Small sample single-channel source separation anti-interference method and device based on multi-task learning

Abstract

The application discloses a kind of based on multi-task learning small sample single channel source separation anti-interference method and device.The method is: in training phase, first create multi-task training dataset, including target communication signal and multiple interference signals;Then build signal separation network, and create joint training objective function;Finally, the multi-task signal is input into network for training, and the optimal pre-training model is saved;In fine-tuning phase, first create fine-tuning dataset and objective function;Then small sample signal is input into pre-training model for fine-tuning, and the optimal model is saved;In application stage, first obtain the interference mixed signal, input fine-tuned network model for signal separation, then restore the bit data of separated signal.The device includes multi-task training module, small sample fine-tuning module and application module.The application can realize effective separation of communication signal and interference signal under single antenna reception condition, and improve the anti-interference ability of wireless communication.

Description

Technical Field

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

[0002] In today's era of rapid development in wireless communication technology, the stability and reliability of communication links have become crucial for information transmission. However, communication signals are inevitably subject to interference from natural noise, technical equipment, and even malicious interference attacks during transmission. These interferences severely affect communication quality and may even lead to communication failure. Therefore, researching how to recover useful signals amidst various interferences has become a core issue in the field of communication anti-interference.

[0003] Traditional anti-interference techniques, such as frequency hopping spread spectrum, direct sequence spread spectrum, adaptive filtering, and beamforming, can identify and suppress interference to some extent, but they often fall short of ideal results in complex and ever-changing interference environments, especially when facing unknown or novel interference. Furthermore, traditional methods struggle to effectively formulate interference characteristic analysis and anti-interference strategies when only a small number of interference samples are available. Specifically, frequency hopping spread spectrum relies on frequency hopping patterns to evade interference, but may be ineffective against broadband interference; direct sequence spread spectrum spreads signal energy across a wider frequency range to combat interference, but is more suitable for scenarios with high reliability requirements; adaptive filtering requires continuous parameter adjustments based on the characteristics of the received signal, and in the face of unknown interference, the filter may take a long time to converge to suitable parameter settings, potentially affecting communication quality; moreover, if interference characteristics change rapidly, the filter may struggle to keep up; beamforming uses multiple antennas to directionally transmit and receive signals to enhance the signal and suppress interference, but it is not applicable in single-channel scenarios.

[0004] Blind source separation attempts to recover individual sources from mixed signals and is an effective means of separating co-channel interference. Traditional blind source separation methods include ICA and NMF, but these methods typically employ multi-channel separation, with single-channel separation yielding poor results. Single-channel blind source separation technology has been at the forefront of signal processing research. It aims to distinguish and recover multiple mixed source signals using only a single receiving antenna under resource constraints. This technology has significant advantages, such as reducing the required hardware and lowering costs, but its challenges are equally apparent, especially in environments with low signal-to-noise ratios. Traditional algorithms often fail to provide satisfactory separation results, while also incurring a huge computational burden.

[0005] To overcome these problems, deep learning technology has been introduced into the research of single-channel blind source separation. Deep learning methods, leveraging their powerful feature extraction and representation learning capabilities, have achieved breakthroughs in speech signal separation. These methods are mainly divided into two categories: one category operates in the time-frequency domain, such as deep clustering learning and permutation-invariant training, which capture the frequency components of the signal by transforming the speech signal from the time domain to the time-frequency domain; the other category processes directly in the time domain, such as using Long Short-Term Memory (LSTM) recurrent networks and Convolutional Recurrent Networks (CRNs), which can process time-series data and learn the temporal dynamic characteristics of the signal. With the rapid development of these speech separation techniques, their applications in wireless communication have also shown great potential. Deep learning models can not only handle complex signal mixing problems but also adapt to changing communication environments, thereby improving the accuracy and robustness of signal separation. Furthermore, the adaptive learning capability of deep learning models allows them to maintain high separation performance even when facing unknown or dynamically changing signal sources.

[0006] While the application of deep learning in single-channel blind source separation is full of potential, it also faces some challenges, especially in anti-interference communication scenarios. Because it is difficult to obtain enough interference samples for newly emerging interference types, there are often only a few samples available for training the model, which limits the learning and adaptation capabilities of deep learning models.

[0007] One of the core goals of deep learning is to mimic the human perceptual system. The human visual system can quickly identify new categories with very few labeled samples, a capability that has inspired the rapid development of few-shot transfer learning in image classification. Significant progress has been made in this field, involving various transfer learning strategies, including parameter initialization, distance metrics, and domain adaptation. Among parameter initialization strategies, multi-task learning (MTL) has attracted considerable attention due to its potential to improve model generalization ability and efficiency. MTL enhances model performance through learning strategies such as hard parameter sharing and soft parameter sharing. Summary of the Invention

[0008] The purpose of this invention is to provide a method and apparatus for small-sample single-channel source separation and anti-interference that can effectively separate communication signals and interference signals under single-antenna reception, effectively cope with interference from unknown or small sample sizes, significantly improve the generalization ability of the separation model, and thus enhance the anti-interference capability of wireless communication.

