A noisy label screening and modulation identification method based on JS divergence and prototype distribution

Abstract

The application discloses a noisy label screening and modulation identification method based on JS divergence and prototype distribution. In view of the label pollution problem of actual wireless communication signal samples, firstly, the category prediction probability of the same signal is output by using two independently initialized backbone networks, and the reliability of the label is evaluated based on JS divergence. Then, according to the divergence distribution, the samples are selected to form a clean sample set, and the remaining samples form a noise sample set. Then, the clean samples are used to update the category prototype, the similarity of the noise samples to all category prototypes is converted into a prototype distribution, which is fused with the double backbone average prediction distribution, and then the distribution stability and confidence interval across training rounds are used to confirm the pseudo label. Finally, reliable training samples are screened out for training the downstream modulation identification decision module. The method can adaptively distinguish clean samples from noise samples under unknown noise rate conditions, improve the quality of pseudo labels, and enhance the discrimination robustness of similar modulation modes by using signal physical priori.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication signal processing technology, and relates to modulation recognition based on noisy tags. Specifically, it relates to a signal modulation recognition method that combines Jensen-Shannon divergence sample partitioning, prototype distribution calibration pseudo-tag generation, and physical feature fusion under the condition that the training tags are contaminated. Background Technology

[0002] Automatic modulation identification is a key technology in wireless communication sensing, spectrum management, jamming localization, and cognitive electronic warfare. Deep learning methods can learn high-dimensional discriminative features from I / Q sequences, but they typically rely on a large number of accurately labeled samples. In real electromagnetic environments, sample labels may come from manual annotation, automatic annotation systems, or semi-automatic acquisition processes, and are easily affected by factors such as boundary sample confusion, low signal-to-noise ratio fading, modulation category similarity, and differences in the expertise of annotation personnel, resulting in noisy labels.

[0003] Because deep neural networks tend to learn simple patterns first and then memorize labels, if noisy labels are used directly for supervised training, the network may fit the patterns corresponding to incorrect labels into the decision boundary, leading to a decrease in generalization performance.

[0004] Existing sample selection methods often rely on cross-entropy loss or manually set thresholds. In modulation recognition scenarios, samples with low signal-to-noise ratios and similar modulation schemes often have high losses, but they may be difficult to label correctly. Simply discarding them weakens the model's ability to learn critical boundary samples. Therefore, a scheme is needed that can adaptively identify clean samples and generate reliable pseudo-labels for noisy samples under unknown noise rates. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a noisy tag screening and modulation recognition method based on JS divergence and prototype distribution.

[0006] In a first aspect, the present invention provides a noisy label filtering method based on JS divergence and prototype distribution, the implementation steps of which are as follows:

[0007] Step 1: Obtain Data

[0008] Obtain the signal modulation training set containing the given labels. , where x i For the i-th I / Q signal sample, Given labels, N is the number of samples; Convert to C-dimensional one-hot label vector y i =[y i,1 ,y i,2 ,…y i,C ], where yi,c =1 indicates that the given label of the sample is the first... Class, otherwise y i,c =0, c=1,2,…C, where C is the number of modulation categories;

[0009] Step 2: Sample selection based on Jensen-Shannon divergence

[0010] Construct two backbone networks N with identical structures and independently initialized parameters. A and N B For sample x respectively i Output category prediction probability and And calculate the average prediction probability. , where p i =[p i,1 ,p i,2 ,…p i,C ], p i,c Indicates sample x i The average predicted probability of belonging to class c.

[0011] Further calculation of sample x i The unique hot label vector y i With average predicted probability p i The Jensen-Shannon divergence between d i The dynamic cutoff threshold d is determined based on the divergence distribution of all samples in the training set D. cut Based on the dynamic cutoff threshold d cut Determine the sample retention ratio R for the current training round. Within each class, select the sample with the lowest divergence according to the sample retention ratio R and add it to the clean sample set D. clean The remaining samples are added to the noisy sample set D. noisy .

[0012] Step 3: Generating pseudo-labels based on prototype distribution calibration

[0013] Set a high-confidence consistency condition for the dual-network system and select a clean sample set D. clean Samples that meet the conditions are used to update the prototypes of each category.

[0014] For D noisy For each sample in the dataset, its similarity to all class prototypes is calculated and converted into a prototype distribution. This prototype distribution is then compared with the backbone network N. A and N B The average predicted probabilities are adaptively fused using entropy to obtain the calibration distribution. The pseudo-label sample set D is then determined based on the cross-round stability, maximum confidence, and inter-class confidence interval of the calibration distribution. pseudo .

