Signal modulation identification method based on Gabor transform and Hilbert transform

Abstract

The invention relates to the technical field of signal processing, and discloses a signal modulation identification method based on Gabor transform and Hilbert transform, which comprises the following steps: acquiring a target digital modulation signal; performing discrete Gabor transform on the target digital modulation signal to obtain a discrete Gabor transform feature matrix; performing Hilbert transform on the target digital modulation signal to obtain signal IQ path data; taking the discrete Gabor transform feature matrix and the signal IQ path data as input features; performing sliding window clipping on the input features to obtain K target prediction samples; inputting the K target prediction samples into a pre-trained signal category identification network to obtain K prediction results; and voting the K prediction results to determine the category of the target digital modulation signal. According to the method, the recognition accuracy can be remarkably improved, the noise can be effectively suppressed, and the recognition performance of the FSK signal in a low signal-to-noise ratio and complex electromagnetic environment can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and particularly to a signal modulation recognition method based on the Gabor transform and the Hilbert transform. Background Art

[0002] Frequency Shift Keying (FSK) signals are widely used, and their efficient and accurate recognition is crucial. Traditional FSK recognition methods rely on time-frequency analysis or artificial features, but in a complex electromagnetic environment, there are problems such as limited time-frequency resolution, feature extraction relying on experience, poor performance at low signal-to-noise ratios, and difficulty in adapting to the variability of FSK.

[0003] In recent years, deep learning has been introduced into the field of signal recognition. Chinese invention with publication number CN119471578A discloses a radar signal detection method, which obtains I / Q sequences through the Hilbert transform as part of the time-domain feature input, and processes them in combination with frequency-domain features and a cross-attention mechanism. Chinese invention with publication number CN118413422A discloses a general underwater acoustic signal demodulation method, which constructs four-channel samples through multiple Hilbert transforms to enhance features, and processes them using a specifically designed convolutional neural network.

[0004] Although the above-mentioned prior arts have explored signal processing using the Hilbert transform or feature fusion, none of them clearly propose or suggest using the precise time-frequency localization features provided by the Gabor Transform and the signal I / Q path data (reflecting amplitude and phase information) extracted by the Hilbert Transform as key and complementary feature inputs, and combining them with a deep learning network (especially a network with an attention mechanism) for the refined recognition of specific digital modulation signals such as FSK. Therefore, how to fully utilize the synergistic advantages of these two transforms to improve the recognition performance of FSK signals in a complex environment remains the focus of current research. Summary of the Invention

[0005] To solve the problems existing in the above-mentioned prior art, the technical solutions provided by the present invention include: A signal modulation recognition method based on the Gabor transform and the Hilbert transform, comprising the steps of: S1. Obtain a target digital modulation signal; S2. Perform a discrete Gabor transform on the target digital modulation signal to obtain a discrete Gabor transform feature matrix; perform a Hilbert transform on the target digital modulation signal to obtain signal I / Q path data; S3. Use the discrete Gabor transform feature matrix and the signal I / Q path data as input features; perform sliding window clipping on the input features to obtain K target prediction samples; S4. Input the K target prediction samples into a pre-trained signal category recognition network to obtain K prediction results; S5. Perform voting processing on the K prediction results to determine the category of the target digital modulation signal.

[0006] Preferably, the training method of the signal category recognition network includes the following steps: S41. Obtain digital modulation signal simulation samples including FSK signals and non-FSK signals; S42. Perform discrete Gabor transform and Hilbert transform on the simulation samples respectively to obtain corresponding discrete Gabor transform feature matrices and signal IQ path data, and form training feature data; S43. Perform sliding window clipping on the training feature data, and label it by one-hot encoding according to its signal type to obtain a labeled training data set; S44. Use the labeled training data set to train the signal category recognition network.

[0007] Preferably, the method of sliding window clipping includes: setting the length of the sliding window to M and the sliding step to M / 2; discarding the clipping objects with a length less than M after sliding window clipping.

[0008] Preferably, in step S43, the one-hot encoding method includes: setting the label corresponding to the simulation sample of the non-FSK signal type to a predetermined all-zero encoding.

[0009] Preferably, after the discrete Gabor transform and the Hilbert transform, it further includes the step of performing maximum-minimum normalization processing on the discrete Gabor transform feature matrix and the signal IQ path data respectively.