[0009] The technical solution to achieve the purpose of this invention is: a small-sample single-channel source separation anti-interference method based on multi-task learning, comprising a multi-task training stage, a fine-tuning stage, and an application stage, wherein:

[0010] The multi-task training phase includes the following steps:

[0011] Step A1: Create a training dataset, including communication signals, various known interference signals, and mixed signals of the two;

[0012] Step A2: Build a signal separation network, including encoder, separator and decoder modules;

[0013] Step A3: Create a joint training objective function for multi-task learning;

[0014] Step A4: Based on the joint training objective function, use the signal dataset obtained in step A1 to train the signal separation network in step A2, and obtain the trained pre-trained signal separation network.

[0015] The small-sample fine-tuning phase includes the following steps:

[0016] Step B1: Create a fine-tuning dataset, including communication signals and small sample target interference signals, as well as a mixture of the two;

[0017] Step B2: Create the objective function for the fine-tuning stage;

[0018] Step B3: Based on the objective function of the fine-tuning stage, the pre-trained signal separation network obtained in step A4 is fine-tuned using the fine-tuning dataset obtained in step B1 to obtain the fine-tuned signal separation network.

[0019] The application phase includes the following steps:

[0020] Step C1: Acquire the small sample target interference aliasing signal to be separated using a single antenna;

[0021] Step C2: Using the signal separation network fine-tuned in step B3, the target communication signal and interference signal are separated from the aliased signal to complete the blind separation of the aliased signal;

[0022] Step C3: Send the separated signal obtained in step C2 into the demodulator for demodulation to recover the bit data of the separated signal.

[0023] Furthermore, the creation of the training dataset described in step A1 can be performed 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 an interference signal ratio (ISR) of 0-8dB. Noise is added to the aliased signal to generate an aliased communication signal with an SNR of -5dB to 20dB.

[0026] The aliasing communication signal is shown in the following formula:

[0027]

[0028] 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;

[0029] The multi-task training dataset contains target communication signals and various interference signals, with the target communication signals and each type of interference signal mixed separately.

[0030] The second method involves sampling the actual communication equipment and jamming equipment to obtain a dataset of the target communication signal, jamming signal, and aliasing signal.

[0031] Furthermore, the construction of the signal separation network described in step A2 is as follows:

[0032] The goal of a single-channel source separation task is to separate the mixed signal x∈R T N source signals s1, s2, ..., s are separated from the signal. N ∈R T The modified SepFormer model is used, featuring a Transformer-based architecture with a recurrent neural network (RNN), including an encoder, decoder, and separator. Modifications were made to the encoder and separator modules, including the configuration of convolutional layers and activation functions. Furthermore, the convolutional kernel and stride were reduced during convolution, as detailed below:

[0033] (1) Encoder

[0034] The encoder employs a bottleneck design, replacing one-dimensional convolution. It uses a double convolutional layer, initially increasing and then decreasing the number of kernels to accurately extract the encoded features of the signal. The encoder receives a temporal mixed signal x and generates a learned two-dimensional feature representation h∈R. F×T′ Where F is the feature dimension and T′ is the new length of the signal, a convolutional layer is used: h = conv1d(ReLU(conv1d(x))), where conv1d() represents one-dimensional convolution and ReLU() represents a non-linear activation function;

[0035] (2) Separator

[0036] The separator receives the encoded two-dimensional feature representation h and estimates the mask m1,...,m for each source in the mixed signal. N That is: m k = separator(h), k = 1, ..., N; where separator() represents the separator and N represents the number of information sources;

[0037] During the mask estimation process, the separator first processes the two-dimensional feature representation h into overlapping three-dimensional feature blocks h′, and then inputs h′ into the separator transform block: Where F represents the feature dimension, C is the length of each block, and T is the length of the feature dimension. c It refers to the number of blocks;

[0038] The splitting transform block is the core of the splitter. It uses a dual-scale method to simulate short-term and long-term dependencies. It consists of two types of Transformers: an inner loop Transformer (ntraT) and an outer loop Transformer (InterT). IntraT simulates short-term dependencies within each block, while InterT simulates long-term dependencies across blocks after arranging the dimensions.

[0039] The overall transformation definition of the separated transform block is:

[0040]

[0041] Among them, f intra and f inter These represent IntraT and InterT transformations, respectively; P represents permutation operation; h′ represents a 3D feature block; and h″ represents the feature after separation transformation.

[0042] (3) Decoder

[0043] The decoder uses a mask predicted by a transposed convolutional layer and a masking network to reconstruct the separated signals in the time domain:

[0044]

[0045] in, Let m represent the k-th source separated from the convolution, conv1d-transpose() represents one-dimensional deconvolution, and m k This represents the mask of the k-th source.

