Automatic modulation identification core set selection method based on forgetting quality score

Abstract

The invention discloses an automatic modulation identification core set selection method based on forgetting quality scoring, and belongs to the technical field of wireless communication signal processing and deep learning crossing. In order to solve the problems that a signal time-frequency structure and signal-to-noise ratio distribution are ignored in an automatic modulation identification task, an evaluation index is single and the like in the existing method, a prediction track of a sample in multiple rounds of training is recorded, and a forgetting quality score (FoQuS) of a fusion correctness change frequency, a continuous error round, accuracy and a loss value is calculated; and dynamically hierarchically selecting a core set based on the score. And the model is retrained by using the core set, so that the identification performance can be maintained while the training data scale is remarkably reduced, and the method has strong generalization ability and high practical value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication signal processing and deep learning, and particularly relates to a core set selection method for automatic modulation recognition based on a forgetting quality score. BACKGROUND

[0002] With the wide application of deep learning technology in the field of wireless communication, automatic modulation recognition technology based on deep learning has become the mainstream solution in scenarios such as spectrum monitoring, cognitive radio and electronic countermeasures. This method can directly learn discriminative features from raw signal data and achieve high recognition accuracy in complex environments. However, the excellent performance of deep learning models depends on large-scale labeled data sets, and the training process requires multiple iterations through the entire training set and tedious hyperparameter tuning, resulting in huge consumption of computing resources and long model development cycle. This problem is particularly prominent in scenarios such as scientific research iteration and edge device deployment, where storage and computing capabilities are limited.

[0003] To reduce the training cost, core set selection technology is proposed to select a representative subset from the full data. Common strategies include random uniform sampling, gradient norm-based sampling, sample forgetting number-based sampling and feature space coverage-based sampling. These methods perform well in image classification tasks, but directly migrating to automatic modulation recognition tasks has obvious shortcomings. First, radio frequency signals have unique time-frequency structure and physical properties, often accompanied by channel effects such as noise, multipath and interference, which are fundamentally different from the spatial locality and semantic continuity of natural images. Second, the signal-to-noise ratio carried by the signal sample has a decisive influence on the model generalization ability, and the feature distribution of samples under different signal-to-noise ratios is significantly different. Existing methods do not fully consider the signal-to-noise ratio distribution characteristics, and only rely on single indicators such as loss value or gradient size to select samples, which can easily select too many simple samples with high signal-to-noise ratio and ignore difficult samples with low signal-to-noise ratio, resulting in a lack of diversity in the selected core set and a sharp decline in the performance of the trained model under complex channel conditions.

[0004] Some data selection methods in the prior art mainly focus on feature-level selection rather than sample importance evaluation, and the difference in feature selection can cause large deviations in the selection results. Although some methods consider the weighted importance of multiple features, they use a greedy strategy of sequentially eliminating low-importance features, which can easily ignore different features and lose the diversity of the data set. More importantly, these methods do not incorporate dynamic training information, making it impossible to identify key samples that are repeatedly forgotten or difficult to fit during training, resulting in a lack of sufficient support for model generalization ability in the selected core set.

[0005] Therefore, in the face of the unique nature of signal data and the dual challenges of training efficiency in the automatic modulation recognition task, there is an urgent need for a solution that can address the above problems. SUMMARY

[0006] To solve the above technical problems, the present application provides an automatic modulation recognition core set selection method based on forgetting quality score to solve the problems existing in the prior art.

[0007] In the first aspect, to achieve the above object, the present application provides an automatic modulation recognition core set selection method based on forgetting quality score, comprising the following steps:

[0008] S1, pre-processing the original training data set, including signal amplitude normalization and hierarchical random division, to obtain a training set and a test set;

[0009] S2, using a deep learning model to perform multiple rounds of training on the training set, and recording the predicted label, correctness indication and cross-entropy loss of each sample at each training round;

[0010] S3, based on the recorded training trajectory, calculating the forgetting quality score FoQuS of each sample, the FoQuS being obtained by fusing three complementary statistics, including the number of changes in sample prediction correctness, the number of consecutive prediction errors of the sample, and the combination of sample prediction accuracy and loss value;

[0011] S4, sorting all samples in descending order according to the FoQuS, and dynamically dividing them into multiple selection intervals, and adaptively determining the number of samples selected from each selection interval according to the target sampling rate, and selecting samples from each selection interval to form a core set;

[0012] S5, retraining the model using the core set and evaluating the model performance on the test set.

[0013] Optionally, in S1, the pre-processing process includes:

[0014] Amplitude normalization is performed on the IQ data of each signal sample, and standardization is achieved by calculating the mean and standard deviation of the sample;

[0015] The hierarchical random sampling strategy is used to divide the data set into a training set and a test set, ensuring that the distribution of each modulation class and each signal-to-noise ratio level in the training set and the test set is consistent with the original data set.

[0016] Optionally, in S2, the training and recording trajectory process includes:

[0017] Multiple rounds of training are performed, and for each sample at each training round, the predicted label of the model is recorded;

[0018] The correctness indication of the sample is calculated based on the predicted label and the true label, and the correctness indication is a binary variable indicating whether the prediction result is correct;

[0019] record the cross-entropy loss of the sample in the current training round;

[0020] accumulate the number of correct predictions and the sum of cross-entropy losses of each sample.

