CN-122016050-A - Low signal-to-noise interference spectrum denoising and robust demodulation method based on self-supervision diffusion model

CN122016050ACN 122016050 ACN122016050 ACN 122016050ACN-122016050-A

Abstract

The invention discloses a low signal-to-noise ratio interference spectrum denoising and robust demodulation method based on a self-supervision diffusion model, which aims at the problem that demodulation is unstable due to the fact that interference spectrum measurement is susceptible to factors such as noise, pseudo-peak interference, multipath interference, temperature drift and the like under the condition of low signal-to-noise ratio. The method combines the modeling capability of the diffusion model on the real noise statistical characteristic, keeps the phase continuity and the frequency domain structural consistency of interference fringes while suppressing noise interference, and further combines the interference spectrum after denoising reconstruction with the SPF (spectral peak fitting) of the traditional spectrum peak fitting method to realize stable demodulation and high-precision ranging of the interference spectrum under the condition of low signal-to-noise ratio. The invention reduces the dependence on high-quality labeling data and ideal reference spectrum, remarkably improves the robustness and engineering applicability of interference spectrum demodulation, and has good application value.

Inventors

CHENG YUNYONG
YUAN JIAQI
ZHAO LIXIANG
LU JIAJUN
LIU KAICHENG

Assignees

西北工业大学

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (6)

