Single-channel source separation anti-interference method and device based on meta-learning under small sample condition

Abstract

This invention discloses a single-channel source separation and anti-interference method based on meta-learning under small sample conditions. The method comprises the following steps: In the training phase, a meta-learning training dataset is first created, including the target communication signal and various interference signals; then, a subset is randomly extracted from the dataset to construct a task set, which is divided into a support set and a query set; next, a signal separation network is built, and a meta-learning objective function and optimization strategy are created, including inner and outer loops; finally, the task set signals are input into the network for training, and the optimal pre-trained model is saved. In the fine-tuning phase, a fine-tuning dataset is first created and tasks are extracted; then, a fine-tuning objective function and optimization strategy are created; finally, small sample signals are input into the pre-trained model for fine-tuning, and the optimal model is saved. This invention achieves effective separation of communication signals and interference signals under unknown or limited sample conditions, realizes effective interference identification and suppression, improves communication anti-interference capability, and ensures communication stability and reliability.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a single-channel source separation anti-interference method based on meta-learning under small sample conditions. Background Technology

[0002] With the popularization of mobile internet and various mobile terminals and applications, more and more people are using mobile terminals for various audio and video applications in their daily lives. Most of these application scenarios are complex and change at any time. Therefore, the first thing to consider is how to achieve voice separation and enhancement in various different scenarios.

[0003] Currently, mainstream speech separation and enhancement methods primarily employ deep learning-based separation techniques, which can be broadly categorized into two directions: time-domain signal separation and time-frequency domain signal separation. Time-domain signal separation focuses on processing and separating signals directly in the time domain, avoiding information loss that may result from time-frequency conversion. Representative network structures include Conv-Tasnet and Wave-U-Net.

[0004] Time-frequency domain signal separation techniques typically begin by transforming the time-domain signal to the time-frequency domain, such as through a short-time Fourier transform (STFT), and then use a deep learning model for separation. Representative network structures include DCUNET, DCCRN, and Sepformer. Sepformer is a Transformer-based time-frequency domain separation network that utilizes a self-attention mechanism to capture long-range dependencies, thereby improving separation performance.

[0005] With the significant success of deep learning technology in fields such as image recognition and speech processing, source separation techniques based on deep learning have also been introduced into signal separation and anti-interference tasks. Deep learning models typically require a large amount of training data, including samples of communication signals and interference signals. In many practical applications, it is difficult to obtain a large amount of labeled data, especially for newly emerging interference signals, where only a small number of samples are available for learning. This limits the model's generalization ability and rapid adaptability.

[0006] Traditional machine learning research involves acquiring large datasets for a specific task and then training a model from scratch using these datasets. This differs significantly from how humans quickly learn from limited samples based on past experience. Few-Shot Learning (FSL) is a machine learning method that addresses the problem of effective learning with limited data. Its main challenges lie in the scarcity of data and the model's generalization ability. Data scarcity means that only a small amount of labeled data is available for training, making it difficult for the model to learn enough features to distinguish between signals and interference. Furthermore, the model's generalization ability requires it to adapt to constantly changing communication environments and interference types.

[0007] Common few-shot learning methods include data augmentation, generative models, transfer learning, and meta-learning. Data augmentation increases the size of the dataset by transforming existing data, helping the model learn more variations and features, thereby improving generalization ability. Generative models, such as variational autoencoders (VAEs) and generative adversarial networks (GANs), can generate new data samples to expand the training set. Transfer learning utilizes a model to apply knowledge learned on one task to another different but related task, typically involving using a pre-trained model and fine-tuning it to adapt to the new task. Meta-learning, also known as "learning to learn," is a method for constructing learning algorithms that enable them to quickly adapt to new tasks. In few-shot learning, the goal of meta-learning is to train a model that can quickly learn with only a small number of samples when faced with a new task. A good meta-learning model can generalize well to new tasks and environments never encountered during training.

[0008] Chinese patent CN 115985344 A discloses a method and apparatus for evaluating the quality of small-sample, no-reference-source speech based on meta-learning. By learning from different speech quality evaluation training tasks, the meta-learner acquires good generalization ability, thus enabling it to quickly and effectively update the model when faced with new and unknown small-sample speech distortion data. It can obtain a high correlation between subjective and objective evaluations under small-sample conditions. It is mainly applied to objective evaluation of speech quality without reference sources and does not involve applications in speech separation or enhancement. Summary of the Invention

[0009] The purpose of this invention is to provide a single-channel source separation anti-interference method that can quickly adapt to newly emerging unknown interference, achieve effective interference identification and suppression, and ensure the stability and reliability of communication.