[0021] Optionally, in S3, the process of calculating the FoQuS comprises:

[0022] calculating a first statistic, which is the number of changes in the correctness of the sample prediction, counting the number of times the sample changes from correct prediction to incorrect prediction;

[0023] calculating a second statistic, which is the number of consecutive incorrect predictions of the sample, counting the maximum number of consecutive incorrect predictions of the sample;

[0024] calculating a third statistic, which is the combination of the prediction accuracy and the loss value of the sample, calculated based on the accumulated number of correct predictions and the average loss of the sample;

[0025] normalizing the first statistic, the second statistic, and the third statistic respectively;

[0026] adding the three normalized statistics to obtain the FoQuS value.

[0027] Optionally, in S4, the process of dynamically selecting the core set comprises:

[0028] sorting all samples in descending order according to the FoQuS value;

[0029] dividing the sorted sample sequence into three selection intervals, including a first selection interval, a second selection interval, and a third selection interval;

[0030] determining the number of samples selected from each selection interval according to the target sampling rate, wherein the sampling proportion of each selection interval is different when the target sampling rate is different;

[0031] selecting samples from each selection interval according to the determined number of sample selections to form the core set.

[0032] Optionally, in S5, the process of model training and performance verification comprises:

[0033] retraining the deep learning model using the selected core set;

[0034] evaluating the performance of the retrained model on the test set;

[0035] verifying the effectiveness of the core set selection method through comparative experiments.

[0036] In a second aspect, the present application further provides an automatic modulation recognition core set selection system based on forgetting quality score, for implementing an automatic modulation recognition core set selection method based on forgetting quality score, the system comprising:

[0037] a preprocessing module for preprocessing the original training dataset, including signal amplitude normalization and hierarchical random partitioning, to obtain a training set and a test set;

[0038] a training trajectory recording module for using a deep learning model to perform multiple rounds of training on the training set, and recording the predicted label, correctness indication and cross-entropy loss of each sample at each training round;

[0039] a FoQuS calculation module for calculating the forgetting quality score FoQuS of each sample based on the recorded training trajectory, the FoQuS being obtained by fusing three complementary statistics, including the number of changes in sample prediction correctness, the number of consecutive prediction errors of the sample, and the combination of sample prediction accuracy and loss value;

[0040] a core set selection module for sorting all samples in descending order according to the FoQuS, dynamically dividing them into multiple selection intervals, adaptively determining the number of samples to be selected from each selection interval, and selecting samples from each selection interval to form a core set;

[0041] a model training and verification module for retraining the model using the core set and evaluating the model performance on the test set.

[0042] In a third aspect, the present application further provides a computer terminal device, comprising:

[0043] one or more processors;

[0044] a memory coupled to the processor, for storing one or more programs;

[0045] when the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the automatic modulation recognition core set selection method based on forgetting quality score in the first aspect described above.

[0046] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the automatic modulation recognition core set selection method based on forgetting quality score in the first aspect described above.

[0047] In a fifth aspect, the present application further provides a computer program product comprising a computer program, the computer program being executed by a processor to implement the steps of the automatic modulation recognition core set selection method based on forgetting quality score in the first aspect described above.

[0048] Compared with the prior art, the present application has the following advantages and technical effects:

[0049] The automatic modulation recognition core set selection method based on the forgetting quality score provided by the present application can effectively overcome the limitations of the existing core set selection method in the automatic modulation recognition task. The present application fully considers the time-frequency structure characteristics and signal-to-noise ratio distribution characteristics of signal data and is specially designed for the automatic modulation recognition task. By fusing the three complementary statistics of the number of sample prediction correctness changes, the continuous prediction error rounds and the prediction accuracy and loss value, a multi-dimensional comprehensive evaluation of the importance of the sample is realized, and the one-sidedness of single statistical evaluation is avoided. The hierarchical dynamic sampling strategy adopted ensures that the selected core set contains key samples near the decision boundary, difficult samples that are difficult to fit and stable samples that are easy to learn, ensuring the balanced distribution and overall diversity of the samples under different signal-to-noise ratio conditions. This method significantly reduces the data size and computing resource overhead required for model training, while maintaining similar recognition performance to full data training, showing good cross-model architecture generalization ability and important practical value in resource-constrained edge computing environments. BRIEF DESCRIPTION OF DRAWINGS

[0050] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application and their

[0051] Figure 1 The AMR core set selection flowchart based on the forgetting quality score of the embodiments of the present application;

[0052] Figure 2 The FoQuS score calculation schematic diagram of the embodiments of the present application. DETAILED DESCRIPTION

[0053] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0054] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0055] Embodiment one

[0056] The application provides a data processing-based method, the core idea of which is: by recording the predicted trajectory of each sample in the training process, three complementary intermediate statistics are obtained, namely, a first statistic of the number of changes in the prediction correctness of the sample, a second statistic of the number of rounds of continuous prediction errors of the sample, and a third statistic of the combination of the prediction accuracy and the loss value of the sample. Based on these statistics, the method adopts a dynamic sample selection strategy to form a diversified core set. The method can significantly reduce the size of the required training data while maintaining the performance close to that of the full data model.

[0057] The mathematical goal of core set selection is to construct a compact and effective subset (Wherein ), so that the performance of the model trained on the subset is close to that of the model trained on the full data set D. The optimization problem can be formally represented as:

[0058] (1)

[0059] (2)

[0060] (3)

[0061] Wherein, and represent the model parameters obtained by training on the full data D and the core set S respectively; represents the loss function; k is the data sampling rate, ; is the acceptable performance loss threshold.