1. The low signal-to-noise interference spectrum denoising and robust demodulation method based on the self-supervision diffusion model is characterized by comprising the following steps of: And 1, collecting a real interference spectrum and constructing a sample. Repeatedly acquiring the same measured object or the same measuring position for a plurality of times under the same measuring condition to obtain a plurality of real interference spectrum signals with consistent physical information and different noise components; and 2, constructing an interference spectrum denoising network based on a diffusion model. Constructing a diffusion model network SIDNet for interference spectrum denoising reconstruction, wherein the network adopts U-Net as a reference structure; and 3, self-supervision training based on the conditional diffusion model. Performing self-supervision training on the diffusion model network constructed in the step2 by utilizing the interference spectrum sample set constructed in the step 1; And 4, phase-preserving denoising reconstruction of the interference spectrum with low signal-to-noise ratio. After the self-supervision training is finished, inputting the low signal-to-noise interference spectrum to be processed into a diffusion model network after the training is finished; And 5, frequency domain robust demodulation based on the SPF method. And (3) inputting the denoising interference spectrum output in the step (4) into a traditional frequency domain demodulation method (SPF) to perform distance inversion processing.
2. The method for denoising and robust demodulation of interference spectrum with low signal to noise ratio based on self-supervision diffusion model according to claim 1, wherein in step 1, a set of real interference spectrum samples for self-supervision training is obtained, and the process comprises repeated collection, preprocessing and spectrum sample pair construction of the real interference spectrum. And repeatedly collecting the same measured object or the same measuring position for a plurality of times under the same measuring condition to obtain a plurality of real interference spectrum signals. The same measurement conditions at least comprise structural parameters of an optical system, working states of a light source, configuration of a detector and position states of a measured object, so that the repeated acquisition of a plurality of interference spectrums corresponding to the same interference information in physical sense is ensured. The mth real interference spectrum obtained by repeated acquisition can be expressed as follows: Wherein, I (m) (k) represents the real interference spectrum obtained by the mth acquisition, I 0 (k) represents the ideal interference spectrum signal corresponding to the same measurement condition, n (m) (k) represents the random noise component introduced in the mth acquisition process, and k represents the wave number variable. Because of the randomness of the noise sources, the noise components n (m) (k) corresponding to different acquisition moments are independent or approximately independent of each other, and the ideal interference spectrum I 0 (k) remains consistent during multiple acquisitions. The interference spectrum signals obtained through repeated acquisition are subjected to unified preprocessing, wherein the preprocessing comprises, but is not limited to, wavelength axis or wavenumber axis alignment, sampling point rearrangement, intensity normalization and outlier rejection, so that the influence of system drift and sampling inconsistency on subsequent training is eliminated. After the pretreatment is completed, a spectrum sample pair for self-supervision training is constructed from the interference spectrum repeatedly collected. Each set of spectral sample pairs consists of two true interference spectra corresponding to the same interference information but containing different noise components, the relationship of which can be expressed as: Wherein I (1) (k) and I (2) (k) form a set of self-supervising training sample pairs, and n (1) (k) and n (2) (k) are noise components independent or approximately independent of each other. The spectrum sample pair does not depend on an ideal reference spectrum or a clean spectrum label, does not contain a corresponding distance parameter label, and only utilizes real measurement data to construct a training sample.
3. The method for denoising and robust demodulation of low signal-to-noise interference spectrum based on self-supervised diffusion model according to claim 1, wherein in step 2, a diffusion model network SIDNet for denoising reconstruction of low signal-to-noise interference spectrum is constructed. The SIDNet network adopts a U-Net structure consisting of an encoding network, a bottleneck network and a decoding network, and is used for carrying out multi-scale feature modeling and step-by-step reconstruction on interference spectrum signals in the diffusion denoising process. Specifically, the coding network is composed of a multi-stage residual convolution module and is used for carrying out step-by-step feature extraction on an input interference spectrum signal, and expanding a receptive field while reducing feature resolution so as to obtain local structure information and medium-low frequency feature representation of interference fringes. The coding features of each level are reserved through jump connection and used for feature fusion in the subsequent decoding stage, so that key information is prevented from being lost in the downsampling process. The bottleneck network is located between the encoding network and the decoding network and is used for carrying out global semantic modeling on interference spectrum characteristics at the lowest resolution level. And a bottleneck gating module based on Softmax is arranged in the bottleneck network, and the high-level characteristics output by the coding network are selectively modulated to highlight the effective characteristics related to the interference fringe main structure and inhibit the ineffective response of noise dominance, so that the stability of the network under the condition of low signal-to-noise ratio is enhanced. The decoding network structure is symmetrical to the encoding network and is used for recovering the high-resolution representation of the interference spectrum step by step. In the decoding process, the characteristics of the corresponding layers in the coding network are introduced into the decoding network through multi-stage jump connection, so that multi-scale characteristic fusion is realized. In order to further enhance the matching between the coding features and the decoding features, a diffusion condition attention module is arranged on the jump connection path, and the features from the coding network are subjected to self-adaptive weighting and alignment, so that the decoding network can reconstruct by using the condition features more effectively in the diffusion denoising process.
4. The method for denoising and robust demodulation of interference spectrum with low signal to noise ratio based on self-supervision diffusion model according to claim 1, wherein in step 3, self-supervision training is performed on the interference spectrum denoising network SIDNet constructed in step 2 by using the real interference spectrum sample set constructed in step 1 to learn the noise statistics of the interference spectrum with low signal to noise ratio under given conditions. Specifically, a spectrum sample pair obtained by repeated collection is selected from the spectrum sample set, one real interference spectrum is taken as a condition input, and the other real interference spectrum is taken as a target output. In the training process, noise disturbance is gradually introduced into a target spectrum through a diffusion forward process, and noise is predicted and restrained through a SIDNet network guided through a back diffusion process, so that the noise distribution characteristics of a real interference spectrum are learned under the condition that the network does not depend on an ideal reference spectrum and a distance parameter label. The training process can be understood as modeling the real interference spectrum noise distribution under the constraint of the conditional spectrum, and the aim of the training process can be represented as learning the following conditional probability distribution relation: Where I (1) represents a true interference spectrum as a conditional input, and I (2) represents a true interference spectrum corresponding thereto, including different noise components. Through the learning of the condition distribution, the SIDNet network can recover the corresponding denoising spectrum representation under the condition that any real interference spectrum is taken as the condition input.
5. The method for denoising and robust demodulation of low signal-to-noise interference spectrum based on self-supervised diffusion model according to claim 1, wherein in step 4, the interference spectrum denoising network SIDNet completed by the self-supervised training in step 3 is used to perform denoising reconstruction processing on the low signal-to-noise interference spectrum to obtain an interference spectrum signal with higher physical demodulation consistency. Specifically, a low signal-to-noise interference spectrum to be processed is input into a SIDNet network after training. In the diffusion denoising process, SIDNet networks conduct multi-scale feature extraction on input spectrums through coding networks, and selectively enhance high-level semantic features by combining a global feature modulation mechanism in bottleneck networks, so that invalid responses of noise dominance are restrained. Meanwhile, in the process of recovering the spectrum resolution step by step, the decoding network carries out self-adaptive weighting and alignment on the coding features through the diffusion condition attention module, so that the effective information related to the interference fringe structure can be fully utilized in the reconstruction process under the condition constraint.
6. The method for denoising and robust demodulation of low signal-to-noise interference spectrum based on self-supervision diffusion model according to claim 1, wherein in step 5, the denoising reconstructed interference spectrum output in step 4 is input into a traditional frequency domain demodulation method SPF for distance inversion processing, so as to realize robust demodulation under the condition of low signal-to-noise ratio. Specifically, the interference spectrum after denoising is subjected to frequency domain transformation, main peak characteristics related to the distance parameter in the spectrum are extracted, and the position of the main peak is determined in a spectral peak fitting mode, so that inversion calculation of the distance parameter is completed. Because the interference spectrum processed by the step 4 is obviously improved in the aspects of signal-to-noise ratio, spectrum peak stability and spectrum structure consistency, the SPF method can more stably identify the main peak and inhibit the interference of the false peak in the demodulation process, thereby reducing the risk of misjudgment.

