An Unsupervised Blind Recognition Method for Radio Frequency Fingerprint Based on Deep Contrastive Learning

Abstract

The present invention belongs to the field of Internet of Things security technology, and specifically relates to an unsupervised blind recognition method for radio frequency fingerprints based on deep contrast learning. The present invention can automatically complete the blind recognition of radio frequency fingerprints without the prior information of the source signal and the intervention of expert experience, greatly reducing the time and labor costs consumed by manual annotation. When faced with spoofing attacks where multiple transmitters disguise themselves as the same identity identifier or a single device simulates multiple identities in a Sybil attack, the present invention can perform unsupervised clustering analysis on radio frequency fingerprints, effectively identifying the number of source devices, thereby effectively detecting the existence of the above attacks. In addition, the model can also maintain a good blind recognition effect in an actual harsh and complex electromagnetic environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet of Things security, and specifically relates to an unsupervised blind recognition method for radio frequency fingerprint based on deep contrast learning. Background Art

[0002] With the rapid development of the Internet of Things (IoT) and the fifth-generation mobile communication technology (5G), the number and density of wireless devices are increasing rapidly. Some studies estimate that there are more than one million wireless devices interconnected per square kilometer in some urban areas. The communication security in wireless networks is usually based on user passwords, encryption algorithms, and various network communication protocols, which are vulnerable to network attacks such as eavesdropping, monitoring, and replay. However, the hardware-specific differences between devices are difficult to simulate. Even for wireless devices of the same model, various subtle differences will occur in the manufacturing and use processes of their internal components. By analyzing and extracting these inherent errors, a unique feature of the mobile device, namely radio frequency fingerprint (RFF), can be formed. Radio frequency fingerprint has characteristics such as uniqueness and long-term invariance, and can be used as the physical layer feature of the device. Moreover, radio frequency fingerprint recognition works directly on the physical layer, which can operate independently or be combined with and enhanced various traditional wireless network security mechanisms above the physical layer.

[0003] Currently, most of the research on radio frequency fingerprint recognition focuses on the basis of prior knowledge. It is necessary to collect fingerprints for each device or each type of device, store the obtained fingerprints in a fingerprint database, and use the fingerprint database for model training in a supervised learning manner; while there is very little research on the blind recognition of radio frequency fingerprints in the case of no labels and no prior information.

[0004] Although a large number of studies and applications have confirmed the reliability of supervised radio frequency fingerprint technology, in the real electromagnetic space environment, especially in fields such as intrusion detection and military confrontation, it is not always feasible to collect and analyze target signals in advance, manually label them, and enter the relevant labeled signals into the fingerprint database. In practice, it is often difficult to obtain the true information of the target device. In addition, in the field of wireless network security, when multiple transmitters with different fingerprints disguise themselves as the same identity for spoofing attacks, or a single device simulates multiple identities for Sybil attacks, unsupervised clustering analysis of radio frequency fingerprints can effectively identify the number of source devices, thereby effectively detecting the existence of the above attacks. In summary, the research on unsupervised radio frequency fingerprint blind recognition technology has important practical application significance. Secondly, there are a large number of noises and channel interferences in the actual electromagnetic environment, such as co-channel interference of multiple similar frequency devices and channel fading, which often have a great impact on the recognition performance. Summary of the Invention

[0005] In view of the above problems, the present invention provides an unsupervised blind recognition method for RF fingerprints based on deep contrast learning, which can automatically complete the blind recognition of RF fingerprints without the prior information of source signals and the intervention of expert experience, greatly reducing the time and labor costs consumed by manual annotation.