[0062] As Figure 1 shown, in this embodiment, an automatic modulation recognition core set selection method based on forgetting quality score is provided, comprising:

[0063] S1, pre-processing the original training data set, including signal amplitude normalization and hierarchical random division, to obtain a training set and a test set;

[0064] S2, using a deep learning model to perform multi-round training on the training set, and recording the predicted label, correctness indication and cross-entropy loss of each sample at each training round;

[0065] S3, based on the recorded training trajectory, calculating the forgetting quality score FoQuS of each sample, the FoQuS being obtained by fusing three complementary statistics, including the number of changes in the prediction correctness of the sample, the number of rounds of continuous prediction errors of the sample, and the combination of the prediction accuracy and the loss value of the sample;

[0066] S4, sort all samples in descending order according to the FoQuS, and dynamically divide into multiple selection intervals, adaptively determine the number of sample selection of each selection interval according to the target sampling rate, and select samples from each selection interval to form a core set;

[0067] S5, retrain the model using the core set and evaluate the model performance on the test set.

[0068] The specific process includes:

[0069] S1: data set division and preprocessing. Given the original training data set , wherein is the signal sample (usually IQ dual-channel data), is the corresponding modulation category label. The preprocessing includes signal amplitude normalization, and the data set is divided into training set and test set according to the preset proportion, ensuring the balanced distribution of each modulation category and different signal-to-noise ratio samples in the two groups.

[0070] S2: full data set training and trajectory recording. Use the deep learning model to train T rounds on the complete training set, record the prediction label, correctness, cross-entropy loss and other information of each sample in each training round, and form a complete training dynamic trajectory.

[0071] S3: calculation of comprehensive quantitative value. Based on the sample prediction trajectory recorded in the training process, three intermediate statistical quantities are calculated:

[0072] 1) the first statistical quantity based on the number of changes in sample prediction correctness;

[0073] 2) the second statistical quantity based on the number of continuous prediction error rounds;

[0074] 3) the third statistical quantity based on the sample prediction accuracy and loss value;

[0075] After standardizing the above three statistical quantities, the final comprehensive quantitative value is obtained by weighted fusion, which is used as the sample selection basis.

[0076] S4: dynamic core set selection based on comprehensive quantitative value. Arrange all samples in the training set in descending order according to their comprehensive quantitative values. According to the preset division rule, the sorted samples are divided into multiple selection intervals. Based on the preset target sampling rate, the sample selection quantity of each selection interval is adaptively determined according to the built-in sampling configuration strategy. According to the determined number, samples are selected from each selection interval to form a core set for model training.

[0077] S5: Model training and performance verification based on the selected core set. Re-train the model using the core set dynamically selected in S4 and evaluate the model performance on the test set. Through comparative analysis, verify the effectiveness of the core set selection method, and further conduct generalization experiments across model architectures to investigate its universality.

[0078] As an implementation in the embodiment, in S1, the pre-processing process includes:

[0079] The IQ data of each signal sample is amplitude normalized, and the standardization is achieved by calculating the mean and standard deviation of the sample.

[0080] The data set is divided into training set and test set by using hierarchical random sampling strategy, ensuring that the distribution of each modulation category and each signal-to-noise ratio level in the training set and test set is consistent with the original data set.

[0081] Further, the data preprocessing in step S1 includes normalizing the IQ data of each signal sample, and the normalization formula is:

[0082] (4)

[0083] Where x is the original IQ data, and are the mean and standard deviation of the sample, respectively. The data set is divided by using hierarchical random sampling strategy, ensuring that the proportion of each modulation category and each signal-to-noise ratio level in the training set and test set is consistent.

[0084] As an implementation in the embodiment, in S2, the process of training and recording the trajectory includes:

[0085] Multiple rounds of training are performed, and for each sample in each training round, the predicted label of the model is recorded;

[0086] The correctness indicator of the sample is calculated based on the predicted label and the true label, and the correctness indicator is a binary variable indicating whether the prediction result is correct;

[0087] The cross-entropy loss of the sample in the current training round is recorded;

[0088] The number of correct predictions and the total cross-entropy loss of each sample are accumulated.

[0089] Further, the training trajectory recording in step S2 specifically includes:

[0090] For each sample , the predicted label of the model is recorded after the t-th training round , and the correctness indicator of the sample is calculated:

[0091] (5)

[0092] where is an indicator function that takes value 1 if the condition is true and 0 otherwise. The cross-entropy loss and the number of correct predictions are accumulated for each sample.

[0093] Further, the model training parameter setting in the step S2 comprises:

[0094] The training round T is set to 10-50 epochs; the optimizer is Adam or SGD, and the learning rate is in the range of 0.0001-0.01; the batch size batch_size is set to 128-256; and the loss function is cross-entropy loss.

[0095] As an embodiment in the present embodiment, in S3, the process of calculating FoQuS comprises:

[0096] calculating a first statistical quantity, the first statistical quantity being the number of changes in sample prediction accuracy, and counting the number of times the sample is changed from correct prediction to incorrect prediction;

[0097] calculating a second statistical quantity, the second statistical quantity being the number of consecutive incorrect prediction rounds of the sample, and counting the maximum number of consecutive incorrect predictions of the sample;

[0098] calculating a third statistical quantity, the third statistical quantity being a combination of sample prediction accuracy and loss value, and being calculated based on the cumulative number of correct predictions and the average loss of the sample;

[0099] normalizing the first statistical quantity, the second statistical quantity and the third statistical quantity respectively;

[0100] adding the three normalized statistical quantities to obtain the FoQuS value.

