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

Abstract

The invention discloses a single-channel source separation anti-interference method based on meta-learning under a small sample condition, and the method comprises the steps: firstly creating a meta-learning training data set which comprises a target communication signal and a plurality of interference signals in a training stage; then randomly extracting subsets from the data set to construct a task set, and dividing the task set into a support set and a query set; then, building a signal separation network, and creating a meta-learning objective function and an optimization strategy, including internal circulation and external circulation; finally, the task set signals are input into the network for training, and the optimal pre-training model is stored; in the fine tuning stage, firstly, a fine tuning data set is created and tasks are extracted; then creating a fine tuning objective function and an optimization strategy; and finally, inputting the small sample signals into the pre-training model for fine tuning, and storing the optimal model. According to the method, effective separation of the communication signal and the interference signal under the unknown or few sample condition is realized, effective interference identification and suppression are realized, the communication anti-interference capability is improved, and the communication stability and reliability are guaranteed.

Description

Technical Field

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

[0002] With the popularization of mobile Internet and various mobile terminals and applications, more and more people use 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 scenarios.

[0003] At present, mainstream speech separation and enhancement mainly use deep learning separation methods, which can be roughly divided into two directions: time domain signal separation technology and time-frequency domain signal separation technology. Time domain signal separation technology mainly focuses on processing and separating signals directly in the time domain, avoiding the information loss that may be caused by time-frequency conversion. Representative network structures include Conv-Tasnet, Wave-U-Net, etc.

[0004] Time-frequency domain signal separation technology usually first converts the time domain signal to the time-frequency domain, such as through short-time Fourier transform (STFT), and then uses a deep learning model for separation. Representative network structures include DCUNET, DCCRN, Sepformer, etc. Sepformer is a Transformer-based time-frequency domain separation network that uses a self-attention mechanism to capture long-distance dependencies to improve separation performance.

[0005] As deep learning technology has achieved remarkable success in image recognition, speech processing and other fields, source separation technology based on deep learning has also been introduced into signal separation and anti-interference tasks. Deep learning models usually require a large amount of training data, including samples of communication signals and interference signals. In many practical application scenarios, 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, which limits the generalization ability and rapid adaptability of the model.

[0006] The traditional model of machine learning research is to obtain a large dataset for a specific task and then use this dataset to train the model from scratch, which is far from the situation where humans use past experience to quickly complete learning with only a small number of samples. Few-Shot Learning (FSL) is a machine learning method that solves the problem of effective learning when the amount of data is limited. The challenges it faces mainly lie in the scarcity of data and the generalization ability of the model. The scarcity of data means that only a small amount of labeled data is available for training, which makes it difficult for the model to learn enough features to distinguish between signals and interference. The generalization ability of the model requires that the model can adapt to the ever-changing communication environment and interference types.

[0007] Common small sample learning methods include data augmentation, generative models, transfer learning, and meta-learning. Data augmentation increases the size of the data set by transforming existing data, helping the model learn more changes and features, thereby improving generalization capabilities; generative models such as variational autoencoders (VAE) and generative adversarial networks (GAN) can generate new data samples and expand the training set; transfer learning is the use of models to apply knowledge learned on one task to another different but related task, usually involving the use of pre-trained models and fine-tuning to adapt to new tasks; meta-learning, also known as "learning to learn", is a method of building learning algorithms that enable learning algorithms to quickly adapt to new tasks. In small sample learning, the goal of meta-learning is to train a model that can quickly learn when faced with new tasks using only a small number of samples. A good meta-learning model can be well generalized to new tasks and new environments that have never been encountered during training.

[0008] Chinese patent CN 115985344 A discloses a small sample reference-free speech quality evaluation method and device based on meta-learning. By learning from different sound quality evaluation training tasks, the meta-learner obtains better generalization ability, so that when facing new and unknown small sample speech distortion data, the model can be updated quickly and effectively, and a higher subjective and objective evaluation correlation can be obtained under small sample conditions. It is mainly used in the objective evaluation of reference-free sound quality, and does not involve the application of speech separation and enhancement. Summary of the invention

[0009] The purpose of the present 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 realize the present invention is: a single-channel source separation anti-interference method based on meta-learning under small sample conditions, including a meta-learning training stage and a fine-tuning stage, wherein:

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

[0012] Step A1, creating a training data set, including a communication signal, a plurality of known interference signals, and a mixed signal of the two;

[0013] Step A2: randomly extract subsets from the data set to build a task set, and divide the task into a support set and a query set, and extract multiple times to form a batch;

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

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

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

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

[0018] Step B1, creating a fine-tuning data set, including a communication signal, a small sample target interference signal, and a mixed signal of the two;

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

[0020] Step B3, create fine-tuning objective function and optimization strategy;

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

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

[0023] Create a training dataset in one of two ways:

[0024] The first method: using computer software to generate communication signals and interference signals of various modulation modes, the modulation mode is digital modulation or analog modulation, and 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 swept frequency 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, the interference signal ratio ISR is set to 0-8dB, and noise is added to the aliased signal to generate an aliased communication signal, and the signal-to-noise ratio SNR is -5dB to 20dB; the aliased communication signal is shown in the following formula:

[0026]

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

[0028] The meta-learning training dataset contains the target communication signal and multiple interference signals, and the communication signal and each interference signal are mixed separately;

[0029] The second method is to sample the actual communication equipment and the interference equipment to obtain the data sets of the target communication signal, the interference signal and the aliased signal.

[0030] Furthermore, in step A2, a subset is randomly extracted from the data set to construct a task set, and the task is divided into a support set and a query set. Multiple extractions form a batch, as follows:

[0031] First, construct a training dataset X b , where the target communication signal is represented by S and each interference J i Generate an interference signal set A i ={J i,s}, J i,s represents interference samples;

[0032] Then by b Extract a subset from si , source task T si It is considered as known during the training phase of meta-learning;

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

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

[0035] Since the meta-learning MAML algorithm is adopted, the signal separation network adopts the known separation network Conv-TasNet network 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, which are as follows:

[0037] First, in the inner loop, task-specific learning is performed for each task, using gradient descent to find the parameter ω b , the parameters in the separation network minimize the loss on the support set, as follows:

[0038]

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

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

[0041]

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

[0043] Finally, the entire separation network parameters are updated by the gradient descent method, namely:

[0044]

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

[0046] Among them, the loss function uses the scale-invariant signal-to-noise ratio of the 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 targetrepresents the projection component of the estimated signal in the direction of the source signal, e noise represents the vertical component of the estimated signal in the direction of the source signal, s k and represent the source signal and the estimated signal respectively, <·,·> represents the dot product of two vectors, and ||·||2 represents the two-norm.

[0049] Furthermore, based on the objective function and optimization strategy described in step A5, the separation network of step A3 is trained using the task set obtained in step A2 to obtain a trained pre-trained separation network model, as follows:

[0050] The task batches extracted from the meta-learning training set are sent to the separation network, the network parameters are initialized, and the Adam optimizer is used. The inner and outer loop learning rates are set to σ = 0.01 and ξ = 0.05 respectively; a total of 200 rounds of training are performed; after the training, the optimal parameters of the pre-trained separation network model are saved.

[0051] Furthermore, the creation of the fine-tuning data set described in step B1 includes the communication signal and the small sample target interference signal and a mixed signal of the two, as follows:

[0052] Use any of the following two methods to obtain the target communication signal and the small sample target interference signal:

[0053] The first method: computer software is used 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 swept frequency interference.

[0054] Based on the generated target communication signal and interference signal, the target communication signal and the target interference signal are randomly mixed to form an aliased signal, and 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 is to sample the actual communication equipment and the interference equipment to obtain the data set of the target communication signal, the target interference signal and the aliased signal;

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

[0057] Furthermore, in step B2, tasks are randomly extracted from the fine-tuning dataset to construct a task set, and multiple extractions are made into a batch, as follows:

[0058] Extract task T from the fine-tuning dataset t Construct 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 fine-tuning objective function and optimization strategy described in step B3 is as follows:

[0060] The SI-SNR of uPIT loss is used as the fine-tuning loss function. In the fine-tuning stage, the separation network model uses the initialization parameters from the meta-learning training stage to adapt to the target task. The target task-specific separation network fine-tuning model ω is obtained by performing the gradient descent method on the model ω that has passed the meta-training stage. 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, the device is used to implement the 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 data set creation unit, a task set construction unit, a separation network construction unit, a meta-learning creation unit, and a training unit, among which:

[0065] A data set creation unit, which creates a training data set, including a communication signal, a plurality of known interference signals, and a mixed signal of the two;

[0066] The task set construction unit randomly extracts subsets from the data set to construct the task set, and divides the task into a support set and a query set, and extracts multiple times to form a batch;

[0067] Separation network building unit, building signal separation network;

[0068] Meta-learning creation unit, creating meta-learning objective function and optimization strategy, including inner loop and outer loop;

[0069] The training unit trains the separation network of step A3 based on the objective function and the optimization strategy using the task set obtained in step A2 to obtain a trained 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, where:

[0071] A fine-tuning data set creation unit, which creates a fine-tuning data set including a communication signal and a small sample target interference signal and a mixed signal of the two;

[0072] The fine-tuning task set construction unit randomly extracts tasks from the fine-tuning dataset to construct a task set, and multiple extractions form a batch;

[0073] A fine-tuning objective function creation unit is used to create a fine-tuning objective function and an optimization strategy;

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

[0075] Compared with the prior art, the present invention has the following significant advantages: (1) By adopting the MAML algorithm, through the learning stage and the fine-tuning stage, the training model can quickly adapt to new tasks and can effectively deal with the interference of unknown or small samples. It can work effectively even when the number of samples is limited, thereby improving the generalization ability of the model; (2) Through meta-learning, the model can quickly adapt to the emergence of new unknown interference, and effectively separate the communication signal and the interference signal under the condition of small samples under the condition of single antenna reception, realize effective interference identification and suppression, improve the communication anti-interference ability, and ensure the stability and reliability of communication. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a flow chart of the single-channel source separation and anti-interference method based on meta-learning under small sample conditions of the present invention.

[0077] Figure 2 Schematic diagram of the structure of the meta-learning training and fine-tuning network in the present invention. DETAILED DESCRIPTION

[0078] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can imagine various implementation methods of the present invention. Therefore, the following specific implementation methods and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the whole of the present invention or as a limitation or restriction of the technical solution of the present invention.

[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 the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention unless otherwise specifically stated.

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

[0081] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered as part of the specification.

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

[0083] Combination Figure 1 The present invention provides a single-channel source separation anti-interference method based on meta-learning under small sample conditions, including a meta-learning training stage and a fine-tuning stage, wherein:

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

[0085] Step A1, creating a training data set, including a communication signal, a plurality of known interference signals, and a mixed signal of the two;

[0086] Step A2: randomly extract subsets from the data set to build a task set, and divide the task into a support set and a query set, and extract multiple times to form a batch;

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

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

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

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

[0091] Step B1, creating a fine-tuning data set, including a communication signal, a small sample target interference signal, and a mixed signal of the two;

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

[0093] Step B3, create fine-tuning objective function and optimization strategy;

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

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

[0096] Create a training dataset in one of two ways:

[0097] The first method: using computer software to generate communication signals and interference signals of various modulation modes, the modulation mode is digital modulation or analog modulation, and 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 swept frequency 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, the interference signal ratio ISR is set to 0-8dB, and noise is added to the aliased signal to generate an aliased communication signal, and the signal-to-noise ratio SNR is -5dB to 20dB; the aliased communication signal is shown in the following formula:

[0099]

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

[0101] The meta-learning training dataset contains the target communication signal and multiple interference signals, and the communication signal and each interference signal are mixed separately;

[0102] The second method is to sample the actual communication equipment and the interference equipment to obtain the data sets of the target communication signal, the interference signal and the aliased signal.

[0103] As a specific example, in step A2, a subset is randomly extracted from the data set to construct a task set, and the task is divided into a support set and a query set. Multiple extractions form a batch, as follows:

[0104] First, construct a training dataset X b , where the target communication signal is represented by S and each interference J i Generate an interference signal set A i ={J i,s}, J i,s represents interference samples;

[0105] Then by b Extract a subset from si , source task T si It is considered as known during the training phase of meta-learning;

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

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

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

[0109] As a specific example, the creation of the meta-learning objective function and optimization strategy described in step A4 includes an inner loop and an outer loop, which are as follows:

[0110] First, in the inner loop, task-specific learning is performed for each task, using gradient descent to find the parameter ω b , the parameters in the separation network minimize the loss on the support set, as follows:

[0111]

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

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

[0114]

