A data enhancement method for signal recognition deep learning model training

Abstract

The application discloses a data enhancement method for signal recognition deep learning model training, which comprises the following steps: in an electromagnetic signal recognition scene, determining a frequency point and bandwidth of signal occurrence through spectrum visualization software and controlling a signal acquisition and processing device to acquire baseband data of an actual signal; performing frequency domain and time domain extraction and other preprocessing operations on the IQ baseband data according to the occurrence time of the signal to obtain single-signal baseband time domain waveform data; performing data enhancement and expansion on the time domain waveform data by means of resampling, jitter frequency conversion, time domain translation, noise increase and decrease and the like; cutting or supplementing the enhanced signal time domain waveform data according to the length required by an intelligent recognition algorithm to form a fixed-length signal time domain data sample set; and finally, performing transform domain transformation to convert the signal time domain waveform data into time-frequency matrix data through a short-time Fourier transform. The application can effectively improve the performance of a signal recognition deep learning algorithm.

Description

Technical Field

[0001] This invention relates to the field of signal recognition technology, and in particular to a data augmentation method for training deep learning models for signal recognition. Background Technology

[0002] Signal recognition, which involves identifying signal systems and network types, is of great significance for determining target attributes and can also be applied in fields such as cognitive radio and adaptive communication. Traditional signal recognition algorithms involve detailed analysis of individual signals, manually extracting features such as high-order cumulants, instantaneous parameters, and cyclic stationary spectra to develop single-class signal recognition algorithms, which is difficult and time-consuming to develop. With the rapid development of communication technology, the types of signals in the electromagnetic environment are iterating and updating rapidly, are numerous, and have complex systems, placing more stringent requirements on the speed of signal recognition algorithm development and the accuracy of recognition.

[0003] Deep learning artificial intelligence algorithms based on neural networks have attracted attention in various fields in recent years, demonstrating performance far exceeding traditional algorithms in applications such as computer vision, speech recognition, and natural language processing. Intelligent signal recognition has become an inevitable trend. Given that applications in computer vision are the most mature, and that time-frequency data simultaneously reflects two-dimensional information about signals in both time and frequency, making it more conducive to neural networks extracting multi-dimensional features, most current intelligent signal recognition algorithms use time-frequency matrix data as input to the neural network algorithm. Multiple studies have shown that the performance of neural networks is strongly correlated with model size, and larger models require more sample data for training and parameter optimization. However, for signal recognition tasks, especially for target signals in complex electromagnetic environments, particularly short bursts and non-cooperative signals, the time and manpower costs for data sample collection, analysis, and annotation are enormous, making the construction of large-scale sample datasets quite difficult. Summary of the Invention

[0004] The main objective of this invention is to provide a data augmentation method for training deep learning models for signal recognition. This method aims to solve the technical problems in the development of artificial intelligence signal recognition algorithms based on time-frequency matrices as network inputs, such as the difficulty in obtaining large quantities of data samples, the high level of professionalism and difficulty in data annotation, and the inability to directly apply mainstream dataset construction and data augmentation methods.

[0005] To achieve the above objectives, the present invention provides a data augmentation method for training deep learning models for signal recognition, the method comprising the following steps:

[0006] S1: Use signal acquisition and processing equipment to collect baseband signal sample data under actual electromagnetic environment;

[0007] S2: Preprocess the signal to form a single-signal time-domain sample set and confirm the sample label information;

[0008] S3: Perform multi-method data augmentation on a single signal sample set to form an expanded time-domain sample set;

[0009] S4: Based on the number of data points required by the signal recognition algorithm, the expanded time-domain sample set is sliced ​​or truncated according to the signal attributes to form a fixed-length time-domain sample set.

[0010] S5: Based on the input time-frequency map size of the signal recognition algorithm, perform a short-time Fourier transform on the fixed-length time-domain sample set to form a data-augmented time-domain image dataset that can be directly used for training.

[0011] Optionally, in step S2, the signal is preprocessed by acquiring the actual target signal in the frequency domain and time domain, respectively, specifically including:

[0012] When acquiring the actual data in the frequency domain, down-conversion, filtering, and decimation are performed to form baseband IQ data with a signal center frequency of 0Hz, a filtering bandwidth equal to the signal bandwidth, and a decimation signal sampling rate of four times the signal bandwidth, ensuring that each sample in the dataset contains only a single frequency signal.

