Method for detecting im signal based on machine learning and wavelet decomposition coefficient features

Abstract

The application discloses an IM signal detection method based on machine learning and wavelet decomposition coefficient features. The method is characterized in that: the received IM signal is subjected to N-point fast Fourier transform to obtain N subcarriers, and the activation condition of the marked subcarriers is used as a sample label. The preprocessed subcarriers R(k) are subjected to wavelet decomposition, and statistical features are extracted from the high-frequency detail components and low-frequency approximation components obtained through the decomposition, which are used as sample features of the subcarriers R(k) and input into a machine learning algorithm for binary classification. The activation condition of the subcarriers is distinguished according to the sample features, compared with the sample label, and the hyperparameters of the machine learning algorithm are trained. For the IM signal to be detected, statistical features of the wavelet coefficients are extracted after wavelet decomposition, and the trained machine learning algorithm is input to obtain the activation condition of the subcarriers, thereby completing the detection. The method uses the optimization ability of the model to determine the threshold, and can learn the features of the signal more comprehensively.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology and relates to the detection of indexed modulated signals, specifically to an IM signal detection method based on machine learning and wavelet decomposition coefficient features. Background Technology

[0002] Index modulation is an emerging modulation technique that uses the index information of a signal to transmit data. Common carriers of index information include subcarriers and antennas. It can be used to improve the spectral efficiency and energy efficiency of wireless communication systems.

[0003] Existing signal detection methods based on subcarrier energy, such as maximum likelihood detection, are simple and effective at high signal-to-noise ratios (SNRs) and can quickly search for the energy of each subcarrier to determine whether the subcarrier modulates information. However, they are susceptible to noise. At low SNRs, the signal energy and noise energy are close, and the energy-based detection methods have low signal discrimination and reduced detection probability.

[0004] Wavelet transform uses time-frequency analysis to process signals, decomposing each subcarrier into low-frequency approximate components and high-frequency detail components. Since the wavelet decomposition coefficients are related to the signal's time-frequency characteristics, they can reflect the characteristics of the modulated signal itself. Therefore, compared to traditional subcarrier energy-based detection methods, wavelet transform-based signal detection exhibits better performance at low signal-to-noise ratios. However, traditional wavelet transform-based IM signal detection methods directly use the average of all wavelet decomposition coefficients of the subcarriers as the threshold for determining the subcarrier's activation state. This threshold determination method lacks theoretical support and cannot accurately reflect the signal's inherent information, leading to poor detection results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes an IM signal detection method based on machine learning and wavelet decomposition coefficient features. The method decomposes the IM signal based on time-frequency analysis wavelet transform, uses the statistical feature values ​​of the obtained wavelet decomposition coefficients as sample features of the IM signal, and uses whether the subcarrier is activated as a classification variable for model training to achieve high-resolution IM signal detection.

[0006] The IM signal detection method based on machine learning and wavelet decomposition coefficient features has the following specific steps:

[0007] S1: The received IM signal undergoes an N-point Fast Fourier Transform to obtain N subcarriers. Necessary preprocessing, including denoising and normalization, is then performed to improve the effectiveness of subsequent processing. The activation status of the labeled subcarriers is then used as sample labels for model training.

[0008] S2: Determine the wavelet system and the number of wavelet decomposition levels J. Perform J-level wavelet decomposition on the preprocessed subcarrier R(k). The larger the value of J, the higher the accuracy of signal analysis.

[0009]

[0010] Where k = 0, 1, 2, ..., N-1, j represents the scaling variable, and n represents the translation variable. φ jn (k) represent the wavelet function and the scaling function, respectively.

[0011] As a preferred option, the db5 wavelet system is selected to perform wavelet decomposition on the subcarrier.

[0012] S3: J-group high-frequency detail components d obtained from wavelet transform jn and low-frequency approximate component a jn Statistical features are extracted and used as sample features of subcarrier R(k). These features are then input into a machine learning algorithm for binary classification. The hyperparameters of the machine learning algorithm are trained based on the activation status of the subcarriers in the sample feature region and compared with the sample labels.

[0013] Preferably, the statistical features include mean, variance, standard deviation, median, autocorrelation coefficient, monotony, feature cross-correlation, kurtosis, skewness, and extreme values.

[0014] Preferably, the machine learning algorithm is a Support Vector Machine (SVM), Decision Tree, Random Forest, Gradient Boosting Tree, or Extreme Gradient Boosting (XGBoost).

[0015] S4: For the IM signal to be detected, perform wavelet decomposition and extract the statistical features of the wavelet coefficients. Input these features into the machine learning algorithm trained in S3 to obtain the activation status of the subcarriers and complete the detection.

