Multiband signal identification method based on multitask learning

Abstract

The invention discloses a multiband signal identification method based on multitask learning, and the method comprises the steps: obtaining a to-be-identified multiband signal, and carrying out the preprocessing of the to-be-identified multiband signal; the multiband wireless signal identification network based on multitask learning performs spectrum sensing, modulation identification and signal classification on the preprocessed to-be-identified multiband signal; outputting results of spectrum sensing, modulation identification and signal classification corresponding to the multiband signal to be identified; according to the method, the shared features of different recognition tasks are extracted, the single-task recognition performance is improved, spectrum sensing, modulation recognition and signal classification tasks are completed at the same time, the three tasks are completed in one model at the same time, and the complexity of the model is reduced.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a multi-band signal recognition method based on multi-task learning. Background Art

[0002] Wireless signal recognition technology plays a vital role in modern wireless communications because it can distinguish the modulation type, signal type, and individual identification information of different wireless signals by analyzing the signal waveform. However, with the rapid development of communication technology, the types of communication equipment have increased rapidly and the differences in signal modulation types have become even greater. Therefore, it is urgent to design more efficient wireless signal recognition methods to solve these problems. Traditional wireless signal recognition technology requires manual design and selection of features, which can be divided into methods based on likelihood ratio detection and methods based on feature experience drive. This type of method has a high dependence on features, poor recognition accuracy in complex channels and noise environments, and is difficult to adapt to the current communication environment.

[0003] Data-driven methods such as deep learning are now widely used in single wireless signal recognition tasks such as modulation recognition, signal classification, and individual identification of radiators. In most existing wireless signal recognition technology-related works, the recognition of different parameters carried by the signal (signal type, modulation type, etc.) is often implemented as multiple independent tasks, and each task is solved by a specific neural network model. Most existing signal recognition technologies based on deep learning can only recognize a certain feature of the signal (single-task signal recognition) or recognize narrowband signals (without considering the actual multi-band signal situation). Since the same signal often has multiple information, recognition training for a single parameter does not utilize the correlation between different parameters of the signal, and the utilization rate of the data set is low. In addition, training multiple models for multiple tasks will cause huge training overhead, and multiple models occupy more storage and computing resources of the device, making it difficult to deploy on resource-constrained edge sensor devices in actual environments. With the rapid development of modern communication technology, wireless signals may be distributed in a wider frequency band range, and their specific frequency band positions may also be unknown. Therefore, the existing single-task multi-band signal recognition method cannot meet the requirements of multi-band wireless signal recognition accuracy in multi-task scenarios.

[0004] Therefore, the prior art needs to be improved. Summary of the invention

[0005] The technical problem to be solved by the present invention is that, in view of the defects of the prior art, the present invention provides a multi-band signal recognition method based on multi-task learning to solve the problem of low recognition accuracy of the existing single-task multi-band signal recognition method.

[0006] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0007] In a first aspect, the present invention provides a multi-band signal recognition method based on multi-task learning, comprising:

[0008] Acquire a multi-band signal to be identified, and pre-process the multi-band signal to be identified;

[0009] The multi-band wireless signal recognition network based on multi-task learning performs spectrum sensing, modulation recognition and signal classification on the pre-processed multi-band signals to be identified;

[0010] Output the results of spectrum sensing, modulation recognition and signal classification corresponding to the multi-band signal to be identified.

[0011] In one implementation, the multi-band wireless signal recognition network based on multi-task learning includes: a shared feature extraction network, a spectrum sensing classification head, a modulation recognition classification head, and a signal classification classification head.

[0012] In one implementation, the adaptive loss function of the multi-band wireless signal recognition network based on multi-task learning is:

[0013]

[0014] in, and They represent the loss functions of spectrum sensing, modulation recognition and signal classification tasks respectively, σ1, σ2 and σ3 represent the homoscedastic uncertainty parameters corresponding to the three tasks respectively, Θ s represents the parameters of the shared feature extraction network, Θ T1 , Θ T2 , Θ T3 Represent the parameters of the three task classification heads respectively.

[0015] In one implementation, the multi-band wireless signal recognition network based on multi-task learning performs spectrum sensing, modulation recognition, and signal classification on the pre-processed multi-band signal to be recognized, including:

[0016] Extracting shared features from the preprocessed multi-band signal to be identified by using the shared feature extraction network;

[0017] The shared features are respectively input into the spectrum sensing classification head, the modulation identification classification head and the signal classification classification head to obtain corresponding classification results.

[0018] In one implementation, the shared feature extraction network includes: a multi-task feature extraction convolutional layer, a plurality of multi-task basic residual blocks, and a plurality of multi-task special residual blocks;

[0019] The step of extracting the shared features from the preprocessed multi-band signal to be identified by the shared feature extraction network includes:

[0020] Inputting the preprocessed multi-band signal to be identified into the multi-task feature extraction convolution layer and a maximum pooling layer for feature extraction;

[0021] The features extracted from the two layers are passed through a network formed by stacking multiple multi-task basic residual blocks and multiple multi-task special residual blocks, and finally through a global average pooling layer to calculate the shared features; wherein each of the multiple multi-task basic residual blocks and each multi-task special residual block is calculated as follows:

[0022] F o =R(F i )+M(F i );

[0023] Among them, F i represents the input feature, F o represents the output feature, R(·) represents the residual function, M(·) represents the skip connection, and in the multi-task basic residual block, M(F i )=F i , in the multi-task special residual block, M(F i )=Conv1D(BN(F i )), where Conv1D represents the convolutional layer operation and BN represents batch normalization.

