Data enhancement method based on dual-tree complex wavelet transform

Abstract

The invention discloses a data enhancement method based on dual-tree complex wavelet transform, and mainly solves the problems that a non-stationary signal data enhancement method in the prior art lacks multi-scale analysis capability, cannot fully retain local key features and is insufficient in model generalization capability. According to the scheme, the method comprises the following steps: 1) preprocessing an original signal to enable the original signal to meet the input requirement of dual-tree complex wavelet transform (DTCWT); 2) performing multi-scale decomposition on the signal by adopting DTCWT, and separating a low-frequency approximate component and a high-frequency detail component; 3) selecting a random zero sequence to replace the RZSR or a random noise sequence to replace the RNSR according to an enhancement requirement; 4) generating a random sequence, replacing high-frequency detail components, and meanwhile, retaining low-frequency components; and 5) reconstructing the signal through the inverse DTCWT, generating an enhanced signal, and outputting an enhanced signal. The method can effectively enhance the diversity of data, retains the global characteristics and local detail characteristics of signals, and can be used in the technical field of wireless communication.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, and further relates to data augmentation technology, specifically a data augmentation method based on dual-tree complex wavelet transform, which can be used for modulation signal recognition, non-stationary signal processing and deep learning model training in wireless communication. Background Technology

[0002] With the advancement of big data and computing power, deep learning has been widely applied in numerous fields such as natural language processing, audio and speech signal processing, and object detection, achieving unprecedented rapid development. However, deep learning-based methods require a large number of training samples, and in practice, obtaining a large amount of reliable and high-quality data is extremely difficult. Introducing data augmentation, while maintaining the labels, and manually expanding the training set data is a very effective approach.

[0003] Data augmentation is a technique used to increase the diversity of training data. It is widely used not only in deep learning but also in non-deep learning fields. In signal processing, image processing, and other fields, data augmentation can generate new data points by adding noise, filtering, resampling, or transforming. In signal processing, especially in tasks such as target tracking, data augmentation can help models better handle situations with limited data or discontinuous signals, improving the model's robustness and generalization ability. Data augmentation strategies include common image operations such as horizontal flipping, vertical flipping, random rotation, cropping, scaling, and adjusting brightness and contrast. These operations increase data diversity and improve the model's generalization ability.

[0004] In modern communication environments, small-sample data augmentation is a key technology to ensure that systems can accurately identify communication information in non-cooperative communication scenarios. Traditional methods, such as rotation, flipping, and scaling, can expand the size of the dataset, but the types of sample changes they can add are relatively limited. For certain specific application scenarios, they may not be able to fully simulate the complex changes in the real environment, resulting in augmented samples lacking sufficient representativeness. Furthermore, traditional augmentation methods are often based on simple image transformations or signal alterations, ignoring the intrinsic characteristics of the signal. For example, simple image flipping or rotation cannot guarantee the protection of signal features, especially when dealing with time-series data, non-stationary signals, or high-frequency noise. This may result in generated samples lacking high-quality feature representation, meaning they cannot effectively preserve the temporal or spectral characteristics of the data.

[0005] In existing technologies, the Shenzhen Institutes of Advanced Technology, in its patent document "A Data Augmentation Method Based on High-Dimensional Space Sampling" (application number CN20190123936.6), proposed a method combining dictionary matrices and Monte Carlo sampling to map training data to a high-dimensional space and then perform random sampling and correction to generate more diverse training samples. While this method can improve model training performance to some extent, it still lacks in-depth analysis of multi-scale features and a fine-grained mechanism for preserving local key features when dealing with non-stationary time-series signals. This can easily lead to the loss of details or excessive perturbation of complex signals, thereby affecting the model's generalization ability.

[0006] On the other hand, the 29th Research Institute of China Electronics Technology Group Corporation, in its patent document "A Method and Device for Enhancing Small Sample Training Data of Broadband Radiation Source Signals" (application number CN202410824315.1), employed techniques such as Doppler frequency shift, random truncation, frequency offset processing, and noise superposition to simulate different receiving scenarios through random sampling, generating diverse training samples. Although this scheme effectively improves data diversity under small sample conditions, it mainly targets the time truncation and frequency shifting operations of the signal. It still has shortcomings in dealing with multi-scale feature extraction and local detail preservation of the signal. For highly non-stationary or multi-component complex signals, relying solely on the above-mentioned random strategies is insufficient to fully capture subtle changes in the signal, and the enhancement quality is easily affected by noise and parameter settings. Summary of the Invention

[0007] This invention primarily addresses the technical problems of existing data augmentation methods when processing non-stationary time-series signals, such as the lack of effective multi-scale analysis tools, the inability to fully preserve key local features, and the tendency to lead to insufficient model generalization ability. Traditional methods, when dealing with non-stationary signals (such as EEG, speech signals, and modulated signals), often only perform simple transformations or noise injection at a single scale or wide bandwidth, making it difficult to effectively capture local features across different frequency bands. Furthermore, excessive or incorrect noise addition can destroy key information in the original signal, thus degrading the classification or recognition performance of deep learning models.

