A signal recognition method and system based on multi-class feature learning, an electronic device and a storage medium

Abstract

The application discloses a signal recognition method and system based on multi-class feature learning, electronic equipment and storage medium, belonging to the field of closed set recognition, the method comprises the following steps: extracting multi-class features of the original signal, judging whether to analyze the importance of each dimension of the features according to the dimension height of the features, evaluating the importance of each dimension of the high-dimensional features, and selecting the important dimensions in the high-dimensional features; the important high-dimensional features or low-dimensional features are sent into the feature learning network, the category of the signal is recognized, the recall rate of the signal features under different signal categories is calculated, and a recall rate matrix is constructed; the whole test set is divided into two parts, one part of the test signals is used to correct the recall rate matrix, and the category of the other part of the test signals is recognized through a multi-feature decision algorithm. The problem that different signals have different distinguishability on different features and that a high-quality signal multi-classification task cannot be completed through a single feature is solved, and the signal recognition capability and accuracy are further improved.

Description

Technical Field

[0001] This invention belongs to the field of closed set recognition, specifically relating to a signal recognition method, system, electronic device, and storage medium based on multi-class feature learning. Background Technology

[0002] Signal features are diverse. When dealing with multi-class signal problems, some signal categories may be highly discriminative on one feature but poorly discriminative on another, while other categories may exhibit the opposite behavior. This makes traditional AI classification methods, which learn and classify the entire dataset based on the same or multiple types of features, less than satisfactory. Therefore, designing a signal recognition method or system specifically for multi-class signals with such characteristics, based on multi-class feature learning, is of great significance.

[0003] Patent application CN110674725B proposes a device signal type identification method based on the combination of multidimensional feature vectors of detection signals. However, this paper only discusses the selection of the optimal feature from multiple low-dimensional features. The combination of feature vectors based on feature statistics in this paper does not solve the task of selecting the appropriate feature vector from multiple high-dimensional features.

[0004] The patent application document with publication number CN109446877B proposes to directly splice the multidimensional features of the signal and then feed them into a deep neural network for learning. This method can achieve good results on a radar signal set composed of nine radar signals. However, this method does not specifically extract the best features of different types of signals. In some scenarios, the multi-class recognition effect may be poor. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a signal recognition method, system, electronic device, and storage medium based on multi-class feature learning. By exploring the relationship between prior categories and features, it focuses on solving the technical problem that some categories of signals may have good distinguishability with certain features, while other categories of signals may have good distinguishability with other features. It also addresses the issue that high-quality multi-class signal classification tasks cannot be completed using only a single feature, thereby further improving the signal recognition capability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A signal recognition method based on multi-class feature learning includes the following steps:

[0008] Step 1: Extract multi-class features from the original signal. Determine whether to analyze the importance of each dimension of the feature based on its dimensionality. Evaluate the importance of each dimension of the high-dimensional features and select the important dimensions among the high-dimensional features. No feature selection is performed for low-dimensional features.

[0009] Step 2: Design a feature learning network, feed important high-dimensional or low-dimensional features into the feature learning network, identify the categories of different signal samples, obtain the multi-class recognition results under each signal feature, calculate the recall rate of signal features under different signal categories, and construct a recall matrix.

[0010] Step 3: Divide the entire test set into two parts, use one part of the test signal samples to correct the recall matrix, and use a multi-feature decision algorithm to identify the category to which the other part of the test signal samples belong.

[0011] To optimize the above technical solution, the specific measures also include:

[0012] Furthermore, in step 1, the evaluation of the importance of each dimension of the high-dimensional features and the selection of the important dimensions among the high-dimensional features specifically involves:

[0013] Step 1.1: Construct I decision trees using the bootstrap algorithm;

[0014] Step 1.2: Calculate the Gini impurity of the split node k in the i-th decision tree.