[0101] Further, the calculation of the FoQuS comprehensive quantitative value in the step S3 comprises three intermediate statistical quantities:

[0102] 1) the first statistical quantity based on the number of changes in sample prediction accuracy, counting the number of times the sample is changed from correct prediction to incorrect prediction:

[0103] (6)

[0104] 2) the second statistical quantity based on the number of consecutive incorrect prediction rounds of the sample, counting the number of consecutive incorrect predictions of the sample:

[0105] (7)

[0106] 3) the third statistical quantity based on the sample prediction accuracy and the loss value, comprehensively considering the correct classification stability and the loss size:

[0107] (8)

[0108] in, For example The cumulative number of correct predictions, Let be the cross-entropy loss in round t. This is the balance factor (default is 0.1).

[0109] Ultimately, FoQuS was obtained through comprehensive quantification:

[0110] (9)

[0111] in This indicates a normalization operation.

[0112] Furthermore, the physical meaning of the three intermediate statistics in step S3 is as follows:

[0113] The first statistic, based on the number of times the sample prediction accuracy changes, typically consists of samples located near the class decision boundary. These samples are sensitive to small perturbations in the model parameters, and including such samples helps improve the model's generalization ability and boundary stability. The second statistic, based on the number of consecutive incorrect prediction rounds, typically consists of difficult samples with low signal-to-noise ratios or complex features, which are crucial for the model's robustness in noisy environments. The third statistic, based on the sample prediction accuracy and loss value, typically consists of samples with moderate SNR and clear features, which can provide the model with clear and consistent learning signals, ensuring stable training convergence.

[0114] As one implementation method in this embodiment, in S4, the process of dynamically selecting the core set includes:

[0115] Sort all samples in descending order based on their FoQuS values;

[0116] The sorted sample sequence is divided into three selection intervals, including the first selection interval, the second selection interval, and the third selection interval;

[0117] The number of samples selected for each selection interval is adaptively determined based on the target sampling rate, wherein the sampling ratio of each selection interval is different when the target sampling rate is different;

[0118] Samples are selected from each selection interval according to a predetermined sample selection number to form the core set.

[0119] Furthermore, the dynamic core set selection strategy for the comprehensive quantized value in step S4 is as follows:

[0120] According to the descending order of the comprehensive quantitative values (FoQuS) of all samples, the sample sequence is dynamically divided into three continuous selection intervals based on a preset division threshold:

[0121] 1) The first selection interval: the top 30% of the samples in the ordered sequence;

[0122] 2) The second selection interval: the middle 40% of the samples in the ordered sequence;

[0123] 3) The third selection interval: the last 30% of the samples in the ordered sequence;

[0124] The division threshold and the subsequent sampling ratio of each selection interval are configurable parameters preset by the method, which can be adaptively adjusted according to the target sampling rate and other actual technical requirements, forming a dynamic interval division and quota allocation mechanism based on the ordered position, which together constitutes the final core set S.

[0125] Further, the dynamic core set selection strategy of the comprehensive quantitative value in the step S4 can be dynamically adjusted according to the target sampling rate k:

[0126] When k≤1%, the first selection interval accounts for 10%, the second selection interval accounts for 40%, and the third selection interval accounts for 50%;

[0127] When 1%<k≤5%, the first selection interval accounts for 15%, the second selection interval accounts for 45%, and the third selection interval accounts for 40%;

[0128] When 5%<k≤20%, the first selection interval accounts for 25%, the second selection interval accounts for 50%, and the third selection interval accounts for 25%;

[0129] When k>20%, the first selection interval accounts for 35%, the second selection interval accounts for 50%, and the third selection interval accounts for 15%.

[0130] This adaptive strategy ensures that key samples are preferentially retained at very low sampling rates and sample diversity is increased at higher sampling rates.

[0131] As an embodiment in this embodiment, in S5, the process of model training and performance verification includes:

[0132] retraining the deep learning model using the selected core set;

[0133] evaluating the performance of the retrained model on the test set;

[0134] The effectiveness of the core set selection method is verified through comparative experiments.

[0135] Based on this, the embodiment of the present application provides a kind of automatic modulation recognition core set selection method based on forgetting quality score, and the present application has the following beneficial effects compared with prior art:

[0136] 1) Strong pertinence: specially designed for AMR tasks, fully considering the special properties of signal data such as time-frequency structure, signal-to-noise ratio distribution, etc., overcoming the limitations of direct transfer of image field methods;

[0137] 2) Comprehensive evaluation: through the combination of three complementary statistics, the importance of samples is explained from multiple angles such as decision boundary stability, difficult sample coverage and basic learning support, avoiding the one-sidedness of single statistics;

[0138] 3) Diversity guarantee: a stratified sampling strategy is adopted to ensure that the core set contains key samples and covers auxiliary samples, avoiding excessive bias towards a single type and ensuring balanced distribution of selected samples under different signal-to-noise ratios;

[0139] 4) Significant effect: verified on multiple public AMR datasets, only 1%-30% of the data can maintain an accuracy close to that of the full amount of data, for example, using 1% of the data on the RML2016.10a dataset achieves an accuracy of 54.10%, significantly better than other benchmark methods;

[0140] 5) Strong generalization ability: the selected core set maintains stable performance on multiple models such as CNN1D, CNN2D and LSTM, proving the universality of the method;

[0141] 6) High practical value: can significantly reduce computing and storage overhead, accelerate model development iteration, and has important application value in resource-constrained edge computing environments, providing an effective solution for efficient development of AMR tasks.