[0046] Furthermore, the joint training objective function for multi-task learning described in step A3 is as follows:

[0047] During the training phase, the parameters of the encoder, decoder, and separator are shared, and a unified objective function is used to minimize the combined loss of all tasks, i.e.:

[0048]

[0049] Where J(θ) represents the total loss of the training task, θ represents the parameter set shared by all tasks, and α j The loss L is assigned to the j-th task. j The weights, if all tasks are equally important, are α. j =1 / M;

[0050] The parameters θ of the separation model b Optimize in an alternating manner across multiple tasks, i.e.:

[0051]

[0052] Where β represents the learning rate, This indicates gradient descent along θ;

[0053] The loss L per task i Using the sentence-level permutation-invariant uPIT loss with the scale-invariant signal-to-noise ratio (SI-SNR) as the loss function, the scale-invariant signal-to-noise ratio (SI-SNR) is expressed as:

[0054]

[0055] Among them, s k and Let represent the source signal and the estimated signal, <·,·> represent the dot product of two vectors, ||·||2 represent the L2 norm, and s target e represents the estimated projection component of the signal in the direction of the source signal. noise This represents the vertical component of the estimated signal in the direction of the source signal.

[0056] Furthermore, in step A4, based on the joint training objective function, the signal dataset obtained in step A1 is used to train the signal separation network from step A2, resulting in a pre-trained signal separation network, as detailed below:

[0057] The multi-task training set data obtained in step A1 is packaged, the sequences are combined into a batch, and fed into the signal separation network in step A2 for training. The network parameters are set, the Adam optimizer is used, and the initial learning rate is β = 0.0125; a total of 100 rounds of training are conducted; after training, the optimal parameter model is saved.

[0058] 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:

[0059] The target communication signal and the small sample target interference signal are acquired through either of two methods:

[0060] 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.

[0061] 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.

[0062] 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.

[0063] 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.

[0064] Furthermore, the creation of the objective function for the fine-tuning stage described in step B2 is as follows:

[0065] SI-SNR of uPIT loss is used as the fine-tuning loss function.

[0066] Furthermore, in step B3, based on the objective function of the fine-tuning stage, the pre-trained signal separation network obtained in step A4 is fine-tuned using the fine-tuning dataset obtained in step B1 to obtain the fine-tuned signal separation network, as detailed below:

[0067] Fine-tuning of the pre-trained model is performed using a small sample dataset with 20 rounds of fine-tuning. During the fine-tuning process, the encoder, decoder, or separator of the pre-trained model can be frozen. After the fine-tuning is completed, the optimal model is saved.

[0068] A few-sample single-channel source separation anti-interference device based on multi-task learning is provided. This device is used to implement the aforementioned few-sample single-channel source separation anti-interference method based on multi-task learning. The device includes a multi-task training module, a few-sample fine-tuning module, and an application module, wherein:

[0069] The multi-task training module includes a dataset creation unit, a signal separation network construction unit, an objective function creation unit, and a network training unit, wherein:

[0070] The dataset creation unit creates a multi-task training dataset, including communication signals and various interference signals;

[0071] The signal separation network construction unit is used to build encoder, separator and decoder modules;

[0072] Joint training objective function creation unit: Create a custom multi-task learning training objective function;

[0073] The network training unit, based on the joint training objective function, uses the signal dataset obtained by the dataset creation unit to train the network obtained by the signal separation network construction unit, and obtains the pre-trained separation network.

[0074] The small-sample fine-tuning module includes a small-sample dataset creation unit, a fine-tuning objective function creation unit, and a fine-tuning network unit, wherein:

[0075] The small sample dataset creation unit creates fine-tuned datasets, including communication signals and small sample target interference signals, as well as mixed signals of the two.

[0076] The fine-tuning objective function creation unit creates the objective function for the fine-tuning stage.

[0077] The network units are fine-tuned by using the fine-tuning dataset to fine-tune the pre-trained separation model based on the objective function of the fine-tuning stage, resulting in the fine-tuned separation network.

[0078] The application module includes a signal acquisition unit, a signal separation unit, and a demodulation unit, wherein:

[0079] The signal acquisition unit acquires the time-frequency aliasing signal to be separated through a single antenna;

[0080] The signal separation unit uses a finely tuned separation network to separate the target communication signal and the interference signal from the aliased signal, thus completing the separation of the aliased signal.

[0081] The demodulation unit sends the separated communication signal obtained by the signal separation unit into the demodulator for demodulation to recover the bit data of the signal.

[0082] A mobile terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned small-sample single-channel source separation anti-interference method based on multi-task learning.

[0083] Compared with the prior art, the significant advantages of this invention are: (1) It adopts a strategy of multi-task learning combined with few-sample transfer learning, and builds a model that can handle multiple types of interference at the same time. By taking advantage of multi-task learning, it shares knowledge to improve the generalization ability of the model and solves the communication anti-interference problem when the number of target interference samples is insufficient; (2) It adopts a separation model as the basic framework and guides the model to learn how to identify and adapt to various interferences through multi-task learning pre-training method, which improves the robustness and flexibility of the model in practical applications, and can effectively cope with the interference of unknown or few samples, significantly improving the generalization ability of the separation model. Attached Figure Description

[0084] Figure 1 This is a flowchart illustrating the small-sample single-channel source separation anti-interference method based on multi-task learning according to the present invention.

[0085] Figure 2 This is a schematic diagram of a scenario for the small-sample single-channel source separation anti-interference method based on multi-task learning in the embodiments.

[0086] Figure 3 This is a schematic diagram of the structure of the multi-task learning training and fine-tuning network in an embodiment of the present invention.

