Signal classification enhancement method based on multistage wavelet decomposition and hybrid expert model

Abstract

A signal feature compression method based on codebook discrete quantization and multi-task learning comprises the following steps: S1, inputting an original IQ signal, and obtaining discrete feature representation through an encoder and a codebook; and S2, taking the discrete feature representation obtained in the step S1 as an input, performing L-level adaptive wavelet decomposition, generating a low-frequency component and a high-frequency component in each level of decomposition, and taking the k-level low-frequency component as an input of the k + 1-level decomposition, thereby forming a decomposition pyramid. And S3, designing a plurality of expert classification networks, for the kth expert, constructing the input of the kth expert into splicing of the first k high-frequency components and the kth-level low-frequency component, and in order to ensure that the time dimensions of different components are aligned, adjusting all the components to the minimum length by adopting a linear interpolation method. And S4: hybrid expert decision making: an expert pool containing L experts is constructed, and each expert is a lightweight convolutional neural network and is responsible for classifying the corresponding input constructed in S3. Meanwhile, a gating network is constructed, original discrete features are expressed as input, and an L-dimensional probability distribution weight vector is output to represent the importance of each expert on a current input signal. And S5, carrying out weighted fusion and output. And performing weighted summation on the classification output of each expert and the corresponding weight generated by the gating network to obtain a final classification prediction result. The method can effectively improve the accuracy and robustness of signal classification in a complex electromagnetic environment.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of signal processing and artificial intelligence, specifically to a signal classification enhancement method based on multi-level wavelet decomposition and a hybrid expert model. This method is applicable to complex signal classification tasks in scenarios such as communication signal modulation recognition, radar signal recognition, and unmanned system signal perception. By introducing a learnable wavelet multi-scale decomposition structure and a multi-expert collaborative decision-making mechanism, it achieves adaptive modeling and fusion of different frequency characteristics and noise environments, thereby effectively improving the classification accuracy and robustness of the model under multiple signal-to-noise ratios and multiple channel conditions. Background Technology

[0002] With the rapid development of wireless communication, radar detection and Internet of Things systems, signal classification technology is playing an increasingly important role in fields such as intelligent communication, spectrum monitoring and electronic countermeasures [1]. "Gardner W A. Signalinterception: A unifying theoretical framework for feature detection [J]. IEEE Transactions on communications, 2002, 36(8): 897-906." In typical tasks such as modulation recognition and target recognition, the system needs to accurately determine the type of the received IQ signal in order to support subsequent signal demodulation, interference cancellation and resource scheduling. However, in complex electromagnetic environments, signals are easily affected by noise, fading and nonlinear distortion, which leads to a significant decrease in classification performance. Therefore, it is of great engineering significance to study a highly robust and accurate signal classification method.

[0003] Early signal classification methods were mainly based on manual feature extraction and traditional machine learning algorithms. Common methods include using statistical measures such as higher-order cumulants, cyclostationary features and instantaneous features to construct signal discrimination indices [2] "Nandi AK, Azzouz E E. Algorithms for automatic modulation recognition of communication signals [J]. IEEE Transactions on communications, 2002, 46(4): 431-436." These methods usually input manually designed features into classifiers such as support vector machines (SVM), decision trees or K-nearest neighbors (KNN) for recognition. Although such methods perform well in high signal-to-noise ratio scenarios, their feature design relies on empirical knowledge, is sensitive to noise and lacks generalization, and is difficult to adapt to low signal-to-noise ratio and multipath fading environments.