[0008] The approach to achieving the objective of this invention is as follows: First, non-stationary time-series signals are decomposed into multiple scales using DTCWT to obtain detail coefficients and approximation coefficients at different scales. Then, methods such as noise and detail coefficient replacement are used to selectively perturb and reconstruct the coefficients at each scale, simulating signal variations in different real-world scenarios while effectively preserving the main energy and key features of the original signal. By decoupling amplitude and phase information at different scales and flexibly controlling the perturbation intensity, diverse yet still detail-rich augmented data can be generated. Compared to traditional augmentation methods based on single-scale perturbations in the time or frequency domains, this invention effectively preserves key local features of the signal while increasing data diversity, improving the generalization ability and robustness of deep learning models in tasks such as signal recognition and modulation recognition.

[0009] To achieve the above objectives, the present invention includes the following:

[0010] A data augmentation method based on dual-tree complex wavelet transform, the implementation steps of which include:

[0011] (1) The system receives the original signal and preprocesses it to obtain the preprocessed signal, which is used to make the signal meet the input requirements of the dual-tree complex wavelet transform (DTCWT).

[0012] (2) Determine whether the preprocessed signal is a two-dimensional signal. If so, directly perform DTCWT two-dimensional decomposition to obtain low-frequency approximate components and multiple high-frequency detail components; otherwise, perform DTCWT one-dimensional decomposition to obtain low-frequency approximate components and one-dimensional high-frequency detail components.

[0013] (3) Based on the signal enhancement requirements and target scenario, select one data enhancement method from random zero sequence replacement RZSR and random noise sequence replacement RNSR;

[0014] (4) Generate a corresponding random sequence according to the selected enhancement method, wherein: for RZSR, generate a random zero sequence with the same length as the high-frequency component of the signal; for RNSR, calculate the energy distribution of the high-frequency detail component and generate a random noise sequence that matches it;

[0015] (5) Replace the decomposed high-frequency detail components with the random sequence generated in step (4), while keeping the low-frequency approximate components unchanged, in order to avoid global signal distortion and ensure that the global characteristics of the original signal are preserved in the enhanced signal.

[0016] (6) Combine the replaced high-frequency detail components with the low-frequency approximation components, perform inverse dual-tree complex wavelet transform IDTCWT, reconstruct the decomposed signal into a complete enhanced signal, and output it.

[0017] Applications of a data augmentation method based on dual-tree complex wavelet transform in wireless communication systems, biosignal processing, industrial IoT, smart manufacturing, and deep learning.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] First, the invention employs the dual-tree complex wavelet transform (DTCWT) to decompose signals at multiple scales, enabling precise separation of low-frequency approximate features and high-frequency detail features, and independent enhancement at different scales and frequency bands. This effectively preserves the global characteristics of the signal while enhancing local detail information, significantly improving the adaptability and recognition accuracy of deep learning models in non-stationary signal processing.

[0020] Secondly, this invention employs multiple enhancement strategies, including random zero-sequence replacement (RZSR) and random noise sequence replacement (RNSR). RZSR increases the randomness and variability of the signal by replacing detail coefficients with random zero sequences; RNSR, on the other hand, maintains the signal's energy characteristics and enhances its diversity by introducing noise that matches the signal's detail energy. This allows the signal to simulate local signal loss and noise interference in the high-frequency detail region through local perturbations, thereby generating more diverse and realistic enhanced data. This effectively avoids global signal distortion and improves data diversity and the model's anti-interference capability.

[0021] Third, this invention avoids signal distortion caused by noise by replacing the detail coefficients of the signal in the wavelet domain, rather than directly adding noise in the time or frequency domain. By maintaining the amplitude information of the detail coefficients and controlling the phase perturbation, the enhancement process can accurately simulate the changes of the signal under different environments, avoiding the feature loss problem common in traditional methods.