[0024] In one implementation, the step of inputting the preprocessed multi-band signal to be identified into the multi-task feature extraction convolution layer and a maximum pooling layer for feature extraction includes:

[0025] For the input feature F of the multi-task feature extraction convolution layer, the feature is obtained through the convolution feature extraction branch The channel attention map P obtained by the channel attention branch C (F), the spatial attention map P obtained by the spatial attention branch S (F);

[0026] The characteristics Respectively with the channel attention map P C (F) and the spatial attention map P S (F) Perform element-by-element multiplication to obtain features and Features

[0027] The characteristics feature And features Add them together to get the output multi-task extraction features.

[0028] In one implementation, the inputting the shared features into the spectrum sensing classification head, the modulation identification classification head, and the signal classification classification head respectively to obtain corresponding classification results includes:

[0029] The shared features extracted by the shared feature extraction network from the preprocessed multi-band signal to be identified are input into the spectrum sensing classification head, a spectrum sensing result vector is output, and the spectrum sensing result vector is further processed to obtain a spectrum sensing result;

[0030] The shared features extracted by the shared feature extraction network from the pre-processed multi-band signal to be identified are input into the modulation recognition classification head, a modulation recognition result vector is output, and the modulation recognition result vector is further processed to obtain a modulation recognition result;

[0031] The shared features extracted by the shared feature extraction network from the preprocessed multi-band signal to be identified are input into the signal classification head, and a signal classification result vector is output. The signal classification result vector is further processed to obtain a signal classification result.

[0032] In a second aspect, the present invention provides a multi-band signal recognition device based on multi-task learning, comprising:

[0033] A multi-band signal preprocessing module, used for acquiring a multi-band signal to be identified and preprocessing the multi-band signal to be identified;

[0034] A multi-band signal recognition module, which is used for performing spectrum sensing, modulation recognition and signal classification on the pre-processed multi-band signal to be recognized by a multi-band wireless signal recognition network based on multi-task learning;

[0035] The classification result output module is used to output the results of spectrum sensing, modulation recognition and signal classification corresponding to the multi-band signal to be identified.

[0036] In a third aspect, the present invention provides a terminal, comprising: a processor and a memory, the memory storing a multi-band signal recognition program based on multi-task learning, and the multi-band signal recognition program based on multi-task learning, when executed by the processor, is used to implement the operation of the multi-band signal recognition method based on multi-task learning as described in the first aspect.

[0037] In a fourth aspect, the present invention further provides a medium, which is a computer-readable storage medium, and which stores a multi-band signal recognition program based on multi-task learning. When the multi-band signal recognition program based on multi-task learning is executed by a processor, it is used to implement the operation of the multi-band signal recognition method based on multi-task learning as described in the first aspect.

[0038] The present invention adopts the above technical solution to achieve the following effects:

[0039] The present invention preprocesses the multi-band signal to be identified, and can perform spectrum sensing, modulation recognition and signal classification on the preprocessed multi-band signal to be identified based on a multi-band wireless signal recognition network of multi-task learning, and output the spectrum sensing, modulation recognition and signal classification results corresponding to the multi-band signal to be identified; the present invention improves the single-task recognition performance by extracting shared features of different recognition tasks, and simultaneously completes the spectrum sensing, modulation recognition and signal classification tasks, and simultaneously completes the three tasks in one model, thereby reducing the complexity of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0041] Figure 1 It is a flow chart of the multi-band signal recognition method based on multi-task learning in the present invention.

[0042] Figure 2 It is a structural diagram of a multi-band wireless signal recognition network based on multi-task learning in the present invention.

[0043] Figure 3 Schematic diagram of the structure of the multi-task basic residual block in the present invention.

[0044] Figure 4 It is a structural diagram of the multi-task special residual block in the present invention.

[0045] Figure 5 It is a schematic diagram of the structure of the multi-task feature extraction convolutional layer in the present invention.

[0046] Figure 6 It is a functional principle diagram of a terminal in one implementation of the present invention.

[0047] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0049] Exemplary Methods

[0050] Data-driven methods such as deep learning are now widely used in single wireless signal recognition tasks such as modulation recognition, signal classification, and individual identification of radiators. In most existing wireless signal recognition technology-related works, the recognition of different parameters carried by the signal (signal type, modulation type, etc.) is often implemented as multiple independent tasks, and each task is solved by a specific neural network model. Most existing signal recognition technologies based on deep learning can only recognize a certain feature of the signal (single-task signal recognition) or recognize narrowband signals (without considering the actual multi-band signal situation). Since the same signal often has multiple information, recognition training for a single parameter does not utilize the correlation between different parameters of the signal, and the utilization rate of the data set is low. In addition, training multiple models for multiple tasks will cause huge training overhead, and multiple models occupy more storage and computing resources of the device, making it difficult to deploy on resource-constrained edge sensor devices in actual environments. With the rapid development of modern communication technology, wireless signals may be distributed in a wider frequency band range, and their specific frequency band positions may also be unknown. Therefore, the existing single-task multi-band signal recognition method cannot meet the requirements of multi-band wireless signal recognition accuracy in multi-task scenarios.