[0015]

[0016] In the formula, C represents the number of signal categories. Let k be the probability that the split node k belongs to the c-th type of signal;

[0017] Step 1.3: Calculate the feature dimension X j Feature importance of split node k in the i-th decision tree

[0018]

[0019] In the formula, This represents the Gini impurity of the new node l after branching. This represents the Gini impurity of the new node r after branching;

[0020] Step 1.4: Calculate the feature dimension X j Importance of VIM in the whole forest j ;

[0021]

[0022] In the formula, K is the total number of nodes in the decision tree;

[0023] Step 1.5: Normalize the importance of all feature dimensions;

[0024]

[0025] In the formula, VIM j 'X is the normalized feature dimension' j The importance of J in the entire forest, where J is the total number of feature dimensions;

[0026] Step 1.6: Filter out the dimensions whose feature importance is higher than the threshold, and arrange these dimensions in the order of the index in which the feature dimension was extracted to obtain the important feature dimensions of the high-dimensional feature. The threshold is selected by the average value of the importance scores of all feature dimensions.

[0027] Further, in step 2, the feature learning network includes weighted convolutional layers and fully connected layers; from the one-dimensional data input end to the multi-classification result output end, it sequentially passes through a 1*7 convolutional layer, a batch normalization layer, a ReLU layer, a 1*3 max pooling layer, a Res_2 layer, a Res_1 layer, an adaptive average pooling layer, a fully connected layer, and a linear layer; the Res_2 layer includes a 1*3 convolutional layer, a 1*1 convolutional layer, a ReLU layer, and a batch normalization layer; the Res_1 layer includes a 1*3 convolutional layer, a ReLU layer, and a batch normalization layer.

[0028] Further, in step 2, the calculation of the recall rate of signal features under different signal categories and the construction of the recall rate matrix are specifically as follows:

[0029] The formulas for calculating the recall R(m,n) of signal features under different signal categories are as follows:

[0030]

[0031] In the formula, TP mn FN represents a sample whose true signal class is class n and whose predicted class is also class n, given the m-th feature. mn This represents a sample whose true signal category is the nth category but is predicted as another category under the m-th feature.

[0032] The original recall matrix is ​​obtained by plotting feature category m on the x-axis and signal category n on the y-axis.

[0033] Furthermore, step 3 specifically involves:

[0034] Step 3.1: Divide the entire test set into two parts, and use a portion of the test signal samples to build the corrected recall matrix;

[0035] Design several alternative thresholds Thre Different thresholds T were obtained. hre Multiple recall matrices R + (m,n):

[0036]

[0037] Find the threshold that maximizes recall.

[0038]

[0039] The corrected recall matrix R * (m,n) is:

[0040]

[0041] Step 3.2: Select a test signal sample from another part of the test data as input, output the multi-feature decision result, and identify the signal category to which the test signal belongs;

[0042] Specifically, a test signal sample is input from another part of the test data. If most of the features of the signal sample indicate that the signal sample belongs to a certain category of signal, then the sample is considered to belong to that category of signal. Otherwise, the corrected recall matrix needs to be queried, and the signal sample is considered to belong to the category corresponding to the maximum recall value.

[0043] The test signal sample category set C = {1,2,...,c}, the number of categories determined after the test signal samples are identified is M0, and the signal categories c1, c2,...,c are determined based on signal characteristics. z ∈C, Multi-feature decision result C + The formula is expressed as follows:

[0044]

[0045] In the formula, z represents the total number of feature categories.

[0046] The present invention also proposes a signal recognition system based on multi-class feature learning, the system comprising a feature selection module, a feature learning module, and a feature decision module;

[0047] The feature selection module is used to extract multiple features from the original signal, determine whether to analyze the importance of each dimension of the feature based on the dimensionality of the feature, evaluate the importance of each dimension of the high-dimensional feature, and select the important dimensions among the high-dimensional features.

[0048] The feature learning module is used to design a feature learning network, feed important high-dimensional or low-dimensional features into the feature learning network, identify the categories of different signal samples, obtain multi-class recognition results under each signal feature, calculate the recall rate of signal features under different signal categories, and construct a recall matrix.