[0087] Figure 4 This is a schematic diagram of the structure of the separation network used in multi-task learning in this embodiment of the invention. Detailed Implementation

[0088] 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.

[0089] 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.

[0090] 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.

[0091] 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.

[0092] 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.

[0093] Combination Figure 1 This invention discloses a small-sample single-channel source separation anti-interference method based on multi-task learning, comprising a multi-task training stage, a fine-tuning stage, and an application stage, wherein:

[0094] The multi-task training phase includes the following steps:

[0095] Step A1: Create a training dataset, including communication signals, various known interference signals, and mixed signals of the two;

[0096] Step A2: Build a signal separation network, including encoder, separator and decoder modules;

[0097] Step A3: Create a joint training objective function for multi-task learning;

[0098] Step A4: Based on the joint training objective function, use the signal dataset obtained in step A1 to train the signal separation network in step A2, and obtain the trained pre-trained signal separation network.

[0099] The small-sample fine-tuning phase includes the following steps:

[0100] Step B1: Create a fine-tuning dataset, including communication signals and small sample target interference signals, as well as a mixture of the two;

[0101] Step B2: Create the objective function for the fine-tuning stage;

[0102] Step B3: Based on the objective function of the fine-tuning stage, the pre-trained signal separation network obtained in step A4 is fine-tuned using the fine-tuning dataset obtained in step B1 to obtain the fine-tuned signal separation network.

[0103] The application phase includes the following steps:

[0104] Step C1: Acquire the small sample target interference aliasing signal to be separated using a single antenna;

[0105] Step C2: Using the signal separation network fine-tuned in step B3, the target communication signal and interference signal are separated from the aliased signal to complete the blind separation of the aliased signal;

[0106] Step C3: Send the separated signal obtained in step C2 into the demodulator for demodulation to recover the bit data of the separated signal.

[0107] As a specific example, the creation of the training dataset described in step A1 can be done in either of the following two ways:

[0108] 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.

[0109] 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 an interference signal ratio (ISR) of 0-8dB. Noise is added to the aliased signal to generate an aliased communication signal with an SNR of -5dB to 20dB.

[0110] The aliasing communication signal is shown in the following formula:

[0111]

[0112] 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;

[0113] The multi-task training dataset contains target communication signals and various interference signals, with the target communication signals and each type of interference signal mixed separately.

[0114] The second method involves sampling the actual communication equipment and jamming equipment to obtain a dataset of the target communication signal, jamming signal, and aliasing signal.

[0115] As a specific example, the construction of the signal separation network described in step A2 is as follows:

[0116] The signal separation network can use known classic networks such as Conv-TasNet and SepFormer, or a self-built network model.

[0117] The goal of a single-channel source separation task is to separate the mixed signal x∈R T N source signals s1, s2, ..., s are separated from the signal. N ∈R TThis invention employs a modified Separator Transformer model, namely the SepFormer model. This model is a classic speech separation network with a Transformer-based architecture of a recurrent neural network (RNN), including an encoder, decoder, and separator. To better adapt the network to the characteristics of communication signals, modifications were made to the encoder and separator modules, including the configuration of convolutional layers and activation functions. Due to the low information packing density of communication signals, smaller convolutional kernels and strides were implemented during the convolution process, as detailed below:

[0118] (1) Encoder

[0119] The encoder employs a bottleneck design, replacing the previous one-dimensional convolution. Through dual convolutional layers, the number of kernels is initially increased and then decreased, thereby extracting the encoded features of the signal more accurately. The encoder receives a temporal mixed signal x and generates a learned two-dimensional feature representation h∈R. F×T′ Where F is the feature dimension and T′ is the new length of the signal, a convolutional layer is used: h = conv1d(ReLU(conv1d(x))), where conv1d() represents one-dimensional convolution and ReLU() represents a non-linear activation function;

[0120] (2) Separator

[0121] The separator receives the encoded two-dimensional feature representation h and estimates the mask m1,...,m for each source in the mixed signal. N That is: m k = separator(h), k = 1, ..., N; where separator() represents the separator and N represents the number of information sources;

[0122] During the mask estimation process, the separator first processes the two-dimensional feature representation h into overlapping three-dimensional feature blocks h′, and then inputs h′ into the separator transform block: Where F represents the feature dimension, C is the length of each block, and T is the length of the feature dimension. c It refers to the number of blocks;

[0123] The splitting transform block is the core of the splitter. It uses a dual-scale method to simulate short-term and long-term dependencies. It consists of two types of Transformers: an inner loop Transformer (ntraT) and an outer loop Transformer (InterT). IntraT simulates short-term dependencies within each block, while InterT simulates long-term dependencies across blocks after arranging the dimensions.

[0124] The overall transformation definition of the separated transform block is:

[0125]

[0126] Among them, f intra and f inter These represent IntraT and InterT transformations, respectively; P represents permutation operation; h′ represents a 3D feature block; and h″ represents the feature after separation transformation.