[0022] Fourth, because this invention employs multi-scale collaborative enhancement and precise control of the perturbation range, the signal enhancement operation can be focused on local characteristics without disrupting the global structure, thereby significantly improving the model's classification performance under low signal-to-noise ratio conditions. Test results show that the overall accuracy of the model improved by 0.93%, and the highest accuracy improved by 1.87%, effectively reducing overfitting and enhancing the model's generalization ability and practical application effectiveness.

[0023] Fifth, by combining DTCWT with RZSR and RNSR methods, this invention enables the enhanced signal to exhibit stronger robustness and adaptability in complex environments. In particular, in multi-target environments, high noise interference, and dynamically changing scenarios, this invention can effectively capture the multi-scale features of the signal, generate more adaptable training data, and greatly improve the performance of the model in complex environments. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0025] Figure 2 This is a schematic diagram of the three-level decomposition of wavelet transform in this invention;

[0026] Figure 3 This is a schematic diagram of the dual-tree complex wavelet 3-layer decomposition of the present invention;

[0027] Figure 4 This is a diagram of QPSK signal reconstruction data in this invention;

[0028] Figure 5 This is a comparison chart of simulation results before and after enhancement using the method of the present invention. Detailed Implementation

[0029] The present invention will now be further described with reference to the accompanying drawings.

[0030] Example 1: Refer to Appendix Figure 1 The present invention proposes a data augmentation method based on dual-tree complex wavelet transform, comprising the following steps:

[0031] Step 1: Receive signal and perform preprocessing

[0032] The system receives the original signal and preprocesses it to obtain a preprocessed signal, which is used to make the signal meet the input requirements of the dual-tree complex wavelet transform (DTCWT). The preprocessing includes normalization and sampling operations.

[0033] Step 2: Determine the signal type and select the decomposition method

[0034] After receiving the preprocessed signal, it is determined whether it is a two-dimensional signal. If it is a two-dimensional signal, a two-dimensional decomposition using Dual-Tree Complex Wavelet Transform (DTCWT) is performed directly to separate low-frequency approximate components and multiple high-frequency detail components. These high-frequency detail components include those in the horizontal, vertical, and diagonal directions. If it is a non-two-dimensional signal, a one-dimensional decomposition using DTCWT is performed to obtain low-frequency approximate components and one-dimensional high-frequency detail components. This step aims to select an appropriate decomposition method based on the characteristics of the signal's dimension, ensuring that the enhancement operation captures the signal's detailed characteristics across multiple scales and directions.

[0035] Step 3: Select a data augmentation method

[0036] After decomposition, a specific data augmentation method is selected based on the signal enhancement requirements and target scenario. Specifically, one method is chosen between random zero sequence replacement (RZSR) and random noise sequence replacement (RNSR).

[0037] Random Zero Sequence Replacement (RZSR) is used to replace the zero sequence of high-frequency detail components, simulating the local loss of the signal; Random Noise Sequence Replacement (RNSR) generates a random noise sequence based on the energy of the high-frequency detail components, simulating noise interference scenarios.

[0038] Step 4: Generate a random sequence

[0039] Based on the selected enhancement method, a corresponding random sequence is generated. For RZSR, a random zero sequence with the same length as the high-frequency components of the signal is generated; for RNSR, the energy distribution of the high-frequency detail components is calculated, and a random noise sequence matching it is generated. This step, based on traditional wavelet transform, achieves the locality and randomness of the enhancement operation by generating random sequences.

[0040] Step 5: Replace the detail component

[0041] The high-frequency detail components after decomposition are replaced by the random sequence generated in step 4, while the low-frequency approximate components are retained unchanged. In this embodiment, for RZSR, a random zero sequence is used to replace the detail components to simulate the loss of local information in the signal; for RNSR, a random noise sequence is used to replace the detail components to simulate the interference of noise on the signal in a real-world scenario. This operation avoids the problem of global signal distortion while enhancing the diversity of the signal, ensuring that the enhanced signal still retains the global characteristics of the original signal.

[0042] Step 6: Reconstruct the signal by reversing DTCWT

[0043] The replaced high-frequency detail components are combined with the original low-frequency approximation components, and then inverse dual-tree complex wavelet transform (IDTCWT) is performed to reconstruct the decomposed signal into a complete enhanced signal. This process, through the reconstruction characteristics of DTCWT, ensures that the enhanced signal retains the diversity of local variations while preserving the original structure.

[0044] Step 7: Output enhanced signal

[0045] The enhanced signal is output and used for subsequent deep learning model training or other signal processing tasks. The enhanced signal retains the main characteristics of the original signal while also possessing diversity and interference resistance.