[0016] The present invention has the following beneficial effects:

[0017] Applying wavelet decomposition technology to IM-based signal detection, compared to signal energy-based detection methods, effectively reduces the missed detection probability of active subcarriers under low signal-to-noise ratio (SNR). Furthermore, addressing the common practice in traditional wavelet transform-based IM signal detection methods that simply use the mean of all wavelet decomposition coefficients as the threshold for determining subcarrier activation, this method uses wavelet decomposition coefficients as feature values ​​and subcarrier activation as a classification variable for model training. By leveraging the model's optimization capability to determine the threshold, it achieves a more comprehensive learning of signal characteristics. This results in an improvement in IM signal detection probability of approximately 0.017 compared to traditional wavelet transform-based methods and approximately 0.03 compared to signal energy-based methods. Attached Figure Description

[0018] Figure 1 This is a flowchart of an IM signal detection method based on machine learning and wavelet decomposition coefficient features;

[0019] Figure 2 This is the index modulation design diagram in the embodiment;

[0020] Figure 3 This is a schematic diagram of wavelet decomposition;

[0021] Figure 4 This refers to the subcarrier detection probability in the embodiment.

[0022] Figure 5 This represents the probability of missed detection when the subcarrier is activated in the embodiment. Detailed Implementation

[0023] The present invention will be further explained below with reference to the accompanying drawings;

[0024] like Figure 1 As shown, the improved IM signal detection method based on the XGBoost algorithm and wavelet transform specifically includes the following steps:

[0025] Step S1: Index modulation is performed on the speech signal from the public dataset LJSpeech, and after transmission through the channel, it serves as the received IM signal in this embodiment. The index modulation design method is as follows: Figure 2 As shown, every four system subcarriers form a subblock, and the first subcarrier in the subblock is always active and used to carry pilot symbols. The remaining three subcarriers are selected to be active according to the index information.

[0026] The subcarriers of the IM received signal after fast Fourier transform are subjected to Kalman filtering to effectively filter out high-frequency noise, retain useful low-frequency components, and provide a smooth output to improve the effect of subsequent processing.

[0027] Step S2: Due to its excellent orthogonality and tight support, it is particularly suitable for signal processing. Therefore, this embodiment uses the db5 wavelet system to perform a 5-level wavelet decomposition on the subcarrier signal, such as... Figure 3 As shown, a set of wavelet decomposition coefficients [d] is obtained. 1n a 1n d 2n a 2n , ..., d 5n a 5n ]. Using variable y i ={-1 or 1} indicates whether each subcarrier is active.

[0028] Step S3: Extract statistical features from the wavelet decomposition coefficients of subcarrier R(k), including mean, variance, standard deviation, median, autocorrelation coefficient, monotony, characteristic cross-correlation, kurtosis, skewness, and extreme values, as the features of subcarrier R. i Sample features X of (k) ik ={z1,z2,z3,...,z 10}, variable y i As a sample label, 1 indicates that the subcarrier is activated and is recorded as a positive sample, while -1 indicates that the subcarrier is silent and is recorded as a negative sample.

[0029] Step S4: Divide the original data into a training set D and a test set T in a 1:1 ratio. Input the sample features of subcarrier R(k) in the training set D into the machine learning algorithm for binary classification. Based on the activation status of the subcarriers in the sample feature region and comparing it with the sample labels, train the hyperparameters of the machine learning algorithm.

[0030] In this embodiment, Extreme Gradient Boosting (XGBoost) is selected to perform the binary classification task. The logistic function is used as the classification model, and the Gini coefficient G is calculated using a probabilistic model. ini As a basis for determining the importance of features, a greedy search algorithm is applied to place important features in the first decision-making position. The specific steps are as follows:

[0031] S4.1: The Gini coefficient represents the probability that a randomly selected sample from a dataset will be misclassified into another class. A higher Gini coefficient indicates greater uncertainty in the sample set. First, calculate the Gini coefficient G of the sample set D. ini (D), representing the uncertainty of the sample set D:

[0032]

[0033] Where |D| represents the number of samples in dataset D, and |C q| represents the number of samples belonging to class q in dataset D; Q represents the total number of sample classes in dataset D. In this embodiment, Q = 2, C1 represents the set of positive samples, and C2 represents the set of negative samples.

[0034] For the 10 statistical characteristics of subcarriers in the sample set D, calculate each statistical characteristic z in turn. A The Gini coefficient G after segmenting the sample set D using the eigenvalue a ini (D,A,a):

[0035]

[0036] Gini coefficient G ini (D,A,a) represents the statistical characteristics z A The value of z A <a and z A The uncertainty of sample set D after dividing sample set D into subsets D1 and D2 is ≥a.

[0037] The difference in Gini coefficients after segmenting dataset D based on feature A at the value a is defined as the Gini gain G. ain (D,A,a):

[0038] G ain (D,A,a)=G ini (D)-G ini (D,A,a)

[0039] Gini gain G ain The larger the value of (D,A,a), the better the splitting scheme using the value a as the splitting point reduces the uncertainty of the set.

[0040] S4.2: Generate a decision tree of depth 0. The goal of constructing the decision tree is to minimize the sum of the objective functions of all leaf nodes. Based on the statistical importance of the features obtained from prior information, left and right leaf nodes are generated in the following splitting order: variance, standard deviation, autocorrelation coefficient, mean, feature cross-correlation, extreme values, monotonically increasing property, kurtosis, skewness, and median. Then, the feature ranking is forward-adjusted according to the model's detection probability to achieve the optimal detection probability. The optimal splitting order is: mean, variance, standard deviation, autocorrelation coefficient, feature cross-correlation, extreme values, kurtosis, monotonically increasing property, median, and skewness.

