on the decoder corresponds to the gradient at θ. R value at

[0066] Step 6: Calculate the corresponding gradient at encoder θ based on the feedback loop. T value at

[0067] Step 7: Update the temporary model parameters in the inner loop.

[0068] Step 8, utilize The encoder and decoder parameters are updated together until the iteration converges and the loop ends.

[0069] Specific implementation method four: Combination Figures 1 to 7 This embodiment describes an adaptive anti-interference algorithm based on autoencoders and meta-learning. In step B, the Bi-Pilot MetaNet system selects six common interference types: single-tone interference (RS), multi-tone interference (RM), linear sweep interference (RL), tracking interference (RF), band-stop interference (RB), and impulse interference (RP). During simulation, the system generates corresponding interference waveforms in numerical simulation software based on the mathematical model of the interference and samples them to obtain the interference dataset in the baseband.

[0070] Specific Implementation Method Five: Combining Figures 1 to 7 This embodiment describes an adaptive anti-interference algorithm based on autoencoders and meta-learning. Step C, meta-training, validation, and testing consist of different task sets, each with a support set and a query set. Therefore, in terms of scope, the dataset of traditional machine learning is equivalent to the support set or query set of a single task in the meta-learning task set. The difference between tasks lies mainly in the different interference data. This system uses the six different interference types mentioned in step B, along with different interference powers, to obtain different meta-learning task set samples.

[0071] Specific Implementation Method Six: Combination Figures 1 to 7 This embodiment describes the online learning and fine-tuning step in step two of the adaptive anti-interference algorithm based on autoencoders and meta-learning.

[0072] Step a: The receiver encodes a random pilot sequence P1 and transmits it to the transmitter. It can be assumed that the pilot sequence is designed to contain all the combinations of numbers to avoid the generalization problem of neural network training, and that both the receiver and the transmitter know it. The channel state information and the interference received are transmitted to the transmitter through the encoded pilot.

[0073] Step b: After receiving the pilot sequence affected by the channel, the transmitter first trains the channel to reproduce the current channel information, then fixes the channel layer and fine-tunes a set of autoencoders to adapt to the current channel. The encoder and decoder at the transmitter only adjust the weights of the last few layers of neurons. This fine-tuning only requires a small batch of samples, and a pilot sequence is sufficient to meet the requirements. After the transmitter fine-tunes to adapt to the current channel state information, the transmitter also transmits a pilot sequence P2 to the receiver through the new encoder.

[0074] Step c: After receiving the pilot signal from the transmitter, the receiver updates the decoder parameters only according to the pilot signal information to match the encoder of the transmitter.

[0075] Step d: Normal communication. The transmitter sends data, and the receiver performs corresponding demodulation.

[0076] The biggest difference between the Bi-Pilot MetaNet system in this embodiment and traditional autoencoder-based wireless communication systems lies in the improvement of online fine-tuning. Traditional autoencoder communication networks can only fine-tune the receiver's decoder network by injecting pilot signals into the forward communication link, because the transmitter's encoder cannot obtain channel and interference information in the current communication link. To overcome this deficiency, the Bi-Pilot MetaNet system adds an interference simulation layer at the transmitter and a reverse pilot signal link from the receiver to the transmitter. The purpose is to feed back interference characteristics to the transmitter through pilot information in this link, and extract real-time interference characteristics in the interference simulation layer using end-to-end training, thereby enabling the transmitter's encoder network to be fine-tuned and updated online.

[0077] like Figure 3 As shown, both the transmitter and receiver include an autoencoder communication framework. The transmitter updates the entire encoder and decoder using pilot information transmitted from the other party, which has passed through the encoder and channel. The receiver, in turn, updates only the decoder, not the encoder, using pilot information transmitted in the same manner. For an autoencoder, the encoder modulates the source information, while the decoder demodulates it. In this system architecture, the transmitter primarily uses the encoder to implement its functions, while the receiver primarily uses the decoder. The former still needs a complete encoder-decoder to match the modulation and demodulation of the transmitter and receiver during online fine-tuning; that is, the decoder retrained by the transmitter should be as consistent as possible with the decoder retrained by the receiver.

[0078] Specific implementation method seven: Combination Figures 1 to 7 This embodiment describes a general algorithm for online learning and fine-tuning in step two of an adaptive anti-interference algorithm based on autoencoders and meta-learning.

