Signal classification method based on multi-domain feature fusion full complex valued network
By adopting the full complex neural network architecture and multi-domain transformation technology in electromagnetic signal processing, the problem of insufficient complex data analysis capabilities in the existing technology is solved, and efficient classification of electromagnetic signals and comprehensive mining of multi-domain features is achieved.
Patent Information
- Application Number
- CN202411481259.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-23
- Publication Date
- 2025-02-28
AI Technical Summary
Existing real-number neural network technologies are difficult to effectively mine and analyze complex data in electromagnetic signals, which can easily cause information loss or discarding.
The signal classification method based on multi-domain features fusion full complex value network is adopted, and the multi-domain features are comprehensively mined through the full complex operation architecture, custom normalized operators and activation functions promoted in complex planes, combined with invariant transformation, short-time Fourier transform and Wigner-Ville transformation, multi-domain features in electromagnetic signals are comprehensively mined.
It realizes efficient classification of electromagnetic signals, significantly improves the performance of signal classification, avoids information loss and makes full use of the characteristics of complex data.
Smart Images
Figure CN119538025A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of signal processing and is a method for realizing efficient signal classification. Background Art
[0002] At present, deep learning technology is mainly driven by computer vision, natural language processing and other fields. The neural network architecture obtained from the above two fields has strong feature mining and analysis capabilities in the corresponding fields. However, due to the low demand for complex data analysis in related fields, there is still a lack of efficient neural network feature mining and analysis methods for complex data in the electromagnetic field. Compared with natural language and computer vision data, on the one hand, the information distribution of electromagnetic signal data is often more sparse, and it is more difficult to extract semantic information; on the other hand, the characteristics of complex data provide an additional dimension for signal analysis, and the use of existing real neural network technology often easily causes information loss or even discarding of this dimension. Summary of the invention
[0003] The present invention proposes a signal classification method based on a multi-domain feature fusion fully complex-valued network, which realizes efficient classification of electromagnetic signals and can comprehensively mine multi-domain features in electromagnetic signals.
[0004] The technical solution to realize the present invention includes:
[0005] The neural network used for signal classification adopts a full complex operation architecture. In addition to the weights being complex numbers and satisfying the complex operation rules, it uses a custom normalization operator and an activation function generalized in the complex plane. The data is transformed into multiple domains before being sent to the neural network. The backbone feature extraction network of the neural network completes feature extraction in multiple domains respectively, and multiple features are spliced before the classification layer of the neural network to finally achieve signal classification.
[0006] Furthermore, the custom normalization operator performs operations according to the following steps: (1) calculating the complex conjugate of the input operator data; (2) calculating the modulus square value of the input operator data; (3) calculating the sum of all modulus square values to obtain a normalization operator; (4) dividing each data modulus square value by the normalization operator to obtain a normalized result.
[0007] Furthermore, the activation function is:
[0008]
[0009] Where z is the data input to the activation function, and φ is the principal value of the input data.
[0010] Furthermore, the multi-domain transformations performed before inputting into the neural network include: invariant transformation, short-time Fourier transformation and Wigner-Ville transformation.
[0011] Furthermore, the backbone feature extraction network of the complex-valued neural network performs feature mining on the multi-domain transformation respectively, and after obtaining three sets of feature vectors, they are concatenated to form a complete feature vector, which is sent to the neural network classification layer to realize signal classification.
[0012] The beneficial effects of the present invention include:
[0013] Using a fully complex neural network architecture, starting from the normalization method, activation function and data multi-domain transformation, a normalization method suitable for complex data is proposed. A ReLU activation function generalized in the complex plane is used, and the characteristics of the signal in the time domain, local frequency domain, and multi-fractional Fourier transform domain are revealed through invariant transformation, short-time Fourier transform, and Wigner-Ville transform, showing certain performance improvements in signal classification tasks. . BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 The figure is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION
[0015] The present invention will be further described below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited by the implementing regulations.
[0016] The present invention proposes a signal classification method based on a multi-domain feature fusion fully complex-valued network. In addition to the weights being complex numbers and satisfying the complex number operation rules, a custom normalization operator and an activation function generalized in the complex plane are used. The data is multi-domain transformed before being sent to the neural network. The feature extraction on the multi-domain is completed respectively through the backbone feature extraction network of the neural network, and the multiple features are spliced before the classification layer of the neural network to finally achieve signal classification.