[0010] The solution to achieve the present invention is: a single-channel source separation anti-interference method based on meta-learning under small sample conditions, comprising a meta-learning training stage and a fine-tuning stage, wherein:

[0011] The meta-learning training phase includes the following steps:

[0012] Step A1: Create a training dataset, including communication signals, various known interference signals, and mixed signals of the two;

[0013] Step A2: Randomly extract a subset from the dataset to construct a task set, and divide the tasks into a support set and a query set. Extract multiple subsets to form a batch.

[0014] Step A3: Build a signal separation network;

[0015] Step A4: Create the meta-learning objective function and optimization strategy, including the inner loop and the outer loop;

[0016] Step A5: Based on the objective function and optimization strategy, use the task set obtained in step A2 to train the separation network in step A3 to obtain a pre-trained separation network model.

[0017] The small-sample fine-tuning phase includes the following steps:

[0018] Step B1: Create a fine-tuning dataset, including communication signals and small sample target interference signals, as well as a mixed signal of the two;

[0019] Step B2: Randomly extract tasks from the fine-tuning dataset to construct a task set, and extract multiple tasks to form a batch.

[0020] Step B3: Create the fine-tuning objective function and optimization strategy;

[0021] Step B4: Based on the fine-tuning objective function and optimization strategy, use the task set obtained in step B2 to fine-tune the pre-trained separation network model obtained in step A5, and obtain the optimal separation network model after fine-tuning.

[0022] Furthermore, the creation of the training dataset described in step A1 is as follows:

[0023] Create the training dataset using either of the following two methods:

[0024] The first method involves using computer software to generate various communication signals and interference signals with different modulation methods, including digital modulation or analog modulation. The interference signals include digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb spectrum interference signals, broadband noise interference, and frequency sweep interference.

[0025] Based on the generated target communication signal and interference signal, the target communication signal and multiple interference signals are randomly mixed to form an aliased signal, with the signal-to-interference ratio (ISR) set to 0-8dB. Noise is added to the aliased signal to generate an aliased communication signal, with a signal-to-noise ratio (SNR) of -5dB to 20dB. The aliased communication signal is shown in the following formula:

[0026]

[0027] In the formula, x(t) is the time-frequency aliasing communication signal, and s i (t)∈C 1×T Let T represent the source signal, i = 1, 2, ..., λ, where i represents the i-th source signal, λ represents the total number of source signals, and a i Let represent the unknown mixing coefficients corresponding to the i-th source signal, and n(t) represent Gaussian white noise;

[0028] The meta-learning training dataset contains the target communication signal and various interference signals, with the communication signal and each interference signal mixed separately.

[0029] The second method involves sampling the actual communication equipment and jamming equipment to obtain a dataset of the target communication signal, jamming signal, and aliased signal.

[0030] Furthermore, step A2 involves randomly extracting a subset from the dataset to construct a task set, dividing the tasks into a support set and a query set, and then extracting multiple subsets to form a batch, as detailed below:

[0031] First, construct the training dataset X. b The target communication signal is denoted as S, and each interference signal is J. i Generate a set of interference signals A i ={J i,s}, J i,s Indicates interfering samples;

[0032] Then via X b Extract a subset to construct the source task T si Source task T si During the training phase of meta-learning, it is considered known;

[0033] Finally, from the source task T si A batch of tasks {τ1,...,τ} are sampled. B The total number of tasks is B, corresponding to one batch; for each task τ b Given b = 1, 2, ..., B, partition the data into support sets τ. sup,b and query set τ que,b .

[0034] Furthermore, the construction of the signal separation network described in step A3 is as follows:

[0035] Because the meta-learning MAML algorithm is used, the signal separation network adopts the known separation network Conv-TasNet or SepFormer network.

[0036] Furthermore, the creation of the meta-learning objective function and optimization strategy described in step A4 includes an inner loop and an outer loop, as detailed below:

[0037] First, in the inner loop, task-specific learning is performed for each task, and the parameter ω is found using gradient descent. b The parameters in the separating network minimize the loss on the support set, as shown in the following formula:

[0038]

[0039] Where σ is the learning rate of the inner loop, and ω represents the parameters in the separation network. b This represents the parameters of the decoupled network after the inner loop ends. This indicates gradient descent. Represents the set of losses on the support set;

[0040] Then, in the outer loop, the task-specific models derived from the inner loop, i.e., the network models trained on each support set, are used to update the initial parameters of the general model, i.e., the entire separating network. This is done by defining a meta-loss function on the query set, as shown in the following formula:

[0041]

[0042] Among them, L meta () represents the meta-loss function for the entire meta-learning process. ω represents the set of parameters trained on each support set. b Loss performance on the query set;

[0043] Finally, the parameters of the entire separation network are updated using gradient descent, i.e.:

[0044]

[0045] Where ξ is the learning rate of the outer loop;

[0046] The loss function uses the scale-invariant signal-to-noise ratio (SI-SNR) of sentence-level permutation invariant loss as the loss function, and the scale-invariant signal-to-noise ratio SI-SNR is expressed as:

[0047]

[0048] Among them, s targete represents the estimated projection component of the signal in the direction of the source signal. noise s represents the vertical component of the estimated signal in the direction of the source signal. k and Let represent the source signal and the estimated signal, respectively; <·,·> represent the dot product of two vectors; and ||·||2 represent the L2 norm.