[0004] With the rise of deep learning technology, researchers began to explore end-to-end signal classification models. O'Shea et al. first applied convolutional neural networks (CNN) to the task of wireless signal modulation recognition and significantly improved the classification accuracy on the RML2016.10A dataset [3] “O'Shea TJ, Corgan J, Clancy T C. Convolutional radiomodulation recognition networks [C] / / International conference on engineering applications of neural networks. Cham: Springer International Publishing, 2016: 213-226.” Subsequently, various models based on recurrent neural networks (RNN), gated recurrent units (GRU) and Transformer structures were proposed to enhance the model’s ability to model temporal correlation and long-range dependence [4] “Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need [J]. Advances in neural information processing systems, 2017, 30.” These methods do not require manual feature design and achieve end-to-end feature learning, but they still have problems such as fixed structure, insufficient utilization of frequency domain information, and unstable performance in cross signal-to-noise ratio scenarios.

[0005] To enhance the multi-scale feature representation capability of the model, researchers have introduced time-frequency analysis methods such as wavelet decomposition. Unlike Fourier transform, wavelet decomposition has good time-frequency localization characteristics and can simultaneously characterize the low-frequency trend and high-frequency details of the signal at different scales [5] "Mallat S G. A theory for multiresolution signal decomposition: the wavelet representation [J]. IEEE transactions on pattern analysis and machine intelligence, 2002, 11(7): 674-693." In modulation recognition and fault diagnosis, wavelet decomposition can effectively extract envelope features, transient features and local frequency changes, thereby improving classification robustness. However, traditional wavelet decomposition relies on a fixed wavelet basis and is difficult to adapt to different signal types. In recent years, some studies have proposed the Learnable Wavelet Transform (LWT) method, which automatically optimizes wavelet filters through neural networks to achieve adaptive decomposition of signals and has performed well in speech and vibration signal tasks [6] “Zhang J, Wang T, Feng Z, et al. Toward the automatic modulation classification with adaptive wavelet network[J]. IEEE Transactions on Cognitive Communications and Networking, 2023, 9(3): 549-563.” However, in the field of communication signal classification, this type of method is mostly used only as feature preprocessing and is not jointly optimized with the task objective, resulting in the decomposition structure being disconnected from the classification task.

[0006] Meanwhile, Mixture of Experts (MoE) has received widespread attention in the field of deep learning as a dynamic structured modeling method. MoE processes different feature subspaces of the input through multiple expert networks, and the gating network assigns weights according to the input features to achieve collaborative learning among experts [7] "Jacobs RA, Jordan MI, Nowlan SJ, et al. Adaptive mixtures of local experts[J]. Neuralcomputation, 1991, 3(1): 79-87." This structure can improve the expressive power of the model while ensuring computational efficiency. In recent years, the MoE structure has achieved significant results in large-scale language models and image recognition. However, in the field of communication signals, existing MoE research is mostly based on unified feature input and lacks differentiated expert design for multi-frequency and multi-scale signal structures, making it difficult for the model to fully explore the complementary information of features in different frequency bands.

[0007] Furthermore, deep learning models still face challenges in signal classification tasks, including high computational complexity, sensitivity to noise, and poor generalization across signal-to-noise ratios. Especially in low signal-to-noise ratio environments, models are prone to learning noise patterns, leading to overfitting. While some studies enhance model robustness through multi-task learning or contrastive learning, they generally lack multi-scale signal structure modeling mechanisms, making it difficult to achieve effective decomposition and fusion at the feature level.

[0008] In summary, existing signal classification methods still have the following shortcomings: (1) Traditional manual feature methods are unstable in low signal-to-noise ratio environments; (2) Although deep learning models have end-to-end capabilities, they lack a mechanism for fusing frequency domain and multi-scale information. (3) Traditional wavelet decomposition methods have a fixed structure and cannot achieve adaptive decomposition and joint task optimization; (4) The hybrid expert model failed to combine multi-scale signal features for differentiated modeling; (5) The model as a whole is not robust enough under cross-channel and low signal-to-noise ratio conditions.