[0046] This invention employs Dual-Tree Complex Wavelet Transform (DTCWT) for multi-scale signal decomposition. Compared to traditional enhancement methods, DTCWT exhibits better translation invariance and direction selectivity during signal processing, enabling precise capture of local signal features across multiple directions and scales. During enhancement, DTCWT separates low-frequency approximation features from high-frequency detail features, allowing enhancement operations to be performed independently for different frequency bands. This preserves global characteristics while enhancing local signal details. In wireless communication modulation recognition tasks, DTCWT effectively extracts instantaneous frequency changes and amplitude features, resulting in more diverse and realistically detailed enhanced data that closely resembles complex real-world scenarios. This significantly improves the adaptability and recognition accuracy of deep learning models for non-stationary signals. Furthermore, this invention combines multiple enhancement strategies, including Random Zero Sequence Replacement (RZSR) and Random Noise Sequence Replacement (RNSR), by introducing zero sequence replacement and noise interference into the high-frequency detail portion of the DTCWT decomposition, simulating local signal loss and noise perturbation. This enhancement method precisely controls the perturbation range, avoiding global signal distortion while enhancing signal diversity and anti-interference capabilities.

[0047] Example 2: The overall implementation steps of the data augmentation method provided in this example are the same as in Example 1. Please refer to the appendix for details. Figure 2-4 The implementation process of this invention is further described in detail from several aspects, including the modulation signal model, wavelet transform principle, dual-tree complex wavelet transform, decomposition / reconstruction process, and coefficient replacement:

[0048] (I) Modulation Signal Model

[0049] Modulation identification involves identifying the modulation type used to transmit the signal based on the received signal. When a receiver receives a baseband signal x(n) through a modulated time-varying wireless channel, the process can be described mathematically as follows:

[0050]

[0051] Where L is the length of the received signal r(n), * denotes convolution operation, h(n) is the channel impulse response, f0 represents the carrier frequency offset, θ represents the random phase deviation, and g(n) represents the presence of additive white Gaussian noise (AWGN) with zero mean and variance σ².

[0052] To facilitate subsequent processing of the received signal, the received signal r(n) is represented using IQ components:

[0053]

[0054] Where real and immag represent the extracted real and imaginary parts of the received signal r(n). Data augmentation refers to expanding a dataset by applying a series of transformations to the original samples. Given that the modulation type of signal sample s is ξ, an augmented dataset is generated by applying k different transformations:

[0055] {(s,ξ),(s1,ξ),…,(s k ,ξ)} (4-3)

[0056] Therefore, by applying data augmentation methods, a single sample can be expanded to (k+1) samples, thus effectively expanding the dataset. This augmented dataset is then used to train a neural network, thereby improving the accuracy of modulation recognition.

[0057] (II) Wavelet Transform

[0058] Wavelet transform was first proposed in the analysis of seismic waves, and it includes discrete wavelet transform and continuous wavelet transform.

[0059] Definition: Given a function Ψ(t) that is either a basis wavelet or a mother wavelet, a family of functions {Ψ} is obtained by scaling a and shifting b. a,b (t)} is called a wavelet.

[0060]

[0061] When a > 1, the waveform will stretch; conversely, when a < 1, the waveform will contract.

[0062] The wavelet transform of signal x(t) is to shift the basis wavelet Ψ(t) by a time b, and then perform an inner product with the signal x(t) to be analyzed using different scales a. The expression is as follows:

[0063]

[0064] If the scaling factor 'a' increases, the time width of the wavelet function increases; conversely, if the scaling factor 'a' decreases, the time width of the wavelet function decreases, while the bandwidth increases. It can be seen that wavelet transform has multi-resolution characteristics. Discrete wavelet transform is obtained by discretizing the continuous wavelet transform. Therefore, discrete wavelet transform possesses the orthogonality of the basis wavelets, and the time-domain function of the transform can better represent the characteristics of the signal itself, effectively eliminating the information redundancy defects brought about by the continuous wavelet transform, making it more suitable for processing non-stationary signals. In signal processing, wavelet transform only subdivides the low-frequency part of the signal; for example... Figure 2As shown, this is a three-level wavelet decomposition with a binary tree structure. S is the initial signal, H1, H2, and H3 are the low-frequency decomposition components, and D1, D2, and D3 are the high-frequency decomposition components.

