An incremental learning method for signal recognition with decoupling of feature extraction and classification

Abstract

The present invention relates to a signal recognition incremental learning method in which feature extraction and classification are decoupled. The method is implemented based on a complex neural network, comprises a decoupled feature extractor and a classifier, and can obtain signal samples and divide them into baseline stage categories and incremental stage categories according to the stage to which the categories belong. First, the decoupled feature extractor is optimized by weighted summation of center loss, cross entropy loss and reconstruction loss in the baseline stage. Then, in the incremental stage, the feature extractor parameters are frozen, the feature representation of the incremental stage samples is obtained, and the cross entropy loss is calculated to optimize the incremental classifier until convergence. Finally, the node aggregation idea is used to obtain the fused classifier after incremental update. The signal recognition incremental learning method based on the decoupling of feature extraction and classification described in the present invention can effectively reduce the forgetting of previous classification capabilities and expand its classification and recognition capabilities at a relatively low cost.

Description

Technical Field

[0001] The present invention relates to the field of wireless signal recognition, and in particular to a signal recognition method and device based on deep class incremental learning. Background Art

[0002] Radio wave modulated signals are widely used in real life. Receivers identify and classify received radio wave modulated signals and demodulate the underlying signal. Wireless signal transmission undoubtedly significantly increases the rate of information transmission. However, both in practical applications and theoretical research, technical and resource limitations often prevent the provision of training data for some categories during model training. Deep learning network models rely on complete and sufficient training data. Their recognition domains only remain consistent with the data used during training, making them unable to classify newly added categories. This is a problem stemming from the "closed set assumption" underlying most machine learning models. Training and updating neural network models solely on newly added data leads to the so-called "catastrophic forgetting" problem, whereby the model rapidly forgets knowledge of old categories and loses its classification ability. In engineering, after modifying the model's output layer structure, the model is often retrained on a newly constructed dataset containing samples from both old and new categories. This incurs significant human, time, and computing resource expenditures.

[0003] After searching the prior art, it was found that the inventor Qu Zhiyu [Invention Patent Application Publication Number: CN113111774A] converted the radar signal modulation mode into a two-dimensional time-frequency image, and then used the incremental learning technology in the field of image recognition to increase the amount of training samples, especially the use of unlabeled samples. However, this method is only for sample incremental tasks in signal recognition, and is not suitable for category incremental tasks. Yinjun Wang et al. proposed a category incremental learning framework for bearing fault signal data in "An Efficient Incremental Learning of Bearing Fault Imbalanced Data Set via Filter StyleGAN" IEEE Transactions on Instrumentation and Measurement, 2021, 70: 1-10. (i.e., online machine learning of bearing fault imbalanced datasets based on Filter styleGAN) and proved it in experiments based on datasets in this field. However, this method lacks relevance in the field of wireless signal recognition, especially in category incremental tasks based on complex form communication signals, and cannot fully utilize the natural complex characteristics of communication signals. Therefore, the present invention proposes a deep learning method that can complete category incremental tasks for wireless signal recognition, especially complex form communication signals. Summary of the Invention

[0004] The purpose of the present invention is to address the deficiencies of the prior art and provide a signal recognition incremental learning method that decouples feature extraction and classification.

[0005] The object of the present invention is achieved through the following technical solution: a signal recognition incremental learning method that decouples feature extraction and classification, comprising the following steps:

[0006] ●Step 1: Create a complex neural network model, including a decoupled feature extractor and classifier, obtain signal samples and classify them into baseline and incremental phase categories according to the phase they belong to. For the complex signal vector s = I + Qi, the complex convolution kernel weight matrix W = A + Bi, the complex neural network is constructed based on the following method:

[0007] W*s=(A*IB*Q)+i(B*I+A*Q).

[0008] ● Step 2: Baseline phase: Using the sufficient baseline phase category training data and computing resources, we train the feature extractor using the automatic encoder-decoder concept and ensemble loss function, and optimize the parameters of the complex neural network model until convergence. This specifically includes the following sub-steps:

[0009] ■Step 2.1: The sample is represented by a complex feature using the feature extractor. The center loss is calculated based on the sample labels in the baseline phase, and the cross entropy loss is calculated based on the output vector in the classification layer.

[0010] ■Step 2.2: Based on the network structure of the feature extractor, build a decoder with the same structure but in the opposite order. Input the features obtained in step 2.1 to obtain the reconstructed sample and calculate the reconstruction loss value.