[0010] Preferably, the category recognition network is configured to receive the I-channel data and Q-channel data in the signal IQ path data, and the discrete Gabor transform feature matrix respectively as three independent inputs for processing; Among them, the category recognition network includes a first input path for processing the I-channel data and a second input path for processing the Q-channel data, and at least one attention mechanism module is provided after the convolution operation inside the feature extraction module of the first input path and the second input path respectively.

[0011] Beneficial effects 1. Superior Feature Representation and Recognition Accuracy: The present invention creatively combines the discrete Gabor transform feature matrix (providing precise time-frequency localization information) and the I / Q path data of the Hilbert transform (reflecting amplitude and phase information) as the dual inputs of a deep learning network. This combination provides richer and complementary signal features compared to single transforms or simple I / Q data. Particularly for the core characteristics of the frequency variation of FSK signals, it can significantly improve the recognition accuracy rate.

[0012] 2. Enhanced Robustness and Anti-Interference Ability: The excellent time-frequency concentration of the Gabor transform is combined with the structural information of the I / Q data, and then through a deep learning network (especially a network containing an attention mechanism) for adaptive feature weighting and fusion, it can effectively suppress noise and significantly improve the recognition performance of FSK signals in low signal-to-noise ratio and complex electromagnetic environments. The specific combination of the front-end features of the present invention provides a better foundation for the robustness of the back-end network. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic flow chart of a signal modulation recognition method based on the Gabor transform and the Hilbert transform provided in a preferred embodiment of the present invention; Figure 2 It is a schematic flow chart of a signal category recognition network training method provided in a preferred embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0014] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings. In the description of the present invention, it should be understood that the orientation or positional relationships indicated by the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. are based on the orientation or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention.

[0015] The present invention relates to the field of digital signal processing. Specifically, it relates to a signal modulation recognition method based on the Gabor transform and the Hilbert transform, aiming to improve the recognition accuracy rate and robustness of digital modulation signals, especially frequency shift keying (FSK) signals, in complex electromagnetic environments.

[0016] In many application scenarios such as wireless communication, spectrum monitoring, and electronic countermeasures, accurately identifying the type of received digital modulation signal is a fundamental and crucial task. Traditional signal modulation recognition methods, such as those based on specific statistical features or simple time-frequency analysis, tend to experience a significant decline in performance when faced with complex situations such as low signal-to-noise ratio (SNR), multipath interference, or unknown signal parameters. For example, in a non-cooperative communication scenario, the receiving party may not be able to obtain the precise carrier frequency or symbol rate of the signal, which poses a huge challenge to traditional recognition algorithms that rely on precise parameters. Another example is that in a dense wireless environment, the target signal may be submerged in a large amount of background noise or other interfering signals, making it difficult for traditional methods to effectively extract discriminative features, resulting in a high false positive rate or false negative rate. These problems limit the application effect of traditional methods in modern complex electromagnetic environments.

[0017] To overcome the deficiencies in the above-mentioned existing technologies, the present invention proposes a novel signal modulation recognition method. The core innovative idea is to jointly utilize the fine time-frequency localization features provided by the Gabor Transform and the signal envelope and phase information (manifested as I / Q path data) extracted by the Hilbert Transform, and use these two significantly complementary features as inputs together, and feed them into a specially designed and pre-trained deep learning network (i.e., the signal category recognition network) for processing, so as to achieve efficient and accurate discrimination of the target digital modulation signal category. Through this unique combination of features and deep learning, the present invention is expected to significantly improve the performance of the signal recognition system in low SNR and complex interference environments, and enhance the adaptability and generalization ability to signals with different modulation parameters. The present invention can be widely applied to product forms such as wireless communication receivers, spectrum monitoring devices, and signal intelligence analysis systems.

[0018] Embodiment 1 To make the objectives, technical solutions, and advantages of the present invention clearer, as Figure 1 shown, the following will elaborate in detail on a signal modulation recognition method based on the Gabor Transform and the Hilbert Transform proposed by the present invention.

[0019] Step S1: Obtain the target digital modulation signal. The target digital modulation signal can be an actual signal received through a wireless channel or a simulation signal generated for verifying or applying this method. The acquisition method can include but is not limited to capturing through an antenna and a radio frequency front end, or reading from a stored signal data file. The purpose of this step is to provide the original input data for the subsequent feature extraction and recognition processes.