[0013] When acquiring actual data in the time domain, the start and end times of the signal are found by the change in the amplitude of the time-domain waveform signal. These times are then extracted to form multiple single-burst signal baseband IQ data, ensuring that each sample in the dataset contains only a single-burst signal at a single frequency point.

[0014] Optionally, step S3 specifically includes: using one or more of the following methods to enhance and expand the baseband IQ data of the signal and perform fixed-length processing, and then using short-time Fourier transform to convert the time and frequency domains through feature domain transformation.

[0015] Optionally, when performing data enhancement on the baseband IQ data of the signal, frequency domain jitter based on digital up-conversion and down-conversion is adopted, and the specific expression is as follows:

[0016] f shift =αBw, α∈[-0.25,0.25]

[0017]

[0018] Where x(n) is the original baseband IQ data of the signal, Bw is the signal bandwidth, α is the frequency domain jitter coefficient, which is a random number between -0.25 and 0.25, x fshift (n) represents the baseband IQ data of the signal after random frequency domain jitter, and is the length of the baseband IQ data.

[0019] Optionally, when performing data enhancement on the baseband IQ data of the signal, time-domain jitter based on adding noise points is adopted;

[0020] If the time-domain jitter direction is to the right, then it is a time-domain delay, and the specific expression is:

[0021] N shift =βN, β∈[-0.25,0.25]

[0022]

[0023] If the time-domain jitter direction is to the left, it is time-domain advance, and the specific expression is:

[0024]

[0025] Where β is the time-domain jitter coefficient, which is a random number between -0.25 and 0.25, and N is the baseband IQ data length.

[0026] Optionally, when performing data enhancement on the baseband IQ data of the signal, artificial noise addition based on energy calculation is adopted, and the specific expression is as follows:

[0027]

[0028] P addnoise =P x(n) ,γ∈(0,0.1]

[0029] x addnoise ()=G(n)×P addnoise +x()

[0030] Where P x(n) denoted as the original baseband data power, γ is the ratio coefficient of the added noise power to the original baseband data power, and G(n) is a normally distributed random sequence with a mean of 0 and a variance of 1 of the same length as the original baseband data.

[0031] Optionally, when performing data enhancement on the baseband IQ data of the signal, artificial denoising based on wavelet transform is adopted, specifically as follows: selecting the wavelet basis function as the Haar basis and the number of decomposition layers as 2, performing wavelet decomposition on the original baseband data of the signal, and calculating the wavelet coefficients of each layer; setting a global threshold for each decomposition layer using a fixed threshold estimation method, and performing soft thresholding on the higher-order wavelet coefficients of each layer according to the set threshold; reconstructing the signal based on the processed wavelet coefficients to obtain the denoised and enhanced signal data.

[0032] Optionally, when enhancing the baseband IQ data, a signal resampling based on a polyphase filter is employed. Specifically, a fractional-multiplier sampling process is used, involving interpolation, filtering, and decimation. Let the ratio of the resampled sampling rate to the original signal sampling rate be L / M, where L = 100, M = δL, δ ∈ [-0.3, 0.3]. First, the original baseband signal data is upsampled by L times, then filtered by an FIR low-pass filter, and finally downsampled by M times to obtain the enhanced data after resampling by L / M times. The cutoff frequency of the low-pass filter is... The gain is L.

[0033] Optionally, step S4 specifically includes: based on the relationship between the number of sampling points of the sample data and the number of sampling points required for the input of the intelligent signal recognition algorithm, each sample data needs to be selected to perform one of the following three operations: 1) perform fixed-length segmentation on data samples with a number of data points greater than 5 times the number required for the algorithm input; 2) perform fixed-length truncation on data samples with a number of data points greater than 1 times but less than 5 times the number required for the algorithm input; 3) first pad zeros on data samples with a number of data points less than the number required for the algorithm input to obtain the number of data points required by the algorithm, then add noise, and finally convert all expanded single signal and single burst time-domain waveform data samples into a fixed-length signal time-domain waveform data sample set.

[0034] Optionally, the method further includes step S5: performing feature domain transformation on the processed fixed-length signal time-domain waveform data sample set; specifically including: using short-time Fourier transform operation, performing a sliding window operation with overlap rate from front to back on the data, selecting an appropriate window function to limit the time-frequency domain range of the signal within the window, and calculating the signal power spectrum within the sliding window at each time step; wherein, the number of FFT points is 1024, the window function is selected as a Hanning window with a length of 256 points, and the overlap rate is set to 0.5; according to the order of arrival of the sliding window, the calculated power spectrum results are spliced ​​along the time domain to form a two-dimensional time-frequency matrix, i.e., time-frequency matrix data, with 256 splicing frames, converting the signal time-domain waveform data into time-frequency matrix data required for training the intelligent signal recognition algorithm, with a size of 256×256.