[0051] In view of the above technical problems, an embodiment of the present invention provides a multi-band signal recognition method based on multi-task learning. The method preprocesses the multi-band signal to be identified, and can perform spectrum sensing, modulation recognition and signal classification on the preprocessed multi-band signal to be identified based on a multi-band wireless signal recognition network based on multi-task learning, and outputs the spectrum sensing, modulation recognition and signal classification results corresponding to the multi-band signal to be identified. The embodiment of the present invention improves the single-task recognition performance by extracting shared features of different recognition tasks, and simultaneously completes the spectrum sensing, modulation recognition and signal classification tasks, and simultaneously completes the three tasks in one model, thereby reducing the complexity of the model.

[0052] like Figure 1 As shown, an embodiment of the present invention provides a multi-band signal recognition method based on multi-task learning, comprising the following steps:

[0053] Step S100: acquiring a multi-band signal to be identified, and preprocessing the multi-band signal to be identified.

[0054] In this embodiment, a multi-band wireless signal recognition network based on multi-task learning is proposed, which can simultaneously learn multiple related tasks, including spectrum sensing, modulation recognition, and signal classification. Through the shared feature extraction network and task-specific classification head in the multi-band wireless signal recognition network based on multi-task learning, the correlation and shared information between different tasks can be effectively captured, thereby improving the recognition performance.

[0055] Specifically, Figure 2 As shown, the multi-band wireless signal recognition network based on multi-task learning provided in this embodiment includes: a shared feature extraction network, a spectrum sensing classification head, a modulation recognition classification head and a signal classification classification head. Among them, the shared feature extraction network is an improved ResNet (i.e., a deep residual network), which replaces part of the traditional convolution in the ResNet network with a multi-task feature extraction convolution (i.e., MTFE-Conv, hereinafter referred to as MTFE-Conv), and the two residual blocks based on MTFE-Conv are respectively called a multi-task basic residual block (i.e., MTL-Basic block, hereinafter referred to as MBB) and a multi-task special residual block (i.e., MTL-Special block, hereinafter referred to as MSB), and these two residual blocks are respectively as shown in Figure 3 As shown and Figure 4 As shown; that is, the shared feature extraction network includes: MTFE-Conv, multiple MBBs and multiple MSBs.

[0056] In one implementation of this embodiment, the spectrum sensing classification head, the modulation recognition classification head, and the signal classification classification head are all composed of some simple fully connected layers (FC) and Dropout layers, and use the ReLU activation function. The inputs are all shared features learned by the shared feature extraction network, and the outputs are the results of their respective tasks. The multi-task multi-band wireless signal recognition network uses an end-to-end multi-task model design to improve the single-task recognition performance by extracting shared features of different recognition tasks, and simultaneously completes multi-band spectrum sensing, modulation recognition, and signal classification tasks, reducing the overall complexity of the model.

[0057] This embodiment provides a new type of MTFE-Conv (i.e., multi-task feature extraction convolution), such as Figure 5As shown. This new convolution combines the channel attention mechanism and the spatial attention mechanism, and by improving the learning effect of useful features, the shared feature extraction network can more efficiently extract shared feature information of multiple tasks, improve the recognition performance of the multi-band wireless signal recognition network based on multi-task learning, and reduce the number of parameters of the multi-band wireless signal recognition network based on multi-task learning. This embodiment also proposes a multi-task adaptive loss function design scheme, which can be adjusted to balance the training progress of different tasks during the training process, improve the generalization of the model and the model's ability to extract features from different tasks.

[0058] In this embodiment, the spectrum occupancy information, modulation type and signal type parameters of wireless communication signals within a wide frequency band are jointly considered, and based on the idea of ​​multi-task learning, an end-to-end single-step training signal recognition network based on multi-task learning is designed. The multi-band wireless signal recognition network based on multi-task learning uses an end-to-end model design so that three tasks can be completed simultaneously in a multi-task learning model without the need for multiple stages of training and testing. At the same time, MTFE-Conv and multi-task adaptive loss function design schemes are introduced, which effectively improves the recognition accuracy of a single task and reduces the overall model complexity and recognition reasoning calculation amount.

[0059] In an implementation of this embodiment, since the method is to identify a multi-band signal, the multi-band signal is composed of a plurality of narrow-band signals with different center frequencies, which can be expressed as:

[0060]

[0061] Where P represents the number of occupied subbands, h p (t) represents the channel response of the subband where the pth signal is located, * represents the convolution operation, and η(t) is the additive white Gaussian noise (AWGN). p (t) is the signal of the pth occupied subband, which is the signal modulated by the transmitter and has a specific modulation type and signal type.