[0127] (3) Decoder

[0128] The decoder uses a mask predicted by a transposed convolutional layer and a masking network to reconstruct the separated signals in the time domain:

[0129]

[0130] in, Let m represent the k-th source separated from the convolution, conv1d-transpose() represents one-dimensional deconvolution, and m k This represents the mask of the k-th source.

[0131] As a specific example, the creation of the joint training objective function for multi-task learning described in step A3 is as follows:

[0132] During the training phase, the parameters of the encoder, decoder, and separator are shared, and a unified objective function is used to minimize the combined loss of all tasks, i.e.:

[0133]

[0134] Where J(θ) represents the total loss of the training task, and θ represents the parameter set shared by all tasks; since some tasks may be more important than others, α is used. j L represents the loss assigned to the j-th task. j The weights, if all tasks are equally important, are α. j =1 / M;

[0135] The parameters θ of the separation model b Optimize in an alternating manner across multiple tasks, i.e.:

[0136]

[0137] Where β represents the learning rate, This indicates gradient descent along θ;

[0138] The loss L per task i The scale-invariant signal-to-noise ratio (SI-SNR) of sentence-level permutation-invariant uPIT loss is used as the loss function.

[0139] Scale-invariant signal-to-noise ratio (SI-SNR) is expressed as:

[0140]

[0141] Among them, s k and Let represent the source signal and the estimated signal, <·,·> represent the dot product of two vectors, ||·||2 represent the L2 norm, and s target e represents the estimated projection component of the signal in the direction of the source signal. noise This represents the vertical component of the estimated signal in the direction of the source signal.

[0142] As a specific example, step A4, based on the joint training objective function, uses the signal dataset obtained in step A1 to train the signal separation network from step A2, resulting in a pre-trained signal separation network, as detailed below:

[0143] The multi-task training set data obtained in step A1 is packaged, the sequences are combined into a batch, and fed into the signal separation network in step A2 for training. The network parameters are set, the Adam optimizer is used, and the initial learning rate is β = 0.0125; a total of 100 rounds of training are conducted; after training, the optimal parameter model is saved.

[0144] 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:

[0145] The target communication signal and the small sample target interference signal are acquired through either of two methods:

[0146] 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.

[0147] 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.

[0148] 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.

[0149] 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.

[0150] As a specific example, the creation of the fine-tuning stage objective function described in step B2 is as follows:

[0151] SI-SNR of uPIT loss is used as the fine-tuning loss function.

[0152] As a specific example, step B3, based on the fine-tuning stage objective function, uses the fine-tuning dataset obtained in step B1 to fine-tune the pre-trained signal separation network obtained in step A4, resulting in the fine-tuned signal separation network, as detailed below:

[0153] Fine-tuning of the pre-trained model is performed using a small sample dataset. The fine-tuning rounds are set to 20 rounds. During the fine-tuning process, some layers of the pre-trained model can be frozen, such as the encoder, decoder, or separator. After the fine-tuning is completed, the optimal model is saved.

[0154] This invention also provides a small-sample single-channel source separation anti-interference device based on multi-task learning, used to implement a small-sample single-channel source separation anti-interference method based on multi-task learning. The device includes a multi-task training module, a small-sample fine-tuning module, and an application module, wherein:

[0155] The multi-task training module includes a dataset creation unit, a signal separation network construction unit, an objective function creation unit, and a network training unit, wherein:

[0156] The dataset creation unit creates a multi-task training dataset, including communication signals and various interference signals;

[0157] The signal separation network construction unit is used to build encoder, separator and decoder modules;

[0158] Joint training objective function creation unit: Create a custom multi-task learning training objective function;

[0159] The network training unit, based on the joint training objective function, uses the signal dataset obtained by the dataset creation unit to train the network obtained by the signal separation network construction unit, and obtains the pre-trained separation network.

[0160] The small-sample fine-tuning module includes a small-sample dataset creation unit, a fine-tuning objective function creation unit, and a fine-tuning network unit, wherein:

[0161] The small sample dataset creation unit creates fine-tuned datasets, including communication signals and small sample target interference signals, as well as mixed signals of the two.

[0162] The fine-tuning objective function creation unit creates the objective function for the fine-tuning stage.

[0163] The network units are fine-tuned by using the fine-tuning dataset to fine-tune the pre-trained separation model based on the objective function of the fine-tuning stage, resulting in the fine-tuned separation network.

[0164] The application module includes a signal acquisition unit, a signal separation unit, and a demodulation unit, wherein:

[0165] The signal acquisition unit acquires the time-frequency aliasing signal to be separated through a single antenna;

[0166] The signal separation unit uses a finely tuned separation network to separate the target communication signal and the interference signal from the aliased signal, thus completing the separation of the aliased signal.

[0167] The demodulation unit sends the separated communication signal obtained by the signal separation unit into the demodulator for demodulation to recover the bit data of the signal.

[0168] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned small-sample single-channel source separation anti-interference method based on multi-task learning.