[0011] ■Step 2.3: Calculate the weighted sum of the center loss, cross entropy loss, and reconstruction loss in step 2.2 as the total loss function, and optimize the parameters of the feature extractor and the benchmark classifier according to the total loss function until convergence;

[0012] ●Step 3: In the incremental phase, freeze the feature extractor parameters, input the incremental phase samples into the feature extractor to obtain their feature representations, and then use the feature representations to train the incremental classifier C i , calculate the cross entropy loss and apply it to the incremental classifier C i Optimize until convergence;

[0013] ●Step 4: Use the class increment adapter to transform the baseline classifier C b and incremental classifier C i Perform fusion to obtain the fusion classifier C m , as the classifier after the model is updated. It specifically includes the following sub-steps:

[0014] ■Step 4.1: Build a smaller query dataset from the baseline category samples and the incremental category samples in a certain proportion, and use the feature extractor to obtain the feature representation query set;

[0015] ■Step 4.2: Baseline classifier C b and incremental classifier C i The weights are concatenated to obtain the matrix W, and the trainable transformation matrix P is used to convert W into the corresponding hidden representation;

[0016] Step 4.3: Aggregate the rows and use the updated weights as a classifier. Calculate the cross entropy loss on the feature representation query set and optimize the transformation matrix P until convergence.

[0017] ●Step 5: Input the sample to be predicted, obtain the feature representation of the signal to be measured through the feature extractor, and then pass it through the fusion classification layer C m An output vector group represented by the signal feature is obtained, and the output vector index of the maximum value in the output vector group is used as the prediction result to complete the output.

[0018] The present invention decouples the feature extractor and the classifier to combat the catastrophic forgetting problem that is common in incremental learning tasks, so as to achieve the purpose of expanding the classification and recognition capabilities of the signal recognition model at a relatively low cost without excessively forgetting the previous classification capabilities. Compared with the prior art, the decoupled feature extractor of the present invention is based on the idea of ​​automatic encoder-decoder, combines the center loss value, cross entropy loss value and KL divergence loss value to calculate the total loss function, and then performs back propagation to update the parameters of the neural network model, which can more fully mine and utilize the complex domain features of wireless signals, and provide conditions for the incremental learning task of signal recognition. The decoupled classifier of the present invention uses a complex-valued incremental adapter designed based on the node-based idea to complete the update in the incremental stage, which can effectively reduce the forgetting of previous classification capabilities and expand its classification and recognition capabilities at a relatively low cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a structural schematic diagram of the present invention;

[0020] Figure 2 It is a description of the process of constructing the query set involved in one embodiment of the present invention;

[0021] Figure 3 It is a schematic flow diagram of the present invention;

[0022] Figure 4 It is a category increment task constructed under two signal recognition datasets according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0024] ●Step 1: Create a complex neural network model, including a decoupled feature extractor, feature representation layer, and classifier. Obtain signal samples and classify them into baseline and incremental phase categories according to the phase they belong to. For the complex signal vector s = I + Qi and the complex convolution kernel weight matrix W = A + Bi, the calculation and implementation of the complex neural network components are as follows:

[0025] W*s=(A*IB*Q)+i(B*I+A*Q);

[0026] Among them, * is the convolution operation.

[0027] ● Step 2: Baseline phase: Using the sufficient baseline phase category training data and computing resources, we train the feature extractor using the automatic encoder-decoder concept and ensemble loss function, and optimize the parameters of the complex neural network model until convergence. This specifically includes the following sub-steps:

[0028] ■Step 2.1: The sample is represented by its complex features using the feature extractor, and the center loss value L is calculated based on the sample label in the baseline stage. ct :

[0029]

[0030] in, Represents category y i The corresponding feature center. Then the cross entropy loss value L is calculated based on the classification layer output ce :

[0031]

[0032] Among them, C(F(x i )) represents the classifier's response to the original input x i predictions.

[0033] ■Step 2.2: Based on the network structure of the feature extractor, build a decoder with the same structure but in the opposite order. Input the features obtained in step 2.1 to obtain the reconstructed sample and calculate the reconstruction loss value.

[0034]

[0035] Among them, D(F(x i )) represents the automatic encoder-decoder module for the original input x iIntuitively, the higher the degree of restoration of the original input x by the decoder when relying solely on the latent semantic vector, the more complete the information "inherited" by the latent semantic vector, and it also plays the role of feature extraction in the traditional sense.