[0049] The feature decision module is used to modify the recall matrix using a portion of the test signal samples and to identify the category of another portion of the test signal samples through a multi-feature decision algorithm.

[0050] The present invention also proposes an electronic device, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the signal recognition method based on multi-class feature learning as described above.

[0051] The present invention also proposes a computer-readable storage medium storing a computer program that enables a computer to execute the signal recognition method based on multi-class feature learning as described above.

[0052] The beneficial effects of this invention are:

[0053] This invention is a data analysis system developed using only the hardware resources and software platform of a computer. It features a simple structure, high flexibility, and low cost. By mining the relationship between prior categories and features, it differs from traditional supervised multi-class classification methods in that it does not directly feed the raw data into a feature learning network to obtain multi-class results. Instead, it designs a signal recognition system integrating multi-class feature extraction, selection, learning, and decision-making. For feature extraction, it utilizes time-domain, frequency-domain, and time-frequency-domain signal analysis techniques to extract multi-class high-dimensional features from the raw signal. For feature selection, an algorithm is designed to autonomously select the most important feature dimensions from each class of high-dimensional features, further mining the unique high-dimensional features of different signal classes and improving the potential for multi-class recall. For feature learning, a one-dimensional ResNet network architecture is designed based on the characteristics of the input signal. For feature decision-making, a feature decision-making scheme is designed to mine the potential recall value of each feature on each class of signal samples. In practical application, under the same feature learning network framework, this invention significantly outperforms traditional supervised multi-class classification methods in terms of recognition performance. Attached Figure Description

[0054] Figure 1 This is a flowchart of the method;

[0055] Figure 2 Score the phase features and their corresponding importance;

[0056] Figure 3 Scoring the three-level wavelet features and their corresponding importance;

[0057] Figure 4a This is the feature learning network proposed in this invention;

[0058] Figure 4b This is a schematic diagram of the Res_1 layer structure of the feature learning network;

[0059] Figure 4c This is a schematic diagram of the Res_2 layer structure of a feature learning network;

[0060] Figure 5 This is a multi-class confusion matrix for the original features and phase features. Detailed Implementation

[0061] The invention will now be described in further detail with reference to the accompanying drawings.

[0062] In one embodiment, the present invention proposes a signal recognition method based on multi-class feature learning, the flowchart of which is as follows: Figure 1 As shown, the specific steps include:

[0063] Step 1: Feature Extraction and Feature Selection. Extract multiple types of features from the original signal. Determine whether to analyze the importance of each dimension of the feature based on its dimensionality. Evaluate the importance of each dimension for high-dimensional features and select the important feature dimensions for each type of high-dimensional feature. No feature selection is performed for low-dimensional features.

[0064] Multiple features include time-domain features, frequency-domain features, and time-frequency features. Time-domain features include root mean square, peak factor, impulse factor, margin factor, kurtosis factor, waveform factor, and skewness, etc. Frequency-domain features include spectrum, energy spectrum, power spectrum, cepstrum, Fourier transform, etc. Time-frequency features include short-time Fourier transform, wavelet transform, Hilbert transform, etc.

[0065] High-dimensional features refer to non-original signal features with a feature dimension greater than 3, while low-dimensional features refer to signal features with a feature dimension less than or equal to 3.

[0066] The evaluation method for important feature dimensions is the Gini impurity of the feature dimension, and the specific steps are as follows:

[0067] Step 1.1: Construct I decision trees using the bootstrap algorithm;

[0068] Step 1.2: Calculate the Gini impurity of the split node k in the i-th decision tree.