[0035] The beneficial technical effects of this invention are as follows:

[0036] (1) It reduces the workload of collecting signal samples for the training dataset. This invention proposes a data augmentation method for expanding datasets used in training deep learning artificial intelligence signal recognition algorithms. Based on existing data samples, a large number of samples can be directly generated through computation, expanding the dataset size and reducing the time and manpower costs of signal acquisition and processing. Furthermore, for supervised neural network algorithms, data augmentation only increases the number of data samples without changing the labels, reducing the workload of manual analysis and labeling. Through resampling data augmentation, signal data at different sampling rates can be directly obtained without changing the sampling rate parameter settings of the signal acquisition and processing equipment. The same signal can be acquired and analyzed multiple times at different sampling rates, and a resampling calculation method based on a polyphase filter structure is used to accelerate the processing transformation speed. Through frequency domain jitter data augmentation, existing data undergoes slight frequency shifting to adjust the frequency offset, replacing the signal acquisition and processing work of the same signal at different acquisition center frequencies. Through time domain jitter data augmentation, different time segments of signal samples can be quickly obtained, avoiding the acquisition of the same signal at different time slots. Through artificial noise addition and artificial noise reduction data augmentation, signal sample data under different environmental noise levels can be simulated to a certain extent, reducing the signal acquisition and processing work for the same signal under different signal-to-noise ratio conditions.

[0037] (2) The mainstream data augmentation methods have been optimized, the quality of data augmentation samples has been improved, and it can be used as the input form of various signal recognition algorithms. This invention improves upon existing image data enhancement methods for signal recognition algorithms by proposing a dedicated data enhancement technique based on electromagnetic signal processing. This technique processes the original time-domain waveform, performs feature domain transformation, and uses short-time Fourier transform calculations. The enhanced time-frequency data not only achieves an enhancement effect similar to image data but also more closely reflects the inherent characteristics of the signal, making it easier for neural network models to extract multi-dimensional features. Regarding existing image data scaling enhancement techniques, this invention alters the relationship between the original data sampling rate and the signal bandwidth, changing the size of the signal portion in the time-frequency data to achieve an image scaling effect. Regarding existing image data translation enhancement techniques, this invention adds a small frequency offset to the original data to achieve the effect of translating the signal target along the frequency axis of the time-frequency matrix, and performs time-domain shifting of the original data to achieve the same effect. Regarding existing image data noise enhancement techniques, this invention uses artificial noise addition or denoising techniques on the time-domain signal to change the signal-to-noise ratio, and then performs time-frequency transformation to achieve the effect of adding or denoising the time-frequency matrix. Unlike directly performing image data augmentation on the signal's time-frequency matrix, this method processes the original time-domain signal and converts it into a time-frequency matrix, ensuring that the essential characteristics of the signal data remain unchanged. This reduces the impact of data augmentation on sample labels and improves the quality of the signal sample set. Furthermore, the proposed signal data augmentation method enhances and expands the original time-domain signal data, making it applicable when the signal recognition algorithm uses baseband time-domain data or other transform-domain data as input, thus demonstrating its versatility.

[0038] (3) Improved performance of intelligent signal recognition algorithms. This invention proposes a dedicated data augmentation optimization method for intelligent signal recognition algorithms. This involves either an offline implementation that pre-enhances and expands the original dataset or an online implementation that randomly selects enhancement methods during training. This allows the network structure parameters to more fully utilize the existing signal dataset, maximizing the dataset's gain for network training. Through this dedicated signal data augmentation method, the enhanced dataset will contain data samples different from the original dataset. Furthermore, the signal data augmentation method proposed in this invention, based on signal processing, simulates the possible scenarios of collected signal samples to a certain extent, making the augmented data more realistic. When training the network model to optimize network parameters, the number of data samples is greatly increased, reducing the tendency for overfitting in networks of the same size. The diversity of data samples, such as different sampling rates, different frequency offsets, and different signal-to-noise ratios, improves the robustness and generalization of the network. In situations where it is impossible to quickly collect more real signal samples, this invention has a good effect on improving the performance of existing network algorithms. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating a data augmentation method for training a deep learning model for signal recognition according to the present invention.