[0169] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0170] Example

[0171] This embodiment provides a small-sample single-channel source separation anti-interference method based on multi-task learning, including a multi-task training stage, a fine-tuning stage, and an application stage, wherein:

[0172] The multi-task training phase includes the following steps:

[0173] Step A1: Create a training dataset, including communication signals, various known interference signals, and mixed signals of the two;

[0174] Step A2: Build a signal separation network, including encoder, separator and decoder modules;

[0175] Step A3: Create a joint training objective function for multi-task learning;

[0176] Step A4: Based on the joint training objective function, use the signal dataset obtained in step A1 to train the signal separation network in step A2, and obtain the trained pre-trained signal separation network.

[0177] The small-sample fine-tuning phase includes the following steps:

[0178] Step B1: Create a fine-tuning dataset, including communication signals and small sample target interference signals, as well as a mixture of the two;

[0179] Step B2: Create the objective function for the fine-tuning stage;

[0180] Step B3: Based on the objective function of the fine-tuning stage, the pre-trained signal separation network obtained in step A4 is fine-tuned using the fine-tuning dataset obtained in step B1 to obtain the fine-tuned signal separation network.

[0181] The application phase includes the following steps:

[0182] Step C1: Acquire the small sample target interference aliasing signal to be separated using a single antenna;

[0183] Step C2: Using the signal separation network fine-tuned in step B3, the target communication signal and interference signal are separated from the aliased signal to complete the blind separation of the aliased signal;

[0184] Step C3: Send the separated signal obtained in step C2 into the demodulator for demodulation to recover the bit data of the separated signal.

[0185] Figure 2 This diagram illustrates a scenario for a small-sample, single-channel source separation anti-interference method based on multi-task learning, designed to combat unknown interference. The transmitter and receiver in the diagram are responsible for signal transmission and reception, respectively, while the interference source simulates potential interference encountered during communication. The goal of source separation anti-interference is to separate the source signal and the interference signal from the received interfered signal to ensure normal communication. Due to a lack of training data for unknown interference, known interference sources are used for training. Through multi-task learning, the model learns how to identify and adapt to interference. Multi-task learning allows the model to handle multiple interference types simultaneously, improving generalization ability through knowledge sharing. This training strategy enables the system to effectively identify and suppress interference when facing unknown interference, utilizing existing knowledge and experience to ensure communication stability and reliability.

[0186] Figure 3 This is a schematic diagram of the network structure for multi-task learning training and fine-tuning. During the multi-task training phase, a dataset X is created. b The dataset includes communication signals and various interference signals. Based on a joint training objective function, the model is trained using this dataset to obtain a pre-trained separation network, as detailed below:

[0187] The data from the multi-task training set is packaged, the sequences are combined into batches, and fed into the split network. The network parameters are set, and the Adam optimizer is used with an initial learning rate of β = 0.0125. The training is conducted for 100 rounds. After training, the optimal parameter model is saved.

[0188] During the fine-tuning phase, a fine-tuning dataset X is created. n This includes communication signals, small-sample target interference signals, and mixed signals of both, including target communication signals and X.b The communication signal is the same as that in X, and the interference signal is the same as that in X. b The types of interference signals vary, and the number of small samples ranges from 0 to 10.

[0189] Based on the objective function of the fine-tuning stage, the pre-trained separator model is fine-tuned using the fine-tuning dataset to obtain the fine-tuned separator network, as follows: The pre-trained model is fine-tuned using a small sample dataset, with 20 fine-tuning rounds. Too many fine-tuning rounds can easily lead to overfitting. During the fine-tuning process, some layers of the pre-trained model can be frozen, such as freezing the encoder, decoder, or separator; after fine-tuning, the optimal model is saved.

[0190] Figure 4 This is a schematic diagram of the separation network used in multi-task learning. Since multi-task learning is a model-independent learning method, the signal separation network can be a known classic network such as Conv-TasNet, SepFormer, etc., or a self-built network model.

[0191] In single-channel source separation tasks, the goal of source separation anti-interference is to isolate the mixed signal x∈R. T N source signals s1, s2, ..., s are separated from the signal. N ∈R T This embodiment employs a modified SepFormer model, which features a Transformer-based architecture of a recurrent neural network (RNN), including an encoder, decoder, and separator. To better adapt the network to the characteristics of communication signals, modifications were made to the encoder and separator modules, including the configuration of convolutional layers and activation functions. Due to the low information packing density of communication signals, smaller convolutional kernels and strides were implemented during convolution; specifically as follows:

[0192] Encoder: The encoder employs a bottleneck design, replacing the previous one-dimensional convolution. Through double convolutional layers, the number of kernels is initially increased and then decreased, thereby extracting the encoded features of the signal more accurately. The encoder receives a temporal mixed signal x and generates a learned two-dimensional feature representation h∈R. F×T′ Where F is the feature dimension and T′ is the new length of the signal, a convolutional layer is used: h = conv1d(ReLU(conv1d(x)));

[0193] Separator: The separator receives the encoded two-dimensional feature representation h and estimates the mask m1,...,m for each source in the mixed signal. N That is: m k =separator(h),k=1,...,N;

[0194] In this process, it first processes the representation h into overlapping blocks h′ and inputs them into the SepFormer block: Where C is the length of each block, and T c It refers to the number of blocks;

[0195] The SepFormer block is the core of the separator, using a dual-scale method to simulate short-term and long-term dependencies. It consists of two types of Transformers: IntraT and InterT. IntraT simulates short-term dependencies within each block, while InterT simulates long-term dependencies across blocks after arranging the dimensions.