Description

Low signal-to-noise interference spectrum denoising and robust demodulation method based on self-supervision diffusion model Technical Field The invention relates to the field of optical measurement and signal processing, in particular to a low signal-to-noise interference spectrum denoising and robust demodulation method based on a self-supervision diffusion model. Background The spectrum interferometry technology is widely applied to the fields of precise displacement measurement, microstructure size detection, industrial online measurement and the like, and is characterized in that high-precision demodulation is carried out on interference spectrum signals so as to invert distance parameters. However, in the actual engineering environment, the interference spectrum is always in a low signal-to-noise ratio state and is accompanied with pseudo-peak interference and spectrum shape distortion due to the influence of factors such as light source intensity fluctuation, detection noise, multipath reflection, environmental temperature drift, system stability and the like, so that demodulation failure easily occurs in the traditional frequency domain demodulation method under the condition of low signal-to-noise ratio. The existing denoising method generally depends on an ideal reference spectrum or a high-quality clean spectrum as a supervision tag, but the cost and difficulty for acquiring the data in a real measurement system are high, and even the method is not feasible. In recent years, a self-supervision learning and diffusion model is advanced in the fields of image and signal denoising, but when the self-supervision learning and diffusion model is directly applied to interference spectrum signals, phase information is easily destroyed, so that the subsequent physical demodulation precision is influenced. Therefore, how to realize the phase-preserving denoising reconstruction of the interference spectrum and improve the demodulation stability and the demodulation precision under the low signal-to-noise ratio scene under the conditions of no real clean spectrum label and no ideal reference spectrum is still a key technical problem to be solved in the field. Disclosure of Invention The invention provides a low signal-to-noise ratio interference spectrum denoising and robust demodulation method based on a self-supervision diffusion model, which realizes phase-preserving denoising reconstruction of interference spectrum signals by learning statistical distribution of real interference spectrum noise under the condition of no need of distance labels and ideal reference spectrum, and obviously improves ranging stability and accuracy under the condition of low signal-to-noise ratio by combining a traditional frequency domain demodulation method. The method comprises the following steps: And 1, collecting a real interference spectrum and constructing a sample. Repeatedly acquiring the same measured object or the same measuring position for a plurality of times under the same measuring condition to obtain a plurality of real interference spectrum signals with consistent physical information and different noise components; and 2, constructing an interference spectrum denoising network based on a diffusion model. Constructing a diffusion model network SIDNet for interference spectrum denoising reconstruction, wherein the network adopts U-Net as a reference structure; and 3, self-supervision training based on the conditional diffusion model. Performing self-supervision training on the diffusion model network constructed in the step2 by utilizing the interference spectrum sample set constructed in the step 1; And 4, phase-preserving denoising reconstruction of the interference spectrum with low signal-to-noise ratio. After the self-supervision training is finished, inputting the low signal-to-noise interference spectrum to be processed into a diffusion model network after the training is finished; And 5, frequency domain robust demodulation based on the SPF method. And (3) inputting the denoising interference spectrum output in the step (4) into a traditional frequency domain demodulation method (SPF) to perform distance inversion processing. In the step 1, a real interference spectrum sample set for self-supervision training is obtained, and the process comprises repeated acquisition, pretreatment and spectrum sample pair construction of the real interference spectrum. Specifically, under the same measurement condition, repeated collection is carried out on the same measured object or the same measurement position for a plurality of times, so as to obtain a plurality of real interference spectrum signals. The same measurement conditions at least comprise structural parameters of an optical system, working states of a light source, configuration of a detector and position states of a measured object, so that the repeated acquisition of a plurality of interference spectrums corresponding to the same interference informa