[0049] Furthermore, based on the objective function and optimization strategy described in step A5, the separation network obtained in step A2 is used to train the separation network model obtained in step A3, resulting in a pre-trained separation network model, as detailed below:

[0050] The task batches extracted from the meta-learning training set are fed into the segregating network. The network parameters are initialized, and the Adam optimizer is used. The learning rates of the inner and outer loops are set to σ = 0.01 and ξ = 0.05, respectively. The network is trained for a total of 200 rounds. After training, the optimal parameters of the pre-trained segregating network model are saved.

[0051] Furthermore, the creation of the fine-tuning dataset described in step B1 includes communication signals, small sample target interference signals, and a mixture of both, as detailed below:

[0052] Acquire the target communication signal and a small sample of target interference signal using either of the following two methods:

[0053] The first method involves using computer software to generate communication signals and small-sample interference signals. The communication signal modulation method is digital modulation or analog modulation. The interference signals include digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb spectrum interference signals, broadband noise interference, and frequency sweep interference.

[0054] Based on the generated target communication signal and interference signal, the target communication signal and target interference signal are randomly mixed to form an aliased signal. Noise is added to the aliased signal to generate an aliased communication signal with a signal-to-noise ratio (SNR) of -5dB to 20dB.

[0055] The second method involves sampling the actual communication equipment and jamming equipment to obtain a dataset of the target communication signal, the target jamming signal, and the aliased signal.

[0056] The target communication signal obtained is the same as the communication signal in step A1, and the type of interference signal is different from that in step A1. The number of small samples is 0-10.

[0057] Furthermore, step B2 involves randomly extracting tasks from the fine-tuning dataset to construct a task set, and then repeatedly extracting tasks to form a batch, as detailed below:

[0058] Extracting task T from the fine-tuning dataset t When constructing the task set, since there is only one target interference class, the support set and query set are no longer divided.

[0059] Furthermore, the creation of the fine-tuning objective function and optimization strategy described in step B3 is as follows:

[0060] Using SI-SNR with uPIT loss as the fine-tuning loss function, in the fine-tuning phase, the separable network model adapts to the target task using initialization parameters from the meta-learning training phase. The target task-specific separable network fine-tuning model ω is obtained by performing gradient descent on the model ω that has already undergone the meta-training phase. n :

[0061] ω n ←ω-σ▽ωL(T t ,ω)

[0062] The optimization process uses the same learning rate σ as the inner loop.

[0063] A single-channel source separation anti-interference device based on meta-learning under small sample conditions is disclosed. This device is used to implement the aforementioned single-channel source separation anti-interference method based on meta-learning under small sample conditions. The device includes a meta-learning training module and a fine-tuning module, wherein:

[0064] The meta-learning training module includes a dataset creation unit, a task set construction unit, a decoupled network construction unit, a meta-learning creation unit, and a training unit, among which:

[0065] The dataset creation unit creates a training dataset, including communication signals, various known interference signals, and mixed signals of both.

[0066] The task set construction unit randomly extracts a subset from the dataset to construct a task set, and divides the tasks into a support set and a query set, and extracts multiple times to form a batch.

[0067] A signal separation network is constructed using a separate network building unit.

[0068] Meta-learning creates units that establish meta-learning objective functions and optimization strategies, including inner and outer loops;

[0069] The training unit, based on the objective function and optimization strategy, uses the task set obtained in step A2 to train the separation network in step A3, thereby obtaining a pre-trained separation network model.

[0070] The fine-tuning module includes a fine-tuning dataset creation unit, a fine-tuning task set construction unit, a fine-tuning objective function creation unit, and a fine-tuning optimization unit, wherein:

[0071] The fine-tuning dataset creation unit creates fine-tuning datasets, including communication signals, small sample target interference signals, and mixed signals of the two.

[0072] Fine-tuning task set construction unit: randomly extract tasks from the fine-tuning dataset to construct task sets, and extract multiple times to form a batch.

[0073] The fine-tuning objective function creation unit creates the fine-tuning objective function and optimization strategy.

