Unknown target classification method based on zero-order learning

Abstract

The invention discloses an unknown target classification method based on zero-order learning. The method comprises the following steps: acquiring known unknown signal data; a semantic space is constructed for the signals through an encoder F, a classifier C and a decoder d; distance measurement is found for the trained semantic space, a discrimination model is established, the discrimination model is adopted to distinguish signal categories of the test signals, and the signal categories comprise n known categories and nR unknown categories. The method can distinguish how many unknown classes exist in the test set, not merely marking the test sample as unknown, and can solve the problem that the existing unknown signal identification method has the defects of generalization ability and resource waste caused by global use of Mahalanobis distance.

Description

Technical Field

[0001] The present invention relates to the field of signal recognition, and in particular to an unknown target classification method based on zero-shot learning. Background Art

[0002] Signal recognition is one of the important and challenging tasks in the field of signal processing and communication. It is often a common situation that some signal classes have no accessible training data to perform the recognition task. Therefore, zero-shot learning, as a learning method widely used in the field of image processing, is also very important in signal recognition. Unfortunately, there is little research on this area due to the unexplainable signal semantics.

[0003] "SR2CNN: Zero-Shot Learning for Signal Recognition" published in "Institute of Electrical and Electronics Engineers" (2021: 2316-2329) can not only identify known category signals, but also unknown category signals by performing distance measurement on the extracted semantic attribute features. However, the global use of Mahalanobis distance wastes a lot of computing time, resulting in a waste of resources. The invention with publication number CN114839613A discloses a method for unknown signal recognition based on zero-shot learning, which proposes improvements on SR2CNN, deepens the network depth, and supplements the recognition effect under low signal-to-noise ratio, but still does not deviate from the traditional convolutional neural network architecture, and the generalization performance is not strong. At the same time, a commonly used loss function is selected, and the loss function is not optimized according to specific tasks. Summary of the invention

[0004] The purpose of the present invention is to propose an unknown target classification method based on zero-shot learning, which can solve the problems of the shortcomings of the existing unknown signal recognition methods in terms of generalization ability and the waste of resources caused by the global use of Mahalanobis distance.

[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0006] A method for classifying unknown targets based on zero-shot learning, the method comprising the following steps:

[0007] S1, obtain known and unknown signal data;

[0008] S2, constructs a semantic space for the signal through an encoder F, a classifier C, and a decoder d; wherein the encoder F is modeled by the transformer encoder architecture, which projects the input signal into a latent semantic space representation; the classifier C is modeled by a fully connected neural network, which takes the latent semantic space representation as input and determines the label of the data; the decoder D is modeled by the transformer decoder, which is used to generate a reconstructed signal, and it is expected that the reconstructed signal is close to the input signal;

[0009] S3, find the distance metric for the semantic space of training, establish the discriminant model, and use the discriminant model to distinguish the signal categories of the test signal. The signal categories include n known classes and n R unknown class; the differentiation process specifically includes the following sub-steps:

[0010] For each known class k, calculate the corresponding semantic center vector S k ;

[0011] For each known class k, calculate the test signal The measurement vector With S k difference;

[0012] The first threshold θ1 is compared with the given minimum distance d1 to distinguish between known and unknown classes:

[0013]

[0014] Among them, c1 is the control parameter, called the discrimination coefficient; is the average distance of known categories; S is the set of known semantic center vectors; express prediction; if d1<Θ1, then otherwise

[0015] If the test signal Belongs to a known class, and the label is obtained by the following formula

[0016] If the test signal Belongs to the unknown class, use R to represent the recorded unknown class, and define S R is the set of semantic center vectors of R; add a new signal label R1 in R, let is the semantic center vector The unknown signal Save in set middle, Compare the second threshold θ2 with the minimum distance d2:

[0017]

[0018] Where λ2 is used to balance the two distances d1 and d med ;d med The candidate set is composed of k' centers with the smallest distance from F(I) K 'The median of the distance calculation;

[0019] Assume n R is the number of signal labels recorded in R. If d2>Θ2, a new signal label is added to R. set up If d2≤Θ2, the label is obtained using the following formula And the signal Save in collection middle.