[0036] ■Step 2.3: Take the weighted sum of the center loss value, cross entropy loss value in step 2.1, and reconstruction loss value in step 2.2 as the total loss function, and perform the weighted summation of the feature extractor and the benchmark classifier C according to the total loss function. b The parameters are optimized until convergence;

[0037] L=L ce +λL ct +θL re

[0038] Where λ and θ are the weights of the other two loss functions, and their sizes can be adjusted to control the model's ability bias. The parameter optimization method of the neural network model in step 2.3 includes back propagation and gradient descent algorithms.

[0039] ●Step 3: In the incremental phase, freeze the feature extractor parameters, input the incremental phase samples into the feature extractor to obtain their feature representations, and then use the feature representations to train the incremental classifier C i , calculate the cross entropy loss and apply it to the incremental classifier C i The optimization is performed until convergence, and the specific calculation method is the same as step 2.1. The parameter optimization method of the neural network model in step 3 includes back propagation and gradient descent algorithm.

[0040] ●Step 4: Use the class increment adapter to transform the baseline classifier C b and incremental classifier C i Perform fusion to obtain the fusion classifier C m , as the classifier after the model is updated. Specifically, it includes the following sub-steps:

[0041] ■Step 4.1: Construct a smaller query set from the baseline category samples and incremental category samples in a certain proportion. The query set is constructed as follows Figure 2 As shown. Subsequently, the feature extractor with frozen parameters is used to obtain the feature representation query set;

[0042] ■Step 4.2: Baseline classifier C b and incremental classifier C i The weights are concatenated to get the matrix W:

[0043]

[0044] In this embodiment, the weight updating process is described by taking the classifier weight vector of the j-th class as an example.

[0045] Use the trainable transformation matrix P to transform W into the corresponding hidden representation and calculate the correlation coefficient between the weight vectors of class i and class j:

[0046]

[0047] Among them, w i Represents the i-th node vector of the classifier weight matrix W, and <·> represents the calculation of the canonical inner product of two complex vectors. For the central node j, all coefficients are normalized using the softmax function:

[0048]

[0049] Among them, |W I | represents the number of classification nodes.

[0050] Normalized relationship coefficient Used to aggregate information from all nodes:

[0051]

[0052] Among them, α is the hyperparameter of the attention learning rate.

[0053] ■Step 4.3: Perform node aggregation on each row vector and use the updated weights as a classifier:

[0054]

[0055] Calculate the cross entropy loss on the feature representation query set in the same manner as described in step 2.1. Optimize the transformation matrix P until convergence. The parameter optimization method of the neural network model in step 4.3 includes back propagation and gradient descent algorithms.

[0056] ●Step 5: Input the sample to be predicted, obtain the feature representation of the signal to be measured through the feature extractor, and then pass it through the fusion classification layer C m An output vector group represented by the signal feature is obtained, and the output vector index of the maximum value in the output vector group is used as the prediction result to complete the output.

[0057] like Figure 4 As shown in FIG, a category increment task was designed on two different signal recognition datasets, and the effectiveness of the method was verified by completing the category increment task using the method described in the present invention. Figure 4 From left to right, the recognition of the baseline category and the incremental category of the same data set without using the method of the present invention and using the method of the present invention are respectively shown. Figure 4 The confusion matrix, which is commonly used in neural network model classification tasks, is used as a way to present recognition performance.

[0058] The above description is only a description of the preferred embodiments of the present application and does not limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical content should be regarded as equivalent valid embodiments and fall within the scope of protection of the technical solution of the present application.