[0040] Figure 2 This is a flowchart illustrating the signal recognition dataset construction and data augmentation process based on signal processing according to the present invention.

[0041] Figure 3 This is a schematic diagram illustrating the preprocessing principle for constructing a signal recognition dataset based on signal processing, as described in this invention.

[0042] Figure 4 This is a schematic diagram illustrating the data augmentation principle of the signal recognition dataset based on signal processing according to the present invention.

[0043] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0044] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0045] Currently, in related technical fields, the development of artificial intelligence signal recognition algorithms based on time-frequency matrices as network input faces challenges such as difficulty in obtaining large quantities of data samples, high requirements and difficulty in data annotation, and the inability to directly apply mainstream dataset construction and data augmentation methods.

[0046] To address this issue, various embodiments of the data augmentation method for training deep learning models for signal recognition are proposed in this invention. The data augmentation method for training deep learning models for signal recognition provided by this invention determines the frequency and bandwidth of the signal output using spectrum visualization software and controls the signal acquisition and processing equipment to acquire baseband data of the actual signal. The IQ baseband data undergoes preprocessing operations such as frequency domain and time domain extraction based on the signal's occurrence time to obtain single-signal baseband time-domain waveform data. Resampling, jitter conversion, time-domain shifting, and noise addition / reduction are used to augment the time-domain waveform data. The augmented signal time-domain waveform data is then segmented or supplemented according to the length required by the intelligent recognition algorithm to form a fixed-length signal time-domain data sample set. Finally, a transform domain transformation is performed, converting the signal time-domain waveform data into time-frequency matrix data through short-time Fourier transform, which effectively improves the performance of deep learning algorithms for signal recognition.

[0047] This invention provides a data augmentation method for training deep learning models for signal recognition, referring to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the data augmentation method for training deep learning models for signal recognition according to the present invention.

[0048] In this embodiment, the data augmentation method for training a deep learning model for signal recognition includes the following steps:

[0049] S1: Use signal acquisition and processing equipment to collect baseband signal sample data under actual electromagnetic environment;

[0050] S2: Preprocess the signal to form a single-signal time-domain sample set and confirm the sample label information;

[0051] S3: Perform multi-method data augmentation on a single signal sample set to form an expanded time-domain sample set;

[0052] S4: Based on the number of data points required by the signal recognition algorithm, the expanded time-domain sample set is sliced ​​or truncated according to the signal attributes to form a fixed-length time-domain sample set.

[0053] S5: Based on the input time-frequency map size of the signal recognition algorithm, perform a short-time Fourier transform on the fixed-length time-domain sample set to form a data-augmented time-domain image dataset that can be directly used for training.

[0054] It should be noted that data augmentation is a mainstream approach to address the problem of insufficient data samples. By using certain methods to process a limited dataset, the distribution of the training dataset can be adjusted based on prior information, thereby enriching the quantity and diversity of samples. This approach increases the implicit constraints of the model in a low-cost and efficient manner, avoids overfitting of the network model, and improves the robustness and generalization of the model algorithm. It can also, to some extent, solve the problem of data imbalance between different categories of samples.

[0055] Data augmentation typically involves adding minor changes to existing data or synthesizing new data from existing data to expand the dataset without affecting the semantic information of the data. This allows the model parameters to better learn the invariant features of the sample data, enabling them to identify target samples that are different from the training samples, are in different background noise, or are in different contexts. It also helps to distinguish non-target negative samples that have similar external appearances but different essential characteristics.

[0056] In the field of computer vision, common data augmentation techniques can be based on geometric transformations or pixel transformations. Image data augmentation methods based on geometric transformations directly manipulate the original data, either locally or globally, changing geometric features such as the position and relative size of the target without affecting its original attributes. Examples include: flipping, rotating, cropping, scaling, translating, and dithering. Image data augmentation methods based on pixel transformations, on the other hand, act on individual pixels, adjusting the distribution of image background or optical features, thus enriching the image's environmental background noise to some extent. Examples include: adding noise (salt-and-pepper noise, Gaussian noise, etc.), applying Gaussian blur, adjusting HSV contrast, brightness, or saturation, and histogram equalization.