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

[0116] Finally, the entire separation network parameters are updated by the gradient descent method, namely:

[0117]

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

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

[0120]

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

[0122] As a specific example, based on the objective function and optimization strategy described in step A5, the separation network of step A3 is trained using the task set obtained in step A2 to obtain a trained pre-trained separation network, as follows:

[0123] The task batches extracted from the meta-learning training set are sent to the separation network, the network parameters are initialized, and the Adam optimizer is used. The inner and outer loop learning rates are set to σ = 0.01 and ξ = 0.05 respectively; a total of 200 rounds of training are performed; after the training, the optimal parameters of the pre-trained separation network model are saved.

[0124] As a specific example, the creation of the fine-tuning data set described in step B1 includes a communication signal and a small sample target interference signal and a mixed signal of the two, as follows:

[0125] Use any of the following two methods to obtain the target communication signal and the small sample target interference signal:

[0126] Use any of the following two methods to obtain the target communication signal and the small sample target interference signal:

[0127] The first method: computer software is used 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 swept frequency interference.

[0128] Based on the generated target communication signal and interference signal, the target communication signal and the target interference signal are randomly mixed to form an aliased signal, and 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 is to sample the actual communication equipment and the interference equipment to obtain the data set of the target communication signal, the target interference signal and the aliased signal;

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

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

[0132] Extract task T from the fine-tuning dataset t Construct 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 the optimization strategy described in step B3 is as follows:

[0134] The SI-SNR of uPIT loss is used as the fine-tuning loss function. In the fine-tuning stage, the separation network model uses the initialization parameters from the meta-learning training stage to adapt to the target task. The target task-specific separation network fine-tuning model ω is obtained by performing the gradient descent method on the model ω that has passed the meta-training stage. 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] The present invention also provides a single-channel source separation anti-interference device based on meta-learning under small sample conditions, which is used to implement the 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 data set creation unit, a task set construction unit, a separation network construction unit, a meta-learning creation unit, and a training unit, among which:

[0139] A data set creation unit, which creates a training data set, including a communication signal, a plurality of known interference signals, and a mixed signal of the two;

[0140] The task set construction unit randomly extracts subsets from the data set to construct the task set, and divides the task into a support set and a query set, and extracts multiple times to form a batch;

[0141] Separation network building unit, building signal separation network;

[0142] Meta-learning creation unit, creating meta-learning objective function and optimization strategy, including inner loop and outer loop;

[0143] The training unit trains the separation network of step A3 based on the objective function and the optimization strategy using the task set obtained in step A2 to obtain a trained 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, where:

[0145] A fine-tuning data set creation unit, which creates a fine-tuning data set including a communication signal and a small sample target interference signal and a mixed signal of the two;

[0146] The fine-tuning task set construction unit randomly extracts tasks from the fine-tuning dataset to construct a task set, and multiple extractions form a batch;

[0147] A fine-tuning objective function creation unit is used to create a fine-tuning objective function and an optimization strategy;

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

[0149] The present invention is 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, creating a training data set, including a communication signal, a plurality of known interference signals, and a mixed signal of the two;

[0154] Step A2: randomly extract subsets from the data set to construct a task set, and divide the task into a support set and a query set, and extract multiple times to form a batch;

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

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

[0157] Step A5: Based on the objective function and the optimization strategy, the separation network of step A3 is trained using the task set obtained in step A2 to obtain a trained pre-trained separation network;

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

[0159] Step B1, creating a fine-tuning data set, including a communication signal, a small sample target interference signal, and a mixed signal of the two;

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

[0161] Step B3, create fine-tuning objective function and optimization strategy;

[0162] Step B4: Based on the fine-tuning objective function and optimization strategy, the task set obtained in step B2 is used to fine-tune the pre-trained separation model obtained in step A5 to obtain a 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 the present invention. In wireless communications, signals are often interfered by a small amount of unknown interference signals, and meta-learning provides an effective method to solve this interference.

[0164] First, construct a training dataset X b , where the target communication signal is represented by S and each interference J i Generate an interference signal set A i ={J i,s}; Then by b Extract a subset from si , the task is considered known during the training phase of meta-learning.

[0165] In the fine-tuning phase, by sampling the dataset X n To obtain the target task T t , which consists of the target interference signal and the target communication signal. Since the source task and the target task involve different interference sources, the 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 stage and for performance fine-tuning of specific tasks in the fine-tuning stage.