[0006] The technical solution of the present invention is as follows:

[0007] An unsupervised blind recognition method for RF fingerprints based on deep contrast learning, comprising the following steps:

[0008] S1. Perform data augmentation processing on the acquired I / Q signals of radio devices. The data augmentation processing method is to randomly select two methods from the following three data augmentation methods:

[0009] ① Additive Gaussian white noise augmentation method: Add additive Gaussian noise with random intensity to the I / Q signal samples;

[0010] ② Rayleigh channel fading augmentation method: Use a Rayleigh fading channel to augment the I / Q signal samples;

[0011] ③ Random dropout augmentation method: Randomly select a continuous region with a set length in the I / Q signal samples and set the I / Q data in this region to 0;

[0012] Define the sample x i After augmentation, the augmented sample is obtained and The superscripts a and b represent two of the three data augmentation methods. After augmenting the N acquired samples, 2N I / Q samples are obtained

[0013] S2. Construct a deep residual shrinkage neural network, including a convolutional layer and a residual shrinkage layer. The convolutional layer uses the non-linear function LeakyReLU as the activation function, and the residual shrinkage layer is composed of stacked residual shrinkage stacks. The residual shrinkage stack is composed of a convolutional layer, two residual shrinkage units, and a pooling layer connected in sequence; the residual shrinkage unit is a structure with a residual skip connection. First, the input feature map P is convolved twice to obtain x, and after global average pooling of x, a 1×1×C-dimensional feature vector is obtained, where C is the number of channels. The obtained feature vector is input into two fully connected layers. The first fully connected layer uses ReLU activation and performs BN operation, and the second fully connected layer uses the Sigmoid function for activation to obtain α:

[0014]

[0015] where z is the output of the second fully connected layer, and α is the learned scaling vector, and the values in the vector are between (0,1); multiply α by the feature vector to obtain the dynamic threshold λC :

[0016] λ C = a c ·GAP(C)

[0017] where GAP represents the global average pooling operation, and λ C represents the threshold corresponding to channel C, and α C represents the scaling factor corresponding to channel C. All dynamic thresholds are represented by λ. Inputting λ and x into the dynamic threshold function F(·) gives:

[0018]

[0019] Finally, adding F(x, λ) to P gives the output P′ of the residual shrinkage unit:

[0020] p' = P + F(x, λ)

[0021] Finally, the output of the residual shrinkage stack passes through batch normalization and the ReLU activation function to obtain the final sample features;

[0022] S3. Utilize the augmented samples to train the constructed deep residual shrinkage neural network. Design the loss function to include instance-level contrast loss and cluster-level contrast loss. Define the features obtained by passing two types of augmented samples through the deep residual shrinkage neural network as and The specific loss function design method is as follows:

[0023] Design of the instance-level contrast loss function: Form sample pairs from the samples obtained by the two augmentation methods, select to form positive sample pairs, and regard the remaining 2N - 2 pairs of samples as negative sample pairs. Use two fully connected layers g I (·) to map the sample feature h to a subspace:

[0024]

[0025] where is the data representation obtained by the augmented sample after passing through the representation learning module. Apply the instance-level contrast loss in the subspace z. The similarity between samples within any sample pair is measured by the cosine distance:

[0026]

[0027] where k1, k2 ∈ {a, b} and i, j ∈ [1, N]. For a certain augmented sample i of x its contrast loss is expressed as:

[0028]

[0029] Among them, τI represents the temperature coefficient, which is used to control the discrimination of the model for negative samples. The numerator part calculates the enhanced samples and x i Another enhanced sample of The similarity between them, and the denominator part calculates The similarity with all 2N samples. For both enhancements of x i , the contrast loss is calculated. Then the instance-level contrast loss of sample x i is:

[0030]

[0031] Cluster-level contrast loss function design: The sample features h are mapped to a subspace of dimension M through a fully connected layer, where M is the total number of categories. In the fully connected layer, the softmax function is used to perform a probability mapping transformation on the output, and the output vector y is regarded as the probability that the sample is assigned to each cluster. Define the matrix as the final output of a batch of samples {x1,…,x N} under the a data augmentation method. Similarly, Y b is the output under the b data augmentation method. The element in the i-th row and k-th column of Y a represents the probability that sample x is assigned to cluster k under the a data augmentation method. The k-th column of Y i is regarded as the representation of cluster k; Define the cluster-level contrast to use two fully connected layers g a (·): C (·):

[0032]

[0033] where represents the assignment probability vector of sample , that is, the i-th row of the Y a matrix; Define as the k-th column of the Y a matrix, that is, the representation of cluster k under the a data augmentation method. Similarly, the representation of cluster k under the b data augmentation method constitutes a positive sample pair with it At the same time and the other 2M - 2 pairs of samples are all negative sample pairs; The cosine distance is used to measure the similarity between cluster pairs:

[0034]

[0035] Among them, k1, k2 ∈ {a, b} and i, j ∈ [1, M]. The contrast loss of cluster i under the a data augmentation method is expressed as:

[0036]

[0037] Among them, τ C is the temperature coefficient, and the cluster-level contrast loss function is obtained as follows:

[0038]

[0039] where H(Y) is the entropy of the clustering assignment probability, which is used to prevent most instances from being assigned to the same cluster;

[0040] The final loss function is:

[0041]

[0042] Minimize the contrast loss through the gradient descent algorithm to update the parameters of the entire network, namely f(·), g I (·), g C (·), and finally obtain the trained network;

[0043] S4. Input the acquired signal data into the trained network for feature extraction to obtain h = f(x), and then calculate the cluster assignment c = arg max g C (h) to obtain the recognition result.

[0044] The beneficial effects of the present invention are as follows: Without the prior information of the source signal and the intervention of expert experience, the blind recognition of the RF fingerprint is automatically completed, greatly reducing the time and labor costs of manual annotation, and the model can maintain a good blind recognition effect in the actual harsh and complex electromagnetic environment. When the present invention faces spoofing attacks where multiple transmitters disguise themselves as the same identity identifier or a single device simulates multiple identities for a Sybil attack, it can perform unsupervised clustering analysis on the RF fingerprint, effectively identifying the number of source devices, thereby effectively detecting the existence of the above attacks. In addition, the model can also maintain a good blind recognition effect in the actual harsh and complex electromagnetic environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the overall network structure of the present invention.

[0046] Figure 2 It is an enhancement schematic diagram after QPSK modulation of the I / Q signal, where (a) is the original signal, (b) is the additive white Gaussian noise enhanced signal, (c) is the Rayleigh fading enhanced signal, and (d) is the randomly lost enhanced signal.

[0047] Figure 3 It is the overall network structure of the learning module.

[0048] Figure 4 It is the structure diagram of the residual shrinkage layer.

[0049] Figure 5 It is the structure diagram of the residual shrinkage unit. Specific implementation manners

[0050] The present invention will be described in detail below with reference to the accompanying drawings.

[0051] The overall network structure of the present invention is as Figure 1 shown, and includes the following steps:

[0052] 1. Use the method of data augmentation to construct positive and negative sample pairs for comparison.

[0053] Data augmentation is a common method to make the neural network robust to specific types of changes and avoid overfitting, that is, to use additional samples after augmented changes to expand the training set. In view of the characteristics of radio signals and for I / Q signal samples, the present invention proposes the following data augmentation methods:

[0054] (1) Additive white Gaussian noise.

[0055] Additive white Gaussian noise is a common wireless noise in the fields of communication, image, audio, etc. In the field of communication, additive white Gaussian noise is defined as white noise whose noise intensity follows a Gaussian distribution.

[0056] In the training of deep learning models, Gaussian noise can be added to the input samples to improve the robustness and generalization ability of the model. By adding noise to the input data, the model is forced to learn the features in the input samples that are robust to small changes, which can help it perform better on unknown data.

[0057] Noise is one of the main reasons for the decline in the accuracy of radio frequency fingerprint recognition. Existing radio frequency fingerprint recognition models often use high-quality data with high signal-to-noise ratio for training, resulting in a sharp decline in model performance as the signal-to-noise ratio of the samples decreases. Therefore, the present invention uses additive white Gaussian noise with random intensity as one of the enhancement means for I / Q samples, aiming to enable the model to learn the noise-independent fingerprints of the device and improve the robustness of the fingerprints.

[0058] (2) Rayleigh channel fading.

[0059] The wireless channel is another important reason for the decline in the accuracy of deep learning-based radio frequency fingerprint recognition. When wireless signals propagate in the actual electromagnetic space, they will be affected by various factors, including the scattering effect of the Earth's atmosphere on electromagnetic waves, and the interference of various ground obstacles such as urban buildings and vehicles on the straight-line propagation path of electromagnetic waves. The source signal will split into multiple electromagnetic wave components, and each component will reach the destination through different electromagnetic paths. Since the paths passed by each electromagnetic wave component are different, the time required to reach the target position will also have slight differences, so their phases will also be different. After each electromagnetic wave component reaches the receiver, the components with different phases will affect each other, and may be superimposed and enhanced or weakened in antiphase. In this way, there may be great changes in the waveform, amplitude, phase, etc. between the signal received by the receiver and the transmitted signal, resulting in serious distortion of the signal, that is, the multipath fading effect occurs.