[0196] The overall transformation definition of the SepFormer block is: Where f intra and f inter These represent IntraT and InterT conversions respectively, with P representing permutation operations;

[0197] Decoder: The decoder uses a mask predicted by transposed convolutional layers and a masking network to reconstruct the separated signals in the time domain.

[0198]

[0199] in This represents the k-th source that has been separated.

[0200] In summary, the small-sample single-channel source separation anti-interference method and device based on multi-task learning proposed in this invention solves the communication anti-interference problem when the number of target interference samples is insufficient. Simulation results show that the designed multi-task learning signal separation method can effectively cope with interference from unknown or small number of samples and significantly improve the generalization ability of the separation model.

[0201] 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 small sample single-channel source separation anti-interference method based on multi-task learning, characterized in that, The method comprises a multi-task training phase, a fine-tuning phase and an application phase, wherein: The multi-task training phase comprises the following steps: Step A1, creating a training data set comprising a communication signal, a plurality of known interference signals and a mixed signal of the two; Step A2, building a signal separation network comprising an encoder, a separator and a decoder module; Step A3, creating a joint training objective function for multi-task learning, specifically as follows: During the training phase, the parameters of the encoder, decoder and separator are shared and a unified objective function is used to minimize the combined loss of all tasks, that is: ； wherein Ltotal represents the total loss of the training task, φ represents the parameter set shared by all tasks, is the weight assigned to the loss of the i-th task, ;​​ Separating parameters of the model Optimizing in multiple tasks in an alternating fashion, i.e.: ； wherein, denotes a learning rate, denotes a gradient descent along is performed. Loss for each task The scale-invariant signal-to-noise ratio SI-SNR is used as the loss function for the sentence-level permutation invariant UPIT loss, and the scale-invariant signal-to-noise ratio SI-SNR is expressed as: ； wherein and denotes the source signal and the estimated signal, denotes the two vector dot products, denotes the two-norm, denotes the projection component of the estimated signal in the direction of the source signal, denotes the orthogonal component of the estimated signal in the direction of the source signal; Step A4, based on the joint training objective function, using the signal data set obtained in step A1, training the signal separation network of step A2 to obtain a trained pre-trained signal separation network; The small sample fine-tuning phase comprises the following steps: Step B1, creating a fine-tuning data set comprising a communication signal and a small sample target interference signal and a mixed signal of the two; Step B2, creating a fine-tuning phase objective function; Step B3, based on the fine-tuning phase objective function, fine-tuning the trained pre-trained signal separation network obtained in step A4 using the fine-tuning data set obtained in step B1 to obtain a fine-tuned signal separation network; The application phase comprises the following steps: Step C1, obtaining a small sample target interference mixed signal to be separated through a single antenna; Step C2, using the fine-tuned signal separation network of step B3 to separate the target communication signal and interference signal from the mixed signal to complete the blind separation of the mixed signal; Step C3, sending the separated signal obtained in step C2 to a demodulator for demodulation to recover the bit data of the separated signal.

2. The multi-task learning based small sample single-channel source separation anti-jamming method according to claim 1, characterized in that, The creation of the training data set in step A1 uses any one of the following two methods: The first method: using computer software to generate communication signals of multiple different modulation methods and interference signals, the modulation method being digital modulation or analog modulation, the interference signals including 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, wideband noise interference and sweep frequency interference; Based on the generated target communication signal and interference signal, the target communication signal and a plurality of interference signals are randomly mixed to form a mixed signal, the interference to signal ratio ISR is set to 0-8dB, noise is added to the mixed signal to generate a mixed communication signal, and the signal to noise ratio SNR is-5dB to 20dB; The mixed communication signal is as follows: ； 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 target communication signal and each interference signal are mixed separately in the multi-task training data set; The second method: sampling actual communication equipment and interference equipment to obtain a data set of target communication signals, interference signals and mixed signals.

3. The multi-task learning based small sample single-channel source separation anti-jamming method according to claim 2, characterized in that, The signal separation network of step A2 is built as follows: The goal of the single-channel source separation task is to separate one source signal from a mixed signal ; the modified separation transformer, namely SepFormer model, has a Transformer-based architecture without recurrent neural network, RNN, including an encoder, a decoder and a separator, the encoder and the separator modules are modified, including the configuration of the convolution layer and the activation function, and the convolution kernel and the step are reduced in the convolution process, as follows: (1) Encoder The encoder adopts a bottleneck design, instead of one-dimensional convolution, to accurately extract the encoding features of the signal by passing through two convolutional layers, initially increasing and then reducing the number of kernels; the encoder receives the time-domain mixed signal and generates a learned two-dimensional feature representation wherein is the feature dimension, is the new length of the signal, using the convolutional layer: wherein conv1d() represents one-dimensional convolution, and ReLU() represents a kind of nonlinear activation function; (2) Separator The separator receives the encoded two-dimensional feature representation and estimates a mask for each source in the mixed signal i.e. ; where separator() represents the separator and N represents the number of sources; In estimating the mask, the separator first transforms the two-dimensional feature representation into overlapping three-dimensional feature blocks and inputs them to a separation transform block: where F denotes the feature dimension, is the length of each block, is the number of blocks; The separation transform block is the core of the separator, which uses a two-scale method to simulate short-term and long-term dependencies, consisting of two Transformers, namely an intra-Transformer (ntraT) and an inter-Transformer (InterT), where IntraT simulates short-term dependencies within each block, and InterT simulates long-term dependencies across blocks after arranging dimensions; The overall conversion of the separation transform block is defined as: ； wherein, and denote IntraT and InterT conversion, respectively, denotes a permutation operation, denotes a three-dimensional feature block, denotes a separated transformed feature; (3) Decoder The decoder uses a transposed convolution layer and a mask predicted by the mask network to reconstruct the separated signal in the time domain: ； wherein, represents the first source separated out source, conv1d-transpose() represents a one-dimensional inverse convolution, represents the mask of the first source.