[0065] (III) Dual-tree complex wavelet transform

[0066] The dual-tree complex wavelet transform, while inheriting the excellent properties of the wavelet transform, possesses approximately translation-invariant characteristics. During signal decomposition and reconstruction, it suppresses frequency aliasing, improving the effectiveness and authenticity of useful feature information in the signal. The basic working principle of the dual-tree complex wavelet transform is as follows: Figure 3 As shown, it consists of two parallel real part trees and imaginary part trees, which decompose and reconstruct the original input signal respectively. Figure 3 The diagram shows the process of decomposition and reconstruction of a three-layer dual-tree complex wavelet transform.

[0067] In the signal decomposition and reconstruction process, the sampling points of the imaginary part tree are always located in the middle of the real part tree. The decomposition coefficients between the two trees can be effectively utilized, reducing the loss of useful information during the entire decomposition and reconstruction process. Furthermore, the wavelet coefficient bisection method is used in each level of the decomposition process of the dual-tree complex wavelet transform to improve computational efficiency and reduce unnecessary complex calculations. Based on the principle of the dual-tree complex wavelet transform, the complex wavelet can be constructed as follows:

[0068]

[0069] In the formula and Represents two real wavelets, where i is the complex unit.

[0070] Based on the theory of wavelet transform, the wavelet coefficients and scaling coefficients of the real part tree are derived as follows:

[0071]

[0072] j = 1, 2, ..., J (4-9)

[0073] in, and These represent the wavelet coefficients and scaling coefficients of the real part tree wavelet transform, respectively. Similarly, the wavelet coefficients and scaling coefficients of the imaginary part tree are calculated according to the theory of wavelet transform:

[0074]

[0075] in, and These are the wavelet coefficients and scaling coefficients of the imaginary tree wavelet transform, respectively.

[0076] Based on the wavelet coefficients and scaling coefficients of the obtained real and imaginary part trees, the wavelet coefficients and scaling coefficients in the dual-tree complex wavelet transform are calculated:

[0077]

[0078] The reconstructed wavelet coefficients and scaling coefficients are then:

[0079]

[0080] Therefore, the signal after decomposition and reconstruction by dual-tree complex wavelet transform can be expressed as:

[0081]

[0082] (iv) Decomposition and Reconstruction Process

[0083] (4a) Decomposition process

[0084] In DTCWT, scaling functions and wavelet functions are used to extract low-frequency (approximate) and high-frequency (detail) information from a signal, respectively.

[0085] φ(x,y)=φ(x)·φ(y),

[0086] w H (x,y)=w(x)·φ(y),

[0087] w V (x,y)=φ(x)·w(y),

[0088] w D (x,y)=w(x)·w(y),

[0089] Where φ(x) is the scaling function and w(x) is the wavelet function.

[0090] For a two-dimensional signal f(x,y), decomposition using DTCWT yields a low-frequency approximate component C. A and multiple high-frequency detail components C T (Including horizontal H, vertical V, and diagonal detail components D):

[0091]

[0092] Where T∈{H,V,D}.

[0093] For a one-dimensional signal f(x), DTCWT calculates the low-frequency approximate components and high-frequency detail components using real part trees and imaginary part trees, respectively; the decomposition results are as follows:

[0094]

[0095] in, is the low-frequency coefficient of the previous layer (the (i-1)th layer); g[f] is the coefficient of the filter, representing the coefficient of the low-pass filter; F is the length of the filter, representing the termination position of the summation; f is the index of the summation, indicating that the calculation starts from the first sample during the filtering process.

[0096] High frequency coefficient The expression is as follows:

[0097]

[0098] Where h[f] are the coefficients of the filter, representing the coefficients of the high-pass filter.

[0099] (4b) Reconstruction process

[0100] A one-dimensional signal can be reconstructed from the original signal using the inverse DTCWT. The formula for the inverse transform is as follows:

[0101]

[0102] For a two-dimensional signal, the process of reconstructing the signal through inverse DTCWT is as follows:

[0103] IQ(n)=C A ·φ(x,y)+C T ·w T (x,y),T∈{H,V,D}

[0104] Among them, C A It is a low-frequency approximation component, C T It is a high-frequency detail component, w T (x,y) are wavelet functions in different directions.

[0105] (V) Coefficient Substitution Method

[0106] To address the poor performance of modulation recognition due to insufficient samples, this invention proposes a wavelet transform-based radio signal data augmentation method to expand the dataset and improve modulation recognition accuracy. First, discrete wavelet transform is applied to decompose the IQ sequence, obtaining approximation coefficients and detail coefficients at different resolutions. Next, a new sequence is generated using a suitable replacement method to replace the detail coefficients, and the replaced detail coefficients and the original approximation coefficients are used to reconstruct new samples. By continuously replacing the detail coefficients using the appropriate replacement method, a large number of new reconstructed samples are obtained. The new reconstructed samples are then merged with the original IQ samples to obtain an enhanced training set.