[0069]

[0070] In the formula, C represents the number of signal categories. Let k be the probability that the split node k belongs to the c-th type of signal;

[0071] Step 1.3: Calculate the feature dimension X j Feature importance of split node k in the i-th decision tree

[0072]

[0073] In the formula, This represents the Gini impurity of the new node l after branching. This represents the Gini impurity of the new node r after branching;

[0074] Step 1.4: Calculate the feature dimension X j Importance of VIM in the whole forest j ;

[0075]

[0076] In the formula, K is the total number of nodes in the decision tree;

[0077] Step 1.5: Normalize the importance of all feature dimensions;

[0078]

[0079] In the formula, VIM j 'X is the normalized feature dimension' j The importance of J in the entire forest, where J is the total number of feature dimensions;

[0080] Step 1.6: Filter out the dimensions whose feature importance is higher than the threshold, and arrange these dimensions in the order of the index in which the feature dimension was extracted to obtain the important feature dimensions of the high-dimensional feature. The threshold is selected by the average value of the importance scores of all feature dimensions.

[0081] The original signal used in this embodiment is the orthogonal I and Q baseband signals obtained from an ADS-B receiver. This signal has 30 categories and a total of 11,937 samples. The extracted signal features include features reconstructed from three-level wavelet decomposition, phase features of the signal, and features of the original signal, totaling three categories of features. The phase features and their corresponding importance scores in this invention are as follows: Figure 2 As shown, the three-level wavelet features of this invention and their corresponding importance scores are as follows: Figure 3As shown, it can be seen that among the phase features, four consecutive segments have relatively high scores, and among the third-level wavelet features, the mid-to-low frequency features have relatively high scores. Therefore, these features are selected to complete the subsequent feature learning and feature decision-making tasks. Calculations show that the original phase feature input dimension is 200 dimensions, and the dimension of the high-importance phase features after feature selection is 74 dimensions; the original third-level wavelet input feature dimension is 459 dimensions, and the dimension of the high-importance third-level wavelet features after feature selection is 167 dimensions.

[0082] Step 2: Feature Learning. Design a feature learning network, and feed the important high-dimensional or low-dimensional features of the extracted original signal into the feature learning network to obtain the categories of different signals, the recall matrix under different signal feature categories, and the multi-class recognition results under each category of signal features.

[0083] The feature learning network consists of 16 weighted convolutional layers and fully connected layers. From the one-dimensional data input to the multi-class classification result output, the layers sequentially pass through a 1*7 convolutional layer, a batch normalization layer, a ReLU layer, a 1*3 max pooling layer, two Res_2 layers, four Res_1 layers, an adaptive average pooling layer, a fully connected layer, and a linear layer. The batch normalization layer has 64 channels; the Res_2 layers consist of two 1*3 convolutional layers, one 1*1 convolutional layer, two ReLU layers, and two batch normalization layers, with both Res_2 layers having 64 channels; the Res_1 layers consist of two 1*3 convolutional layers, two ReLU layers, and two batch normalization layers, with the first two Res_1 layers having 128 channels and the last two Res_1 layers having 256 channels.

[0084] The formulas for calculating the recall R(m,n) of signal features under different signal categories are as follows:

[0085]

[0086] In the formula, TP mn FN represents a sample whose true signal class is class n and whose predicted class is also class n, given the m-th feature. mn This represents a sample whose true signal category is the nth category but is predicted as another category under the m-th feature.

[0087] The original recall matrix is ​​obtained by plotting feature category m on the x-axis and signal category n on the y-axis.

[0088] In this embodiment, the ratio of the training set to the test set is set to 6:4, with 7162 and 4775 samples respectively. During feature learning, the labeled signal samples are randomly shuffled. The feature learning network is shown in Figure 4, and the original recall matrix is ​​shown in Table 1. The overall recall rates of the three-level wavelet features, original data features, and phase features are 86.08%, 84.61%, and 82.26%, respectively. Figure 5 The figure shows the multi-class confusion matrix of the original features and phase features under 30 categories. As can be seen from the figure, in terms of the original signal features, the signal of category 13 is difficult to identify, while the signal of category 14 has better discriminative power. In terms of phase features, some data of the signal of category 13 can be accurately identified, while the signal of category 14 is easily mistaken for category 20 and category 29. The features presented by this group of signals are consistent with the application scenario of the algorithm proposed in this invention: some categories have good discriminative power on one feature but poor discriminative power on another feature, while other categories show the opposite trend.