[0142] S1: Dataset division and preprocessing:

[0143] This embodiment uses the RML2016.10a dataset, which is a widely used benchmark dataset in the field of automatic modulation recognition. The specific parameters of the dataset are as follows:

[0144] Modulation category: contains 11 modulation methods, including 8PSK, BPSK, QPSK, 16QAM, 64QAM, BFSK, CPFSK, PAM4, and 3 analog modulations WB-FM, AM-SSB and AM-DSB;

[0145] Data format: each sample is IQ dual-channel time-domain data with 128 sampling points;

[0146] SNR range: from -20dB to 18dB, uniformly distributed with 2dB interval, totally 20 SNR levels;

[0147] Data size: 220000 samples in total, 1000 samples per modulation class per SNR level.

[0148] The preprocessing steps specifically include:

[0149] 1) Amplitude normalization: normalize the IQ data of each sample to eliminate the influence of amplitude deviation, the normalization formula is .

[0150] where x is the original IQ data, and are the mean and standard deviation of the sample, respectively.

[0151] 2) Dataset division: divide the dataset into training set and test set with a ratio of 4:1. The division process adopts stratified random sampling strategy to ensure that:

[0152] the proportion of each modulation class in the training set and test set is consistent;

[0153] the samples of each SNR level are uniformly distributed in the two sets;

[0154] The training set contains 176000 samples, and the test set contains 44000 samples.

[0155] This stratified division strategy avoids the distribution bias introduced by data division, ensuring the fairness of subsequent core set selection and model evaluation.

[0156] S2: Full dataset training and trajectory recording;

[0157] This step uses the CNN1D model as the basic architecture to train on the complete training set, while recording detailed training dynamic information.

[0158] (1) Model architecture: The CNN1D model contains 4 convolutional layers, each followed by batch normalization and ReLU activation function, and finally outputs the probability distribution of 11 classes through the fully connected layer. The specific structure is:

[0159] Conv1D(64, kernel_size=3)→BatchNorm→ReLU→MaxPool

[0160] Conv1D(128, kernel_size=3)→BatchNorm→ReLU→MaxPool

[0161] Conv1D(256, kernel_size=3)→BatchNorm→ReLU

[0162] Conv1D(256, kernel_size=3)→BatchNorm→ReLU→GlobalAvgPool

[0163] Dense(128)→ReLU→Dropout(0.5)

[0164] Dense(11)→Softmax

[0165] (2) Training parameters:

[0166] Training rounds T = 50 epochs;

[0167] Optimizer: SGD (Stochastic Gradient Descent), learning rate lr = 0.01, momentum momentum = 0.9;

[0168] Batch size batch_size = 128;

[0169] Loss function: Cross-entropy loss;

[0170] Learning rate scheduling: Use cosine annealing strategy to adjust learning rate periodically.

[0171] (3) Trajectory recording mechanism: During the training process, for each sample Record the following information:

[0172] ① Prediction label trajectory: After the end of the t-th epoch, record the predicted label of sample i by the model:

[0173] (10)

[0174] Where represents the probability of the c-th class output by the model, and the arg max operation selects the class with the maximum probability as the predicted label.

[0175] ② Correctness trajectory: Based on the predicted label and the true label, calculate the correctness of the sample in the t-th round:

[0176] (11)

[0177] ③ Loss accumulation: Record the cross-entropy loss of sample in the t-th round , and accumulate the total loss:

[0178] (12)

[0179] Where the cross-entropy loss is defined as:

[0180] (13)

[0181] i.e., the negative log of the model's predicted probability of the true class.

[0182] (4) Cumulative number of correct predictions:

[0183] (14)

[0184] This value reflects the total number of times the sample was correctly classified throughout the entire training process.

[0185] Through the complete 50 rounds of training, each sample generates a complete behavior trajectory containing the prediction trajectory, correctness sequence, and loss record, providing a sufficient data basis for the subsequent calculation of the FoQuS comprehensive quantitative value.

[0186] As shown in Figure 2 , the calculation of the FoQuS comprehensive quantitative value includes the calculation of three intermediate statistical quantities and the standardization combination process.

[0187] S3: Calculation of the FoQuS comprehensive quantitative value:

[0188] Based on the training trajectory recorded in step S2, this step calculates three intermediate statistical quantities and combines them into the final FoQuS comprehensive quantitative value.

[0189] (1) Calculation of the first statistical quantity based on the number of changes in sample prediction correctness:

[0190] The first statistical quantity is used to capture the number of times a sample changes from correct prediction to incorrect prediction during training, reflecting its instability near the decision boundary, and the calculation formula is formula (6). The physical meaning of this index: samples that change from correct prediction to incorrect prediction are usually located near the class decision boundary, and these samples are sensitive to small perturbations in model parameters and are easily forgotten. Including such samples helps to improve the generalization ability and boundary stability of the model.