[0020] Step S2: Extract features of the target digital modulation signal in a two-path manner. Specifically, on the one hand, perform a Discrete Gabor Transform (DGT) on the target digital modulation signal to obtain a discrete Gabor transform feature matrix that can reflect the joint distribution of signal energy in time and frequency. The discrete Gabor transform is an important time-frequency analysis tool. It analyzes the signal by using a window function (such as a Gaussian window) with good localization characteristics in both the time domain and the frequency domain, so as to provide information about the signal in both time and frequency simultaneously. Each element of the obtained discrete Gabor transform feature matrix represents the energy or amplitude of the signal at a specific time point and a specific frequency point, and thus can effectively reveal the pattern of signal frequency variation over time, such as FSK signals. On the other hand, in parallel, perform a Hilbert Transform on the same target digital modulation signal to obtain the IQ path data of the signal. The Hilbert transform can convert a real-valued signal into an analytic signal. The real part of the analytic signal is the original signal itself (I path data, In-phase component), and the imaginary part is the Hilbert transform of the original signal (Q path data, Quadrature-phase component). The I path data and the Q path data together characterize the instantaneous amplitude and instantaneous phase information of the signal, which is crucial for understanding the modulation characteristics of the signal. In this way, the present invention extracts the key information of the target signal from two different dimensions (time-frequency joint distribution and instantaneous amplitude-phase), providing a rich and complementary feature basis for subsequent recognition.

[0021] After performing the discrete Gabor transform and the Hilbert transform, a preferred technical process is to perform min-max normalization on the obtained discrete Gabor transform feature matrix and the signal IQ path data respectively. Min-max normalization is a common data preprocessing technique, whose purpose is to linearly map the original data to a specified range (such as 0 to 1 or -1 to 1). The advantage of doing this is that it can eliminate the problems of non-convergence or performance degradation of model training that may be caused by large differences in dimension or numerical range between different features, and at the same time helps to improve the stability and efficiency of subsequent deep learning network processing. For example, there may be a large order-of-magnitude difference between the energy values in the discrete Gabor transform feature matrix and the amplitude values in the IQ path data. Through normalization, they can be unified to a similar numerical scale.

[0022] Step S3: Use the discrete Gabor transform feature matrix and the signal IQ path data that have undergone the above processing (which may include normalization) as input features. To make full use of the local information of the signal and increase the number of samples to meet the training and prediction requirements of the deep learning model, it is necessary to perform sliding window clipping on these input features to obtain K target prediction samples. The specific method of sliding window clipping can include: setting a fixed sliding window length M and a sliding step (e.g., M / 2). The sliding window slides sequentially on the original discrete Gabor transform feature matrix and the corresponding IQ path data, sliding a distance of one step each time, and intercepting the part within the window as a sample. If the length of the original feature sequence is not sufficient to form a complete window length M, the clipping object that is less than M is discarded to ensure that all samples input to the subsequent network have a unified dimension. Through this sliding window operation, a long target signal can be divided into K (K is an integer greater than or equal to 1) target prediction samples with a certain overlap and a length of M. This not only increases the number of samples for prediction, thereby improving the robustness of the final decision through the subsequent voting mechanism, but also helps the network model capture the local features of the signal at different time periods.

[0023] Step S4: Input the K target prediction samples obtained by sliding window clipping into a pre-trained signal class recognition network one by one or in batches. This signal class recognition network is a deep learning model, and its internal structure and parameters are learned through training on a large number of labeled signal samples, and is specifically used to identify the modulation class of the signal from the input features. After receiving each target prediction sample, the network will perform a series of complex non-linear transformations and feature extractions, and finally output the prediction result corresponding to the sample, that is, the modulation class (e.g., 2FSK, 4FSK, non-FSK, etc.) that the sample is most likely to belong to or the probability distribution of belonging to each class. Therefore, for the K target prediction samples, K corresponding prediction results will be obtained.

[0024] Step S5: Perform voting on the K prediction results to determine the final class of the target digital modulation signal. Voting is a simple and effective ensemble learning strategy, and its basic idea is that the minority obeys the majority. For example, it can be counted which modulation class appears the most times among the K prediction results, and this class is taken as the final recognition result of the entire target digital modulation signal. This processing method can effectively smooth the misjudgment that may be caused by the noise or unclear features of a single sample, thereby improving the overall recognition accuracy and stability.

[0025] To enable the above signal class recognition network to work effectively, it itself requires a rigorous training process. As Figure 2As shown, the training method of the signal category recognition network may specifically include the following steps: Step S41: Obtain a large number of digital modulation signal simulation samples, which should at least include the FSK signal types to be recognized in the target application (for example, binary frequency shift keying BFSK, multi - level frequency shift keying MFSK, continuous phase frequency shift keying CPFSK, Gaussian filtered frequency shift keying GFSK, etc.) and some non - FSK signal types (for example, phase shift keying PSK, quadrature amplitude modulation QAM, amplitude modulation AM, frequency modulation FM, etc.). The purpose of obtaining these simulation samples is to construct a diverse and representative data set for training the deep learning model. These signals can be generated by professional signal simulation software under different signal - to - noise ratios, symbol rates, carrier frequency offsets, etc., to enhance the generalization ability of the model.