[0089] Table 1 Original Recall Matrix

[0090]

[0091]

[0092] Step 3: Feature Decision. The entire test set is divided into two parts. The recall matrix obtained in Step 2 is corrected using signal samples from one part of the test set. A multi-feature decision algorithm is then used to identify the category of the other part of the test signals. Specifically:

[0093] Step 3.1: Divide the entire test set into two parts, and use a portion of the test signal samples to build the corrected recall matrix;

[0094] Design several alternative thresholds T hre Different thresholds T were obtained. hre Multiple recall matrices R + (m,n):

[0095]

[0096] Find the threshold that maximizes recall.

[0097]

[0098] The corrected recall matrix R * (m,n) is:

[0099]

[0100] Step 3.2: Select a test signal sample from another part of the test data as input, output the multi-feature decision result, and identify the signal category to which the test signal belongs;

[0101] A test signal sample is input from another part of the test data. If most of the features of the signal sample indicate that the signal sample belongs to a certain category of signal, then the sample is considered to belong to that category of signal. Otherwise, the corrected recall matrix needs to be queried, and the signal sample is considered to belong to the category corresponding to the maximum recall.

[0102] The test signal sample category set C = {1,2,...,c}, the number of categories determined after the test signal samples are identified is M0, and the signal categories c1, c2,...,c are determined based on signal characteristics. z ∈C, Multi-feature decision result C + The formula is expressed as follows:

[0103]

[0104] In the formula, z represents the total number of feature categories.

[0105] In this embodiment, the test set includes Test Set 1 and Test Set 2, with a ratio of 2:1, containing 3183 and 1592 samples respectively. A corrected recall matrix is ​​built based on the signal samples from Test Set 1, and the category classification is tested for the signals from Test Set 2. When a sample is input from Test Set 2, if most features indicate that the sample belongs to a certain category, then the sample is considered to belong to that category; otherwise, the corrected recall matrix is ​​queried, and the sample is considered to belong to the category corresponding to the highest probability.

[0106] Experiments show that when the selection range of the candidate threshold is a sequence with a step size of 0.01, a minimum value of 0.8 and a maximum value of 1, the optimal threshold thre is determined to be 0.9 based on the signal features and corresponding labels in test set 1. The value of 1 in the original recall matrix is ​​replaced with the optimal threshold, and the corrected recall matrix is ​​shown in Table 2.

[0107] Table 2 Recall Matrix After Threshold Correction

[0108]

[0109]

[0110] For example, if a test sample in test set 2 is classified as class 5, class 23, and class 23 under these three features, the multi-feature decision formula can be used to directly determine that the sample belongs to class 23. Similarly, if a test sample is classified as class 8, class 21, and class 28 under the three-level wavelet feature, the original data feature, and the phase feature, and then consult Table 2, we can find that the three-level wavelet feature has a recall rate of 93.78% for class 8, the original data feature has a recall rate of 31.46% for class 21, and the phase feature has a recall rate of 13.73% for class 28. The class with the highest recall rate is class 8, and this strategy can be used to determine that the sample belongs to class 8.

[0111] Table 3 shows the performance of the three independent features and the proposed method on test set 1 and test set 2. The proposed method deeply mines the information of the three types of features, and its performance is far better than that of a single feature.

[0112] Table 3. Comparison of recall rates for individual features and proposed methods.

[0113]

[0114]

[0115] In another embodiment, the present invention proposes a signal recognition system based on multi-class feature learning, the system comprising a feature selection module, a feature learning module, and a feature decision module;

[0116] The feature selection module is used to extract multiple features from the original signal. It determines whether to analyze the importance of each dimension of the feature based on its dimensionality, evaluates the importance of each dimension of the high-dimensional feature, and selects the important dimensions among the high-dimensional features.