[0015] Step 4: Generation of Reliable Training Set

[0016] With D clean With D pseudo The union of the sets D is used as a reliable training set. reliable It is used to train deep learning networks to identify the modulation type of modulated signals.

[0017] Secondly, the present invention provides a modulation recognition method based on noisy tag screening, the specific steps of which are as follows:

[0018] Step 1: For a given signal modulation training set D, use the above-mentioned noisy label selection method based on JS divergence and prototype distribution to generate a reliable training set D. reliable .

[0019] Step 2: For the reliable training set D reliable The instantaneous amplitude, instantaneous phase, and instantaneous frequency of the original I / Q signal are extracted from the samples.

[0020] Step 3: Construct a modulation recognition decision module based on the instantaneous amplitude, instantaneous phase, instantaneous frequency of the I / Q signal and the backbone network N. A and N B The feature extraction output is used as input, and after fusion, it is input into the classifier, and the output is the modulation recognition result.

[0021] The present invention has the following beneficial effects:

[0022] 1. The Jensen-Shannon divergence is used to compare the given label distribution with the average prediction distribution of the two networks. The divergence is symmetric and has a finite range of values, which can stably characterize the degree of inconsistency between the label and the model prediction.

[0023] 2. The dynamic threshold for determining the sample retention ratio is adaptively determined by the divergence distribution of the current training round, without requiring prior knowledge of the label noise rate; by selecting low-divergence samples by category, the risk of over-selection of simple categories and over-discarding of difficult categories is reduced.

[0024] 3. The pseudo-labels are not determined by a single network prediction or a single candidate category prototype validation, but are jointly calibrated by the prototype distribution of all categories and the average prediction distribution of the dual backbone, and are screened by cross-round stability, maximum confidence and inter-category confidence interval, which improves the reliability of labels when reusing noisy samples.

[0025] 4. When performing modulation identification, the instantaneous amplitude, instantaneous phase and instantaneous frequency of the I / Q signal are explicitly integrated, making full use of the physical priors of different modulation methods in amplitude and phase change laws, so that the model has stronger physical consistency and discrimination ability under similar modulation categories and complex noise conditions. Attached Figure Description

[0026] Figure 1 The flowchart shows a noisy label filtering method based on JS divergence and prototype distribution.

[0027] Figure 2 The flowchart shows the sample selection process based on Jensen-Shannon divergence.

[0028] Figure 3 A flowchart for generating pseudo-labels based on prototype distribution calibration;

[0029] Figure 4 This is a schematic diagram of the modulation recognition decision module in the embodiment;

[0030] Figure 5 Performance of clean sample screening under different label noise rates;

[0031] Figure 6 Performance of pseudo-tag generation under different tag noise rates;

[0032] Figure 7 This diagram illustrates the recognition accuracy at different signal-to-noise ratios under different tag noise rates.

[0033] Figure 8 This is the modulation category confusion matrix under 60% tag noise conditions. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0035] A noisy tag filtering method based on JS divergence and prototype distribution, such as Figure 1 As shown, the implementation steps are as follows:

[0036] Step 1: Obtain Data

[0037] Obtain the signal modulation training set containing the given labels. , where x i For the i-th I / Q signal sample, Give it a label, For the sample size; Convert to C-dimensional one-hot label vector y i =[y i,1 ,y i,2 ,…y i,C ], where y i,c =1 indicates that the given label of the sample is the first... Class, otherwise y i,c =0, c=1,2,…C, where C is the number of modulation categories.

[0038] Step 2: Sample selection based on Jensen-Shannon divergence

[0039] Construct two backbone networks N with identical structures and independently initialized parameters. A and N B For sample x respectively i Output category prediction probability , And calculate the average prediction probability. , where p i =[p i,1 ,p i,2 ,…p i,C ], p i,c Indicates sample x i The average predicted probability of belonging to class c.

[0040] Further calculation of sample x i The unique hot label vector y i With average predicted probability p i The Jensen-Shannon divergence between d i :

[0041]

[0042]

[0043] Where, m i For the one-hot label vector y i With average predicted probability p i The mean distribution, This indicates the calculation of the KL divergence between two probability distributions.

[0044] The dynamic cutoff threshold is obtained based on the divergence distribution of all samples in the training set D. :

[0045]

[0046]

[0047]

[0048] Where, d avg d is the average divergence of all samples in the current training round. min The minimum divergence value, These are the filter coefficients. To adjust the threshold, it is determined before training begins using the average Jensen-Shannon divergence from the last few training epochs of the warm-up phase or preset validation data. The adjusted threshold... This is used to determine whether the overall prediction results of the current dual-backbone network are still in a state of significant difference from the given label. When At that time, based on the minimum divergence d of the current training round min Tighten the cutoff threshold; when This indicates that the dual-backbone network has developed a certain class discrimination capability, using the current average divergence. As a cutoff threshold.