[0060] To counteract the impact of multipath fading on fingerprints and enable the model to be applied to the actual environment, the present invention uses a Rayleigh fading channel to enhance the original I / Q signal samples. The Rayleigh fading channel is a modeling implementation of a multipath channel and is suitable for mobile communication devices. The Rayleigh channel model can be defined as:

[0061]

[0062] where τ k is the delay caused by the propagation of the electromagnetic wave component in the k-th path, T s is the sampling period, α k is the path index, and its definition is as follows:

[0063] α k = A + jB

[0064] where both A and B are random variables subject to a normal distribution and are independent of each other, with a mean of 0 and a variance of σ 2 . σ 2 The value of is:

[0065]

[0066] where T rms is the root mean square (RMS) delay spread of the channel. To sum up, the signal obtained after the modulated signal s(t) passes through Rayleigh fading is as follows:

[0067] r(t) = s(t)*(t, T) + n(t)

[0068] where * is the convolution operator and n(t) represents optional Gaussian noise.

[0069] (3) Random loss.

[0070] In the field of image processing, the Cutout operation selects a square region of a fixed size, fills it with all zeros, and allows the square region to be outside the image. This idea is similar to the Dropout method in neural networks, which reduces the overfitting of the model by ignoring a certain number of features. Inspired by the Cutout method in image enhancement, the present invention designs an enhancement method of random loss, randomly selects a certain length of continuous region in the original I / Q signal, and sets the I / Q data in this region to zero.

[0071] Appendix Figure 2 QPSK modulation is performed on the I / Q signal, and the effects of the above three enhancement methods are visually demonstrated through the constellation diagram. The present invention uses two data enhancement methods to enhance the sample x i to generate two enhanced instances and The present invention sequentially applies three enhancement methods, namely additive white Gaussian noise, Rayleigh channel fading, and random loss, to the I / Q samples. Each enhancement will be applied with a certain probability. The combination of the above three possible enhancement methods together constitutes a sample enhancement T(·).

[0072] 2. Then, the two enhanced samples and are input into the representation learning module. The goal of this module is to learn the sample feature representation that is independent of noise and channels, that is, the inherent radio frequency fingerprint of the device. The present invention designs a deep residual shrinkage neural network, which is shared between the two enhanced samples and is used to extract features from the enhanced samples and to generate two feature maps and

[0073] The overall network structure of the representation learning module is shown in Appendix Figure 3 as shown. The main body of the network includes a convolutional layer and a residual shrinkage layer. The instance-level contrast module and the cluster-level contrast module are mainly used for the construction of the loss function, and the backpropagation algorithm is used to optimize the network parameters.

[0074] In the preprocessing stage, the I / Q samples are divided with a sequence length of 1024, so the dimension of the input samples is 2×1024 and they have been normalized. Before the samples are input into the residual shrinkage module, it is necessary to first use the convolutional layer to perform preliminary feature extraction on the I / Q signal. In addition to the convolution operation, the convolutional layer usually also includes a pooling operation and the activation of the output. The specific parameters of the convolutional layer are shown in Table 1:

[0075] Table 1 Convolutional layer parameters

[0076]

[0077]

[0078] The present invention selects the non-linear function Leaky ReLU as the activation function. The expression of Leaky ReLU is:

[0079]

[0080] It can be seen from this that the Leaky ReLU function no longer completely discards the negative part, but gives a small gradient, so that the neuron will not be completely inactivated when outputting a negative number. Through experiments, alternating the Leaky ReLU function with the linear activation function Linear can greatly alleviate the difficulty of model training, and the situation where the model cannot converge no longer occurs.