Claims

1. A signal recognition incremental learning method with decoupling of feature extraction and classification, characterized by: The following steps are involved: (1) Create a complex neural network model, including a decoupled feature extractor and classifier, obtain signal samples and classify them into baseline stage categories and incremental stage categories according to the stage to which they belong; (2) In the benchmark phase, sufficient benchmark phase category training data and computing resources are used to train the feature extractor using the automatic encoder-decoder concept and the integrated loss function, and the parameters of the complex neural network model are optimized until convergence. Specifically, the following sub-steps are included: (2.1) The sample uses the feature extractor to obtain its complex feature representation, calculates the center loss value based on the sample label in the baseline stage, and calculates the cross entropy loss value based on the output vector in the classification layer; (2.2) Based on the network structure of the feature extractor, a decoder with the same structure but opposite order is built. The features obtained in step (2.1) are input to obtain the reconstructed samples and the reconstruction loss value is calculated. (2.3) taking a weighted sum of the center loss value, the cross entropy loss value, and the reconstruction loss value in step (2.2) as a total loss function, and optimizing the parameters of the feature extractor and the benchmark classifier according to the total loss function until convergence; (3) In the incremental stage, the feature extractor parameters are frozen, the incremental stage samples are input into the feature extractor to obtain their feature representations, and then the incremental classifier is trained using the feature representations. The cross entropy loss is calculated and the incremental classifier is optimized until convergence. (4) Use the incremental classifier to fuse the baseline classifier and the incremental classifier to obtain the fused classifier as the updated classifier. The specific steps include the following: (4.1) A query dataset is constructed proportionally from the baseline category samples and the incremental category samples, and a feature extractor is used to obtain a feature representation query set; (4.2) Concatenate the baseline classifier and the incremental classifier weights, convert them into corresponding hidden representations using a trainable transformation matrix, and calculate the correlation between any two classification nodes; (4.3) performing node aggregation on each row vector, using the updated weights as a classifier, computing the cross entropy loss on the feature representation query set, and optimizing the transformation matrix until convergence; (5) Input the sample to be predicted, obtain the feature representation of the signal to be measured through the feature extractor, and then obtain the output vector group of the signal feature representation through the fusion classification layer, and use the output vector index of the maximum value in the output vector group as the prediction result to complete the output.

2. The incremental learning method for signal recognition with decoupling of feature extraction and classification according to claim 1, characterized in that: The decoupled feature extraction layer in step (1) includes three complex convolutional layers and one complex fully connected layer; for the complex signal vector , complex convolution kernel weight matrix , the complex neural network component is calculated and implemented as follows: (1)。 3. The incremental learning method for signal recognition with decoupling of feature extraction and classification according to claim 1, characterized in that: The decoupled classifier includes a complex fully connected layer, and the calculation and implementation of the complex neural network component are as shown in formula (1).

4. The incremental learning method for signal recognition with decoupling feature extraction and classification according to claim 1, characterized in that: The total loss function described in step (2.3) consists of three parts, namely the cross entropy loss function , center loss and reconstruction losses ; The three loss functions are: (2) (3) (4) in, is the number of samples involved in calculating the loss, represents sample data, represents the sample label, Represents sample labels The corresponding category center, is the eigenvector, is the classifier output vector, Input for the decoder The reconstructed signal; the total loss function is as follows: (5) in, is the total sample loss, and They are the weights of the center loss function and the reconstruction loss function, respectively, which are used to adjust the model ability bias.

5. The incremental learning method for signal recognition with decoupling of feature extraction and classification according to claim 1, characterized in that: The classifier update for incremental learning is implemented by an incremental adapter independent of the classifier.

6. The incremental learning method for signal recognition with decoupling of feature extraction and classification according to claim 1, characterized in that: The class incremental adapter utilizes a transfer matrix Obtain the hidden representation of the original classification node and calculate the correlation between any two classification nodes. The calculation process is as follows: (6) in, Represents the classifier weight matrix No. node vectors, It represents the calculation of the canonical inner product of two complex vectors; the softmax function is then used to normalize the correlation between classification nodes, as shown below: (7) in, Indicates the number of classification nodes.

7. The incremental learning method for signal recognition with decoupling of feature extraction and classification according to claim 6, characterized in that: The incremental adapter completes the single-step update of the hidden representation of the original classification node through neighbor aggregation. The update process is as follows: (8) in, is the hyperparameter of the attention learning rate.

8. The incremental learning method for signal recognition with decoupling of feature extraction and classification according to claim 1, characterized in that: The classifier update process is completed by training a dataset called the query set, which is a subset of the dataset including the baseline categories and the incremental categories.

9. The incremental learning method for signal recognition with decoupling feature extraction and classification according to claim 7, characterized in that: The updated weights are used as classifiers to participate in the final incremental learning task. The classifier acquisition process is expressed as: (9) In the prediction stage, the feature representation of the signal to be measured is obtained through the feature extractor, and then the output vector group of the signal feature representation is obtained through the fusion classification layer, and the output vector index of the maximum value in the output vector group is used as the prediction result to complete the output.

Images