[0107] (5a) RZSR method

[0108] RZSR (Random Zero Sequence Replacing) increases the randomness and diversity of a signal by replacing its wavelet transform coefficients with a randomly generated sequence of zeros. The specific steps are as follows:

[0109] Step 1: Wavelet Transform and Detail Coefficient Extraction. First, perform DTCWT on the original signal f(t). After DTCWT, the low-frequency coefficients C of signal f(t) are obtained. A and high frequency coefficient C H C V C D .

[0110] Step 2: Generate a detail coefficient C H C V C D A random zero sequence Z of the same length, which is generated randomly from a uniform distribution:

[0111] Z = {z1, z2, ..., z} n},

[0112] Step 3: Combine the generated random zero sequence Z with the detail coefficients C H C V C D Element-wise multiplication yields new detail coefficients:

[0113]

[0114] If z i =0, then the corresponding detail coefficients will be nulled (high-frequency components of the signal are lost), while if z = 0, then the corresponding detail coefficients will be nulled (high-frequency components of the signal are lost), and ...ified (high-frequency components of the signal are lost), and if z = 0, then the corresponding detail coefficients will be i If the coefficient is 1, the original detail coefficients remain unchanged. This process introduces randomness and enhances the diversity of the dataset.

[0115] Step 4: Use the modified detail coefficients and the original low-frequency approximation coefficients C A Signal reconstruction is performed. The reconstruction process uses the inverse dual-tree complex wavelet transform (IDTCWT) to reconstruct the signal by adding new detail coefficients. Compared with the original low-frequency coefficient C A Perform an inverse transform to obtain the enhanced signal.

[0116] (5b) RNSR method

[0117] RNSR (Random Noise Sequence Replacing) generates new training samples by replacing the detailed coefficients of a signal with a sequence of random noise. The specific steps are as follows:

[0118] Step 1: Perform wavelet transform on the original signal to obtain the detail coefficients and approximation coefficients of the signal.

[0119] Step 2: Calculate the energy of the detail coefficients of the signal to ensure that the generated noise sequence has the same energy characteristics.

[0120] Step 3: Use a random process to generate a noise sequence with the same energy as the original detail coefficients.

[0121] Step 4: Replace the detail coefficients of the original signal with the generated noise sequence to obtain a new signal sample.

[0122] Step 5: Combine the replaced detail coefficients with the original approximation coefficients, and use inverse wavelet transform to reconstruct the signal to obtain new training samples.

[0123] The specific method is similar to RZSR, except that it requires calculating the energy of the detail coefficients of the original signal and generating a new noise sequence with the same energy as the original detail coefficients.

[0124] Example 3: This example provides an application of the data augmentation method mentioned in Example 1 or 2, including applications in wireless communication systems, biosignal processing, industrial IoT, smart manufacturing, and deep learning. The following provides specific practical applications of the method of this invention, further elaborating on its application prospects:

[0125] 1. Wireless Communication and Modulation Identification:

[0126] In the field of wireless communication, modulation recognition is a key technology, especially in complex environments (such as multipath propagation, noise, and clutter). Traditional data augmentation methods, such as image rotation and scaling, are often insufficient to capture the characteristics of wireless signals. This invention, through multi-scale analysis of DTCWT and enhancement strategies of RZSR and RNSR, can accurately capture the local features and diversity of signals in the time and frequency domains, thereby improving the robustness and accuracy of modulation recognition systems in high-noise and high-interference environments. This method can effectively enhance training data and improve the modulation recognition accuracy of wireless communication systems (such as 5G and IoT), and is particularly suitable for applications such as radio monitoring, interference source localization, and signal recognition.

[0127] 2. Biosignal Processing:

[0128] The signal enhancement method proposed in this invention has significant application value in biological signal processing (such as medical imaging data like electroencephalograms (EEG) and electrocardiograms (ECG)). In biological signal processing, signal acquisition is often affected by noise and artifacts, leading to signal quality degradation and impacting subsequent diagnosis and analysis. Utilizing dual-tree complex wavelet transform for multi-scale processing can effectively extract the instantaneous frequency and amplitude features of the signal. Furthermore, RZSR and RNSR enhancement methods can simulate noise variations under different environments while preserving signal characteristics, thus enhancing data diversity. This enhancement method can improve the accuracy of biological signal analysis in applications such as medical image analysis, disease prediction, and anomaly detection, helping doctors make more precise diagnoses.