[0081] The stacking of Residual Shrinkage Stacks (RSS) constitutes the residual shrinkage layer, and its structure is as shown in the appendix Figure 4 shown. RSS first uses a 1×1 convolutional kernel to perform convolution on the input feature map, and then concatenates two RSU and a max pooling layer. The Residual Shrinkage Unit (RSU) is a basic module of the residual shrinkage layer designed inspired by the channel attention model SENet and combined with the idea of the adaptive threshold function in wavelet denoising. The structure of RSU is as shown in the appendix Figure 5 shown.

[0082] Overall, RSU is a structure with a residual skip connection. An adaptive threshold shrinkage module is added on the basis of the ordinary residual unit. The direct mapping part of the residual skip connection structure directly uses the input information as the output, which can protect the integrity of the information to a certain extent, and make the network easier to optimize and be able to stack more layers. The adaptive threshold shrinkage module first performs global average pooling (GAP) on the feature map. GAP averages the feature map of each channel C. After the GAP operation, a 1×1×C-dimensional feature vector is obtained, which is input into two fully connected layers. The first fully connected layer uses ReLU activation and performs BN operation. The second fully connected layer uses the Sigmoid function for activation. The expression of the Sigmoid function is:

[0083]

[0084] where z is the output of the second fully connected layer, and α is the learned scaling vector, and the values in the vector are between (0,1). The finally obtained dynamic threshold is:

[0085] λ C = a c ·GAP(C)

[0086] Where λ is the soft threshold vector, λ C represents the threshold corresponding to channel C, α C represents the scaling factor corresponding to channel C. Then the output of the entire RSU can be expressed as:

[0087] p'=P+F(x,λ)

[0088] Where P is the input feature map, which is both input to the residual module and directly mapped to the output. x is the output of P after two convolutions, and P′ is the output feature map after RSU. F(·) represents the dynamic threshold function, which is expressed as:

[0089]

[0090] The representation learning model proposed in the present invention learns thresholds through deep neural networks, effectively applies the idea of adaptive threshold denoising, can suppress noise in I / Q signals, and extract the inherent fingerprint of the device.

[0091] 3. Finally, the model will calculate the contrast loss of the sample and perform back propagation based on it in turn to optimize the overall parameters of the network. The overall contrast loss of this model includes instance-level contrast loss and cluster-level contrast loss. The instance-level contrast loss mainly compares samples, and its goal is to narrow the distance between positive sample pairs and separate positive samples from negative samples. The cluster-level contrast loss is specially designed for clustering tasks. Its comparison target is clusters. The same cluster under different data enhancements is regarded as a positive sample, and different clusters are regarded as negative samples. This constitutes a clustering loss, which enables the model to be optimized for clustering tasks and obtain better clustering performance.

[0092] First, the instance-level contrast module is described, whose goal is to calculate the instance-level contrast loss. For a batch of data of size N, the present invention first calculates each instance x i Perform two different data augmentations, a and b, to generate a total of 2N I / Q samples For samples The remaining 2N-1 samples can form sample pairs with it, and the model selects x i Another enhanced sample of Form a positive sample pair with it The remaining 2N-2 pairs of samples are considered negative pairs.

[0093] In order to reduce the information loss caused by contrast loss, the instance-level contrast module does not directly perform contrastive learning on the feature matrix, but stacks two fully connected layers g I (·), the fully connected layer can map the feature map h to a subspace:

[0094]

[0095] Among them, is the enhanced sample which is the data representation obtained through the representation learning module. Instance-level contrastive loss is applied in the subspace z. For any pair of samples, the similarity between samples can be measured by cosine distance, that is:

[0096]

[0097] where k1, k2 ∈ {a, b} and i, j ∈ [1, N]. For an enhanced sample i of x its contrastive loss can be expressed as:

[0098]

[0099] where τ I represents the temperature coefficient, which is used to control the discrimination ability of the model for negative samples. In the above formula, the numerator part calculates the similarity between the enhanced sample and another enhanced sample i of x , and the denominator part calculates the similarity between and all 2N samples. The contrastive loss is calculated for both enhancements of x i , so the instance-level contrastive loss of sample x i can be expressed as:

[0100]

[0101] Then, the cluster-level contrastive module is described, and its goal is to calculate the cluster-level contrastive loss. In the cluster-level contrastive module, the model of the present invention maps the sample representation h to a subspace with dimension M through a fully connected layer, where M is the total number of categories. In the fully connected layer, the softmax function is used to perform a probability mapping transformation on the output, and the output vector y can be regarded as the probability that the sample is assigned to each cluster. Define the matrix as the final output of a batch of samples {x1, …, x n} under enhancement a (Y b is the output of this batch under enhancement b). The element a in the i-th row and k-th column of Y represents the probability that sample x i is assigned to cluster k under enhancement a, and the k-th column of Y a can be regarded as the representation of cluster k.