[0057] Unlike image data in computer vision, time-frequency matrix data used in signal recognition has its own unique characteristics. Simply applying image data augmentation techniques can cause the training data to deviate from the signal characteristics, misleading the neural network in learning the original features of the signal, causing the parameter iteration update direction of the network model to shift, and resulting in a decrease in the efficiency of the model algorithm. Signal time-frequency matrices have the following characteristics compared to optical image data: 1) Signal time-frequency matrices do not contain pixels. A signal time-frequency matrix is ​​not an optical image; each color block only represents the magnitude of the instantaneous frequency component energy at a certain time resolution and frequency resolution. Its color itself has no specific physical meaning; it is merely a representation of energy strength contrast. Furthermore, it does not contain multiple components such as HSV and RGB. Most image data augmentation methods based on pixel transformation cannot be directly applied to signal time-frequency matrix data augmentation; 2) Signal time-frequency matrices are not the original representation of the signal. The signal time-frequency matrix is ​​a special representation of a signal in the transform domain. Unlike optical images, which directly reflect the inherent characteristics of an object such as shape, texture, and color, the signal time-frequency matrix cannot directly express the inherent characteristics of a signal, such as frequency, bandwidth, sampling rate, and code rate. Directly enhancing the dataset by transforming the time-frequency matrix cannot guarantee that the essential characteristics of the signal will not be affected. For example, after directly rotating the time-frequency matrix, obtaining the signal time-domain waveform data through operations such as inverse Fourier transform and demodulating it to obtain enhanced signal code stream data or demodulated waveform data, the flipping of the time-frequency matrix changes the correspondence between the signal's time and frequency domains. Its symbol arrangement or waveform changes are altered compared to the original signal, meaning the essence of the signal changes, altering its category label and violating the principles of data augmentation. 3) Signal diversity cannot be achieved by directly enhancing the time-frequency matrix. The essence of data diversity is the artificial simulation of different manifestations of a target in real-world situations. For optical images, changes in the target's posture, distance, angle, etc., can be enriched through simple image flipping, scaling, and other operations. The diversity of signals is usually caused by different sampling rates, frequency shifts, and signal-to-noise ratios, making it difficult to directly simulate various situations through time-frequency matrix transformations, thus hindering the completion of training datasets and optimization of model generalization performance. 4) Signals are time series, and causal relationships cannot be ignored. Unlike image data, which is merely an expression of spatial information with only relative spatial relationships and absolute temporal causal relationships, the time-frequency matrix represents the signal's transformed domain expression in the original domain. That is, it simultaneously represents frequency domain information in the time domain, and the temporal causal relationship is paramount and cannot be reversed. Therefore, data augmentation techniques such as image domain flipping and rotation cannot be used for data augmentation of the signal's time-frequency matrix.

[0058] Based on this, in order to explain this application more clearly, a specific example of a data augmentation method for training a deep learning model for signal recognition is provided below.

[0059] See Figure 2 In electromagnetic signal identification scenarios, spectrum visualization software is used to determine the outgoing frequency and bandwidth of the signal. The signal acquisition and processing equipment is then controlled to perform operations such as analog frequency conversion, matched filtering, amplification, automatic AD sampling, digital channel preprocessing, digital down-conversion, and filtering on the RF signal received by the antenna, acquiring the IQ baseband data at the corresponding frequency and bandwidth. The IQ baseband data obtained by the signal acquisition and processing equipment must meet the following conditions: within the acquisition time and bandwidth range, only a single frequency point (single channel) of the signal appears, and the sampling rate is a fixed ratio to the bandwidth. Otherwise... To meet this requirement, signal processing operations such as digital down-conversion and filtering decimation need to be performed again. The baseband time-domain waveform of the signal is observed, and the start and end points of the signal are selected for time-domain truncation based on the signal's appearance and disappearance times. This yields time-domain waveform data for a single signal and a single burst, ensuring that each sample data point contains exactly one signal in both the time and frequency domains. The signals are then analyzed, and their types are manually labeled. Techniques from signal processing fields, such as resampling, frequency dithering, time dithering, artificial noise addition, and artificial noise reduction, are used to further process the signal's time-domain waveform data. The system enhances and augments data across multiple dimensions, scales, and scenarios. Based on different enhancement coefficients and combinations of enhancement methods, original single-signal and single-burst data samples can be augmented to any multiple. The required number of time-domain data points is calculated based on the FFT points and time frames set in the time-frequency matrix input to the intelligent recognition algorithm. The length of the enhanced single-signal and single-burst time-domain waveform data samples is compared one by one with the required number of points input to the algorithm. Different data truncation or addition methods are used to form a standard augmented dataset of single-signal and single-burst time-domain waveform data of fixed length. Finally, feature domain transformation is performed using a Short-Time Fourier Transform (STFT) operation. A sliding window operation with overlap is applied to the data from front to back. An appropriate window function is selected to limit the time-frequency domain range of the signal within the window. The signal power spectrum within the sliding window at each moment is calculated. Based on the order in which the sliding windows arrive, the calculated power spectrum results are concatenated along the time domain to form a two-dimensional time-frequency matrix, i.e., time-frequency matrix data. This achieves the effect of converting the signal time-domain waveform data into an augmented sample set of time-frequency matrix data required for training the intelligent signal recognition algorithm.