[0041] Calculate each statistical feature z sequentially A Using the eigenvalue a as the gain coefficient O corresponding to the split point bj gain The eigenvalue 'a' corresponding to the optimal gain coefficient is used as the splitting point for the current feature.

[0042]

[0043] Where λ is the regularization coefficient; It is the square of the sum of the first derivatives of the samples contained in the left leaf node; H is the square of the sum of the first derivatives of the samples contained in the right leaf node; L H is the sum of the second derivatives of the samples contained in the left leaf node; R It is the sum of the second derivatives of the samples contained in the right leaf node; The objective function value of the left leaf node; The objective function value of the right leaf node; This represents the objective function value when no other features are selected.

[0044] Construct the loss function L oss :

[0045]

[0046] Where l(·) is the logistic function; m is the total number of samples. Let be the predicted label of the model for the i-th sample.

[0047] S4.3: The training set D was modeled 100 times using the XGBoost model. Bayesian optimization was selected to train the model hyperparameters, including the tree depth, number of trees, regularization term parameters, and learning rate of the XGBoost model, so that the model could achieve good performance on the training set.

[0048] Step S5: Using the model trained in Step S4, perform discriminative predictions on samples in the 50,000 test set T under different signal-to-noise ratio conditions, and compare the results with energy-based detection methods and traditional wavelet transform-based detection methods. Figure 4 and Figure 5 The detection probabilities and missed detection probabilities of different methods are presented. At a signal-to-noise ratio of -4dB, the detection probability of this method is 0.989692, which is an improvement of about 0.017 compared to the traditional wavelet transform-based detection method and about 0.03 compared to the energy-based detection method. The missed detection probability of this method is about 0.8%, which is an improvement of about 1.2 percentage points compared to the traditional wavelet transform-based detection method and about 2.7 percentage points compared to the energy-based detection method. This method is improved compared to the other two existing methods, which shows that this method is superior to the energy-based detection method and the traditional wavelet transform-based detection method.

[0049] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents. The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art can make various modifications or additions to the described specific embodiments or use similar methods to substitute them, but without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. An IM signal detection method based on machine learning and wavelet decomposition coefficient features, wherein the received IM signal is subjected to an N-point Fast Fourier Transform to obtain N subcarriers, and the activation status of the labeled subcarriers is used as sample labels. A machine learning algorithm is then used to detect the subcarrier activation status of the IM signal. The method is characterized by: The specific steps are as follows: S1: Denoise and normalize the subcarrier signal; S2: Determine the wavelet system and the number of wavelet decomposition layers J, and perform wavelet decomposition on the preprocessed subcarrier R(k); S3: High-frequency detail components d obtained from wavelet decomposition J group jn and low-frequency approximate component a jn Statistical features are extracted and used as sample features of subcarrier R(k). These features are then input into a machine learning algorithm for binary classification. The hyperparameters of the machine learning algorithm are trained based on the activation status of the subcarriers in the sample feature region and compared with the sample labels. S4: For the IM signal to be detected, perform wavelet decomposition and extract the statistical features of the wavelet coefficients. Input these features into the machine learning algorithm trained in S3 to obtain the activation status of the subcarriers and complete the detection.

2. The IM signal detection method based on machine learning and wavelet decomposition coefficient features as described in claim 1, characterized in that: Kalman filtering is applied to the subcarrier signal to retain low-frequency components and filter out high-frequency noise.

3. The IM signal detection method based on machine learning and wavelet decomposition coefficient features as described in claim 1, characterized in that: Wavelet decomposition of the subcarriers was performed using the db5 wavelet system.

4. The IM signal detection method based on machine learning and wavelet decomposition coefficient features as described in claim 3, characterized in that: Set the wavelet decomposition level J = 5.

5. The IM signal detection method based on machine learning and wavelet decomposition coefficient features as described in claim 1, characterized in that: The statistical features include mean, variance, standard deviation, median, autocorrelation coefficient, monotony, cross-correlation, kurtosis, skewness, and extreme values.

6. The IM signal detection method based on machine learning and wavelet decomposition coefficient features as described in claim 1, characterized in that: The machine learning algorithm is a support vector machine, decision tree, random forest, gradient boosting tree, or extreme gradient boosting tree.

7. The IM signal detection method based on machine learning and wavelet decomposition coefficient features as described in claim 6, characterized in that: The logistic function is used as the classification model for the extreme gradient boosting tree.

8. The IM signal detection method based on machine learning and wavelet decomposition coefficient features as described in claim 6, characterized in that: The Gini coefficient is used as the basis for the importance of features, and a greedy search algorithm is applied to place important features in the first decision place to determine the feature splitting order in the extreme gradient boosting tree.

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

Images