[0020] Furthermore, in step S2, the encoder F consists of 6 encoder layers, each of which includes a self-attention mechanism layer and a feedforward neural network layer; the encoder F is represented by a mapping from the input space X to the latent semantic space Z; the central loss of the encoder F is expressed as:

[0021]

[0022] Where N is the batch size, x i ∈X,y i For x i The label, z i =F(x i )∈Z; Represents class y in Z i The semantic center vector of With class y i The semantic features are updated accordingly.

[0023] Furthermore, in step S2, the classifier C uses the cross entropy loss L ce Control the error of the classifier C and assign adaptive weights w to different categories or samples i ; Cross entropy loss L of classifier C ce for:

[0024]

[0025] Where C(F(x i )) is x i The predicted value of .

[0026] Furthermore, in step S2, the decoder D uses reconstruction loss to evaluate the difference between the original signal data and the reconstructed signal data.

[0027]

[0028] Where D(F(x i )) is the signal x i reconstruction.

[0029] Furthermore, the total loss function of the semantic space is:

[0030] L t =L ce +λ ct L ct +λ r L r .

[0031] Furthermore, the first threshold θ1 is dynamically updated using the following formula:

[0032] Θ1=β×Θ1 prev +(1-β)×θ1 new

[0033] where β is the smoothing factor.

[0034] Furthermore, the second threshold is dynamically updated using the following formula:

[0035]

[0036] Where δ is the learning rate.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] First, the unknown target classification method based on zero-shot learning of the present invention introduces regularization for the center loss to prevent overfitting, avoid excessive deviation of the center point, and enhance the generalization ability of the model; for the cross entropy loss, adaptive weights are given to different categories or samples, the weights can be dynamically adjusted according to the difficulty of the sample or the category imbalance, and adversarial training is added to improve the robustness of the model to noise.

[0039] Second, the unknown target classification method based on zero-shot learning of the present invention introduces a new discrimination criterion, reduces the number of samples for median calculation through pre-screening, reduces computational complexity, enhances adaptability, and improves classification efficiency.

[0040] Third, the unknown target classification method based on zero-shot learning of the present invention is different from the traditional open set recognition method which only divides the test samples into n+1 classes, but divides the test samples into n+n R Class, where n R The number of unknown classes identified for the discriminator, so that it can distinguish how many unknown classes are in the test set, rather than just labeling the test samples as unknown. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a schematic diagram of an unknown target classification method based on zero-shot learning according to an embodiment of the present invention;

[0042] Figure 2 This is a diagram of the transformer model architecture of an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0044] The present invention discloses a method for classifying unknown targets based on zero-shot learning, the method comprising the following steps:

[0045] S1, obtain known unknown signal data.

[0046] S2, constructs a semantic space for the signal through an encoder F, a classifier C and a decoder d; wherein the encoder F is modeled by the transformer encoder architecture, which projects the input signal into a latent semantic space representation; the classifier C is modeled by a fully connected neural network, which takes the latent semantic space representation as input and determines the label of the data; the decoder D is modeled by the transformer decoder, which is used to generate a reconstructed signal, and it is expected that the reconstructed signal is close to the input signal.

[0047] S3, find the distance metric for the semantic space of training, establish the discriminant model, and use the discriminant model to distinguish the signal categories of the test signal. The signal categories include n known classes and n R An unknown class.

[0048] Signals are a special type of data, which is very different from images. Although the semantic attributes of images are given from the perspective of visual information, it is almost impossible to extract the semantic features of signals without relying on any calculations. Therefore, a natural way to automatically extract the semantic information of signal data is to use a feature extractor network. Considering the unique characteristics of the signal, the input shape of the feature extractor network should be a 1-dimensional signal instead of a square matrix. In our scheme, the encoder F consists of 6 encoder layers, each of which contains a self-attention mechanism layer and a feedforward neural network layer such as Figure 2 The encoder F is represented by a mapping from the input space X to the latent semantic space Z. In order to minimize the intra-class variation in the space Z while keeping the semantic features between classes well separated, a center loss is used. Let x i ∈X,y i For x i The label of z i =F(xi )∈Z. In order to prevent the center point from being overly biased towards a certain type of sample or individual sample, an L2 regularization constraint is added to the center point to prevent overfitting, avoid excessive deviation of the center point, and enhance the generalization ability of the model. Assuming the batch size is N, the center loss is expressed as:

[0049]

[0050] in Represents class y in Z i The semantic center vector of Need to follow class y i The semantic features are updated based on the changes in the training data set. Ideally, the entire training dataset should be considered, and the features of each class need to be averaged in each iteration. In practice, it can be updated based on For each batch Learning rate update, where α is the update parameter;

[0051] and Calculated by the following formula:

[0052]

[0053] Among them, if the condition inside () is met, then δ(·)=1, otherwise δ(·)=0.

[0054] The classifier C will distinguish the labels of samples based on semantic features. It consists of several fully connected layers. Further use the cross entropy loss L ce Control the error of the classifier C and assign adaptive weights w to different categories or samples i , the weights can be dynamically adjusted according to the difficulty or class imbalance of the sample, and adversarial training is added to improve the robustness of the model to noise. A small perturbation δ is applied to the input, requiring the model to still give the same prediction results under the perturbed input. The loss is defined as:

[0055]

[0056] Where C(F(x i )) is x i The predicted value of .

[0057] In order to retain the effective semantic information in Z, this embodiment uses an autoencoder. Specifically, a decoder D is used to reconstruct X from Z. The decoder D is composed of a transformer decoder with the following structure: Figure 2 In addition, the reconstruction loss is used to evaluate the difference between the original signal data and the reconstructed signal data. The reconstruction loss is expressed as:

[0058]

[0059] Where D(F(x i )) is the signal x i The more complete the reconstructed signal is, the more effective information Z carries. Therefore, the autoencoder greatly helps the model generate appropriate semantic features.

[0060] The total loss function combines the cross entropy loss, center loss, and reconstruction loss as:

[0061] L t =L ce +λ ct L ct +λ r L r

[0062] The weight λ ct and λ r To balance these three loss functions. The present invention carefully designs the total loss function. The cross entropy loss is used to learn information from the label. The center loss minimizes the change of the semantic space within the class while keeping the semantic features between classes well separated, which also helps to separate the unknown classes. The reconstruction loss enables the model to learn more information about the signal data because the data is well reconstructed.

[0063] The total loss function L is summarized in Algorithm 1 t The entire learning process, where θ F ,θ C ,θ D They represent the model parameters of encoder F, classifier C and decoder D respectively.

[0064]

[0065]

[0066] The encoder F is the tail end of the proposed framework, but it is the core. It distinguishes known and unknown classes based on the latent semantic space Z. For each known class k, the encoder F extracts and calculates the corresponding semantic center vector S k for:

[0067]

[0068] Where m is the number of all training samples. get When each k measurement vector With S k The difference between F and S k The generalized distance is defined as follows:

[0069]

[0070] Among them A k is the transformation matrix associated with class k, Represents the matrix A k The inverse of k When A is the covariance matrix ∑ of the semantic features of k-type signals, d(·,·) is called the Mahalanobis distance. k When A is the unit matrix I, d(·,·) is simplified to the Euclidean distance. k It can also be Λ and σ 2 I, where Λ is composed of ∑ and The diagonal matrix composed of the diagonal elements of S k The dimension of A k = Λ and A k =σ 2 The distances corresponding to I are called the second distance and the third distance. Note that when applying the Mahalanobis distance, the second distance, and the third distance, the covariance matrix of each known class needs to be calculated in advance.

[0071] With the above distance metric, we can build a discriminant model. The workflow of the discriminant model is divided into two steps: the first step is to distinguish known classes from unknown classes; the second step is to distinguish which known classes or unknown classes the test signal belongs to. The first step is completed by comparing the first threshold θ1 with the given minimum distance d1.

[0072]

[0073] Where S is a set of known semantic center vectors. express If d1<Θ1, then otherwise Since the center loss is used in training, it is assumed that the semantic features of k-type signals follow a multivariate Gaussian distribution. Based on historical classification data, the average distance of known categories is calculated. And standard deviation σk, set the first threshold Θ1 as follows:

[0074]

[0075] Where c1 is the control parameter, called the discrimination coefficient.