[0060] See Figure 3When constructing datasets for intelligent signal recognition algorithms, the actual acquired data needs to be preprocessed, acquiring the target signals in both the frequency and time domains. For frequency domain acquisition, down-conversion, filtering, and decimation are used to generate baseband IQ data with a signal center frequency corresponding to 0Hz, a filtering bandwidth equal to the signal bandwidth, and a decimation sampling rate four times the signal bandwidth. This ensures that each sample in the dataset contains only a single frequency signal. For non-continuous communication signals (i.e., target signals whose duration is shorter than a single acquisition time), time-domain burst detection is performed based on the baseband IQ data acquired in the frequency domain. The start and end times of the signal are identified by changes in the amplitude of the time-domain waveform, and then truncated to form multiple single-burst signal baseband IQ data sets. This ensures that each sample in the dataset contains only a single-burst signal at a single frequency. This time-frequency domain truncation removes non-signal time-frequency features, making it easier for the neural network structure to learn the detailed features of the signal itself and greatly avoiding biases introduced by the background environment into the network algorithm.

[0061] See Figure 4 When performing data augmentation on datasets used for training intelligent signal recognition algorithms, techniques such as signal resampling, frequency domain jittering, time domain jittering, artificial noise addition, and artificial noise reduction are employed. Each technique is processed independently and does not affect the others. Augmentation techniques can be randomly selected and combined, increasing the diversity of augmentation styles. Subsequently, fixed-length sample truncation is performed to reduce the constraints on the neural network structure. Then, feature domain transformation is performed, and time-frequency domain conversion is achieved through short-time Fourier transform to achieve data augmentation of the time-frequency matrix dataset.

[0062] When using frequency domain jitter technology based on digital up-conversion to enhance baseband IQ data, the calculation formula is as follows:

[0063] f shift =αBW, α∈[-0.25,0.25]

[0064]

[0065] Where x(n) is the original baseband IQ data of the signal, Bw is the signal bandwidth, α is the frequency domain jitter coefficient, which is a random number between -0.25 and 0.25, x fshift (n) represents the baseband IQ data of the signal after random frequency domain jitter, and is the length of the baseband IQ data.

[0066] When using time-domain jittering based on adding noise points to enhance baseband IQ data, if the time-domain jitter direction is to the right, i.e., the time-domain delay, is calculated using the following formula:

[0067] N shift=βN, β∈[-0.25,0.25]

[0068]

[0069] If the time-domain jitter direction is to the left, which means the time-domain advance, the calculation formula is as follows:

[0070]

[0071] Where β is the time-domain jitter coefficient, which is a random number between -0.25 and 0.25, and N is the baseband IQ data length.

[0072] When using energy-based artificial noise enhancement techniques to augment baseband IQ data, the calculation formula is as follows:

[0073]

[0074] P addnoise =P x(n) ,γ∈(0,0.1]

[0075] x addnoise ()=G(n)×P addnoise +(n)

[0076] Where P x(n) Let be the original baseband data power, γ be the ratio coefficient between the added noise power and the original baseband data power, and G(n) be a random sequence of the same length as the original baseband data. In an optional embodiment, a normally distributed random sequence with a mean of 0 and a variance of 1 is selected.

[0077] In an optional embodiment, when using wavelet transform-based artificial denoising technology to enhance the baseband IQ data of the signal, firstly, the wavelet basis function is selected as the Haar basis and the number of decomposition layers is 2, and wavelet decomposition is performed on the original baseband data of the signal to calculate the wavelet coefficients of each layer; secondly, a global threshold is set for each decomposition layer using a fixed threshold estimation method, and soft thresholding is performed on the higher-order wavelet coefficients of each layer according to the set threshold; finally, the signal is reconstructed based on the processed wavelet coefficients to obtain the denoised and enhanced signal data.

[0078] When using signal resampling technology based on polyphase filters to enhance baseband IQ data, an interpolation filtering sampling process is employed. First, the signal data is interpolated, then filtered by a low-pass filter, and finally decimated. Compared to standalone interpolation and decimation filters, this method is simpler in structure and easier to calculate. Let L / M be the ratio of the resampled sampling rate to the original signal sampling rate, and...