[0062] Since complex-valued signals cannot be directly input into convolutional neural networks, signal preprocessing is required to decompose the signal into real and imaginary parts:

[0063] X re =Real(x(t));

[0064] X im =Imag(x(t));

[0065] Among them, X re is the real part of x(t), X im is the imaginary part of x(t).

[0066] X re and X im are put into two different tensor channels:

[0067]

[0068] In addition, when training a neural network, when the value of the network input data is very large or very small, problems such as gradient disappearance and slow convergence may occur. Therefore, before training or testing the network, data normalization is used to map the real and imaginary parts of the input signal to the interval [-1,1]. This process can be expressed as:

[0069]

[0070] and Represent the maximum and minimum values ​​of all input samples respectively.

[0071] In this embodiment, before the multi-task multi-band wireless signal recognition network analyzes the pre-processed multi-band signal to be recognized, it needs to be trained. The specific process is as follows:

[0072] 1. Generate a multi-band signal data set and divide it into training set and test set;

[0073] 2. Input the training set into the multi-band wireless signal recognition network based on multi-task learning, and perform offline training on the multi-task network;

[0074] 3. Input the test set into the multi-band wireless signal recognition network based on multi-task learning, and perform online testing on the multi-task network. At this time, the input of the network is the multi-band wireless signal, and the output is the spectrum sensing result, modulation recognition result and signal classification result.

[0075] The process of generating a multi-band signal data set and dividing it into a training set and a test set includes: first, generating multi-band signal samples including spectrum occupancy, modulation type and signal type, and marking the multi-band signal samples with corresponding multi-task labels; then, randomly shuffling the generated multi-band signals, using one part as a training set and the other part as a test set, and preprocessing the signals.

[0076] The training set is input into a multi-band wireless signal recognition network based on multi-task learning, and the multi-task network is trained offline, including: using a labeled data set to train the parameters of the model. During offline training, the model will optimize and update the parameters through a back propagation algorithm and an optimization algorithm to minimize the loss function on the training data. After multiple iterations, the model converges or reaches a predetermined stop condition. During the offline training stage, the model can learn the characteristics and representation of the data so as to make accurate predictions in subsequent tests or practical applications.

[0077] Regarding the design of the loss function during offline training: The loss function of multi-task learning is in the form of a weighted summation, and the weights of the loss functions of different sub-tasks need to be manually adjusted. The performance of the multi-band wireless signal recognition network based on multi-task learning is also affected by the weights of the loss function. Improper weight settings may even prevent the multi-task multi-band wireless signal recognition network from obtaining relatively good results for multiple tasks. However, manually setting these weight hyperparameters is very time-consuming and labor-intensive, and a more convenient and quick method needs to be designed to learn the weights that best suit the network. Taking into account the task-dependent uncertainty of each task, the model of this embodiment can automatically learn the weights of the loss functions of the modulation recognition, signal classification, and spectrum sensing sub-tasks, and can adjust itself during the training process, which can effectively improve the performance of the multi-task model.

[0078] In one implementation of this embodiment, the adaptive loss function of the multi-band wireless signal recognition network based on multi-task learning is:

[0079]

[0080] in, and They represent the loss functions of spectrum sensing, modulation recognition and signal classification tasks respectively, σ1, σ2 and σ3 represent the homoscedastic uncertainty parameters corresponding to the three tasks respectively, Θ s represents the parameters of the shared feature extraction network, Θ T1 , Θ T2 , Θ T3 Represent the parameters of the three task classification heads respectively.

[0081] During the training process, σ1, σ2 and σ3 are defined as trainable parameters and are adaptively adjusted during the training process.

[0082] The online test in the above training process means that after the model is trained, it is applied to the data set for real-time prediction or inference. During the online test, the model will receive input data and make predictions or inferences based on the parameters and rules obtained from the previous training. The multi-band signal samples are input into the multi-band wireless signal recognition network based on multi-task learning. The shared feature extraction network extracts the shared features between the three tasks, and then the shared features are input into the multi-task classification head to output the spectrum sensing results, modulation recognition results and signal classification results.

[0083] In this embodiment, the above training process can obtain a trained multi-band wireless signal recognition network based on multi-task learning. Based on this trained network, spectrum perception, modulation recognition and signal classification can be performed on the pre-processed multi-band signal to be identified, thereby realizing multi-task multi-band wireless signal recognition.

[0084] like Figure 1 As shown, an embodiment of the present invention provides a multi-band signal recognition method based on multi-task learning, comprising the following steps:

[0085] Step S200: a multi-band wireless signal recognition network based on multi-task learning performs spectrum sensing, modulation recognition and signal classification on the pre-processed multi-band signal to be recognized.

[0086] Specifically, in an implementation of this embodiment, step S200 includes the following steps:

[0087] Step S201: extracting shared features from the preprocessed multi-band signal to be identified through the shared feature extraction network.

[0088] Specifically, in an implementation of the present embodiment, the shared features in the preprocessed multi-band signal to be identified are extracted through the shared feature extraction network, including: inputting the preprocessed multi-band signal to be identified into the MTFE-Conv and a maximum pooling layer for feature extraction; passing the features extracted from the two layers through a network formed by stacking multiple MBBs and multiple MSBs, and finally passing through a global average pooling layer to calculate the shared features.