[0074] The fine-tuning optimization unit, based on the fine-tuning objective function and optimization strategy, uses the task set obtained in step B2 to fine-tune the pre-trained separation network model obtained in step A5, and obtains the optimal separation network model after fine-tuning.

[0075] Compared with the prior art, the significant advantages of this invention are: (1) It adopts the MAML algorithm, and through the learning stage and the fine-tuning stage, the training model can quickly adapt to new tasks and effectively cope with the interference of unknown or small number of samples. Even when the number of samples is limited, it can work effectively and improve the generalization ability of the model; (2) Through meta-learning, the model can quickly adapt to the newly emerging unknown interference, and achieve effective separation of communication signals and interference signals under small sample conditions in the case of single antenna reception, realize effective interference identification and suppression, improve the anti-interference ability of communication, and ensure the stability and reliability of communication. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating the single-channel source separation and anti-interference method based on meta-learning under small sample conditions according to the present invention.

[0077] Figure 2 This is a schematic diagram of the structure of the meta-learning training and fine-tuning network in this invention. Detailed Implementation

[0078] It is readily understood that, based on the technical solution of this invention, those skilled in the art can conceive of various embodiments of this invention without altering its essential spirit. Therefore, the following specific embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of this invention or as limitations or restrictions on its technical solution.

[0079] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0080] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0081] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0082] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0083] Combination Figure 1 This invention discloses a single-channel source separation anti-interference method based on meta-learning under small sample conditions, comprising a meta-learning training phase and a fine-tuning phase, wherein:

[0084] The meta-learning training phase includes the following steps:

[0085] Step A1: Create a training dataset, including communication signals, various known interference signals, and mixed signals of the two;

[0086] Step A2: Randomly extract a subset from the dataset to construct a task set, and divide the tasks into a support set and a query set. Extract multiple subsets to form a batch.

[0087] Step A3: Build a signal separation network;

[0088] Step A4: Create the meta-learning objective function and optimization strategy, including the inner loop and the outer loop;

[0089] Step A5: Based on the objective function and optimization strategy, use the task set obtained in step A2 to train the separation network in step A3 to obtain a pre-trained separation network model.

[0090] The small-sample fine-tuning phase includes the following steps:

[0091] Step B1: Create a fine-tuning dataset, including communication signals and small sample target interference signals, as well as a mixture of the two;

[0092] Step B2: Randomly extract tasks from the fine-tuning dataset to construct a task set, and extract multiple tasks to form a batch.

[0093] Step B3: Create the fine-tuning objective function and optimization strategy;

[0094] Step B4: Based on the fine-tuning objective function and optimization strategy, use the task set obtained in step B2 to fine-tune the pre-trained separation network model obtained in step A5, and obtain the optimal separation network model after fine-tuning.

[0095] As a specific example, the creation of the training dataset described in step A1 is as follows:

[0096] Create the training dataset using either of the following two methods:

[0097] The first method involves using computer software to generate various communication signals and interference signals with different modulation methods, including digital modulation or analog modulation. The interference signals include digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb spectrum interference signals, broadband noise interference, and frequency sweep interference.

[0098] Based on the generated target communication signal and interference signal, the target communication signal and multiple interference signals are randomly mixed to form an aliased signal, with the signal-to-interference ratio (ISR) set to 0-8dB. Noise is added to the aliased signal to generate an aliased communication signal, with a signal-to-noise ratio (SNR) of -5dB to 20dB. The aliased communication signal is shown in the following formula:

[0099]

[0100] In the formula, x(t) is the time-frequency aliasing communication signal, and s i (t)∈C 1×T Let T represent the source signal, i = 1, 2, ..., λ, where i represents the i-th source signal, λ represents the total number of source signals, and a i Let represent the unknown mixing coefficients corresponding to the i-th source signal, and n(t) represent Gaussian white noise;

[0101] The meta-learning training dataset contains the target communication signal and various interference signals, with the communication signal and each interference signal mixed separately.

[0102] The second method involves sampling the actual communication equipment and jamming equipment to obtain a dataset of the target communication signal, jamming signal, and aliased signal.

[0103] As a specific example, step A2 involves randomly extracting a subset from the dataset to construct a task set, dividing the tasks into a support set and a query set, and then extracting multiple subsets to form a batch, as detailed below:

[0104] First, construct the training dataset X. b The target communication signal is denoted as S, and each interference signal is J. i Generate a set of interference signals A i ={J i,s}, J i,s Indicates interfering samples;

[0105] Then via X b Extract a subset to construct the source task T si Source task T si During the training phase of meta-learning, it is considered known;

[0106] Finally, from the source task T si A batch of tasks {τ1,...,τ} are sampled. B The total number of tasks is B, corresponding to one batch; for each task τ b Given b = 1, 2, ..., B, partition the data into support sets τ. sup,b and query set τ que,b .