[0049] Set the sample retention ratio for the current training round. In each category of sample subset D (c) In the middle, all samples x i ∈D (c) According to the divergence according to d i Sort by size from smallest to largest, based on the percentage of samples retained. Select several samples with small divergence and add them to the clean sample set D. clean Sample subset D (c) The remaining samples are added to the noisy sample set D. noisy This ensures that the clean sample set remains balanced across different modulation categories, such as Figure 2 As shown.

[0050] Step 3: Generating pseudo-labels based on prototype distribution calibration

[0051] like Figure 3 As shown, for a clean sample set D clean The samples in the backbone network N A and N B The predicted categories for this sample are the same, and the maximum predicted probability is not lower than the confidence threshold. Then the sample satisfies the high-confidence consistency condition of the two-network system. Based on the clean sample set D clean Sample x that satisfies the high-confidence consistency condition of the dual network i Calculate the sample average feature z i :

[0052]

[0053] Among them, f A (x i ) and f B (x i ) are the backbone network N A and N B Feature extraction output.

[0054] Then update the class c prototype P using an exponential moving average method. c :

[0055]

[0056] Where t is the training round and β is the momentum coefficient. The set of sample indices in the current training round that satisfy the high confidence consistency condition of the two networks and are predicted as class c; when When empty, keep .

[0057] For clean samples that do not meet the high-confidence consistency condition of the dual network, they are still retained in the clean sample set D in the current training round. clean It participates in the supervised training of the dual backbone network and the downstream modulation recognition decision module with its given label, but does not participate in the category prototype update of the current training round.

[0058] For the noise sample set D noisy Sample x in j First calculate the average feature z j With each category prototype P c cosine similarity :

[0059]

[0060] Where |||2 represents the calculation of the 2-norm, and T represents the transpose. Then, the temperature coefficient T is used. p Convert all category similarities into a prototype distribution q. j :

[0061] .

[0062] The prototype distribution q j With backbone network N A and N B average predicted probability p j Perform entropy adaptive fusion to obtain the calibration distribution u j :

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] in, To prevent constants with a denominator of zero, g p To predict the normalized information certainty of the distribution, g q The normalized information certainty of the prototype distribution. For entropy adaptive fusion weights, u j For calibration distribution; for u j Perform an exponential moving average across rounds:

[0069]

[0070] in, This represents the moving average coefficient of the calibration distribution. The highest confidence value is selected from all C categories. ,when Not lower than the false label confidence threshold ,and The difference between the two category confidence components with the largest and second largest median values ​​is not less than the interval threshold. ,and If it remains unchanged over M consecutive training rounds, As noise sample x j pseudo-tags, and (x j , Included in pseudo-label sample set D pseudo . This represents the category number that gives the highest confidence level.

[0071] Step 4: Generation of Reliable Training Set

[0072] With D clean With D pseudo The union of the sets D is used as a reliable training set. reliable The reliable training set D reliable It can be used to train deep learning networks to identify the modulation type of modulated signals.

[0073] A modulation recognition method based on noisy tag screening, the specific steps of which are as follows:

[0074] Step 1: For a given signal modulation training set D, use the above-mentioned noisy label selection method based on JS divergence and prototype distribution to generate a reliable training set D. reliable .

[0075] Step 2: For the reliable training set D reliable Extract the original I / Q signal x from the samples. i =[(I i,1 Q i,1 ),(I i,2 Q i,2 ),…(I i,L Q i,L The instantaneous amplitude a] i,k Instantaneous phase φ i,k and instantaneous frequency v i,k :

[0076]

[0077]

[0078]

[0079] Where L represents the length of the I / Q sequence. Instantaneous frequency v i,k It is derived from the phase difference and solves the phase entanglement problem.

[0080] Step 3: Construct the modulation recognition decision module using a reliable training set D. reliable Conduct training.

[0081] like Figure 4 As shown, the modulation recognition decision module includes a physical feature extractor, a deep consistency feature extractor, a feature fusion layer, and a classifier. The physical feature extractor includes three one-dimensional convolutional branches, with inputs of instantaneous amplitude, instantaneous phase, and instantaneous frequency, respectively. The outputs of the three branches are concatenated along the channel dimension to form the physical feature h. phy The input to the deep consistency feature extractor is the backbone network N. A and N B Feature extraction output f A (x i ) and f B (x i The average of the two features forms the deep consistency feature h. con The feature fusion layer will integrate the deep consistency features h. con and h phy splicing into fusion feature h fus =[h con ;h phy The input is then fed into the classifier to obtain the final modulation category probability.