[0089] In this embodiment, a multi-band wireless signal recognition network based on multi-task learning is composed of a shared feature extraction network and three multi-task classification heads, which realize spectrum perception, modulation recognition and signal classification of multi-band signals. The multi-task shared feature extraction network is an improved ResNet (i.e., deep residual network). ResNet is a typical deep learning architecture that solves the gradient degradation and vanishing problems in the neural network training process by introducing jump connections. This design allows the network to learn residual functions, that is, the network can update the output features through the learned residuals without directly learning complex output features. The shared feature extraction network replaces some traditional convolutions in the ResNet network with MTFE-Conv. The three multi-task classification heads include: spectrum perception classification head, modulation recognition classification head and signal classification classification head. The inputs of the three classification heads are all shared features learned by the shared feature extraction network, and the outputs are the results of their respective tasks. The multi-task classification head consists of some simple fully connected layers.

[0090] In this embodiment, the calculation process of each MBB and each MSB is as follows:

[0091] F o =R(F i )+M(F i );

[0092] Among them, F i represents the input feature, F o represents the output feature, R(·) represents the residual function, that is, the mapping relationship that the residual block needs to learn. In MBB and MSB, R(·) consists of MTFE-Conv, batch normalization, etc. M(·) represents the skip connection. In MBB, since no dimension matching operation is required, M(F i )=F i In MSB, a dimension matching operation is performed, not only to match the input and output feature dimensions, but also to reduce the spatial size of the input features to a certain extent, further improving the performance and efficiency of the network. i )=BN(Conv1D(F i )), Conv1D represents the convolutional layer operation, and BN represents batch normalization.

[0093] Specifically, in an implementation of this embodiment, the preprocessed multi-band signal to be identified is input into the MTFE-Conv and a maximum pooling layer for feature extraction, and both MBB and MSB are stacked by multiple MTFE-Convs. The MTFE-Conv includes: for the input feature F of the MTFE-Conv, a feature is obtained by a convolution feature extraction branch. The channel attention map P obtained by the channel attention branchC (F), the spatial attention map P obtained by the spatial attention branch S (F); the feature Respectively with the channel attention map P C (F) and the spatial attention map P S (F) Perform element-by-element multiplication to obtain features and Features The characteristics feature And features Add them together to get the output multi-task extraction features.

[0094] like Figure 5 As shown, in this embodiment, MTFE-Conv can automatically learn the importance of each convolution kernel, select useful features and weaken the extraction of unimportant features, and includes four main branches: convolution feature extraction branch, channel attention branch, spatial attention branch and fusion branch.

[0095] 1) Convolutional feature extraction branch: This branch uses one-dimensional grouped convolution for feature extraction. In terms of feature extraction, grouped convolution is the same as standard convolution. Through different convolution kernels, the convolution operation can extract different features from the input data. Assume that the input feature is The output features are Among them C i and C0 represent the number of feature channels of input and output respectively, L i and L0 represent the lengths of the input and output feature vectors respectively. The calculation process of this branch is:

[0096]

[0097] in, Represents grouped convolution operation. In convolutional neural networks, by stacking multiple convolutions, more abstract and complex features can be learned, but this also leads to an increase in the number of parameters and computation. Grouped convolution effectively reduces the amount of computation and calculation, but there is also the problem of isolating information exchange between different groups. Introducing the channel shuffle operation before the grouped convolution operation can randomly rearrange the channels within each group, disrupt the order of the channels within the group, further increase the interaction between features, and help the network learn more robust and generalized feature representations.

[0098] 2) Channel attention branch: The channel attention mechanism can automatically learn the importance of each channel, and then weight each channel, which improves the model's attention to the feature information of different channels. This branch allows the convolution feature extraction branch to pay more attention to the feature channels that are useful for the current task during convolution operations. Figure 5As shown in the figure, after the input features are shuffled, the feature map is first compressed based on two dimensions through the global maximum pooling layer (GMP) and the global average pooling layer (GAP) to obtain feature descriptions of two different dimensions to reduce information loss in the pooling process. Then, a one-dimensional convolutional layer is used for learning, and finally a sigmoid activation function is used to obtain the channel feature map P C (F). In each MSB, a downsampling is required before the Sigmoid activation function to ensure the consistency of the dimension. Assume that the input signal is F∈C i ×L i , the branch calculation process can be expressed as:

[0099] P C (F)=Sigmoid(Conv1D(GMP(F)+GAP(F)))∈C0×1;

[0100] Among them, f 1D (·) represents convolution operation, GMP(·) represents global maximum pooling layer operation, GAP(·) represents global average pooling layer operation, and Sigmoid(·) represents Sigmoid activation function.