[0009] Therefore, there is an urgent need for a signal classification enhancement method that combines multi-level wavelet decomposition with a hybrid expert mechanism. This method achieves adaptive multi-scale decomposition of the signal by introducing a learnable wavelet decomposition structure, and then uses a gating network to realize dynamic weighting and collaborative decision-making among multiple experts, thereby fully integrating frequency information and feature patterns at different scales and effectively improving the accuracy and robustness of signal classification in complex electromagnetic environments. Summary of the Invention

[0010] To overcome the problems of poor robustness, limited feature representation ability, and insufficient model generalization performance of existing signal classification methods in low signal-to-noise ratio, multi-channel, and non-stationary environments, this invention proposes a signal classification enhancement method based on multi-level wavelet decomposition and hybrid expert model.

[0011] This invention constructs a learnable wavelet decomposition module to perform multi-scale decomposition of the features of the input discrete signal, achieving adaptive feature extraction and hierarchical representation of different frequency components. Simultaneously, it combines a hybrid expert structure, designing multiple expert classification networks for features at different scales, and dynamically assigning weights through a gating network to achieve adaptive collaboration and decision fusion among experts. This innovative method jointly optimizes multi-scale time-frequency decomposition and expert selection mechanisms within a unified end-to-end framework, preserving the multi-scale analysis advantages of wavelet transform while fully utilizing the nonlinear modeling capabilities of deep networks. In this way, this invention can effectively improve the accuracy and robustness of signal classification in complex electromagnetic environments, providing efficient and reliable signal recognition solutions for scenarios such as intelligent communication, radar identification, and wireless sensing.

[0012] The technical solution adopted by this invention to solve its technical problem is as follows: A signal classification enhancement method based on multi-level wavelet decomposition and hybrid expert model includes the following steps: S1: Input the original IQ signal, and obtain discrete feature representations through the encoder and codebook.

[0013] S2: Using the discrete feature representation obtained in step S1 as input, perform L-level adaptive wavelet decomposition, with each level of decomposition generating a low-frequency component. and a high-frequency component The low-frequency component of the kth order This will be used as the input for the (k+1)th level of decomposition, thus forming the decomposition pyramid.

[0014] S3: Design multiple expert classification networks. For the k-th expert, construct its input as a concatenation of the first k high-frequency components and the k-th low-frequency component. To ensure the alignment of the time dimensions of different components, use linear interpolation to adjust all components to the minimum length among them.

[0015] S4: Hybrid Expert Decision Making: Construct an expert pool containing L experts, each of whom is a lightweight convolutional neural network responsible for classifying the corresponding input constructed in S3. Simultaneously, construct a gating network that takes the original discrete feature representation as input and outputs an L-dimensional probability distribution weight vector, representing the importance of each expert to the current input signal.

[0016] S5: Weighted Fusion and Output. The classification output (logits) of each expert is weighted and summed with the corresponding weights generated by the gating network to obtain the final classification prediction result.

[0017] Preferably, step S1 includes the following: S1.1: Encode the original signal to obtain high-dimensional continuous features.

[0018] Input raw IQ signal (N is the number of sampling points), through the encoder Extracting high-dimensional continuous features: (1) Where D is the feature dimension.

[0019] S1.2: Perform vector quantization on high-dimensional continuous features to output discrete features.

[0020] Using codebook ( (Codebook size) for high-dimensional continuous features Discrete quantization, computation With all codebook vectors Euclidean distance, select nearest neighbor index : (2) Output discrete features .

[0021] Preferably, step S2 includes the following: S2.1: Adaptive wavelet decomposition is implemented using a learnable lifting scheme. Let the input of the k-th level decomposition be... (Regulation Its decomposition process includes the following sub-steps: S2.1.1: Split, input The parity index is split into two subsets based on the time dimension: the even part. and odd part .

[0022] S2.1.2: Prediction, using a prediction operator consisting of convolutional layers. P according to predict Calculate detail coefficients (high-frequency components). The formula is: (3) S2.1.3: Update, using an update operator consisting of convolutional layers. According to detail factor renew This yields the approximation coefficients (low-frequency components). The formula is: (4) S2.2: Approximation coefficients of the previous stage output As input for the next level of decomposition, repeat step S2.1 until the L-level decomposition is completed, yielding the decomposed sequence. .