[0079] Step (1): Initialize encoder and decoder parameters based on MAML offline training results. This represents the parameters of the transmitter encoder neural network. This represents the parameters of the transmitter decoder neural network. This represents the parameters of the receiver encoder neural network. E represents the neural network parameters of the receiver decoder. TD represents the transmitter encoder. T E represents the transmitter decoder. R D represents the receiver encoder. R This represents the receiver decoder, initializing learning rates α1, α2, and α3, and randomly initializing the interference simulation layer φ. I ;

[0080] Step (2), repeat steps (3) to (7);

[0081] Step (3): The receiver transmits pilot P1, and the transmitter receives R1 = h[E]. R (P1)]+inter;

[0082] Step (4): Fixed encoder E at the transmitter end T Training the interference simulation layer:

[0083] Step (5): Fix the interference simulation layer I at the transmitter and fine-tune the training encoder and decoder:

[0084] Let the random training bit sequence be m, then the decoder output is:

[0085] Encoder update:

[0086] Decoder update:

[0087] Step (6): The transmitter transmits pilot P2 through the updated encoder, and the receiver receives it as R2 = h[E]. T (P2)]+inter;

[0088] Step (7): Fine-tune the training decoder at the receiving end. The loop ends and normal communication resumes when the receiving end's loss converges or exceeds the maximum number of iterations.

[0089] Table 3-3 Single-carrier autoencoder network parameters of the Bi-Pilot MetaNet system

[0090]

[0091] Table 3-4 Parameters of the Multi-Carrier OFDM Autoencoder Network in the Bi-Pilot MetaNet System

[0092]

[0093]

[0094] Working principle

[0095] This invention designs a neural network communication system based on the meta-learning MAML algorithm and pilot bidirectional fine-tuning—the Bi-Pilot MetaNet communication system, and designs a matching adaptive anti-interference algorithm flow under this model.

[0096] The Bi-Pilot MetaNet system is divided into three parts: the uplink transmitter, the downlink receiver, and the enemy jamming source. Both the uplink transmitter and the downlink receiver are communication systems based on autoencoders. The enemy jamming source is a traditional jamming transmitter that adds different types of interference to the communication link to affect the normal communication between the transmitter and receiver.

[0097] Both the sender and receiver possess the same complete set of autoencoder neural networks, which have undergone offline meta-learning training and are correctly matched for use. During normal communication, the sender modulates and encodes the source bits of the information through the encoder neural network to the transmit antenna. After passing through the channel, the bits arrive at the receiver's receive antenna (the blue link in the diagram below). After processing, the data is sent to the receiver's decoder for demodulation and decoding, and then the recovered information sequence is obtained. This is the normal forward communication process of Bi-Pilot MetaNet.