[0026] Step S42: Perform discrete Gabor transform and Hilbert transform processing similar to the aforementioned step S2 on each of the obtained simulation samples. That is, calculate the corresponding discrete Gabor transform feature matrix and the signal IQ - path data for each simulation sample. In this way, a series of mutually matching training feature data pairs are formed. Similarly, after this stage, it is preferably to perform maximum - minimum normalization processing on these discrete Gabor transform feature matrices and signal IQ - path data respectively. The purpose is the same as the reason for normalizing the target signal before, that is, to ensure the scale consistency of the training data, which is beneficial to model training.

[0027] Step S43: Perform a sliding window clipping process similar to the aforementioned step S3 on the formed training feature data to obtain a large number of training sample segments. Different from the target signal processing, during the training stage, the modulation types of these simulation samples are known. Therefore, after the sliding window clipping, according to the true type of the original simulation signal corresponding to each training sample segment, it needs to be labeled by the one-hot encoding method. One-hot encoding is a common method for converting categorical variables into a numerical form that is easy for machine learning algorithms to process. For example, if there are N possible modulation categories, each category can be represented as an N-dimensional vector, where the position corresponding to the category is 1 and the rest are 0. A noteworthy detail is that in a specific implementation of the present invention, for those simulation samples belonging to non-FSK signal types, their corresponding labels can be set to a predetermined all-zero encoding (i.e., a vector with all dimensions being 0). This special encoding method helps the network better distinguish the FSK signal from the set of all other non-FSK signals, especially in the design of the network output layer and the selection of the loss function. After the sliding window clipping and one-hot encoding labeling, a training data set containing a large number of labeled samples is obtained, and this data set can be further divided into a training set and a validation set for model learning and performance evaluation.

[0028] Step S44: Use the constructed labeled training data set to complete the training of the signal category recognition network. The training process usually involves inputting the training samples into the network in batches. The network calculates the predicted output according to the current parameters, and then by comparing the difference between the predicted output and the true label (i.e., the value of the loss function), uses the backpropagation algorithm to iteratively adjust the parameters such as the weights and biases inside the network, with the goal of minimizing the loss function. When the loss function converges to a certain extent, or reaches the preset number of training epochs, or the performance on the validation set no longer improves, the training process can be stopped. At this time, the obtained network model is considered to be "pre-trained".

[0029] Regarding the internal structure of the signal category recognition network, the present invention provides a preferred configuration method. In this configuration, the category recognition network is designed to be able to receive the I-channel data (a one-dimensional sequence) and the Q-channel data (another one-dimensional sequence) in the signal IQ-channel data respectively, as well as the discrete Gabor transform feature matrix (a two-dimensional matrix), and process these three as three independent input branches in parallel. This multi-branch input structure allows the network to focus on different types of features from the beginning.

[0030] More specifically, the class recognition network may include a first input processing path for specifically processing the I-channel data and a second input processing path for specifically processing the Q-channel data. In these two paths, each feature extraction module (composed of several convolutional layers, pooling layers, etc.) is preferably provided with at least one attention mechanism module after the internal convolutional operation.

[0031] Introducing the attention mechanism in the processing paths of the I-channel and Q-channel data can make the network more focused on those segments or patterns in the signal amplitude and phase changes that are most indicative of the modulation type discrimination, thereby improving the effectiveness and pertinence of feature extraction. Especially in the presence of noise interference, the attention mechanism helps the network to "penetrate" the noise and capture key information. It should be noted that the reason why the present invention only introduces the attention mechanism in the processing paths of the I-channel and Q-channel data and does not consider adding the attention mechanism to the processing path of the feature matrix is as follows: on the one hand, adding the attention mechanism to the feature matrix will cause a substantial increase in the computational complexity of the network and a significant increase in the training time, so convolution is used for global perception as a substitute; on the other hand, the characteristics of the attention mechanism are more sensitive to time series data and are more suitable for feature extraction of IQ data and feature superposition with convolutional features.