[0079] L=100,M=δL,δ∈[-0.3,0.3]

[0080] First, the original baseband signal data is upsampled and interpolated by a factor of L. Then, it is filtered by an FIR low-pass filter to prevent frequency aliasing. Finally, it is downsampled by a factor of M to obtain the enhanced data after resampling by a factor of L / M. The cutoff frequency of the low-pass filter is... The gain is L. Resampling is implemented using a multi-stage filter structure, which further accelerates the processing speed.

[0081] When converting the augmented baseband IQ data into a standard fixed-length dataset, based on the relationship between the number of sampling points in the sample data and the number of sampling points required for the input of the intelligent signal recognition algorithm, each sample data needs to undergo one of the following three operations: 1) Perform fixed-length segmentation on data samples with a number of data points much greater than the algorithm input (greater than 5 times); 2) Perform fixed-length truncation on data samples with a number of data points slightly greater than the algorithm input (greater than 1 and less than 5 times); 3) First, pad the data samples with a number of data points less than the algorithm input to obtain the number of data points required by the algorithm, then add noise, and finally convert all augmented single-signal and single-burst time-domain waveform data samples into a fixed-length signal time-domain waveform data sample set.

[0082] The processed fixed-length signal time-domain waveform data sample set is subjected to feature domain transformation using a short-time Fourier transform (STFT) operation. The data is then subjected to a sliding window operation with overlap from front to back. An appropriate window function is selected to limit the time-frequency domain range of the signal within the window. The signal power spectrum within the sliding window at each time step is calculated. The number of FFT points is 1024, the window function is a Hanning window with a length of 256 points, and the overlap rate is set to 0.5. Based on the arrival order of the sliding window, the calculated power spectrum results are stitched together along the time domain to form a two-dimensional time-frequency matrix, i.e., the instantaneous time-frequency matrix data. The number of stitched frames is 256, converting the signal time-domain waveform data into the time-frequency matrix data required for training the intelligent signal recognition algorithm, with a size of 256×256.

[0083] This embodiment provides a data augmentation method for training deep learning models for signal recognition. The method includes: in an electromagnetic signal recognition scenario, determining the frequency and bandwidth of the signal output using spectrum visualization software and controlling a signal acquisition and processing device to acquire baseband data of the actual signal; preprocessing the IQ baseband data by extracting frequency and time domains based on the signal's occurrence time to obtain single-signal baseband time-domain waveform data; augmenting and expanding the time-domain waveform data using resampling, jitter conversion, time-domain shifting, and noise addition / reduction techniques; segmenting or supplementing the augmented signal time-domain waveform data according to the required length input to the intelligent recognition algorithm to form a fixed-length signal time-domain data sample set; and finally, performing a transform domain transformation, converting the signal time-domain waveform data into time-frequency matrix data using a short-time Fourier transform, which effectively improves the performance of deep learning algorithms for signal recognition.

[0084] The above are merely preferred embodiments of the invention and do not limit the patent scope of the invention. Any equivalent structural or procedural changes made using the contents of the invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the invention.

Claims

1. A data enhancement method for signal recognition deep learning model training, characterized in that, The method comprises the following steps: S1: using a signal acquisition processing device to collect baseband signal sample data in an actual electromagnetic environment; S2: pre-processing the signal to form a single signal time domain sample set and confirming sample label information; S3: performing multi-means data enhancement on the single signal sample set to form an expanded time domain sample set, including: using one or more of signal resampling technology, frequency domain jitter, time domain jitter, artificial noise addition, and artificial noise removal means to perform data enhancement and expansion on the signal baseband IQ data, and then performing fixed-length processing, and then converting the time and frequency domains by using short-time Fourier transform; S4: according to the number of data points required by the signal recognition algorithm, performing slicing or intercepting operations on the expanded time domain sample set according to signal attributes to form a fixed-length time domain sample set, including: according to the relationship between the sampling point number of the sample data and the sampling point number required by the input of the intelligent signal recognition algorithm, each sample data needs to be selected to perform one of the following three operations: 1) fixed-length division of data samples with a data point number greater than 5 times the required input of the algorithm, 2) fixed-length intercepting of data samples with a data point number greater than 1 times and less than 5 times the required input of the algorithm from the starting point, and 3) zero padding of data samples with a data point number less than the required input of the algorithm to obtain the data point number required by the algorithm, and then adding noise, and finally converting all single signal and single burst time domain waveform data samples after expansion into a fixed-length signal time domain waveform data sample set; S5: According to the input size of the time-frequency graph of the signal recognition algorithm, the short-time Fourier transform is performed on the fixed-length time domain sample set to form a data enhanced time domain image data set that can be directly used for training. The processed fixed-length signal time domain waveform data sample set is subjected to feature domain transformation, including: using the short-time Fourier transform operation, performing a sliding window operation with an overlap rate from front to back on the data, selecting a suitable window function to limit the time-frequency domain range of the signal in the window, and calculating the power spectrum of the signal in each sliding window at each time; wherein the FFT point number is 1024, the Hamming window with a length of 256 points is selected as the window function, and the overlap rate is set to 0.5; according to the order of arrival of the sliding window, the calculated power spectrum results are spliced along the time domain to form a two-dimensional time-frequency matrix, i.e. a time-frequency matrix data, the splicing frame number is 256 frames, the signal time domain waveform data is converted into a time-frequency matrix data required for intelligent signal recognition algorithm training, and the size is .