[0107] As a specific example, the construction of the signal separation network described in step A3 is as follows:

[0108] Because the model-independent meta-learning MAML algorithm is used, the signal separation network adopts known classic separation networks, such as Conv-TasNet, SepFormer, etc., or a self-built network model.

[0109] As a specific example, the creation of the meta-learning objective function and optimization strategy described in step A4, including the inner and outer loops, is as follows:

[0110] First, in the inner loop, task-specific learning is performed for each task, and the parameter ω is found using gradient descent. b The parameters in the separating network minimize the loss on the support set, as shown in the following formula:

[0111]

[0112] Where σ is the learning rate of the inner loop, and ω represents the parameters in the separation network. b This represents the parameters of the decoupled network after the inner loop ends. This indicates gradient descent. Represents the set of losses on the support set;

[0113] Then, in the outer loop, the task-specific models derived from the inner loop, i.e., the network models trained on each support set, are used to update the initial parameters of the general model, i.e., the entire separating network. This is done by defining a meta-loss function on the query set, as shown in the following formula:

[0114]

[0115] Among them, L meta () represents the meta-loss function for the entire meta-learning process. ω represents the set of parameters trained on each support set. b Loss performance on the query set;

[0116] Finally, the parameters of the entire separation network are updated using gradient descent, i.e.:

[0117]

[0118] Where ξ is the learning rate of the outer loop;

[0119] The loss function uses the scale-invariant signal-to-noise ratio (SI-SNR) of the sentence-level permutation-invariant training (PIT) loss as the loss function. The scale-invariant signal-to-noise ratio SI-SNR is expressed as:

[0120]

[0121] Among them, s target e represents the estimated projection component of the signal in the direction of the source signal. noise s represents the vertical component of the estimated signal in the direction of the source signal. k and Let represent the source signal and the estimated signal, respectively; <·,·> represent the dot product of two vectors; and ||·||2 represent the L2 norm.

[0122] As a specific example, step A5, based on the objective function and optimization strategy, uses the task set obtained in step A2 to train the separation network in step A3, resulting in a pre-trained separation network, as detailed below:

[0123] The task batches extracted from the meta-learning training set are fed into the segregating network. The network parameters are initialized, and the Adam optimizer is used. The learning rates of the inner and outer loops are set to σ = 0.01 and ξ = 0.05, respectively. The network is trained for a total of 200 rounds. After training, the optimal parameters of the pre-trained segregating network model are saved.

[0124] As a specific example, the creation of the fine-tuning dataset described in step B1 includes communication signals, small sample target interference signals, and a mixture of both, as follows:

[0125] Acquire the target communication signal and a small sample of target interference signal using either of the following two methods:

[0126] Acquire the target communication signal and a small sample of target interference signal using either of the following two methods:

[0127] The first method involves using computer software to generate communication signals and small-sample interference signals. The communication signal modulation method is digital modulation or analog modulation. The interference signals include digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb spectrum interference signals, broadband noise interference, and frequency sweep interference.

[0128] Based on the generated target communication signal and interference signal, the target communication signal and target interference signal are randomly mixed to form an aliased signal. Noise is added to the aliased signal to generate an aliased communication signal with a signal-to-noise ratio (SNR) of -5dB to 20dB.

[0129] The second method involves sampling the actual communication equipment and jamming equipment to obtain a dataset of the target communication signal, the target jamming signal, and the aliased signal.

[0130] The target communication signal obtained is the same as the communication signal in step A1, and the type of interference signal is different from that in step A1. The number of small samples is 0-10.

[0131] As a specific example, step B2, which involves randomly extracting tasks from the fine-tuning dataset to construct a task set and then extracting multiple tasks to form a batch, is as follows:

[0132] Extracting task T from the fine-tuning dataset t When constructing the task set, since there is only one target interference class, the support set and query set are no longer divided.

[0133] As a specific example, the creation of the fine-tuning objective function and optimization strategy described in step B3 is as follows:

[0134] Using SI-SNR with uPIT loss as the fine-tuning loss function, in the fine-tuning phase, the separable network model adapts to the target task using initialization parameters from the meta-learning training phase. The target task-specific separable network fine-tuning model ω is obtained by performing gradient descent on the model ω that has already undergone the meta-training phase. n :

[0135] ω n ←ω-σ▽ωL(T t ,ω)

[0136] The optimization process uses the same learning rate σ as the inner loop.

[0137] This invention also provides a single-channel source separation anti-interference device based on meta-learning under small sample conditions. This device is used to implement the aforementioned single-channel source separation anti-interference method based on meta-learning under small sample conditions. The device includes a meta-learning training module and a fine-tuning module, wherein:

[0138] The meta-learning training module includes a dataset creation unit, a task set construction unit, a decoupled network construction unit, a meta-learning creation unit, and a training unit, among which:

[0139] The dataset creation unit creates a training dataset, including communication signals, various known interference signals, and mixed signals of both.