[0032] The features extracted by these two paths can then be fused with the features extracted by other paths (such as the path for processing the discrete Garbor transform feature matrix) (which can be achieved by splicing, adding, or other common feature fusion methods in the prior art), and then pass through structures such as fully connected layers to finally output the classification result.

[0033] Embodiment 2 To more specifically illustrate the technical solution of the present invention, an exemplary implementation process is given below.

[0034] It is assumed that the target digital modulation signals to be recognized mainly include 2FSK, 4FSK, and some common non-FSK signals (such as BPSK, QPSK, 16QAM).

[0035] 1. Training stage of the signal class recognition network: S41: Data generation and acquisition. Use simulation tools such as MATLAB to generate a large number of BFSK, MFSK, and PSK, QAM, AM signal samples. For example, 1000 signal segments are generated for each modulation type, the noise is distributed from 0 dB to 40 dB, the symbol rate is from 100 to 500 symbols per second, the sampling rate is set to 10 kHz, and the signal length is 30 seconds to simulate the real environment and enhance the generalization ability of the model.

[0036] S42: Feature extraction and preprocessing. For each generated simulation sample: Perform a discrete Gabor transform (for example, select a Gaussian window for the window function, with a window length of σ = 512, and obtain a feature matrix of 512×Nt after the transform, where Nt is related to the signal length).

[0037] Perform a Hilbert transform to obtain I-channel and Q-channel data, and the length of each channel of data is the same as that of the original signal.

[0038] Perform min-max normalization on the obtained discrete Gabor transform feature matrix and I-channel and Q-channel data respectively, and scale the numerical range to [0, 1] or [-1, 1].

[0039] S43: Sliding window clipping and annotation. Set the sliding window length M = 2048 and the sliding step M / 2 = 1024. Perform sliding window clipping on the feature data (Gabor matrix and IQ data) of each simulation sample. For example, for a signal with a length of 10240, the dimension of the normalized Gabor matrix is 512×10240, and the dimension of the IQ data is 2×10240. After sliding the window, for each clipped sample segment, the dimension of the Gabor matrix is 512×2048, and the dimension of the IQ data is 2×2048. Perform one-hot encoding annotation according to the type of the original signal corresponding to each segment. Assume the class order is [BFSK, MFSK, PSK, QAM, AM].

[0040] BFSK sample label: [1, 0, 0, 0, 0]; MFSK sample label: [0, 1, 0, 0, 0]; The sample labels of non-FSK signals (PSK, QAM, AM) are uniformly set to a predetermined all-zero encoding: [0, 0, 0, 0, 0]. Or, if it is necessary to distinguish the internal categories of non-FSK, encode them separately and distinguish them in the design of the network output layer. However, in this example, in order to highlight the distinction between FSK and non-FSK, the all-zero encoding is used to represent the large category of "non-FSK". Construct a training data set from all the clipped samples with labels and divide it into a training set and a validation set according to a ratio of, for example, 80%:20%.

[0041] S44: Network training. Construct a deep neural network with three input branches as a signal category recognition network.

[0042] Branch 1 (processing I-channel data): The input dimension is 1×2048. It includes several one-dimensional convolutional layers (for example, the convolutional kernel size is 1×3 or 1×5), activation functions (such as ReLU or GELU), pooling layers (such as max pooling), and a self-attention module is inserted after the convolutional layer.

[0043] Branch 2 (processing Q-channel data): The structure is similar to Branch 1 and also includes an attention mechanism.

[0044] Branch 3 (processing the discrete Gabor transform feature matrix): The input dimension is 256×2048. It includes several two-dimensional convolutional layers (for example, the convolutional kernel size is 3×3), activation functions, and pooling layers. The features extracted by the three branches are fused at a certain intermediate layer (for example, flattened and concatenated, and then passed through a fully connected layer). The Softmax activation function is used in the network output layer to output the probabilities corresponding to various categories. The cross-entropy loss function is used as the loss function. The Adam optimizer is used for parameter update. The training set data is fed into the network in batches for training, and the performance is evaluated on the validation set after each epoch until the model converges.

[0045] 2. Recognition stage of the target signal: S1: Obtain the target signal. Receive an unknown digital modulation signal.

[0046] S2: Feature extraction and preprocessing. Perform exactly the same discrete Gabor transform, Hilbert transform, and maximum-minimum normalization operations on the target signal as in S42 of the training stage to obtain the discrete Gabor transform feature matrix and IQ path data.

[0047] S3: Sliding window clipping. Use the same sliding window length M = 2048 and sliding step M / 2 = 1024 as in S43 of the training stage to perform sliding window clipping on the feature data of the target signal to obtain K target prediction samples.