[0129] 3. Industrial Internet of Things and Smart Manufacturing:

[0130] In the Industrial Internet of Things (IIoT) and smart manufacturing, equipment condition monitoring and fault diagnosis are core applications. Sensor data often contains noise and outliers, which traditional data augmentation methods struggle to handle. This invention enhances the multi-scale features of signals through dual-tree complex wavelet transform and utilizes RZSR and RNSR to simulate various interference scenarios, effectively improving the training performance of equipment health monitoring and fault prediction models. Especially when equipment experiences abnormal fluctuations, the augmented data can be used to better train the model, improving the accuracy of fault diagnosis and enabling early fault prediction, intelligent operation and maintenance, and equipment lifecycle management.

[0131] 4. Deep Learning and Transfer Learning:

[0132] With the widespread application of deep learning technology, the size and quality of datasets directly affect the effectiveness of model training. In deep learning training, especially in few-shot learning, data augmentation plays a crucial role. Traditional data augmentation methods, such as flipping and scaling, often fail to uncover the diversity of data at a deeper level. This invention, by combining wavelet transform with various random sequence replacements, effectively augments the dataset and improves the model's generalization ability. By applying the augmentation methods of this invention in different domains and tasks (such as image recognition, speech processing, and natural language processing), the accuracy of deep learning models can be significantly improved, the risk of overfitting reduced, and the adaptability and performance of models enhanced in transfer learning.

[0133] This invention, by combining dual-tree complex wavelet transform with random sequence substitution techniques, can provide richer and more diverse training data for signal processing tasks in various fields, thereby improving the model's performance, robustness, and generalization ability. It has broad application prospects in wireless communication, biomedical signal processing, industrial IoT, deep learning, and other fields, and possesses significant practical value and wide-ranging application potential.

[0134] The effects of the present invention will be further explained below with reference to simulation experiments.

[0135] 1. Simulation conditions:

[0136] The simulation experiments of this invention were completed in an NVIDIA GeForce GTX 3080 GPU hardware environment, using the TensorFlow framework for data augmentation and model training, with acceleration parts supported by NVIDIA CUDA.

[0137] 2. Simulation content:

[0138] The dataset used is one-tenth of the open-source RML2016.10a modulated signal dataset. Each modulated signal contains 100 samples at each signal-to-noise ratio (SNR), with SNR ranging from -20dB to 18dB, in 2dB increments. The data is split in a 3:1:1 ratio (training set, validation set, test set). In the experiments, the performance differences of the proposed DTCWT-based enhancement method (combining RZSR and RNSR) with traditional enhancement methods (such as flipping, scaling, and noise superposition) in deep learning models were compared. First, the signals in the training set were decomposed using dual-tree complex wavelet transform (DTCWT) to extract low-frequency approximate components and high-frequency detail components. Enhancement was then performed using a mixture of RZSR and RNSR. Subsequently, a CNN-LSTM model was trained using the enhanced data, and its classification accuracy at different SNRs was evaluated. Simulation results are shown below. Figure 5 As shown.

[0139] 3. Simulation results:

[0140] Reference Figure 5 The simulation results before and after data augmentation are shown. Before data augmentation, with 100 samples (60 training samples), the average recognition rate was 50.09%, and the highest recognition rate was 76.77%. After data augmentation, with 160 samples (120 training samples, doubled from 60), the average recognition rate was 51.02%, and the highest recognition rate was 78.64%. It can be seen that both the average and highest recognition rates have increased to varying degrees.

[0141] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0142] The above simulation analysis proves the correctness and effectiveness of the method proposed in this invention.

[0143] The parts of this invention not described in detail are common knowledge to those skilled in the art.