[0102] Since each sample belongs to only one cluster, ideally Y aThe rows of are one - hot vectors, so all columns should be different from each other. Similar to the fully - connected layer g used in the instance - level contrast module I (·), the cluster - level contrast module uses two other fully - connected layers g C (·), and this operation can be expressed as:

[0103]

[0104] where represents the assignment probability vector of sample , that is, the i - th row of the Y a matrix. Define as the k - th column of the Y a matrix, that is, the representation of cluster k under data augmentation a. Then the representation of cluster k under data augmentation b forms a positive sample pair with it At the same time and the other 2M - 2 pairs of samples are all negative sample pairs. Similarly, the model uses cosine distance to measure the similarity between cluster pairs, that is:

[0105]

[0106] where k1, k2 ∈ {a, b} and i, j ∈ [1, M]. The contrast loss of cluster i under augmentation a can be expressed as:

[0107]

[0108] Then the final cluster - level contrast loss can be expressed as:

[0109]

[0110] where H(Y) is the entropy of the clustering assignment probability, which can prevent most instances from being assigned to the same cluster.

[0111] The training of the instance - level contrast module and the cluster - level contrast module is an end - to - end process, and they are optimized simultaneously. The overall contrast loss of the model is the sum of the losses of the two parts:

[0112]

[0113] The detailed process of the training stage is as follows:

[0114] Input: Dataset D, total number of training epochs E(epochs), batch size N of each batch of data,

[0115] number of classes M, temperature coefficient τ I and τ C .

[0116] ​Output: Clustering assignment results.

[0117] Training stage:

[0118] S1: Repeat steps S2 to S8 for E times.

[0119] S2: Select from the dataset D

[0120] S3: Randomly select two data augmentations T a , T b ;

[0121] S4: Generate instance-level and cluster-level representations of the augmented samples:

[0122]

[0123]

[0124]

[0125] S5: Calculate the instance-level contrastive loss by Equation (1)

[0126] S6: Calculate the instance-level contrastive loss by Equation (2)

[0127] S7: Calculate the total contrastive loss by Equation (3)

[0128] S8: Minimize the contrastive loss through the gradient descent algorithm Update the parameters of the entire network, i.e., f(·), g I (·), g C (·);

[0129] S9: End of loop, training completed;

[0130] The recognition stage is:

[0131] For the obtained signal x;

[0132] Extract features through the representation learning module: h = f(x);

[0133] Calculate the cluster assignment c: c = arg max g C (h);

[0134] Thus, the blind recognition of radio frequency fingerprints is completed.