[0140] The task set construction unit randomly extracts a subset from the dataset to construct a task set, and divides the tasks into a support set and a query set, and extracts multiple times to form a batch.

[0141] A signal separation network is constructed using a separate network building unit.

[0142] Meta-learning creates units that establish meta-learning objective functions and optimization strategies, including inner and outer loops;

[0143] The training unit, based on the objective function and optimization strategy, uses the task set obtained in step A2 to train the separation network in step A3, thereby obtaining a pre-trained separation network model.

[0144] The fine-tuning module includes a fine-tuning dataset creation unit, a fine-tuning task set construction unit, a fine-tuning objective function creation unit, and a fine-tuning optimization unit, wherein:

[0145] The fine-tuning dataset creation unit creates fine-tuning datasets, including communication signals, small sample target interference signals, and mixed signals of the two.

[0146] Fine-tuning task set construction unit: randomly extract tasks from the fine-tuning dataset to construct task sets, and extract multiple times to form a batch.

[0147] The fine-tuning objective function creation unit creates the fine-tuning objective function and optimization strategy.

[0148] The fine-tuning optimization unit, based on the fine-tuning objective function and optimization strategy, uses the task set obtained in step B2 to fine-tune the pre-trained separation network model obtained in step A5, and obtains the optimal separation network model after fine-tuning.

[0149] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0150] Example

[0151] This embodiment provides a single-channel source separation anti-interference method based on meta-learning under small sample conditions, including a meta-learning training phase and a fine-tuning phase, wherein:

[0152] The meta-learning training phase includes the following steps:

[0153] Step A1: Create a training dataset, including communication signals, various known interference signals, and mixed signals of the two;

[0154] Step A2: Randomly extract a subset from the dataset to construct a task set, and divide the tasks into a support set and a query set, and extract multiple times to form a batch;

[0155] Step A3: Build a signal separation network;

[0156] Step A4: Create the meta-learning objective function and optimization strategy, including the inner loop and the outer loop;

[0157] Step A5: Based on the objective function and optimization strategy, use the task set obtained in step A2 to train the separation network in step A3 to obtain the trained pre-trained separation network.

[0158] The small-sample fine-tuning phase includes the following steps:

[0159] Step B1: Create a fine-tuning dataset, including communication signals and small sample target interference signals, as well as a mixed signal of the two;

[0160] Step B2: Randomly extract tasks from the fine-tuning dataset to construct a task set, and extract multiple tasks to form a batch.

[0161] Step B3: Create the fine-tuning objective function and optimization strategy;

[0162] Step B4: Based on the fine-tuning objective function and optimization strategy, use the task set obtained in step B2 to fine-tune the pre-trained separation model obtained in step A5 to obtain the fine-tuned separation network.

[0163] Figure 2 This is a schematic diagram of the structure of the meta-learning training and fine-tuning network in this invention. In wireless communication, signals often face interference from a small number of unknown interference signals, and meta-learning provides an effective method to solve this interference.

[0164] First, construct the training dataset X. b The target communication signal is denoted as S, and each interference signal is J. i Generate a set of interference signals A i ={J i,s}; then through from X b Extract a subset to construct the source task T si This task is considered known during the training phase of meta-learning.

[0165] During the fine-tuning phase, the dataset X was sampled. n To obtain the target task T t It consists of target interference signals and target communication signals. Since the source task and target task involve different interference sources, a model trained only on the source task is difficult to adapt well to the target task. In this embodiment, a model-independent meta-learning MAML algorithm is used for parameter initialization in the meta-learning training phase and for performance fine-tuning for specific tasks in the fine-tuning phase.

[0166] In the meta-learning training phase, the algorithm aims to find a set of initialization parameters that readily adapt to the target domain. This phase is accomplished through an optimization process comprising an inner loop for task-specific learning and an outer loop for updating general model parameters based on the task-specific model. First, starting from the source task T... si A batch of tasks {τ1,...,τ} are sampled. B}, of size B; for each task τ b The data is divided into support sets τ sup,b and query set τ que,b .

[0167] In the inner loop, task-specific learning is performed for each task, and the parameter ω is found using gradient descent. b These parameters minimize the loss on the support set, and the meta-learning rate σ is:

[0168]

[0169] Where σ is the learning rate of the inner loop;

[0170] In the outer loop, the task-specific model derived from the inner loop is used to update the initialization parameters of the general model. This is accomplished by defining a meta-loss function on the query set.