[0191] (2) Calculation of the second statistical quantity based on the number of consecutive prediction errors:

[0192] The second statistical quantity counts the number of consecutive incorrect predictions of a sample, which is used to identify stubborn samples that are difficult to fit, and the calculation formula is formula (7). The physical meaning of this index: samples with a high number of consecutive errors are usually difficult samples with low signal-to-noise ratio or complex features. These samples are difficult to learn, but they are crucial for the robustness of the model in a noisy environment, as these samples can avoid the core set containing only easy-to-learn high SNR samples.​

[0193] (3) Calculation of the third statistic based on the sample prediction accuracy and the loss value:

[0194] The third statistic Taking into account the stability of the correct classification of the sample and the average loss in the training process, the overall learning value is described. The calculation formula is formula (8).

[0195] Wherein: is the cumulative number of correct predictions of the sample, The larger the value, the more stable the sample, and the second term is the average loss of the sample. The larger the loss, the more difficult the sample is to learn. is a balance factor, which is set to 0.1 in the present embodiment, and the third statistic index ensures that the core set contains enough easy-to-learn samples to provide stable gradient information for the model. The design idea of the third statistic: not all samples should be "difficult samples", and appropriate easy-to-learn samples help the model converge quickly. Samples with high third statistics are usually samples with medium SNR and clear features, which can provide clear and consistent learning signals for the model.

[0196] (4) FoQuS comprehensive quantitative value combination:

[0197] In order to eliminate the dimensional differences of the three statistics, first, each statistic is normalized to the minimum-maximum:

[0198] (15)

[0199] Then add the three normalized statistics to get the final FoQuS comprehensive quantitative value, as shown in formula (9). The advantage of this comprehensive quantitative value is that the first statistic focuses on samples near the decision boundary, improving the generalization ability, the second statistic focuses on difficult samples, improving the noise robustness. The third statistic focuses on stable learning to ensure training convergence. The three complement each other to ensure the comprehensiveness and diversity of the core set.

[0200] In the present embodiment, after calculating the FoQuS comprehensive quantitative value for 176,000 training samples, the distribution of the comprehensive quantitative value presents an approximate normal distribution, ranging from 0.15 to 2.85, with an average value of about 1.5.

[0201] S4: Hierarchical core set selection:

[0202] This step selects the core set according to the FoQuS comprehensive quantitative value using a hierarchical proportional sampling strategy to ensure the diversity and representativeness of the selected samples.

[0203] (1) Sample sorting and layering:

[0204] All 176000 training samples are arranged in descending order of FoQuS comprehensive quantification value, and then divided into three levels:

[0205] The first selection interval: the first 30% of the samples, i.e. the first 52800 samples. The FoQuS of these samples is between 2.0-2.85, and the samples contained generally have the following characteristics: moderate first statistics (3-6 times); certain second statistics times (10-25 times); higher first statistics; relatively balanced SNR distribution.

[0206] The second selection interval: the middle 40% of the samples, i.e. the middle 70400 samples. The FoQuS of these samples is between 1.2-2.0, the first statistics is less (1-3 times); the second statistics times is moderate (5-15 times); the third statistics is moderate; and it is mainly distributed in the medium-high SNR range.

[0207] The third selection interval: the last 30% of the samples, i.e. the last 52800 samples. The FoQuS of these samples is between 0.15-1.2, which are usually high-SNR easy-to-learn samples; but still have certain value for basic training of the model.

[0208] (2) Stratified proportional sampling:

[0209] Set the target sampling rate k (for example, k=0.05 means selecting 5% of the data), and extract samples from each level in the following proportions:

[0210] Select 30% of the target number from the first selection interval: n_first=0.30×k×|D|;

[0211] Select 50% of the target number from the second selection interval: n_second=0.50×k×|D|;

[0212] Select 20% of the target number from the third selection interval: n_third=0.20×k×|D|;

[0213] Taking k=0.05 as an example, the target core set size is 176000×0.05=8800 samples, and the specific allocation is as follows:

[0214] First selection interval: 2640 samples;

[0215] Second selection interval: 4400 samples;

[0216] Third selection interval: 1760 samples;

[0217] Sampling strategy description: Instead of using the simple "select the global top-k samples" strategy, through stratified proportional sampling, ensure that the core set covers samples of different importance levels. The first selection interval accounts for 30%, ensuring that enough key samples are included; the second selection interval accounts for 50%, serving as the main body of the core set, balancing performance and diversity; the third selection interval accounts for 20%, retaining some simple samples to maintain training stability.

[0218] (3) Core set construction:

[0219] Within each level, samples are selected in order of FoQuS comprehensive quantitative value from high to low, and the core set S is finally formed. In this embodiment, the core set under a 5% sampling rate includes: 11 modulation categories are covered, 20 SNR levels are represented, the signal-to-noise ratio distribution is basically the same as the original training set, and samples of different difficulty levels are evenly distributed.

[0220] S5: Performance verification:

[0221] The selected core set is used to retrain the CNN1D model, and the performance is evaluated on the test set to verify the effectiveness of the core set selection method.

[0222] (1) Core set training:

[0223] The training settings are completely consistent with step S2, including: the same CNN1D model structure; 50 training rounds; SGD optimizer; learning rate of 0.01; momentum is 0.9; batch size is 128; cross-entropy loss function. The only difference is that the training data set is reduced from the complete 176,000 samples to the core set of 8,800 samples (5% sampling rate).

[0224] (2) Performance evaluation:

[0225] The recognition accuracy of the model is evaluated on 44,000 test samples. The evaluation indicators include:

[0226] The accuracy of the baseline and the author's method when 1%, 5%, 10%, 20%, and 30% of the 176,000 samples in the training set are selected.