[0117] The feature learning module is used to design the feature learning network, feed important high-dimensional or low-dimensional features into the feature learning network, identify the categories of different signal samples, obtain the multi-class recognition results under each signal feature, calculate the recall rate of signal features under different signal categories, and construct the recall rate matrix.

[0118] The feature decision module is used to modify the recall matrix using a portion of the test set signal samples, and to identify the category of another portion of the test signals through a multi-feature decision algorithm.

[0119] In another embodiment, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the signal recognition method based on multi-class feature learning as described above.

[0120] In another embodiment, the present invention provides a computer-readable storage medium characterized in that it stores a computer program that causes a computer to perform the signal recognition method based on multi-class feature learning as described above.

[0121] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0122] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0123] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A signal recognition method based on multi-class feature learning, characterized in that, The method comprises the following steps: Step 1: extracting multi-class features of original signals, the original signals being IQ two-way baseband signals obtained in an ADS-B receiver; According to the dimension of the features, it is judged whether to analyze the importance of each dimension of the features, the importance of each dimension of the high-dimensional features is evaluated, and important dimensions in the high-dimensional features are selected; The low-dimensional features are not selected; Step 2: designing a feature learning network, inputting the important high-dimensional features or low-dimensional features into the feature learning network, identifying the categories of different signal samples, obtaining multi-classification recognition results under each signal feature, calculating the recall rates of the signal features under different signal categories, and constructing a recall rate matrix; The feature learning network comprises a convolution layer with weights and a full connection layer; from a one-dimensional data input end to a multi-classification result output end, sequentially passing through a 1*7 convolution layer, a batch normalization layer, a ReLU layer, a 1*3 maximum pooling layer, a Res_2 layer, a Res_1 layer, an adaptive average pooling layer, a full connection layer and a linear layer; the Res_2 layer comprises a 1*3 convolution layer, a 1*1 convolution layer, a ReLU layer and a batch normalization layer; the Res_1 layer comprises a 1*3 convolution layer, a ReLU layer and a batch normalization layer; Step 3: dividing the whole test set into two parts, using a part of test signal samples to correct the recall rate matrix, and identifying the categories of another part of test signal samples through a multi-feature decision algorithm; Step 3 specifically comprises: Step 3.1: dividing the whole test set into two parts, and using a part of test signal samples to establish a corrected recall rate matrix; Several alternative thresholds T are designed hre , obtaining several recall matrices R hre under different thresholds T + (m,n): Finding a threshold that maximizes recall The modified recall matrix R * (m, n) is: Step 3.2: selecting a test signal sample from another part of test data as input, outputting a multi-feature decision result, and identifying the signal category to which the test signal belongs; Specifically, a test signal sample is input from another part of test data, if most features of the signal sample determine that the signal sample belongs to a certain same category signal, it is considered that the sample belongs to the signal category; otherwise, the corrected recall rate matrix is queried, and it is considered that the signal sample belongs to the category corresponding to the maximum recall rate; The test signal sample category set C = {1, 2,..., c}, the category number set determined after the test signal sample is identified is M0, and the signal categories c1, c2,..., c determined according to the signal features z ∈C, multi-feature decision result C + The formula is as follows: In the formula, z is the total number of categories of features.

2. The signal recognition method based on multi-class feature learning according to claim 1, characterized in that, In step 1, the importance of each dimension of the high-dimensional features is evaluated, and important dimensions in the high-dimensional features are selected specifically as follows: Step 1.1: constructing an Ith decision tree by using a bootstrap algorithm; Step 1.2: Calculate the Gini Impurity of the split node k of the i-th decision tree where C is the number of signal classes, is the probability that split node k belongs to the cth signal class. Step 1.3: Calculate feature dimension X j Feature importance at decision tree i split node k wherein Gini impurity of the new node l after branching, Gini impurity of the new node r after branching; Step 1.4: Calculate feature dimension X j Importance in the whole forest VIM j ; In the formula, K is the total number of nodes in the decision tree; Step 1.5: normalizing the importance of all feature dimensions; where VIM j is the normalized feature dimension X j importance in the whole forest, J is the total number of feature dimensions; Step 1.6: screening the dimensions whose feature importance is higher than a threshold value, arranging the dimensions in order according to the index order of the feature dimensions, obtaining important feature dimensions of the high-dimensional features, and the threshold value is selected as the average value of the importance scores of all feature dimensions.