[0171]

[0172] Finally, the model parameters are updated using gradient descent, i.e.:

[0173]

[0174] Where ξ is the learning rate of the outer loop.

[0175] During the fine-tuning phase, the model adapts to the target task using initialization parameters from the meta-learning training phase. This phase is identical to the inner loop process, focusing on using the target task T. t The model parameters are refined using samples; therefore, a task-specific model ω is obtained by performing gradient descent on the initialized model ω. n :

[0176] ω n ←ω-σ▽ωL(T t ,ω)

[0177] The optimization process uses the same learning rate σ as the inner loop.

[0178] In summary, the present invention provides a single-channel source separation and anti-interference method based on meta-learning under small sample conditions. This method utilizes meta-learning MAML to enhance the model's capabilities in signal separation and anti-interference communication, effectively addressing interference situations with unknown or limited samples, and significantly improving the generalization ability of the separation model.

[0179] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A single-channel source separation anti-interference method based on meta-learning under small sample conditions, characterized in that, It includes a meta-learning training phase and a fine-tuning phase, in which: The meta-learning training phase includes the following steps: Step A1: Create a training dataset, including communication signals, various known interference signals, and mixed signals of the two; Step A2: Randomly extract a subset from the dataset to construct a task set, and divide the tasks into a support set and a query set. Extract multiple subsets to form a batch. Step A3: Build a signal separation network; Step A4: Create the meta-learning objective function and optimization strategy, including the inner and outer loops, as detailed below: First, in the inner loop, task-specific learning is performed for each task, using gradient descent to find the parameters. The parameters in the separating network minimize the loss on the support set, as shown in the following formula: ； in It is the learning rate of the inner loop. This represents the parameters in the separate network. This represents the parameters of the decoupled network after the inner loop ends. This indicates gradient descent. Represents the set of losses on the support set; Then, in the outer loop, the task-specific models derived from the inner loop, i.e., the network models trained on each support set, are used to update the initial parameters of the general model, i.e., the entire separating network. This is done by defining a meta-loss function on the query set, as shown in the following formula: ； in, The meta-loss function represents the entire meta-learning process. This represents the set of parameters trained on each support set. Loss performance on the query set; Finally, the parameters of the entire separation network are updated using gradient descent, i.e.: ； in It is the learning rate of the outer loop; The loss function uses the scale-invariant signal-to-noise ratio (SI-SNR) of sentence-level permutation invariant loss as the loss function, and the scale-invariant signal-to-noise ratio SI-SNR is expressed as: ； in, This represents the projection component of the estimated signal in the direction of the source signal. This represents the estimated vertical component of the signal in the direction of the source signal. and These represent the source signal and the estimated signal, respectively. To represent the dot product of two vectors, Represents the L2 norm; Step A5: Based on the objective function and optimization strategy, use the task set obtained in step A2 to train the separation network in step A3 to obtain a pre-trained separation network model. The small-sample fine-tuning phase includes the following steps: Step B1: Create a fine-tuning dataset, including communication signals and small sample target interference signals, as well as a mixed signal of the two; Step B2: Randomly extract tasks from the fine-tuning dataset to construct a task set, and extract multiple tasks to form a batch. Step B3: Create the fine-tuning objective function and optimization strategy; Step B4: Based on the fine-tuning objective function and optimization strategy, use the task set obtained in step B2 to fine-tune the pre-trained separation network model obtained in step A5, and obtain the optimal separation network model after fine-tuning.

2. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1, characterized in that, The creation of the training dataset described in step A1 is as follows: Create the training dataset using either of the following two methods: The first method involves using computer software to generate various communication signals and interference signals with different modulation methods, including digital modulation or analog modulation. The interference signals include digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb spectrum interference signals, broadband noise interference, and frequency sweep interference. Based on the generated target communication signal and interference signal, the target communication signal and multiple interference signals are randomly mixed to form an aliased signal, with the signal-to-interference ratio (ISR) set to 0-8dB. Noise is added to the aliased signal to generate an aliased communication signal, with a signal-to-noise ratio (SNR) of -5dB to 20dB. The aliased communication signal is shown in the following formula: ； In the formula, It is a time-frequency aliasing communication signal. Indicates the source signal, Indicates signal length. , Indicates the first One source signal, Indicates the total number of source signals. Indicates the first The unknown mixing coefficients corresponding to each source signal Indicates Gaussian white noise; The meta-learning training dataset contains the target communication signal and various interference signals, with the communication signal and each interference signal mixed separately. The second method involves sampling the actual communication equipment and jamming equipment to obtain a dataset of the target communication signal, jamming signal, and aliased signal.

3. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 2, characterized in that, Step A2 describes the process of randomly extracting a subset from the dataset to construct a task set, dividing the tasks into a support set and a query set, and then extracting multiple subsets to form a batch, as detailed below: First, construct the training dataset. The target communication signal is represented as Each interference Generate a set of interference signals , Indicates interfering samples; Then through from Extract a subset to construct the source task Source task During the training phase of meta-learning, it is considered known; Finally, from the source task A batch of sampling tasks The total number of tasks is This corresponds to a batch; for each task , The data is divided into support sets. and query set .

4. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 3, characterized in that, The construction of the signal separation network described in step A3 is as follows: Because the meta-learning MAML algorithm is used, the signal separation network adopts the known separation network Conv-TasNet or SepFormer network.

5. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1, characterized in that, Step A5, based on the objective function and optimization strategy, uses the task set obtained in step A2 to train the separation network from step A3, resulting in a pre-trained separation network model, as detailed below: The task batches extracted from the meta-learning training set are fed into the split network. The network parameters are initialized, and the Adam optimizer is used. The learning rates of the inner and outer loops are set to [values ​​to be filled in]. and The training process consists of 200 rounds. After training, the optimal parameters of the pre-trained separable network model are saved.

6. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1, characterized in that, Step B1 describes the creation of a fine-tuning dataset, which includes communication signals, small sample target interference signals, and a mixture of both, as detailed below: Acquire the target communication signal and a small sample of target interference signal using either of the following two methods: The first method involves using computer software to generate communication signals and small-sample interference signals. The communication signal modulation method is digital modulation or analog modulation. The interference signals include digital modulation interference signals, analog modulation interference signals, noise frequency modulation interference signals, noise amplitude modulation interference signals, single-tone interference signals, multi-tone interference signals, comb spectrum interference signals, broadband noise interference, and frequency sweep interference. Based on the generated target communication signal and interference signal, the target communication signal and target interference signal are randomly mixed to form an aliased signal. Noise is added to the aliased signal to generate an aliased communication signal with a signal-to-noise ratio (SNR) of -5dB to 20dB. The second method involves sampling the actual communication equipment and jamming equipment to obtain a dataset of the target communication signal, the target jamming signal, and the aliased signal. The target communication signal obtained is the same as the communication signal in step A1, and the type of interference signal is different from that in step A1. The number of small samples is 0-10.

7. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1, characterized in that, Step B2 describes the random extraction of tasks from the fine-tuning dataset to construct a task set, with multiple extractions forming a batch, as follows: Extracting tasks from the fine-tuning dataset When constructing the task set, since there is only one target interference class, the support set and query set are no longer divided.

8. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1, characterized in that, Step B3 describes the creation of the fine-tuning objective function and optimization strategy, as follows: Using the SI-SNR of uPIT loss as the fine-tuning loss function, in the fine-tuning phase, the separable network model adapts to the target task using the initialization parameters from the meta-learning training phase, through the model that has already undergone the meta-training phase. Gradient descent is used to obtain a task-specific fine-tuned model of the segregation network. : ； The optimization process uses the same learning rate as the inner loop. .

9. A single-channel source separation anti-interference device based on meta-learning under small sample conditions, characterized in that, This device is used to implement the single-channel source separation anti-interference method based on meta-learning under small sample conditions as described in any one of claims 1 to 8. The device includes a meta-learning training module and a fine-tuning module, wherein: The meta-learning training module includes a dataset creation unit, a task set construction unit, a decoupled network construction unit, a meta-learning creation unit, and a training unit, among which: The dataset creation unit creates a training dataset, including communication signals, various known interference signals, and mixed signals of both. The task set construction unit randomly extracts a subset from the dataset to construct a task set, and divides the tasks into a support set and a query set, and extracts multiple times to form a batch. A signal separation network is constructed using a separate network building unit. Meta-learning creates units that establish meta-learning objective functions and optimization strategies, including inner and outer loops; The training unit, based on the objective function and optimization strategy, uses the task set obtained in step A2 to train the separation network in step A3, thereby obtaining a pre-trained separation network model. The fine-tuning module includes a fine-tuning dataset creation unit, a fine-tuning task set construction unit, a fine-tuning objective function creation unit, and a fine-tuning optimization unit, wherein: The fine-tuning dataset creation unit creates fine-tuning datasets, including communication signals, small sample target interference signals, and mixed signals of the two. Fine-tuning task set construction unit: randomly extract tasks from the fine-tuning dataset to construct task sets, and extract multiple times to form a batch. The fine-tuning objective function creation unit creates the fine-tuning objective function and optimization strategy. The fine-tuning optimization unit, based on the fine-tuning objective function and optimization strategy, uses the task set obtained in step B2 to fine-tune the pre-trained separation network model obtained in step A5, and obtains the optimal separation network model after fine-tuning.

Images