[0048] S44: Network prediction. Input these K target prediction samples into the trained signal category recognition network one by one, and the network outputs a predicted category (or the probabilities of various categories) for each sample.

[0049] S5: Voting process. Count the modulation category that appears most frequently among the K prediction results (for example, if K = 10 samples, 7 of which are predicted as BFSK, 2 as MFSK, and 1 as non-FSK, then the final decision is BFSK). If the output is a probability, the output probability vectors of the K samples can also be averaged, and then the category with the highest probability is taken as the final result.

[0050] Through the above method, the present invention can make full use of the advantages of the Gabor transform in time-frequency analysis and the ability of the Hilbert transform to extract amplitude-phase information, combined with the powerful feature learning and classification performance of deep learning, especially the focusing effect of the attention mechanism on key information, so as to realize the efficient and accurate recognition of digital modulation signals, especially FSK signals. This method not only improves the recognition accuracy but also enhances the robustness of the system in complex environments and the adaptability to different signal parameters.

[0051] In summary, the present invention provides a new paradigm for signal modulation recognition that combines advanced signal processing technology and deep learning intelligent decision-making. Its core lies in constructing complementary multi-dimensional feature inputs through Gabor transform and Hilbert transform, and using a specially designed deep neural network containing an attention mechanism for efficient classification. This method successfully solves the key technical problem of limited performance of traditional recognition methods in complex electromagnetic environments, and brings significant technological progress and application value to the field of digital signal modulation recognition. Based on the core principles and technical inspirations disclosed by the present invention, those skilled in the art can reasonably modify, replace or combine details such as the specific structure of the network, parameter selection, sliding window strategy, etc. These workarounds should be deemed to fall within the scope of protection of the present invention as long as they do not deviate from the core technical ideas of the present invention and the technical problems to be solved, and can achieve basically the same technical effects.

[0052] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A signal modulation recognition method based on the Gabor transform and the Hilbert transform, characterized in that Including the steps: S1. Obtain the target digital modulation signal; S2. Perform discrete Gabor transform on the target digital modulation signal to obtain a discrete Gabor transform feature matrix; perform Hilbert transform on the target digital modulation signal to obtain signal IQ path data; S3. Use the discrete Gabor transform feature matrix and the signal IQ path data as input features; Perform sliding window clipping on the input features to obtain K target prediction samples; S4. Input the K target prediction samples into a pre-trained signal class recognition network to obtain K prediction results; S5. Perform voting on the K prediction results to determine the class of the target digital modulation signal.

2. The signal modulation recognition method based on the Gabor transform and the Hilbert transform according to claim 1, wherein The training method of the signal class recognition network includes the following steps: S41. Obtain digital modulation signal simulation samples including FSK signals and non-FSK signals; S42. Perform discrete Gabor transform and Hilbert transform on the simulation samples respectively to obtain corresponding discrete Gabor transform feature matrices and signal IQ path data, and form training feature data; S43. Perform sliding window clipping on the training feature data and label it in one-hot encoding mode according to its signal type to obtain a labeled training data set; S44. Use the labeled training data set to train the signal class recognition network.

3. The signal modulation recognition method based on the Gabor transform and the Hilbert transform according to claim 1 or 2, characterized in that The method of sliding window clipping includes: setting the length of the sliding window to M and the sliding step to M / 2; discarding the clipping objects with a length less than M after sliding window clipping.

4. The signal modulation recognition method based on the Gabor transform and the Hilbert transform according to claim 2, wherein The one-hot encoding mode in step S43 includes: setting the label corresponding to the simulation sample of the non-FSK signal type to a predetermined all-zero encoding.

5. The signal modulation recognition method based on the Gabor transform and the Hilbert transform according to claim 1 or 2, characterized in that: After the discrete Gabor transform and the Hilbert transform, it further includes the step of performing maximum-minimum normalization processing on the discrete Gabor transform feature matrix and the signal IQ path data respectively.

6. The signal modulation recognition method based on Gabor transform and Hilbert transform according to claim 1 or 2, characterized in that: The class recognition network is configured to separately receive the I-channel data and Q-channel data in the signal IQ path data, and the discrete Gabor transform feature matrix, and process them as three independent inputs; Among them, the class recognition network includes a first input path for processing the I-channel data and a second input path for processing the Q-channel data, and at least one attention mechanism module is provided after the convolution operation inside the feature extraction module of each of the first input path and the second input path.

Images