CN-122027043-A - Self-supervision signal denoising method and system based on signal characteristic self-adaption and leachable loss weight

CN122027043ACN 122027043 ACN122027043 ACN 122027043ACN-122027043-A

Abstract

The invention discloses a self-supervision signal denoising method and system based on signal characteristic self-adaption and a leachable loss weight, and belongs to the technical field of modern wireless communication. The method comprises the steps of extracting multi-dimensional statistical characteristics, weighting the multi-dimensional statistical characteristics to generate weighted characteristic vectors, predicting CLIP (Contrastive Language-LMAGE PRETRAINING, CLIP) alignment loss weights and noise target proportions of noise-containing signals based on the weighted characteristic vectors, generating time domain characteristics and frequency domain characteristics based on the noise-containing signals, constructing CLIP alignment loss, constructing a residual signal prediction model, constructing a total loss function based on the CLIP alignment loss and the noise target proportions, training the residual signal prediction model based on the total loss function, inputting the noise-containing signals to be tested into the trained residual signal prediction model to obtain residual signals, subtracting the residual signals from the noise-containing signals to be tested, and obtaining denoised signal output. The self-monitoring signal denoising method can realize self-monitoring signal denoising based on signal characteristic self-adaption and leachable loss weight, and is convenient to use.

Inventors

ZHENG JIANYONG
Wu Bingtou
WEI HONGYU
PENG YAN
PENG YAXIN
KONG HAO
ZHOU YANG
QU DONG

Assignees

上海大学

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (8)

1. A self-supervision signal denoising method based on signal characteristic self-adaption and leachable loss weight is characterized by comprising the following steps: Step 1, extracting multi-dimensional statistical features from noise-containing signals, and weighting the multi-dimensional statistical features through a learnable weight parameter to generate weighted feature vectors, wherein the learnable weight parameter is the CLIP alignment loss weight and the noise target proportion in the historical noise-containing signals; Step 2, inputting the weighted feature vector into a pre-trained prediction network, and predicting the CLIP alignment loss weight and the noise target proportion of the noise-containing signal; Step 3, decomposing the noise-containing signal into an odd sampling sequence and an even sampling sequence, and respectively carrying out time domain and frequency domain feature coding to obtain time domain features and frequency domain features; Step 4, based on the time domain characteristics and the frequency domain characteristics, the CLIP alignment loss weight is constructed; Step 5, constructing a residual signal prediction model based on a U-Net encoder-decoder structure, constructing a total loss function based on the CLIP alignment loss and the noise target proportion, and training the residual signal prediction model based on the total loss function; step 6, inputting the noise-containing signal to be detected into a trained residual signal prediction model to obtain a residual signal; And 7, subtracting the residual signal from the noise-containing signal to be detected to obtain a denoised signal and outputting the denoised signal.
2. The method of claim 1, wherein in step 1, the multi-dimensional statistical features include at least one of Shannon entropy, power spectral entropy, amplitude variance, phase jump, peak-to-average ratio, and standard deviation of frequency domain energy distribution.
3. The method for denoising self-supervision signals based on signal feature self-adaption and learning loss weight according to claim 2, wherein in step 2, the weighted feature vector is input into a pre-trained prediction network, and the chip alignment loss weight and the noise target ratio of the predicted noise-containing signal are specifically: And inputting the weighted feature vector into a deep neural network for coding, and respectively predicting the CLIP alignment loss weight and the noise target proportion of the noise-containing signal through a multi-layer perceptron structure, wherein a logarithmic space parameterization and temperature regulation mechanism is adopted in the prediction process for ensuring the numerical stability of the prediction weight.
4. The method for denoising self-supervision signals based on signal feature self-adaption and learnable loss weights according to claim 3, wherein in step 3, the noisy signals are decomposed into odd sampling sequences and even sampling sequences, and time domain and frequency domain feature codes are respectively performed to obtain time domain features and frequency domain features, specifically: Decomposing the noise-containing signal into an odd sampling sequence and an even sampling sequence, and respectively carrying out time domain feature coding and frequency domain spectrum transformation on each odd sampling sequence and each even sampling sequence to obtain time domain features and frequency domain features, wherein the frequency domain spectrum transformation adopts short-time Fourier transformation and introduces a learnable frequency weight vector to carry out self-adaptive weighting on different frequency components.
5. The method for denoising self-monitored signals based on signal feature adaptation and learnable loss weights according to claim 4, wherein in step 4, CLIP alignment loss is constructed based on time domain features, frequency domain features and CLIP alignment loss weights, specifically: Mapping the time domain features and the frequency domain features to the same embedded space, and respectively constructing CLIP alignment loss for the odd sampling sequence and the even sampling sequence based on the similarity between the time domain and the frequency domain represented by the learnable temperature parameters.
6. The method of claim 5, wherein in step 5, the total loss function includes residual reconstruction loss, consistency loss, CLIP alignment loss, noise estimation loss, and weight regularization term.
7. The method of self-monitoring signal denoising based on signal feature adaptation and learnable loss weights according to claim 6, wherein the noise estimation loss is constructed by noise target ratio.
8. A self-monitoring signal denoising system based on signal feature self-adaptation and learning loss weight, applied to the self-monitoring signal denoising method based on signal feature self-adaptation and learning loss weight as set forth in any one of claims 1 to 7, comprising: The signal feature extraction and weighting module is used for extracting multi-dimensional statistical features from the noise-containing signals and carrying out leachable weight weighting to generate weighted feature vectors; The self-adaptive weight prediction module is used for receiving the weighted feature vector and predicting the CLIP alignment loss weight and the noise target proportion of the noise-containing signal; the time-frequency domain feature coding module is used for decomposing the noise-containing signal into an odd sampling sequence and an even sampling sequence, respectively carrying out time domain feature coding and frequency domain feature coding, and outputting time domain features and frequency domain features; The multi-mode alignment and loss calculation module is used for constructing CLIP alignment loss based on the time domain features, the frequency domain features and the CLIP alignment loss weight; the U-Net residual prediction module is used for constructing a residual signal prediction model based on a U-Net encoder-decoder structure; The multi-objective loss function construction module is used for forming a total loss function; The model training optimization module is used for training the residual signal prediction model by utilizing the total loss function; And the output module is used for calculating the denoised signal output based on the residual signal predicted by the residual prediction model after training and the noise-containing signal to be detected.