[0227] (3) Comparison of experimental results:

[0228] Table 1 shows the comparison results of the method of the present application and 10 benchmark methods on the RML2016.10a data set.

[0229] Table 1

[0230]

[0231] The ablation experiment verifies the complementarity of the three statistics, and the complete FoQuS comprehensive quantitative value performs better than any single statistic or two-by-two combination, which illustrates the importance of diverse statistics.

[0232] Compared with the existing core set selection method, the automatic modulation recognition core set selection method based on the FoQuS comprehensive statistics of the present application considers the special properties of signal data, uses multiple statistics to evaluate sample importance, and adopts a hierarchical sampling strategy to ensure diversity, so that the training data can be significantly reduced while maintaining good recognition performance, thereby providing an effective solution for efficient development of AMR tasks.

[0233] Embodiment two

[0234] In this embodiment, a computer terminal device is provided, comprising:

[0235] one or more processors;

[0236] a memory coupled to the processor(s) for storing one or more programs;

[0237] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the automatic modulation recognition core set selection method based on the forgetting quality score described above.

[0238] In this embodiment, a computer readable storage medium having a computer program stored thereon is also provided, and the computer program is executed by a processor to implement the steps of the automatic modulation recognition core set selection method based on the forgetting quality score described above.

[0239] In this embodiment, an electronic device is also provided, comprising a memory and a processor, the memory having a computer program stored therein, and the processor is configured to run the computer program to perform the steps of the automatic modulation recognition core set selection method based on the forgetting quality score described above.

[0240] In this embodiment, a computer program product is also provided, comprising a computer program, which is executed by a processor to implement the steps of the automatic modulation recognition core set selection method based on the forgetting quality score described above.

[0241] The above-described processes can be implemented on a processor, or can be stored in a memory (or referred to as a computer readable medium), which includes permanent and non-permanent, removable and non-removable media, and can be realized by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device.

[0242] These computer programs can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to generate a computer implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the flow Figure 1 The flow or multiple flows and / or the block Figure 1 The steps of the function specified in one or more blocks can be implemented by different modules corresponding to different steps.

[0243] An apparatus or system is provided in the embodiment. The system is referred to as an automatic modulation recognition core set selection system based on forgetting quality score, which includes:

[0244] A preprocessing module for preprocessing the original training data set, including signal amplitude normalization and hierarchical random division, to obtain a training set and a test set;

[0245] A training trajectory recording module for using a deep learning model to perform multiple rounds of training on the training set, and recording the prediction label, correctness indication and cross-entropy loss of each sample at each training round;

[0246] A FoQuS calculation module for calculating the forgetting quality score FoQuS of each sample based on the recorded training trajectory, wherein the FoQuS is obtained by fusing three complementary statistics, including the number of changes in sample prediction correctness, the number of continuous prediction errors of the sample, and the combination of sample prediction accuracy and loss value;

[0247] The core set selection module is configured to sort all samples in descending order according to the FoQuS, dynamically divide the samples into multiple selection intervals, adaptively determine the number of samples to be selected from each selection interval according to a target sampling rate, and select samples from each selection interval to form a core set.

[0248] The model training and verification module is configured to retrain a model using the core set and evaluate the performance of the model on a test set.

[0249] As an implementation in the embodiment, the preprocessing module includes:

[0250] The normalization unit is configured to perform amplitude normalization on the IQ data of each signal sample, and achieve standardization by calculating the mean and standard deviation of the sample.

[0251] The division unit is configured to divide the data set into a training set and a test set using a hierarchical random sampling strategy, and ensure that the distribution of each modulation category and each signal-to-noise ratio level in the training set and the test set is consistent with that in the original data set.

[0252] As an implementation in the embodiment, the training trajectory recording module includes:

[0253] The predicted label recording unit is configured to perform multiple rounds of training, and for each sample in each training round, record the predicted label of the model.

[0254] The correctness indication calculation unit is configured to calculate the correctness indication of the sample based on the predicted label and the true label, and the correctness indication is a binary value indicating whether the prediction is correct.

[0255] The loss recording unit is configured to record the cross-entropy loss of the sample in the current training round.

[0256] The accumulation unit is configured to accumulate the number of correct predictions and the total cross-entropy loss of each sample.

[0257] As an implementation in the embodiment, the FoQuS calculation module includes:

[0258] The first statistical quantity calculation unit is configured to calculate a first statistical quantity, which is the number of changes in the prediction correctness of the sample, and counts the number of times the sample changes from correct prediction to incorrect prediction.

[0259] The second statistical quantity calculation unit is configured to calculate a second statistical quantity, which is the number of consecutive incorrect predictions of the sample, and counts the maximum number of consecutive incorrect predictions of the sample.

[0260] The third statistical quantity calculation unit is configured to calculate a third statistical quantity, which is a combination of the prediction accuracy and the loss value of the sample, and is calculated based on the cumulative number of correct predictions and the average loss of the sample.

[0261] a normalization processing unit configured to normalize the first statistical quantity, the second statistical quantity, and the third statistical quantity, respectively;

[0262] a fusion unit configured to add the normalized three statistical quantities to obtain a FoQuS value.