[0101] 3) Spatial attention branch: The spatial attention mechanism can automatically learn the importance of each spatial position (corresponding to a feature map or feature vector), and then weight each position, thereby increasing the model's attention to feature information in different spaces. This branch allows the grouped convolution feature extraction branch to pay more attention to spatial positions that are useful for the current task during convolution operations. Figure 5 As shown in Figure 1, this branch first performs maximum pooling and average pooling on the input features and concatenates the results to obtain an effective feature description. Then, the pooled results are passed through a one-dimensional convolution layer, and finally a spatial attention map P is obtained through a Sigmoid activation function. S (F), Similarly, in each MSB, an upsampling needs to be added before the Sigmoid activation function to ensure the consistency of the dimension. The calculation process of this branch can be expressed as:

[0102] P S (F)=Sigmoid(Conv1D([MaxPool(F);AvgPool(F)]))∈1×L0;

[0103] Among them, MaxPool(·) represents the maximum pooling layer operation, and AvgPool(·) represents the average pooling layer operation.

[0104] 4) Fusion branch: The input of this branch is the channel attention map P obtained by the channel attention branch of feature FC (F), the spatial attention map P obtained by the spatial attention branch S (F), and obtained through the convolutional feature extraction branch Then, Respectively with the channel attention map P C (F) and the spatial attention map P S (F) Perform element-by-element multiplication to obtain and The results of the three sub-branches are added together to get the result F of MTFE-Conv out . The branch Figure 5 As shown, the whole process can be expressed as:

[0105]

[0106] in, Represents element-wise multiplication.

[0107] Specifically, in an implementation of this embodiment, step S200 further includes the following steps:

[0108] Step S202: input the shared features into the spectrum sensing classification head, the modulation recognition classification head and the signal classification classification head respectively to obtain corresponding classification results.

[0109] Specifically, in an implementation of this embodiment, the inputting the shared features into the spectrum sensing classification head, the modulation recognition classification head, and the signal classification classification head respectively to obtain corresponding classification results includes:

[0110] The shared features extracted by the shared feature extraction network from the preprocessed multi-band signal to be identified are input into the spectrum sensing classification head, a spectrum sensing result vector is output, and the spectrum sensing result vector is further processed to obtain a spectrum sensing result;

[0111] The shared features extracted by the shared feature extraction network from the pre-processed multi-band signal to be identified are input into the modulation recognition classification head, a modulation recognition result vector is output, and the modulation recognition result vector is further processed to obtain a modulation recognition result;

[0112] The shared features extracted by the shared feature extraction network from the preprocessed multi-band signal to be identified are input into the signal classification head, and a signal classification result vector is output. The signal classification result vector is further processed to obtain a signal classification result.

[0113] As an example, in this embodiment, consider the number of subbands N b = 18 multi-band signals, where the maximum number of occupied sub-bands is No =3, number of modulation types N m =11, number of signal types N c =7. is the spectrum sensing result vector, is the modulation identification result vector, Signal classification result vector, N vm =N o ×N m , N vc =N o ×N c .

[0114] For the spectrum-aware classification head, we use the spectrum-aware label encoding vector Label encoding is performed for spectrum sensing tasks. If a subband contains a signal, the corresponding position in the label encoding is set to 1, indicating that the subband is occupied. If the subband does not contain a signal, the corresponding position in the label encoding is set to 0, indicating that the subband is not occupied. In the inference test phase, a threshold ξ is introduced to determine the number of occupied subbands, ξ = -1. For the output vector of the spectrum sensing head We compare each of its elements with ξ. If only one element is greater than ξ, it means that only one subband is occupied, and The largest element in is set to 1, and the rest are set to 0. If only two elements are greater than ξ, this indicates that only two subbands are occupied. The two largest elements in are set to 1, and the rest are set to 0. o or more elements are greater than ξ, then only N o subbands are occupied, Maximum N o elements are set to 1 and the rest are set to 0.

[0115] For the modulation identification classification head, the modulation identification label encoding vector v is used mr The occupied subband modulation type is encoded by label. mr In [0:N m -1] represents the one-hot coding of the first subband modulation type, located in [N m :2N m -1] represents the one-hot coding of the modulation type of the second subband, located in [N vm -N m :N vm -1] represents the one-hot coding of the modulation type of the last subband. In the test phase, we will With v mr Same length N mTo split:

[0116]

[0117] In vector In the above equation, the subscript of the element with the maximum value represents the modulation type predicted by the modulation recognition classification head. The calculation process can be expressed as:

[0118]

[0119] in Respectively represent the 1st, 2nd, ..., Nth o The numerical index corresponding to the modulation type of the occupied subband.

[0120] For the signal classification head, the signal classification label encoding vector v is used sc The signal type of the occupied subband is encoded by label. sc In [0:N c -1] represents the one-hot coding of the signal type of the first subband, located in [N c :2N c -1] represents the one-hot coding of the signal type of the second subband, located in [N vc -N c :N vc -1] represents the one-hot encoding of the signal type of the last subband. In the testing phase, we will With v sc The label encoding method is the same as the length N c To split:

[0121]

[0122] In vector In , the element subscript of the maximum value represents the signal type predicted by the signal classification head. This calculation process can be expressed as:

[0123]

[0124] in Respectively represent the 1st, 2nd, ..., Nth o The numerical index corresponding to the signal type of the occupied subband.

[0125] like Figure 1 As shown, an embodiment of the present invention provides a multi-band signal recognition method based on multi-task learning, comprising the following steps:

[0126] Step S300: outputting the results of spectrum sensing, modulation recognition and signal classification corresponding to the multi-band signal to be identified.