[0076] When the distribution of input data changes, the model may not be able to adapt to the fixed Θ1. In some tasks, the statistical characteristics of the input data itself may change over time or location. For example, in a multi-sensor system, different sensors may be affected by different interference sources, resulting in an increase or decrease in measurement error. In this case, Θ1 needs to be dynamically adjusted to better adapt to the new environmental conditions. If the comparative evaluation finds that the performance of the model under the current setting is not ideal (for example, the misclassification rate is too high, the prediction is inaccurate, etc.), dynamically updating Θ1 can be used as an adaptive optimization method to help the model improve its accuracy.

[0077] To this end, the following formula is used to dynamically update the first threshold θ1:

[0078] Θ1=β×Θ1 prev +(1-β)×θ1 new

[0079] where β is the smoothing factor.

[0080] In the loss function of this embodiment, a central loss component is provided, the goal of which is to minimize the central loss component relative to the semantic layer.

[0081]

[0082] Due to the monotonicity of the exponential function, we have:

[0083]

[0084] Where t is the dimension of Gaussian distribution and I is the unit matrix. and The above equation can be equivalently written as:

[0085]

[0086] in This indicates that the output of the semantic layer is likely to follow a Gaussian distribution.

[0087] The second step is more complicated. If the test signal belongs to a known class, its label It can be obtained by the following formula:

[0088]

[0089] The main difficulty lies in dealing with In the first step, the case is classified as unknown. Specifically, R represents the recorded unknown class and defines S R is the set of semantic center vectors of R. When , add a new signal label R1 in R, set is the semantic center vector The unknown signal Save in set In when When , the second threshold θ2 is compared with the minimum distance d2, and the minimum distance d2 is calculated by the following formula:

[0090]

[0091] According to F(x) and S k The distance between d1 and d2 can be a good choice for the second threshold value θ2. d1 is the minimum distance used for the first time in the test for selecting the second threshold value θ2. In fact, this embodiment tests a set of choices for the second threshold value θ2 and numerically finds that when the second threshold value θ2 is set between d1 and d med The unknown class can usually be correctly identified when d med is F(x) and each S k The median distance between .

[0092] In order to reduce complexity, we do not directly calculate all d(F(I),S k ), using the following optimization strategy:

[0093] 1. Pre-screening of candidate centers:

[0094] According to the simple Euclidean distance, the k' centers with the smallest distance to F(I) are selected as the candidate set K'.

[0095] Usually |K'|=ceil(α·K),α∈(0,1).

[0096] Compute the approximate median distance:

[0097] 2. For the candidate set K ' to calculate the median distance:

[0098] d med ≈median({d(F(I),S k' )} k'∈K' )

[0099] Where p is the empirical coefficient.

[0100] Therefore, the second threshold θ2 is finally set to

[0101]

[0102] In practical applications, as the signal changes or the training data differs, the static λ2 may not be able to adapt to these changes. Therefore, dynamically adjusting λ2 can enhance the flexibility of the model, allowing the model to adjust its judgment criteria for signal categories according to actual conditions. This adjustment mechanism is often used to balance the ability to identify known categories and distinguish unknown signals in a classification system, especially when the data environment changes or system requirements change.

[0103] Where λ2 is used to balance the two distances d1 and d med , the dynamic adjustment method is as follows:

[0104]

[0105] Where δ is the learning rate.

[0106] Next, let n R is the number of signal labels recorded in R. If d2>Θ2, a new signal label is added to R. set up This embodiment does not have any R The value of (the size of the set R) imposes no prior restrictions, that is, the model of this embodiment can never know the number of unknown classes to be judged. If d2≤θ2, this embodiment is set as follows:

[0107]

[0108] And the signal Save in In. Accordingly, Calculated by the following formula:

[0109]

[0110] in Representing a collection Therefore, as the number of unknown signal predictions increases, the discriminant model will be refined way to gradually improve itself.

[0111] The workflow of the discriminator is summarized in Algorithm 2.