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

[0167] In the inner loop, we perform task-specific learning for each task and use gradient descent to find the parameter ω 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, which is done on the query set by defining a meta-loss function:

[0171]

[0172] Finally, the model parameters are updated by the gradient descent method, namely:

[0173]

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

[0175] In the fine-tuning phase, the model is adapted to the target task using the initialization parameters from the meta-learning training phase. This phase is the same as the inner loop process, focusing on using the target task T t The target 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 single-channel source separation and anti-interference method based on meta-learning under small sample conditions of the present invention uses meta-learning MAML to enhance the model's capabilities in signal separation and anti-interference communication, which can effectively cope with interference from unknown or small samples and significantly improve the generalization ability of the separation model.

[0179] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as 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 the meta-learning training phase and the fine-tuning phase, where: The meta-learning training phase includes the following steps: Step A1, creating a training data set, including a communication signal, a plurality of known interference signals, and a mixed signal of the two; Step A2: randomly extract subsets from the data set to build a task set, and divide the task into a support set and a query set, and extract multiple times to form a batch; Step A3: Building a signal separation network; Step A4: Create a meta-learning objective function and optimization strategy, including an inner loop and an outer loop; Step A5: Based on the objective function and the optimization strategy, the separation network of step A3 is trained using the task set obtained in step A2 to obtain a trained pre-trained separation network model; The small sample fine-tuning phase includes the following steps: Step B1, creating a fine-tuning data set, including a communication signal, a small sample target interference signal, and a mixed signal of the two; Step B2: Randomly extract tasks from the fine-tuning dataset to build a task set, and extract multiple tasks to form a batch; Step B3, create fine-tuning objective function and optimization strategy; Step B4: Based on the fine-tuning objective function and optimization strategy, the task set obtained in step B2 is used to fine-tune the pre-trained separation network model obtained in step A5 to 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 is characterized in that: The creation of the training dataset described in step A1 is as follows: Create a training dataset in one of two ways: The first method: using computer software to generate communication signals and interference signals of various modulation modes, the modulation mode is digital modulation or analog modulation, and 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 swept frequency 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, the interference signal ratio ISR is set to 0-8dB, and noise is added to the aliased signal to generate an aliased communication signal, and the signal-to-noise ratio SNR is -5dB to 20dB; the aliased communication signal is shown in the following formula: Where x(t) is the time-frequency mixed communication signal, s i (t)∈C 1×T represents the source signal, T represents the signal length, i=1,2,...,λ, i represents the i-th source signal, λ represents the total number of source signals, a i represents the unknown mixing coefficient corresponding to the i-th source signal, and n(t) represents Gaussian white noise; The meta-learning training dataset contains the target communication signal and multiple interference signals, and the communication signal and each interference signal are mixed separately; The second method is to sample the actual communication equipment and the interference equipment to obtain the data sets of the target communication signal, the interference signal and the aliased signal.

3. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 2 is characterized in that: In step A2, randomly extract subsets from the data set to construct a task set, and divide the task into a support set and a query set. Multiple extractions form a batch, as follows: First, construct a training dataset X b , where the target communication signal is represented by S and each interference J i Generate an interference signal set A i ={J i,s }, J i,s represents interference samples; Then by b Extract a subset from si , source task T si It is considered as known during the training phase of meta-learning; Finally, from the source task T si A batch of tasks {τ1,...,τ B }, the total number of tasks is B, corresponding to a batch; for each task τ b , b=1,2,...,B, divide the data into support set τ sup,b and the query set τ que,b .