[0082] To verify the effectiveness of the method of this invention under noisy labeling conditions, experiments were conducted on the publicly available modulated signal dataset RML2016.10a. Samples with a signal-to-noise ratio of 0–18 dB were selected, encompassing 11 modulation categories. The dataset was divided into training, validation, and test sets in a 6:2:2 ratio, with the training set containing 66,000 samples, and the validation and test sets each containing 22,000 samples.

[0083] Symmetrical label noise is applied only to the training set labels, with noise rates set to 0%, 10%, 20%, 30%, 40%, 50%, and 60%, respectively. The validation and test sets retain correct labels. Since all sample labels are correct when the noise rate is 0%, the number of false labels generated and the accuracy are not statistically analyzed. During training, a dual-backbone network N is first used. A N BThe sample prediction distribution is obtained, and clean samples are selected based on the Jensen-Shannon divergence between the given label distribution and the average prediction distribution. Then, the unselected samples are calibrated using the class prototype distribution and the dual-backbone average prediction distribution, and pseudo-labels are generated based on the maximum confidence, inter-class confidence interval, and stability across training rounds. Finally, the modulation recognition decision module is trained using clean samples and pseudo-label samples.

[0084] The experimental results are shown in Table 1:

[0085] Table 1

[0086]

[0087] When the label noise rate is 0%–30%, the test accuracy remains at 83.52%–84.62%; when the noise rate is increased to 40%, 50%, and 60%, the test accuracy is 81.69%, 80.68%, and 73.63%, respectively. Meanwhile, the sample retention rate R decreases from 81.44% to 45.87%, indicating that the dynamic JS divergence threshold can adaptively reduce high-risk training samples according to the degree of label contamination.

[0088] like Figure 5 As shown, when the tag noise rate does not exceed 50% and the signal-to-noise ratio reaches 4 dB, the overall recognition accuracy remains at 81%–87%. Under 60% tag noise, the accuracy increases from 59.14% at 0 dB to 77.32% at 10 dB, indicating that the present invention still has good recognition stability when tag noise and channel noise coexist.

[0089] like Figure 6 As shown, as the label noise rate increases from 10% to 60%, the precision of clean sample screening decreases from 99.55% to 82.16%, while the recall of true clean samples increases from 82.95% to 93.76%. This indicates that low JS divergence screening by category can balance sample reliability and clean sample coverage under high noise conditions.

[0090] like Figure 7 As shown, the accuracy of pseudo-label adoption increased from 60.48% to 89.89%, and the number of adoptions generally increased with the noise rate. At a noise rate of 60%, the number of adoptions was 18,699, and the accuracy reached 89.89%, indicating that prototype distribution calibration and cross-round stability constraints can effectively screen for highly reliable pseudo-labels.

[0091] Figure 8The modulation category confusion matrix under 60% tag noise conditions is generally diagonally dominant. Under 60% tag noise, the recognition accuracy of AM-SSB, PAM4, and CPFSK reaches 100.00%, 96.00%, and 95.40%, respectively; misidentification is mainly concentrated in similar categories such as QAM16 and QAM64, and QPSK and 8PSK.

[0092] Summary Table 1 and Figures 5-8 The results show that this invention can adaptively screen clean samples based on the level of label noise and improve the quality of fake labels through prototype distribution calibration. Even with 60% symmetrical label noise, the test accuracy still reaches 73.63%, verifying the feasibility and adaptability of this method in high-noise environments.

Claims

1. A method for filtering noisy tags based on JS divergence and prototype distribution, characterized in that: For a signal modulation dataset D with a given label, firstly, two backbone networks N with identical structures and independently initialized parameters are used. A and N B For the same sample, predict the output class probability and calculate the average prediction probability p of the two backbone networks. i Jensen-Shannon divergence d with respect to the label of the sample in dataset D i ; Then, the sample retention ratio R for the current training round is determined using the divergence distribution of all samples in the signal modulation dataset D. Within each category, the sample with the lowest divergence is selected according to the sample retention ratio R and added to the clean sample set D. clean The remaining samples are added to the noisy sample set D. noisy ; Next, clean sample set D is selected. clean Samples satisfying the high-confidence consistency condition of the dual-network model are used to update the prototypes of each category; for the noisy sample set D noisy For each sample in the dataset, its similarity to all class prototypes is calculated and converted into a prototype distribution. This prototype distribution is then adaptively fused with the average predicted probabilities of the two backbone networks to obtain a calibration distribution. The pseudo-label sample set D is determined based on the cross-round stability, maximum confidence, and inter-class confidence interval of the calibration distribution. pseudo ; Finally, D clean With D pseudo The union of the sets D is used as a reliable training set. reliable It is used to train deep learning networks to identify the modulation type of modulated signals.