[0127] In this embodiment, the output results of the above three classification heads are used as the results of spectrum sensing, modulation recognition and signal classification corresponding to the multi-band signal to be identified, and the identification process of the multi-band wireless signal is implemented through multi-task recognition.

[0128] The multi-band wireless signal recognition network based on multi-task learning in this embodiment can effectively capture the correlation and shared information between different tasks, thereby improving the recognition performance; through experimental verification, the multi-band wireless signal recognition network based on multi-task learning proposed in this embodiment has achieved significant performance advantages in spectrum sensing, modulation recognition and signal classification tasks. Compared with other traditional single-task learning methods, the multi-band wireless signal recognition network based on multi-task learning proposed in this embodiment performs better in recognition accuracy and has lower parameter volume and computational complexity. In general, the multi-task learning multi-band wireless signal recognition network provided in this embodiment provides an effective solution for the recognition of multi-band signals and has broad application prospects.

[0129] This embodiment achieves the following technical effects through the above technical solution:

[0130] This embodiment preprocesses the multi-band signal to be identified, and can perform spectrum sensing, modulation recognition and signal classification on the preprocessed multi-band signal to be identified based on a multi-band wireless signal recognition network with multi-task learning, and outputs the spectrum sensing, modulation recognition and signal classification results corresponding to the multi-band signal to be identified; this embodiment improves the single-task recognition performance by extracting shared features of different recognition tasks, and simultaneously completes the spectrum sensing, modulation recognition and signal classification tasks, and simultaneously completes the three tasks in one model, thereby reducing the complexity of the model.

[0131] Exemplary Devices

[0132] Based on the above embodiments, the present invention further provides a multi-band signal recognition device based on multi-task learning, comprising:

[0133] A multi-band signal preprocessing module, used for acquiring a multi-band signal to be identified and preprocessing the multi-band signal to be identified;

[0134] A multi-band signal recognition module, which is used for performing spectrum sensing, modulation recognition and signal classification on the pre-processed multi-band signal to be recognized by a multi-band wireless signal recognition network based on multi-task learning;

[0135] The classification result output module is used to output the results of spectrum sensing, modulation recognition and signal classification corresponding to the multi-band signal to be identified.

[0136] This embodiment achieves the following technical effects through the above technical solution:

[0137] This embodiment preprocesses the multi-band signal to be identified, and can perform spectrum sensing, modulation recognition and signal classification on the preprocessed multi-band signal to be identified based on a multi-band wireless signal recognition network with multi-task learning, and outputs the spectrum sensing, modulation recognition and signal classification results corresponding to the multi-band signal to be identified; this embodiment improves the single-task recognition performance by extracting shared features of different recognition tasks, and simultaneously completes the spectrum sensing, modulation recognition and signal classification tasks, and simultaneously completes the three tasks in one model, thereby reducing the complexity of the model.

[0138] Based on the above embodiment, the present invention further provides a terminal, whose principle block diagram can be as follows: Figure 6 shown.

[0139] The terminal includes: a processor, a memory, an interface, a display screen and a communication module connected through a system bus; wherein the processor of the terminal is used to provide computing and control capabilities; the memory of the terminal includes a storage medium and an internal memory; the storage medium stores an operating system and a computer program; the internal memory provides an environment for the operation of the operating system and the computer program in the storage medium; the interface is used to connect to external devices; the display screen is used to display corresponding information; and the communication module is used to communicate with a cloud server or other devices.

[0140] When the computer program is executed by a processor, it is used to implement the operation of a multi-band signal recognition method based on multi-task learning.

[0141] It can be understood by those skilled in the art that Figure 6 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the terminal to which the scheme of the present invention is applied. The specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0142] In one embodiment, a terminal is provided, comprising: a processor and a memory, wherein the memory stores a multi-band signal recognition program based on multi-task learning, and the multi-band signal recognition program based on multi-task learning is used to implement the operation of the multi-band signal recognition method based on multi-task learning as described above when executed by the processor.

[0143] In one embodiment, a storage medium is provided, wherein the storage medium stores a multi-band signal recognition program based on multi-task learning, and when the multi-band signal recognition program based on multi-task learning is executed by a processor, it is used to implement the operation of the multi-band signal recognition method based on multi-task learning as described above.

[0144] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and volatile memory.

[0145] In summary, the present invention provides a multi-band signal recognition method based on multi-task learning, including: obtaining a multi-band signal to be recognized, and preprocessing the multi-band signal to be recognized; a multi-band wireless signal recognition network based on multi-task learning performs spectrum sensing, modulation recognition and signal classification on the preprocessed multi-band signal to be recognized; outputting the results of spectrum sensing, modulation recognition and signal classification corresponding to the multi-band signal to be recognized; the present invention improves the single-task recognition performance by extracting shared features of different recognition tasks, and simultaneously completes spectrum sensing, modulation recognition and signal classification tasks, and simultaneously completes three tasks in one model, thereby reducing the complexity of the model.

[0146] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A multi-band signal recognition method based on multi-task learning, characterized in that: include: Acquire a multi-band signal to be identified, and pre-process the multi-band signal to be identified; The multi-band wireless signal recognition network based on multi-task learning performs spectrum sensing, modulation recognition and signal classification on the pre-processed multi-band signals to be identified; Output the results of spectrum sensing, modulation recognition and signal classification corresponding to the multi-band signal to be identified.