[0023] Preferably, step S3 includes the following: S3.1: For the first k An expert classification network is constructed using input features. , its origin k High-frequency components With the k Low-frequency components It is pieced together along the channel dimension, and its mathematical expression is: (5) Before the splicing operation, first... to as well as Adjust to the same length over the time dimension using a linear interpolation method. ,in For this k The minimum value of the time dimension among the +1 components.

[0024] Preferably, step S4 includes the following: S4.1: The gated network consists of a one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer connected sequentially. Its input is the discrete feature representation obtained in step S1. .

[0025] S4.2: The fully connected layer of the gated network outputs an L-dimensional vector. Applying a Softmax function to this vector yields the weight vectors for each expert. ,satisfy .

[0026] Preferably, step S5 includes the following: S6.1: Output of each expert network Corresponding weights generated by the gating network Perform a weighted summation to obtain the final classification result. .

[0027] (6)

[0028] The advantages of this invention are: by constructing a learnable wavelet decomposition module, the features of the input discrete signal are decomposed at multiple scales, achieving adaptive feature extraction and hierarchical representation of different frequency components; simultaneously, by combining a hybrid expert structure, multiple expert classification networks targeting different scale features are designed, and weights are dynamically allocated through a gating network to achieve adaptive collaboration and decision fusion among experts. This method innovatively optimizes multi-scale time-frequency decomposition and expert selection mechanisms within a unified end-to-end framework, preserving the multi-scale analysis advantages of wavelet transform while fully utilizing the nonlinear modeling capabilities of deep networks. In this way, this invention can effectively improve the accuracy and robustness of signal classification in complex electromagnetic environments. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method of the present invention.

[0030] Figure 2 This is a diagram of the model training framework used in this invention. Detailed Implementation

[0031] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0032] This embodiment provides a signal classification enhancement method based on multi-level wavelet decomposition and hybrid expert model, which specifically includes the following steps: Step 1: Input the raw IQ signal, and obtain discrete feature representations through the encoder and codebook; Step 2: Using the discrete feature representation obtained in Step 1 as input, perform L-level adaptive wavelet decomposition. Each level of decomposition generates a low-frequency component and a high-frequency component. The low-frequency component of the k-th level will be used as the input of the (k+1)-th level decomposition, thus forming a decomposition pyramid. Step 3: Design multiple expert classification networks. For the k-th expert, construct its input as a concatenation of the first k high-frequency components and the k-th low-frequency component. To ensure the alignment of the time dimensions of different components, use linear interpolation to adjust all components to the minimum length among them.

[0033] Step 4: Hybrid expert decision-making. An expert pool of L experts is constructed, each being a lightweight convolutional neural network responsible for classifying the input generated in Step 3. Simultaneously, a gating network is constructed, taking the original discrete feature representation as input and outputting an L-dimensional probability distribution weight vector, representing the importance of each expert to the current input signal.

[0034] Step 5: Weighted Fusion and Output. The classification output (logits) of each expert is weighted and summed with the corresponding weights generated by the gating network to obtain the final classification prediction result.

[0035] In step 1, the specific operation process is as follows: (Refer to the appendix) Figure 1 , 2 The system receives a complex IQ signal input containing in-phase and quadrature components, and extracts signal features through a convolutional neural network encoder module. The high-dimensional features output from the encoder are then input to a vector quantization module. This module maintains a learnable embedding space as a codebook, containing a preset number of discrete embedding vectors. By calculating the Euclidean distance between the input features and all codebook vectors, the nearest-neighbor codebook vector is selected to replace the original features. This process generates a discrete index representation, and simultaneously calculates the mean square error between the original features and the selected codebook vector as the quantization loss.