[0098] If the interference source begins to emit interference that affects the normal communication between the transmitter and receiver, the system will enter the online learning and fine-tuning phase. This involves fine-tuning the parameters of the encoder and decoder neural networks of both the transmitter and receiver using pilot feedback to adapt to the current interference and channel conditions. Traditional autoencoder communication networks based on pilot feedback for online fine-tuning are often unidirectional, only adding pilot symbols to the forward communication link to train and fine-tune the receiver online. This leaves the transmitter's encoder network unadjusted, weakening its adaptability. The Bi-PilotMetaNet system's online fine-tuning algorithm improves upon the traditional end-to-end fine-tuning strategy by adding a pilot backpropagation link (the yellow link in the diagram) to the forward communication link. It also incorporates an interference simulation layer composed of a neural network in the transmitter's autoencoder network, allowing the transmitter to perceive the interference characteristics present in the current channel. This enables the transmitter's encoder to fine-tune its neural network parameters based on channel and interference characteristics.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. An adaptive anti-interference algorithm based on autoencoder and meta-learning, characterized in that: The specific steps of the adaptive anti-interference algorithm based on autoencoder and meta-learning are as follows: Step 1: Offline Meta-Learning; specifically includes: Step A: Establish the offline training system model: During offline training, the neural network parameters of the encoder and decoder at the transmitter and receiver are randomly initialized. For each iteration cycle, the transmitter randomly generates a binary bit sequence. After passing through the channel via the encoder, it will be affected by channel contamination and interference. Under different tasks, the types and power of interference added during training are different. After the receiver receives the signal, it is sent to the decoder to obtain the prediction result. Within one iteration cycle, the meta-learner will perform gradient descent based on the support set in each task to obtain the temporarily stored receiver parameters. Then, the actual direction in which the model parameters should be updated is obtained through the query set in the task; after completing the inner loop of all tasks in one iteration, the updated parameters of the decoder model are obtained by combining them. Since the training is offline, the receiver can pass the trained parameters to the transmitter's encoder to jointly update the parameters of the transmitter's encoder. Repeat the above iterative process until the receiver loss gradually converges, completing the offline training process. Step B: Generate interference set; Step C: Construct the meta-training task set; Step Two: Online Learning and Fine-tuning; Specific steps are as follows: Step a: The receiver transmits a random pilot sequence to the transmitter after encoding it with its encoder. Both the sender and receiver are aware of this, and the channel state information and the interference received are transmitted to the transmitter through the encoded pilot signal; Step b: After receiving the pilot sequence affected by the channel, the transmitter first trains the channel to reproduce the current channel information, then fixes the channel layer and fine-tunes an autoencoder to adapt to the current channel. The encoder and decoder at the transmitting end only adjust the weights of the last few layers of neurons. After the transmitter fine-tunes to adapt to the current channel state information, it also transmits a pilot sequence to the receiver through the new encoder. ; Step c: After receiving the pilot signal from the transmitter, the receiver updates the decoder parameters only according to the pilot signal information to match the encoder of the transmitter. Step d: Normal communication, the transmitter sends data, and the receiver performs corresponding demodulation; The general algorithm for fine-tuning online learning is as follows: Step (1): Initialize encoder and decoder parameters based on MAML offline training results. , ; This represents the parameters of the transmitter encoder neural network. This represents the parameters of the transmitter decoder neural network. This represents the parameters of the receiver encoder neural network. This represents the parameters of the receiver decoder neural network. Indicates the transmitter encoder. Indicates transmitter decoder, Indicates the receiver encoder. This indicates the receiver decoder, and the initial learning rate. , , Randomly initialize the interference simulation layer ; Step (2), repeat steps (3) to (7); Step (3): The receiver transmits pilot signals. The transmitter received ; Step (4): Fix the encoder at the transmitter. Training the interference simulation layer: ; Step (5): Fixed interference simulation layer at the transmitter Fine-tuning the training encoder and decoder: Let the random training bit sequence be... The decoder output is , Encoder update: , Decoder update: ; Step (6): The transmitter transmits pilot signals through the updated encoder. The receiving end receives as ; Step (7): Fine-tune the training decoder at the receiving end. The loop continues until the receiving end experiences loss convergence or exceeds the maximum number of iterations, at which point it ends and normal communication resumes.

2. The adaptive anti-interference algorithm based on autoencoder and meta-learning according to claim 1, characterized in that: The steps of the offline training algorithm in step A are as follows: Step 1: Randomly initialize encoder and decoder parameters , , Indicates the inner loop step size. Indicates the outer loop step size; Step 2, repeat steps 3 through 8; Step 3: The current interference task set is sampled as follows ; Indicates the distribution of the interference task set; Step 4: Repeat steps 5 and 6 for all tasks; Step 5: Calculate the current task The loss function on the decoder corresponds to the gradient in the decoder. value at ; Step 6: Calculate the corresponding gradient in the encoder based on the feedback loop. value at ; Step 7: Update the temporary model parameters in the inner loop. ; Step 8, utilize The encoder and decoder parameters are updated together until the iteration converges, at which point the loop ends.

3. The adaptive anti-interference algorithm based on autoencoder and meta-learning according to claim 1, characterized in that: In step B, the Bi-Pilot MetaNet system selects six common types of interference: single-tone interference (RS), multi-tone interference (RM), linear sweep interference (RL), tracking interference (RF), band-stop interference (RB), and pulse interference (RP). Based on the mathematical modeling of the interference, the corresponding interference waveforms are generated in the numerical simulation software, and the interference dataset in the baseband is obtained by sampling.

4. An adaptive anti-interference algorithm based on autoencoder and meta-learning as described in claim 1 or 3, characterized in that: In step C, the meta-training, validation, and testing consist of different task sets, and each task consists of a support set and a query set. The six different types of interference mentioned in step B are used, along with different interference powers, to obtain different meta-training task sets.

Images