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

5. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1 is characterized in that: The creation of the meta-learning objective function and optimization strategy described in step A4 includes an inner loop and an outer loop, as follows: First, in the inner loop, task-specific learning is performed for each task, using gradient descent to find the parameter ω b , the parameters in the separation network minimize the loss on the support set, as follows: Where σ is the learning rate of the inner loop, ω represents the parameters in the separation network, and ω b Represents the separation network parameters after the inner loop ends, represents the gradient descent, represents the loss set on the support set; Then, in the outer loop, the task-specific model derived from the inner loop, i.e., the network model trained on each support set, is used to update the initialization parameters of the general model, i.e., the entire separation network. This is done on the query set by defining a meta-loss function, as follows: Among them, L meta () represents the meta-loss function of the entire meta-learning, Denotes the parameter set ω trained for each support set b Loss performance on the query set; Finally, the entire separation network parameters are updated by the gradient descent method, namely: Where ξ is the learning rate of the outer loop; Among them, the loss function uses the scale-invariant signal-to-noise ratio of the sentence-level permutation invariant loss as the loss function, and the scale-invariant signal-to-noise ratio SI-SNR is expressed as: Among them, s target represents the projection component of the estimated signal in the direction of the source signal, e noise represents the vertical component of the estimated signal in the direction of the source signal, s k and represent the source signal and the estimated signal respectively, <·,·> represents the dot product of two vectors, and ||·||2 represents the two-norm.

6. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1 is characterized in that: Based on the objective function and optimization strategy described in step A5, the separation network of step A3 is trained using the task set obtained in step A2 to obtain a trained pre-trained separation network model, as follows: The task batches extracted from the meta-learning training set are sent to the separation network, the network parameters are initialized, and the Adam optimizer is used. The inner and outer loop learning rates are set to σ = 0.01 and ξ = 0.05 respectively; a total of 200 rounds of training are performed; after the training, the optimal parameters of the pre-trained separation network model are saved.

7. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1 is characterized in that: The creation of the fine-tuning data set described in step B1 includes the communication signal and the small sample target interference signal and the mixed signal of the two, as follows: Use any of the following two methods to obtain the target communication signal and the small sample target interference signal: The first method: computer software is used 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 swept frequency interference. Based on the generated target communication signal and interference signal, the target communication signal and the target interference signal are randomly mixed to form an aliased signal, and 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 is to sample the actual communication equipment and the interference equipment to obtain the data sets of the target communication signal, the target interference signal and the aliased signal; The acquired target communication signal is the same as the communication signal in step A1, the interference signal is of a different type from the interference signal in step A1, and the number of small samples is 0-10.

8. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1 is characterized in that: In step B2, tasks are randomly extracted from the fine-tuning dataset to construct a task set, and multiple extractions form a batch, as follows: Extract task T from the fine-tuning dataset t Construct the task set. Since there is only one target interference class, the support set and query set are no longer divided.

9. The single-channel source separation anti-interference method based on meta-learning under small sample conditions according to claim 1 is characterized in that: The creation of fine-tuning objective function and optimization strategy described in step B3 is as follows: The SI-SNR of uPIT loss is used as the fine-tuning loss function. In the fine-tuning stage, the separation network model uses the initialization parameters from the meta-learning training stage to adapt to the target task. The target task-specific separation network fine-tuning model ω is obtained by performing the gradient descent method on the model ω that has passed the meta-training stage. n : oh n ←ω-σ▽ωL(T t ,oh) The optimization process uses the same learning rate σ as the inner loop.

10. A single-channel source separation and anti-interference device based on meta-learning under small sample conditions, characterized in that: The 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 9, and the device includes a meta-learning training module and a fine-tuning module, wherein: The meta-learning training module includes a data set creation unit, a task set construction unit, a separation network construction unit, a meta-learning creation unit, and a training unit, among which: A data set creation unit, which creates a training data set, including a communication signal, a plurality of known interference signals, and a mixed signal of the two; The task set construction unit randomly extracts subsets from the data set to construct the task set, and divides the task into a support set and a query set. Multiple extractions form a batch. Separation network building unit, building signal separation network; Meta-learning creation unit, creating meta-learning objective function and optimization strategy, including inner loop and outer loop; The training unit trains the separation network of step A3 based on the objective function and the optimization strategy using the task set obtained in step A2 to obtain a trained 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, where: A fine-tuning data set creation unit, which creates a fine-tuning data set including a communication signal and a small sample target interference signal and a mixed signal of the two; The fine-tuning task set construction unit randomly extracts tasks from the fine-tuning dataset to construct a task set, and multiple extractions form a batch; A fine-tuning objective function creation unit is used to create a fine-tuning objective function and an optimization strategy; The fine-tuning optimization unit uses the task set obtained in step B2 to fine-tune the pre-trained separation network model obtained in step A5 based on the fine-tuning objective function and the optimization strategy to obtain the optimal separation network model after fine-tuning.

Images