2. The noisy tag filtering method based on JS divergence and prototype distribution as described in claim 1, characterized in that: Convert the sample labels in the modulation dataset D into a C-dimensional one-hot label vector y. i =[y i,1 ,y i,2 ,…y i,C ], where y i,c =1 indicates that the given label of the sample is the first... Class, otherwise y i,c =0, c=1,2,…C, where C is the number of modulation categories.

3. The noisy tag filtering method based on JS divergence and prototype distribution as described in claim 2, characterized in that: The average prediction probability p i Jensen-Shannon divergence with sample labels d i for: Where, m i For the one-hot label vector y i With average predicted probability p i The mean distribution, This indicates the calculation of the KL divergence between two probability distributions.

4. The noisy tag filtering method based on JS divergence and prototype distribution as described in claim 1, characterized in that: The sample retention ratio R N represents the total number of samples in the signal modulation dataset D, and d cut This represents the dynamic cutoff threshold determined based on the divergence distribution of all samples: Where, d avg d is the average divergence of all samples in the current training round. min The minimum divergence value, These are the filter coefficients. To adjust the threshold, it is determined using the average Jensen-Shannon divergence of the last few training rounds in the warm-up phase or preset validation data before training begins.

5. The noisy tag filtering method based on JS divergence and prototype distribution as described in claim 1, characterized in that: The high-confidence consistency condition for the two backbone networks is that the predicted categories of the two backbone networks are the same, and the maximum predicted probability is not lower than the confidence threshold. .

6. The noisy tag filtering method based on JS divergence and prototype distribution as described in claim 1, characterized in that: The method for updating the category prototype is as follows: Where t is the training round and β is the momentum coefficient. The set of sample indices in the current training round that satisfy the high confidence consistency condition of the two networks and are predicted as class c; when When empty, keep ;z i This indicates that the two backbone networks are connected to a clean sample set D. clean Sample x that satisfies the high-confidence consistency condition of the dual network i The average characteristics.

7. The noisy tag filtering method based on JS divergence and prototype distribution as described in claim 1, characterized in that: For the noisy sample set D noisy The calibration distribution of the samples is subjected to an exponential moving average across training rounds. If, in the t-th training round, the largest confidence value among all C categories is not lower than the pseudo-label confidence threshold, the difference between the largest and second largest confidence components of the two categories is not lower than the interval threshold, and the largest confidence value remains unchanged over M consecutive training rounds, then the largest confidence value is taken as the pseudo-label of the noise sample, and this sample is included in the pseudo-label sample set D. pseudo .

8. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1 to 7.

9. A modulation recognition method based on noisy tag screening, characterized in that: The specific steps are as follows: Step 1: For a given signal modulation training set D, use the noisy label selection method based on JS divergence and prototype distribution as described in any one of claims 1 to 7 to generate a reliable training set D. reliable ; Step 2: For the reliable training set D reliable From the samples, extract the instantaneous amplitude, instantaneous phase, and instantaneous frequency of the original I / Q signal; Step 3: Construct a modulation recognition decision module based on the instantaneous amplitude, instantaneous phase, instantaneous frequency of the I / Q signal and the backbone network N. A and N B The feature extraction output is used as input, and after fusion, it is input into the classifier, and the output is the modulation recognition result.

10. The modulation recognition method based on noisy tag screening as described in claim 9, characterized in that: The modulation recognition decision module includes a physical feature extractor, a deep consistency feature extractor, a feature fusion layer, and a classifier. The physical feature extractor includes three one-dimensional convolutional branches, with inputs of instantaneous amplitude, instantaneous phase, and instantaneous frequency, respectively. The outputs of the three branches are concatenated along the channel dimension to form the physical feature h. phy The input to the deep consistency feature extractor is the backbone network N. A and N B Feature extraction output f A (x i ) and f B (x i The average of the two features forms the deep consistency feature h. con The feature fusion layer integrates deep consistency features h con and h phy splicing into fusion feature h fus =[h con ;h phy The input is then fed into the classifier to obtain the final modulation category probability.

Images