4. The multi-task learning based small sample single-channel source separation anti-jamming method according to claim 1, characterized in that, Based on the joint training objective function described in step A4, the signal separation network of step A2 is trained using the signal data set obtained in step A1, and a trained pre-trained signal separation network is obtained, as follows: The data of the multi-task training set obtained in step A1 is packaged, sequentially combined into a batch, and sent to the signal separation network of step A2 for training. The network parameters are set, the Adam optimizer is used, the initial learning rate is ; and co-training is performed for 100 rounds. After the training is completed, the optimal parameter model is saved.

5. The multi-task learning based small sample single-channel source separation anti-jamming method according to claim 1, characterized in that, 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: The target communication signal and the small sample target interference signal are obtained by any one of the following two methods: The first method: using computer software to generate a communication signal and a small sample interference signal, the communication signal modulation mode is digital modulation or analog modulation, the interference signal includes digital modulation interference signal, analog modulation interference signal, noise frequency modulation interference signal, noise amplitude modulation interference signal, single tone interference signal, multi-tone interference signal, comb spectrum interference signal, wideband noise interference, sweep 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 a mixed signal, and noise is added to the mixed signal to generate a mixed communication signal, and the signal-to-noise ratio SNR is -5dB to 20dB; The second method: sampling the actual communication equipment and interference equipment to obtain the data set of the target communication signal, the target interference signal and the mixed signal; The target communication signal obtained is the same as the communication signal in step A1, and the interference signal is different from the interference signal in step A1, and the number of small samples is 0-10.

6. The small sample single-channel source separation anti-jamming method based on multi-task learning according to claim 1, characterized in that, The fine-tuning stage objective function described in step B2 is as follows: The SI-SNR using uPIT loss is used as the fine-tuning loss function.

7. The multi-task learning based small sample single-channel source separation anti-jamming method according to claim 1, characterized in that, Based on the fine-tuning stage objective function described in step B3, the pre-trained signal separation network obtained in step A4 is fine-tuned using the fine-tuning data set obtained in step B1, and a fine-tuned signal separation network is obtained, as follows: The pre-trained model is fine-tuned using the small sample data set, the fine-tuning round is set to 20 rounds, the encoder, decoder or separator of the pre-trained model is selected during fine-tuning, and the optimal model is saved after fine-tuning.

8. A small sample single-channel source separation anti-jamming device based on multi-task learning, characterized by, The device is used to implement the small sample single-channel source separation anti-jamming method based on multi-task learning according to any one of claims 1-7, and the device comprises a multi-task training module, a small sample fine-tuning module and an application module, wherein: The multi-task training module comprises a data set creation unit, a signal separation network building unit, an objective function creation unit and a network training unit, wherein: The data set creation unit creates a multi-task training data set, including a communication signal and multiple interference signals. The signal separation network building unit builds an encoder, a separator, and a decoder module. The joint training target function creation unit creates a custom multi-task learning training target function. The network training unit trains the network obtained by the signal separation network building unit based on the joint training target function and the signal data set obtained by the data set creation unit, to obtain a trained pre-training separation network. The small sample fine-tuning module includes a small sample data set creation unit, a fine-tuning target function creation unit, and a fine-tuning network unit, wherein: The small sample data set creation unit creates a fine-tuning data set, including a communication signal, a small sample target interference signal, and a mixed signal of the two. The fine-tuning target function creation unit creates a fine-tuning stage target function. The fine-tuning network unit fine-tunes the trained pre-training separation model based on the fine-tuning stage target function and the fine-tuning data set, to obtain a fine-tuned separation network. The application module includes a signal acquisition unit, a signal separation unit, and a demodulation unit, wherein: The signal acquisition unit acquires a time-frequency overlap signal to be separated through a single antenna. The signal separation unit separates the target communication signal and the interference signal from the overlap signal using the fine-tuned separation network, to complete the separation of the overlap signal. The demodulation unit sends the separated communication signal obtained by the signal separation unit to a demodulator for demodulation, to recover the bit data of the signal.

9. A mobile terminal comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the multi-task learning-based small sample single-channel source separation anti-interference method according to any one of claims 1-7.

Images