3. The multi-class feature learning based signal recognition method of claim 1, wherein, In step 2, the recall rate of the signal features under different signal categories is calculated, and the recall rate matrix is constructed specifically as follows: The calculation formula of the recall rate R(m, n) of the signal features under different signal categories is as follows: where TPm,n represents the number of samples whose true class is the nth class and whose predicted class is also the nth class under the mth feature. mn where TPm,n represents the number of samples whose true class is the nth class and whose predicted class is also the nth class under the mth feature. mn where FNm,n represents the number of samples whose true class is the nth class but whose predicted class is other than the nth class under the mth feature. The original recall rate matrix is obtained by taking the feature category m as the horizontal coordinate and the signal category n as the vertical coordinate.

4. A signal recognition system based on multi-class feature learning, characterized by, The system comprises a feature selection module, a feature learning module and a feature decision module; The feature selection module is used for extracting multi-class features of an original signal, the original signal being an orthogonal IQ two-way baseband signal obtained in an ADS-B receiver; judging whether to analyze the importance of each dimension of the features according to the dimension of the features, evaluating the importance of each dimension of the high-dimensional features, and selecting important dimensions in the high-dimensional features; The feature learning module is used for designing a feature learning network, feeding the important high-dimensional features or low-dimensional features into the feature learning network, identifying the classes of different signal samples, obtaining multi-class recognition results under each signal feature, calculating the recall rates of the signal features under different signal classes, and constructing a recall rate matrix; The feature learning network comprises a convolution layer with weights and a full connection layer; from a one-dimensional data input end to a multi-class result output end, sequentially passes through a 1*7 convolution layer, a batch normalization layer, a ReLU layer, a 1*3 maximum pooling layer, a Res_2 layer, a Res_1 layer, an adaptive average pooling layer, a full connection layer and a linear layer; the Res_2 layer comprises a 1*3 convolution layer, a 1*1 convolution layer, a ReLU layer and a batch normalization layer; the Res_1 layer comprises a 1*3 convolution layer, a ReLU layer and a batch normalization layer; The feature decision module is used for correcting the recall rate matrix by using a part of test signal samples, and identifying the classes to which another part of test signal samples belong through a multi-feature decision algorithm; The use of a part of test signal samples to correct the recall rate matrix and the identification of the classes to which another part of test signal samples belong through a multi-feature decision algorithm are specifically as follows: Divide the entire test set into two parts, and use a part of test signal samples to establish a corrected recall rate matrix; Several alternative thresholds T are designed hre , obtaining several recall matrices R at different thresholds T hre + (m, n):​ Finding a threshold that maximizes recall The modified recall matrix R * (m, n) is: Select a test signal sample from another part of test data as input, output a multi-feature decision result, and identify the signal class to which the test signal belongs; Specifically, input a test signal sample from another part of test data, if most features of the signal sample determine that the signal sample belongs to a certain same class signal, it is considered that the sample belongs to the signal class; otherwise, the corrected recall rate matrix needs to be queried, and it is considered that the signal sample belongs to the class corresponding to the maximum recall rate; The test signal sample category set C = {1, 2,..., c}, the category number set determined after the test signal sample is identified is M0, and the signal categories c1, c2,..., c determined according to the signal features z ∈C, multi-feature decision result C + The formula is as follows: In the formula, z is the total number of classes of features.

5. An electronic device, comprising: The computer program is stored in the memory and can be run on the processor, and when the processor executes the computer program, the signal recognition method based on multi-class feature learning is realized. The computer program is stored in the memory and can be run on the processor, and when the processor executes the computer program, the signal recognition method based on multi-class feature learning is realized.

6. A computer-readable storage medium, characterized in that: ​

Images