2. The data augmentation method for signal-oriented deep learning model training according to claim 1, wherein, In the step S2, the actual collected target signal is acquired in the frequency domain and the time domain respectively, specifically including: When acquiring the actual collected data in the frequency domain, baseband IQ data is formed by down-conversion, filtering, and decimation, with the signal center frequency corresponding to 0Hz, the filtering bandwidth being the signal bandwidth, and the sampling rate of the decimated signal being four times the signal bandwidth, so as to ensure that the single sample of the data set only contains a single frequency point signal; When acquiring the actual collected data in the time domain, the start time and the end time of the signal are found through the change of the time domain waveform signal amplitude, and a plurality of single burst signal baseband IQ data is formed by intercepting, so as to ensure that the single sample of the data set only contains a single frequency point single burst signal. 3.The data augmentation method for signal-oriented deep learning model training of claim 1, wherein, When the signal baseband IQ data is enhanced, frequency domain jitter based on digital up-conversion and down-conversion is used, and the expression is specifically as follows: wherein, is the signal raw baseband IQ data, is the signal bandwidth, is the frequency domain dithering coefficient, a random number between -0.25 and 0.25, is the signal baseband IQ data after random frequency domain dithering, is the length of the baseband IQ data.

4. The data augmentation method for signal-oriented deep learning model training of claim 1, wherein, When the signal baseband IQ data is enhanced, time domain jitter based on noise point addition is used. If the time domain jitter direction is to the right, it is time domain delay, and the expression is specifically as follows: If the time domain jitter direction is to the left, it is time domain advance, and the expression is specifically as follows: wherein, is a time domain dithering coefficient, is a random number between -0.25 and 0.25, is a baseband IQ data length. 5.The data augmentation method for signal-oriented deep learning model training of claim 1, wherein, When the signal baseband IQ data is enhanced, artificial noise addition based on energy calculation is used, and the expression is specifically as follows: wherein P is the original baseband data power, is a ratio coefficient of the added noise power to the original baseband data power, is a normal distribution random sequence with a mean of 0 and a variance of 1, which is equal in length to the original baseband data.

6. The data augmentation method for signal-oriented deep learning model training of claim 1, wherein, When the signal baseband IQ data is data enhanced, artificial denoising based on wavelet transform is adopted, specifically: selecting a wavelet base function as a Harr base and a decomposition layer number as 2, wavelet decomposing original baseband data of the signal, and calculating wavelet coefficients of each layer; using a fixed threshold estimation method to set a global threshold for each decomposition layer, and performing soft threshold processing on high-order wavelet coefficients of each layer according to the set threshold; reconstructing the signal according to the processed wavelet coefficients to obtain denoised and enhanced signal data.

7. The data augmentation method for signal-oriented deep learning model training of claim 1, wherein, When the signal baseband IQ data is data enhanced, signal resampling based on a polyphase filter is adopted, specifically, a fractional multiple variable sampling processing flow of first interpolation, then filtering, and finally decimation is adopted, assuming that the ratio relationship between the resampling sampling rate and the original signal sampling rate is , and , the original signal baseband data is first up-sampled by interpolation by a factor of , then filtered through a FIR low-pass filter, and then down-sampled by decimation by a factor of M, to obtain enhanced data after resampling by a factor of ; wherein the cutoff frequency of the low-pass filter is , and the gain is .

Images