[0036] In step 2, the specific operation process is as follows: (Refer to the appendix) Figure 1 , 2 The discrete features obtained in step 1 are used as input, and an L-level adaptive wavelet decomposition is performed using a learnable wavelet decomposition module. Each level of decomposition includes two stages: prediction and update. In the k-th level decomposition, the input features are split into odd and even subsequences along the time dimension. The high-frequency details of the even sequence are predicted by the prediction operator (composed of a one-dimensional convolutional network), thus obtaining the high-frequency components; subsequently, the low-frequency part is smoothed and corrected using the update operator to generate low-frequency approximation components. The generated low-frequency components will be used as input for the next level of decomposition, forming a progressively multi-scale pyramid structure. This learnable wavelet decomposition method can automatically adjust the filter parameters according to the task optimization, thereby adaptively extracting features of the signal in different frequency ranges.

[0037] In step 3, the specific operation process is as follows: (Refer to the appendix) Figure 1 , 2 For the k-th expert, its input is defined as a feature tensor formed by concatenating the first k high-frequency components and the low-frequency component of the current level along the channel dimension. Since the time lengths of different components may differ, to ensure consistency in the concatenated feature dimension, this invention uses linear interpolation to adjust all components to the same minimum time length. This design guarantees the alignment of information at different scales, enabling experts to simultaneously focus on the detailed changes and overall trends of the signal. Each expert network employs a lightweight convolutional neural network structure to extract discriminative features at the corresponding scale and output classification results. Through multi-expert parallel modeling, different experts can focus on different frequency levels and signal-to-noise ratio ranges, thereby achieving feature division of labor and complementary representations.

[0038] In step 4, the specific operation process is as follows: (Refer to the appendix) Figure 1 , 2An expert pool containing L experts is constructed, and a gating network is designed to adaptively allocate expert weights. The input to the gating network is the discrete feature representation obtained in step 1. After passing through a one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer, it outputs an L-dimensional weight vector. A Softmax function constraint is applied to this vector, ensuring that the sum of all weights is 1, representing the importance distribution of each expert under the current input. This gating mechanism enables the model to dynamically select a more suitable expert combination based on the features of the input signal, realizing a structured modeling process from "fixed structure" to "dynamic selection". When the input signal is in a low signal-to-noise ratio environment, the gating network can adaptively increase the weight of low-frequency experts; under high signal-to-noise ratio conditions, it enhances the attention of high-frequency experts to detailed features.

[0039] In step 5, the specific operation process is as follows: (Refer to the appendix) Figure 1 , 2 The classification results (logits) output by all expert networks are weighted and summed with the weights generated by the gating network to obtain the fused global classification prediction result. This fusion mechanism achieves unified integration of features at different scales and decision results from different experts, thereby improving the overall classification accuracy and stability. During the model training phase, this invention adopts an end-to-end optimization approach, simultaneously minimizing classification loss, quantization reconstruction loss, and wavelet decomposition constraint loss to ensure that the feature decomposition module, expert decision module, and gating assignment module can be optimized collaboratively. Through this joint training mechanism, this invention achieves a unified optimization framework from signal feature extraction to final classification prediction.

[0040] The above description illustrates typical applications of this invention in communication signal modulation identification and complex channel identification scenarios. This invention achieves multi-scale extraction and hierarchical representation of signal features by introducing multi-level adaptive wavelet decomposition. Combined with a hybrid expert model and a gated weighting mechanism, it enables adaptive fusion and decision enhancement of signal features under different scales and signal-to-noise ratios. This method can significantly improve signal classification accuracy and robustness under conditions of multiple interferences, multiple paths, and non-stationary channels, demonstrating good scalability and engineering application value. This description is merely illustrative and not restrictive. Those skilled in the art will understand that many changes, modifications, and even equivalents can be made within the spirit and scope defined by the claims, all of which will fall within the protection scope of this invention.