[0144] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, those skilled in the art, after understanding the content and principle of the present invention, may make various modifications and changes in form and detail without departing from the principle and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A data augmentation method based on dual-tree complex wavelet transform, characterized in that, Includes the following steps: (1) The system receives the original signal and preprocesses it to obtain the preprocessed signal, which is used to make the signal meet the input requirements of the dual-tree complex wavelet transform (DTCWT). (2) Determine whether the preprocessed signal is a two-dimensional signal. If so, directly perform DTCWT two-dimensional decomposition to obtain low-frequency approximate components and multiple high-frequency detail components; otherwise, perform DTCWT one-dimensional decomposition to obtain low-frequency approximate components and one-dimensional high-frequency detail components. (3) Based on the signal enhancement requirements and target scenario, select one data enhancement method from random zero sequence replacement RZSR and random noise sequence replacement RNSR; (4) Generate a corresponding random sequence according to the selected enhancement method, wherein: for RZSR, generate a random zero sequence with the same length as the high-frequency component of the signal; for RNSR, calculate the energy distribution of the high-frequency detail component and generate a random noise sequence that matches it; (5) Replace the decomposed high-frequency detail components with the random sequence generated in step (4), while keeping the low-frequency approximate components unchanged, in order to avoid global signal distortion and ensure that the global characteristics of the original signal are preserved in the enhanced signal. (6) Combine the replaced high-frequency detail components with the low-frequency approximation components, perform inverse dual-tree complex wavelet transform IDTCWT, reconstruct the decomposed signal into a complete enhanced signal, and output it.

2. The data augmentation method according to claim 1, characterized in that: The preprocessing described in step (1) includes normalization and sampling operations.

3. The data augmentation method according to claim 1, characterized in that: The high-frequency detail components mentioned in step (2) include horizontal, vertical and diagonal directions.

4. The data augmentation method according to claim 1, characterized in that: The random zero sequence replacement RZSR mentioned in step (3) is used to replace the high-frequency detail components with zero sequences to simulate the local loss of the signal; the random noise sequence replacement RNSR generates a random noise sequence based on the energy of the high-frequency detail components to simulate the noise interference scenario.

5. The data augmentation method according to claim 1, characterized in that: The energy distribution of the high-frequency detail components mentioned in step (4) is calculated according to the following formula: Where N is the number of sampling points for the high-frequency components, k = 1, 2, ... N; C T (k) represents the high-frequency coefficient of the kth sampling point.

6. The data augmentation method according to claim 5, characterized in that: The random noise sequence mentioned in step (4) is generated according to the following steps: (4.1) The RNSR generates random noise that follows a Gaussian distribution with zero mean and variance matching the energy of its high-frequency components: Where Z(0,1) is a standard normal distribution; (4.2) Perform a first-order autoregressive filter on the noise sequence to generate a time-dependent noise sequence: Noise corr (k)=α·Noise corr (k-1)+β·Noise(k), Where the attenuation factor α ∈ [0.6, 0.8], Used to maintain a constant noise energy.

7. The data augmentation method according to claim 1, characterized in that: The two-dimensional decomposition and one-dimensional decomposition of DTCWT in step (2) are implemented as follows: For a two-dimensional signal f(x, y), the scaling function φ(x, y) and the directional wavelet function w are used. T (x, y) is decomposed to obtain the low-frequency approximate component C. A and multiple high-frequency detail components C T : Where L is the signal size, φ(x, y) = φ(x)·φ(y), and the directional wavelet function is defined as follows: W H (x,y)=w(x)·φ(y) W V (x,y)=φ(x)·w(y) W D (x,y)=w(x)·w(y) Where φ represents the scaling function, w is the wavelet function, and H, V, and D represent the horizontal, vertical, and diagonal directions, respectively; For a one-dimensional signal f(x), the low-frequency approximate components are calculated using real part trees and imaginary part trees, respectively. and high-frequency detail components in, These are the low-frequency coefficients of the (i-1)th layer; g[f] represents the low-pass filter coefficients; F is the length of the filter; f is the index of the summation; and h[f] represents the high-frequency coefficients.

8. The data augmentation method according to claim 7, characterized in that: The reconstruction process of the dual-tree complex wavelet inverse transform described in step (6) is as follows: The inverse transform formula for reconstructing the original signal from a one-dimensional signal using IDTCWT is as follows: For a two-dimensional signal, the process of reconstructing the signal using IDTCWT is as follows: f(x,y)=C A ·φ(x,y)+∑ T∈{H，V，D} C T ·w T (x,y)。 9. The application of the data augmentation method based on dual-tree complex wavelet transform as described in claim 1 in wireless communication systems, biosignal processing, industrial IoT, smart manufacturing, and deep learning.

10. The application according to claim 9, characterized in that: The wireless communication system includes 5G and the Internet of Things (IoT); the biosignal processing includes processing medical image data such as electroencephalogram (EEG) and electrocardiogram (ECG); the industrial IoT and smart manufacturing include status monitoring and fault diagnosis of their equipment; and the deep learning includes image recognition, speech processing, and natural language processing.

Images