[0112]

[0113]

[0114] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0115] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0116] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0117] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0118] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0119] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for classifying unknown targets based on zero-shot learning, characterized in that: The method comprises the following steps: S1, obtain known and unknown signal data; S2, constructs a semantic space for the signal through an encoder F, a classifier C, and a decoder d; wherein the encoder F is modeled by the transformer encoder architecture, which projects the input signal into a latent semantic space representation; the classifier C is modeled by a fully connected neural network, which takes the latent semantic space representation as input and determines the label of the data; the decoder D is modeled by the transformer decoder, which is used to generate a reconstructed signal, and it is expected that the reconstructed signal is close to the input signal; S3, find the distance metric for the semantic space of training, establish the discriminant model, and use the discriminant model to distinguish the signal categories of the test signal. The signal categories include n known classes and n R unknown class; the differentiation process specifically includes the following sub-steps: For each known class k, calculate the corresponding semantic center vector S k ; For each known class k, calculate the test signal The measurement vector With S k difference; The first threshold θ1 is compared with the given minimum distance d1 to distinguish between known and unknown classes: Among them, c1 is the control parameter, called the discriminant coefficient; d() represents the generalized distance function; is the average distance of known categories; F() is the output function of encoder F; S is the set of known semantic center vectors, σ k is the standard deviation of the known class k; if d1<Θ1, then otherwise K is the known class set, and U is the unknown class set; If the test signal Belongs to a known class, test signal Label for: If the test signal Belongs to the unknown class, use R to represent the recorded unknown class, and define S R is the set of semantic center vectors of R; when When , add a new signal label R1 in R, set is the semantic center vector The unknown signal Save in set middle, when When , the second threshold θ2 is compared with the minimum distance d2: Where R u is the semantic center vector of the unknown category, and λ2 is used to balance the two distances d1 and d med , d med Calculate the median of the distance of the candidate set K' consisting of the k' centers with the smallest distance to F(I); Assume n R is the number of signal labels recorded in R. If d2>Θ2, a new signal label is added to R. set up If d2≤θ2, the label is obtained using the following formula And the signal Save in collection middle.

2. The unknown target classification method based on zero-shot learning according to claim 1, characterized in that: In step S2, the encoder F consists of 6 encoder layers, each of which contains a self-attention mechanism layer and a feedforward neural network layer; the encoder F is represented by a mapping from the input space X to the latent semantic space Z; the central loss of the encoder F is expressed as: Where N is the batch size, x i ∈X,y i For x i The label, z i =F(x i )∈Z; Represents class y in Z i The semantic center vector of With class y i The semantic features are updated according to the changes, M is the total number of parameters in the model, c j is the jth parameter, and λ is the regularization coefficient, which is used to control the strength of the regularization term.

3. The unknown target classification method based on zero-shot learning according to claim 1, characterized in that: In step S2, the classifier C uses the cross entropy loss L ce Control the error of the classifier C and assign adaptive weights w to different categories or samples i ; Cross entropy loss L of classifier C ce for: Where C(F(x i )) is x i The predicted value of , α is the constraint parameter, and δ1 is the adversarial disturbance.

4. The unknown target classification method based on zero-shot learning according to claim 1, characterized in that: In step S2, the decoder D uses the reconstruction loss to evaluate the difference between the original signal data and the reconstructed signal data: Where D(F(x i )) is the signal x i reconstruction.

5. The unknown target classification method based on zero-shot learning according to claim 1, characterized in that: The total loss function of the semantic space is: L t =L ce +λ ct L ct +λ r L r where λ ct ,λ r is the weight ratio of the corresponding loss function.

6. The unknown target classification method based on zero-shot learning according to claim 1, characterized in that: The first threshold θ1 is dynamically updated using the following formula: Θ1=β×Θ1 prev +(1-β)×Θ1 new where β is the smoothing factor, Θ1 prev is the value of Θ1 at the previous moment, Θ1 new Calculates a new value for the current.

7. The unknown target classification method based on zero-shot learning according to claim 1, characterized in that: The second threshold is dynamically updated using the following formula: Where δ2 is the learning rate, λ max To calculate the upper bound.

Images