[0263] As an implementation in the embodiment, the core set selection module comprises:

[0264] a sorting unit configured to sort all samples in descending order according to the FoQuS value;

[0265] an interval division unit configured to divide the sorted sample sequence into three selection intervals, including a first selection interval, a second selection interval, and a third selection interval;

[0266] a sampling number determination unit configured to adaptively determine the sample selection number of each selection interval according to a target sampling rate, wherein the sampling proportion of each selection interval is different when the target sampling rate is different;

[0267] a sample selection unit configured to select samples from each selection interval according to the determined sample selection number to form a core set.

[0268] As an implementation in the embodiment, the model training and verification module comprises:

[0269] a core set training unit configured to retrain the deep learning model using the selected core set;

[0270] a performance evaluation unit configured to evaluate the performance of the retrained model on a test set;

[0271] a verification unit configured to verify the effectiveness of the core set selection method through a comparative experiment.

[0272] The system or device is used to realize the functions of the methods in the above embodiments. Each module in the system or device corresponds to each step in the method, and has been described in the method and will not be repeated here.

[0273] Through the above implementation, the problem of automatic modulation recognition core set selection based on the forgetting quality score in the related art is solved, thereby ensuring that the problems in the prior art are solved.

[0274] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements within the technical scope disclosed by the present application can be easily thought of by those skilled in the art, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for automatic modulation identification core set selection based on a forgetting quality score, characterized in that, The method comprises the following steps: S1, preprocessing the original data set, including signal amplitude normalization and hierarchical random division, to obtain a training set and a test set; S2, using a deep learning model to perform multiple rounds of training on the training set, and recording the prediction label, correctness indication, and cross-entropy loss of each sample at each training round; S3, based on the recorded training track, calculating the forgetting quality score FoQuS of each sample, which is obtained by fusing three complementary statistics, including the number of changes in sample prediction correctness, the number of consecutive prediction errors, and the combination of sample prediction accuracy and loss value; S4, sorting all samples in descending order according to the FoQuS, and dynamically dividing them into multiple selection intervals, adaptively determining the number of selected samples in each selection interval according to the target sampling rate, and selecting samples from each selection interval to form a core set; S5, retraining the model using the core set and evaluating the model performance on the test set.

2. The method of claim 1, wherein, In S1, the preprocessing process includes: Amplitude normalization processing is performed on the IQ data of each signal sample, and standardization is achieved by calculating the mean and standard deviation of the sample; The hierarchical random sampling strategy is used to divide the data set into a training set and a test set, ensuring that the distribution of each modulation category and each signal-to-noise ratio level in the training set and the test set is consistent with that of the original data set.

3. The method of claim 1, wherein, In S2, the training and recording process includes: Multiple rounds of training are performed, and for each sample at each training round, the prediction label of the model is recorded; Based on the prediction label and the true label, the correctness indication of the sample is calculated, which is a binary variable indicating whether the prediction result is correct; The cross-entropy loss of the sample at the current training round is recorded; The number of correct predictions and the total cross-entropy loss of each sample are accumulated.

4. The method of claim 1, wherein, In S3, the process of calculating FoQuS includes: Calculate the first statistic, which is the number of changes in sample prediction correctness, and count the number of times the sample changes from correct prediction to incorrect prediction; Calculate the second statistic, which is the number of consecutive incorrect predictions, and count the maximum number of consecutive incorrect predictions for the sample; Calculate the third statistic, which is the combination of sample prediction accuracy and loss value, based on the cumulative number of correct predictions and the average loss; Normalize the first, second, and third statistics respectively; Add the normalized three statistics to get the FoQuS value.

5. The method of claim 1, wherein, In S4, the process of dynamically selecting the core set includes: Sort all samples in descending order according to the FoQuS value; Divide the sorted sample sequence into three selection intervals, including the first selection interval, the second selection interval, and the third selection interval; Adaptively determine the number of selected samples in each selection interval according to the target sampling rate, where the sampling proportion of each selection interval is different when the target sampling rate is different; Select samples from each selection interval according to the determined number of selected samples to form a core set.

6. The method of claim 1, wherein, In S5, the process of model training and performance verification includes: Re-train the deep learning model using the selected core set; Evaluate the performance of the retrained model on the test set; The effectiveness of the core set selection method is verified through comparative experiments.

7. An automatic modulation recognition core set selection system based on forgetting quality scoring, characterized in that, The system is used to implement the method of any one of claims 1-6, and the system comprises: a preprocessing module configured to preprocess an original training dataset, including signal amplitude normalization and hierarchical random division, to obtain a training set and a test set; a training trajectory recording module configured to perform multiple rounds of training on the training set using a deep learning model, and record the predicted label, correctness indication, and cross-entropy loss of each sample in each training round; a FoQuS calculation module configured to calculate the forgetting quality score FoQuS of each sample based on the recorded training trajectory, wherein the FoQuS is obtained by fusing three complementary statistics, including the number of changes in sample prediction correctness, the number of consecutive prediction errors of the sample, and the combination of sample prediction accuracy and loss value; a core set selection module configured to sort all samples in descending order according to the FoQuS, dynamically divide the samples into multiple selection intervals, adaptively determine the number of selected samples in each selection interval according to a target sampling rate, and select samples from each selection interval to form a core set; a model training and verification module configured to retrain the model using the core set and evaluate the performance of the model on the test set.

8. A computer terminal device, characterized by comprise: one or more processors; a memory coupled to the processors therefor storing one or more programs; when the one or more programs are executed by the one or more processors, so that the one or more processors implement the steps of the method of any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1-6. The computer program is executed by the processor to implement the steps of the method of any one of claims 1-6.

Images