Claims

1. A signal classification enhancement method based on multi-level wavelet decomposition and hybrid expert model, characterized in that: Comprising the following steps: S1: input the original IQ signal, get the discrete feature representation through the encoder and codebook. S2: taking the discrete feature representation obtained in step S1 as input, carrying out L-level adaptive wavelet decomposition, and generating a low-frequency component and a high-frequency component at each level the low-frequency component at the kth level will be taken as the input of the (k+1)th level decomposition, thus forming a decomposition pyramid.​ S3: design multiple expert classification networks, for the kth expert, its input is constructed as the splicing of the first k high-frequency components and the kth low-frequency component, in order to ensure the time dimension alignment of different components, linear interpolation method is used to adjust all components to the smallest length among them. S4: mixed expert decision: build an expert pool containing L experts, each expert is a lightweight convolutional neural network responsible for classifying the corresponding input constructed in S3. At the same time, a gating network is constructed, which takes the original discrete feature representation as input and outputs an L-dimensional probability distribution weight vector representing the importance of each expert for the current input signal. S5: weighted fusion and output. Weighted sum the classification output (logits) of each expert and the corresponding weight generated by the gating network to get the final classification prediction result.

2. The signal classification enhancement method based on multi-level wavelet decomposition and hybrid expert model according to claim 1, characterized in that: Step S1 includes the following: S1.1: encode the original signal to get high-dimensional continuous features. Input raw IQ signal (N is the number of sampling points), through the encoder Extract high-dimensional continuous features: (1) Where D is the feature dimension. S1.2: vector quantization of high-dimensional continuous features, output discrete features. Utilizing codebooks ( For high dimensional continuous features Discrete quantization, compute Euclidean distance to all codebook vectors Select nearest neighbor index : (2) Output discrete features .

3. The method of claim 1, wherein the multi-level wavelet decomposition and hybrid expert model based signal classification enhancement method is characterized by: Step S2 includes the following: S2.1: The adaptive wavelet decomposition is implemented using a learnable lifting scheme, let the input of the k-th level decomposition be (Definition ), the decomposition process contains the following sub-steps: S2.1.1: Split, input Splitting the even part and the odd part of the time dimension into two subsets: even and odd. S2.1.2: prediction, using a prediction operator consisting of a convolution layer P According to Prediction , the detail coefficients (high frequency components) are calculated with the formula: (3) S2.1.3: update, using an update operator consisting of a convolution layer According to the detail coefficients Update , to obtain the approximation coefficients (low frequency components) The formula is: (4) S2.2: output the approximation coefficients of the previous step As input for the next step of decomposition, step S2.1 is repeated until the L-level decomposition is completed, resulting in a decomposition sequence .

4. The method of claim 1, wherein the multi-level wavelet decomposition and hybrid expert model based signal classification enhancement method is characterized by: Step S3 includes the following content: constructing input features for the first k expert classification network , which are spliced along the channel dimension by the first k high-frequency component and the first k low-frequency component , and the mathematical expression is: (5) Before the stitching operation, first of all, the to and adjust to the same length in the time dimension by the linear interpolation method where is the minimum value of the time dimension in the k +1 components.

5. The multi-stage wavelet decomposition and hybrid expert model based signal classification enhancement method of claim 1, wherein: Step S4 includes the following: S4.1 : the gating network is composed of a one-dimensional convolutional layer, a global average pooling layer and a fully connected layer sequentially connected, and the input is the discrete feature representation obtained in step S1 . S4.2: The fully connected layer of the gating network outputs an L-dimensional vector, on which a Softmax function is applied to obtain a weight vector for each expert , satisfying .

6. The multi-stage wavelet decomposition and hybrid expert model based signal classification enhancement method of claim 1, wherein: The step S5 comprises the following content: weighting and summing the outputs of each expert network with the corresponding weights generated by the gating network to obtain the final classification result . (6)。

Images