2. The multi-band signal recognition method based on multi-task learning according to claim 1, characterized in that: The multi-band wireless signal recognition network based on multi-task learning includes: a shared feature extraction network, a spectrum sensing classification head, a modulation recognition classification head and a signal classification classification head.

3. The multi-band signal recognition method based on multi-task learning according to claim 1, characterized in that: The adaptive loss function of the multi-band wireless signal recognition network based on multi-task learning is: in, and They represent the loss functions of spectrum sensing, modulation recognition and signal classification tasks respectively, σ1, σ2 and σ3 represent the homoscedastic uncertainty parameters corresponding to the three tasks respectively, Θ s represents the parameters of the shared feature extraction network, Θ T1 , Θ T2 , Θ T3 Represent the parameters of the three task classification heads respectively.

4. The multi-band signal recognition method based on multi-task learning according to claim 2, characterized in that: The multi-band wireless signal recognition network based on multi-task learning performs spectrum sensing, modulation recognition and signal classification on the pre-processed multi-band signal to be recognized, including: Extracting shared features from the preprocessed multi-band signal to be identified by using the shared feature extraction network; The shared features are respectively input into the spectrum sensing classification head, the modulation identification classification head and the signal classification classification head to obtain corresponding classification results.

5. The multi-band signal recognition method based on multi-task learning according to claim 4 is characterized in that: The shared feature extraction network includes: a multi-task feature extraction convolutional layer, a plurality of multi-task basic residual blocks and a plurality of multi-task special residual blocks; The step of extracting the shared features from the preprocessed multi-band signal to be identified by the shared feature extraction network includes: Inputting the preprocessed multi-band signal to be identified into the multi-task feature extraction convolution layer and a maximum pooling layer for feature extraction; The features extracted from the two layers are passed through a network formed by stacking multiple multi-task basic residual blocks and multiple multi-task special residual blocks, and finally through a global average pooling layer to calculate the shared features; wherein each of the multiple multi-task basic residual blocks and each multi-task special residual block is calculated as follows: F o =R(F i )+M(F i ); Among them, F i represents the input feature, F o represents the output feature, R(·) represents the residual function, M(·) represents the skip connection, and in the multi-task basic residual block, M(F i )=F i , in the multi-task special residual block, M(F i )=Conv1D(BN(F i )), where Conv1D represents the convolutional layer operation and BN represents batch normalization.

6. The multi-band signal recognition method based on multi-task learning according to claim 5, characterized in that: The step of inputting the preprocessed multi-band signal to be identified into the multi-task feature extraction convolution layer and a maximum pooling layer for feature extraction comprises: For the input feature F of the multi-task feature extraction convolution layer, the feature is obtained through the convolution feature extraction branch The channel attention map P obtained by the channel attention branch C (F), the spatial attention map P obtained by the spatial attention branch S (F); The characteristics Respectively with the channel attention map P C (F) and the spatial attention map P S (F) Perform element-by-element multiplication to obtain features and Features The characteristics feature And features Add them together to get the output multi-task extraction features.

7. The multi-band signal recognition method based on multi-task learning according to claim 5, characterized in that: The step of inputting the shared features into the spectrum sensing classification head, the modulation identification classification head, and the signal classification classification head respectively to obtain corresponding classification results includes: The shared features extracted by the shared feature extraction network from the preprocessed multi-band signal to be identified are input into the spectrum sensing classification head, a spectrum sensing result vector is output, and the spectrum sensing result vector is further processed to obtain a spectrum sensing result; The shared features extracted by the shared feature extraction network from the pre-processed multi-band signal to be identified are input into the modulation recognition classification head, a modulation recognition result vector is output, and the modulation recognition result vector is further processed to obtain a modulation recognition result; The shared features extracted by the shared feature extraction network from the preprocessed multi-band signal to be identified are input into the signal classification head, and a signal classification result vector is output. The signal classification result vector is further processed to obtain a signal classification result.

8. A multi-band signal recognition device based on multi-task learning, characterized in that: include: A multi-band signal preprocessing module, used for acquiring a multi-band signal to be identified and preprocessing the multi-band signal to be identified; A multi-band signal recognition module, which is used for performing spectrum sensing, modulation recognition and signal classification on the pre-processed multi-band signal to be recognized by a multi-band wireless signal recognition network based on multi-task learning; The classification result output module is used to output the results of spectrum sensing, modulation recognition and signal classification corresponding to the multi-band signal to be identified.

9. A terminal, characterized in that: include: A processor and a memory, wherein the memory stores a multi-band signal recognition program based on multi-task learning, and when the multi-band signal recognition program based on multi-task learning is executed by the processor, it is used to implement the operation of the multi-band signal recognition method based on multi-task learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a multi-band signal recognition program based on multi-task learning, and when the multi-band signal recognition program based on multi-task learning is executed by a processor, it is used to implement the operation of the multi-band signal recognition method based on multi-task learning as described in any one of claims 1-7.

Images