[0017] This embodiment takes a one-dimensional complex signal s(t) with a length of 128 as an example to demonstrate the data flow in the signal classification method based on multi-domain feature fusion full complex valued network. First, a custom normalization operator is used to perform the operation in the following steps:
[0018] (1) Calculate the complex conjugate of the input data;
[0019] (2) Calculate the modulus value of the input data;
[0020] (3) Calculate the sum of the modulus square values of all elements to obtain a normalization operator;
[0021] (4) Each data modulus value is divided by the normalization operator to obtain the normalized result.
[0022] According to the above steps, a normalized one-dimensional complex signal s(t) is obtained, and the length is still 128. At the same time, it is agreed that in this embodiment, except for the classification layer, the activation function used is defined by the following formula:
[0023]
[0024] Where z is the data input to the activation function, and φ is the principal value of the input data.
[0025] Next, in the embodiment, it is required that the signal as described above is subjected to multi-domain transformation before being sent to the neural network, including: invariant transformation, short-time Fourier transformation and Wigner-Ville transformation. The three transformations are defined by the following formulas respectively:
[0026] Invariant[s(t)]=s(t)
[0027]
[0028] Subsequently, feature mining is performed on the data after the multi-domain transformation according to the backbone feature extraction network of the complex-valued neural network. After obtaining three sets of feature vectors, they are concatenated into a complete feature vector and sent to the neural network classification layer to realize signal classification.
[0029] In the original signal domain after the invariant transformation, a five-layer one-dimensional complex convolutional neural network is used to downsample the signal. The convolution kernel size is 3, the step size is 2, the number of zero padding is 1, the number of output channels is 4, 8, 16, 32, and 64, respectively, the activation function is CardioidReLU, and the output of the neural network is a feature vector with a length of 256. In the time-frequency analysis data obtained after the short-time Fourier transform, a five-layer two-dimensional complex convolutional neural network is used to downsample the signal. The convolution kernel size is 3, the step size is 2, the number of zero padding is 1, the number of output channels is 4, 8, 16, 32, and 64, respectively, the activation function is Cardioid ReLU, and the output feature map is stretched to obtain a feature vector with a length of 1024. On the two-dimensional data after Wigner-Ville transformation, a five-layer two-dimensional complex convolutional neural network is used to downsample the signal. The convolution kernel size is 3, the step size is 2, the number of zero padding is 1, the number of output channels is 4, 8, 16, 32, and 64 respectively, and the activation function is CardioidReLU. The output feature map is stretched to obtain a feature vector with a length of 1024.
[0030] After simply concatenating the three feature vectors, we get a fused feature vector with a length of 2304, which is followed by two layers of complex fully connected layer networks. After the last layer, the model is taken and the classification result is obtained through the Softmax layer to realize signal classification.
Claims
1. A signal classification method based on multi-domain feature fusion fully complex valued network, characterized by: The neural network used for signal classification adopts a full complex operation architecture. In addition to the weights being complex numbers and satisfying the complex operation rules, it uses a custom normalization operator and an activation function generalized in the complex plane. Before the data is sent to the neural network, it is first transformed into multiple domains, and then the feature extraction on multiple domains is completed separately through the backbone feature extraction network of the neural network. Then, the multiple features are spliced before the classification layer of the neural network, and finally sent to the classification layer to realize signal classification.
2. The signal classification method based on multi-domain feature fusion fully complex valued network according to claim 1 is characterized by: The custom normalization operator performs operations according to the following steps: (1) calculating the complex conjugate of the input operator data; (2) calculating the modulus square value of the input operator data; (3) calculating the sum of all modulus square values to obtain a normalization operator; (4) dividing each data modulus square value by the normalization operator to obtain a normalized result.
3. The signal classification method based on multi-domain feature fusion fully complex valued network according to claim 1 is characterized by: The activation function is: Where z is the data input to the activation function, and φ is the principal value of the input data.
4. The signal classification method based on multi-domain feature fusion fully complex valued network according to claim 1 is characterized in that: The multi-domain transformations performed before being sent to the neural network include: invariant transformation, short-time Fourier transformation and Wigner-Ville transformation.
5. The signal classification method based on multi-domain feature fusion fully complex valued network according to claim 4 is characterized in that: The backbone feature extraction network of the complex-valued neural network performs feature mining on the multi-domain transformation respectively, and after obtaining three sets of feature vectors, they are spliced to form a complete feature vector, which is sent to the neural network classification layer to realize signal classification.