Description

Self-supervision signal denoising method and system based on signal characteristic self-adaption and leachable loss weight Technical Field The invention relates to the technical field of modern wireless communication, in particular to a self-supervision signal denoising method and system based on signal characteristic self-adaption and leachable loss weight. Background In modern wireless communication systems, signals are inevitably subject to interference from various types of noise during transmission, including thermal noise, interference noise, multipath fading, and the like. These noise can severely degrade the signal quality, affecting the bit error rate, spectral efficiency, and coverage of the communication system. Therefore, the signal denoising technology has been an important research direction in the field of communication. The traditional signal denoising method mainly comprises the following steps: Filter-based methods such as wiener filtering, kalman filtering, and the like. Such methods require accurate noise statistical models, are sensitive to model mismatch, and have limited performance in complex channel environments. A wavelet transform-based method separates the signal and noise by wavelet decomposition. But the choice of wavelet basis functions depends on experience and has limited effect on non-stationary signals. A supervised approach based on deep learning uses a large number of clean signal-noise-containing signal pairs for training. However, in practical applications, it is extremely difficult and costly to obtain a clean signal that is a perfect match. In addition, the disadvantages of the prior art mainly include: The performance of the filter-based method and the wavelet transform-based method is highly dependent on the accuracy of the a priori model. In the face of noise in the real world, which is complex in source, unknown in characteristics and non-stationary, the performance of these methods tends to be drastically reduced, and parameters need to be manually carefully designed and debugged, and the adaptive capability is lacking. The supervised denoising method based on deep learning requires large-scale and high-quality paired training data. In many practical applications, it is extremely difficult, costly, or even completely impossible to obtain an absolutely pure "Ground Truth" signal. This "data-on-dependency" problem becomes the biggest obstacle restricting the deployment and application of supervised denoising methods in the real world. In view of the foregoing, there is a need for a self-monitoring signal denoising method and system based on signal feature adaptation and learning loss weights, which solve the foregoing problems of the conventional method. Disclosure of Invention The invention aims to provide a self-supervision signal denoising method and system based on signal characteristic self-adaption and learnable loss weight, which can realize sample level self-adaption without self-supervision denoising of a clean signal label, fuse time-frequency domain information and strengthen generalization capability through learnable loss weight. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: a self-supervision signal denoising method based on signal characteristic self-adaption and leachable loss weight comprises the following steps: Step 1, extracting multi-dimensional statistical features from noise-containing signals, and weighting the multi-dimensional statistical features through a learnable weight parameter to generate weighted feature vectors, wherein the learnable weight parameter is the CLIP alignment loss weight and the noise target proportion in the historical noise-containing signals; Step 2, inputting the weighted feature vector into a pre-trained prediction network, and predicting the CLIP alignment loss weight and the noise target proportion of the noise-containing signal; Step 3, decomposing the noise-containing signal into an odd sampling sequence and an even sampling sequence, and respectively carrying out time domain and frequency domain feature coding to obtain time domain features and frequency domain features; Step 4, based on the time domain characteristics and the frequency domain characteristics, the CLIP alignment loss weight is constructed; Step 5, constructing a residual signal prediction model based on a U-Net encoder-decoder structure, constructing a total loss function based on the CLIP alignment loss and the noise target proportion, and training the residual signal prediction model based on the total loss function; step 6, inputting the noise-containing signal to be detected into a trained residual signal prediction model to obtain a residual signal; And 7, subtracting the residual signal from the noise-containing signal to be detected to obtain a denoised signal and outputting the denoised signal. Further, in step 1, the multi-dimensional statistical feature includes at least one of