Claims

1. An unsupervised blind recognition method for RF fingerprints based on deep contrast learning, characterized in that, It includes the following steps: S1. Perform data augmentation processing on the obtained I / Q signals of the radio device. The data augmentation processing method is to randomly select two methods from the following three data augmentation methods: ① Additive white Gaussian noise augmentation method: Add additive Gaussian noise with random intensity to the I / Q signal samples; ② Rayleigh channel fading augmentation method: Use a Rayleigh fading channel to augment the I / Q signal samples; ③ Random loss augmentation method: Randomly select a continuous region with a set length in the I / Q signal samples and set the I / Q data in this region to 0; Define the sample x i After enhancement, the enhanced sample is obtained and The superscripts a and b represent two of the three data augmentation methods. After augmenting the N obtained samples, 2N I / Q samples are obtained S2. Construct a deep residual shrinkage neural network, including a convolutional layer and a residual shrinkage layer. The convolutional layer uses the non-linear function LeakyReLU as the activation function. The residual shrinkage layer is stacked by residual shrinkage stacks. The residual shrinkage stack is composed of a convolutional layer, two residual shrinkage units, and a pooling layer connected in sequence. The residual shrinkage unit is a structure with a residual skip connection. First, the input feature map P is convolved twice to obtain x. After global average pooling of x, a 1×1×C-dimensional feature vector is obtained, where C is the number of channels. The obtained feature vector is input into two fully connected layers. The first fully connected layer uses ReLU activation and performs BN operation. The second fully connected layer uses the Sigmoid function for activation to obtain α: where z is the output of the second fully connected layer, and α is the learned scaling vector, and the values in the vector are between (0,1); Multiply α by the eigenvector to obtain the dynamic threshold λ C : λ C = α C ·GAP(C) where GAP represents the global average pooling operation, λ C represents the threshold corresponding to channel C, and α C represents the scaling factor corresponding to channel C. Denote all dynamic thresholds by λ, and input λ and x into the dynamic threshold function F(·) to obtain: Finally, add F(x,λ) and P to obtain the output P′ of the residual shrinkage unit: P ′ = P + F(x, λ) Finally, the output of the residual shrinkage stack passes through batch normalization and the ReLU activation function to obtain the final sample features; S3. Utilize enhanced samples Train the constructed deep residual shrinkage neural network. Design the loss function to include instance-level contrast loss and cluster-level contrast loss. Define the features obtained by passing two types of enhanced samples through the deep residual shrinkage neural network as and The specific method for designing the loss function is as follows: Instance-level contrastive loss function design: Samples obtained by two augmentation methods are formed into sample pairs, and are composed into positive sample pairs, and the remaining 2N - 2 pairs of samples are all regarded as negative sample pairs. Two fully connected layers g I (·) are used to map the sample feature h to a subspace: Among them is the enhanced sample The data representation obtained through the representation learning module, where an instance-level contrastive loss is applied in the subspace z. For any pair of samples, the similarity between samples is measured by the cosine distance: where \(k_1,k_2\in\{a,b\}\) and \(i,j\in[1,N]\), for a certain augmented sample of \(x\) i its contrastive loss is expressed as: for a certain augmented sample Among them, τ I represents the temperature coefficient, which is used to control the discrimination ability of the model for negative samples. The numerator calculates the similarity between the augmented sample and another augmented sample of x i . The denominator calculates the similarity between and all 2N samples. The contrast loss is calculated for both augmentations of x . Then the instance-level contrast loss of sample x i is as follows: i ​ Cluster-level contrast loss function design: The sample feature h is mapped to a subspace of dimension M through a fully connected layer, where M is the total number of categories. In the fully connected layer, the softmax function is used to perform probability mapping transformation on the output. The output vector y is regarded as the probability of the sample being assigned to each cluster. The matrix is defined For a batch of samples {x1,…,x N The final output of the a data enhancement method. Similarly, Y b is the output under the b data enhancement method, Y a The element in row i and column k of Represents a data enhancement method, sample x i The probability of being assigned to cluster k, Y a The kth column of is regarded as the representation of cluster k; the cluster-level comparison is defined using two fully connected layers g C (·): where represents the distribution probability vector of the sample , that is, the i-th row of the Y a matrix; define as the k-th column of the Y a matrix, that is, the representation of cluster k under the a data augmentation method. Similarly, the representation of cluster k under the b data augmentation method then forms a positive sample pair with it At the same time and the other 2M - 2 pairs of samples are all negative sample pairs; use the cosine distance to measure the similarity between cluster pairs: where k1,k2∈{a,b} and i,j∈[1,M], and the contrast loss of cluster i under the a data augmentation method is expressed as: Among them, τ C is the temperature coefficient, and the cluster-level contrast loss function is obtained as follows: where H(Y) is the entropy of the clustering assignment probability, which is used to avoid most instances being assigned to the same cluster; The final loss function is: Minimize the contrastive loss through the gradient descent algorithm Update the parameters of the entire network, i.e., f(·), g I (·), g C (·), and finally obtain the trained network; S4. Input the acquired signal data into the trained network for feature extraction to obtain h = f(x), and then calculate the cluster assignment c = arg max g C (h) to obtain the recognition result.

Images