Low signal-to-noise ratio interference spectrum denoising and robust demodulation method based on self-supervised diffusion model
By using the SIDNet network with a self-supervised diffusion model and combining it with traditional frequency domain demodulation methods, the demodulation failure problem of low signal-to-noise ratio interferometric spectra was solved. Phase-preserving denoising reconstruction and robust demodulation were achieved under the condition of no ideal reference spectrum, thereby improving the stability and accuracy of ranging.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHWESTERN POLYTECHNICAL UNIV
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-12
AI Technical Summary
In practical engineering environments, interferometric spectra are often in a low signal-to-noise ratio state. Affected by factors such as light source intensity fluctuations, detection noise, multipath reflection, and environmental temperature drift, traditional methods are prone to demodulation failure. Moreover, existing denoising methods rely on ideal reference spectra, which are costly and difficult. Self-supervised learning directly destroys phase information and affects demodulation accuracy.
By constructing a self-supervised diffusion model, the SIDNet network is trained using repeatedly acquired data from real interference spectra to learn the noise distribution. Combined with traditional frequency domain demodulation methods, it achieves phase-preserving denoising reconstruction and robust demodulation. The network adopts a U-Net structure, a bottleneck gating module, and a diffusion conditional attention module to suppress noise and spurious peak interference.
In the absence of ideal reference spectra and distance labels, it significantly improves ranging stability and accuracy under low signal-to-noise ratio conditions, reduces dependence on high-quality labeled data, maintains phase continuity and frequency domain structure consistency, and improves demodulation stability and accuracy.
Smart Images
Figure CN122016050A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of optical measurement and signal processing, specifically to a low signal-to-noise ratio interferometric spectral denoising and robust demodulation method based on a self-supervised diffusion model. Background Technology
[0002] Interferometric spectral measurement technology is widely used in precision displacement measurement, microstructure size detection, and industrial online measurement. Its core lies in the high-precision demodulation of interferometric spectral signals to retrieve distance parameters. However, in real-world engineering environments, factors such as light source intensity fluctuations, probe noise, multipath reflection, environmental temperature drift, and system stability often result in low signal-to-noise ratios (SNR) interferometric spectra, accompanied by spurious peak interference and spectral distortion. This leads to demodulation failures of traditional frequency domain demodulation methods under low SNR conditions. Existing denoising methods typically rely on ideal reference spectra or high-quality clean spectra as supervisory labels, but acquiring such data in real measurement systems is costly, difficult, and sometimes infeasible. In recent years, self-supervised learning and diffusion models have made progress in image and signal denoising, but directly applying them to interferometric spectral signals can easily destroy phase information, thus affecting the subsequent physical demodulation accuracy. Therefore, how to achieve phase-preserving denoising reconstruction of interferometric spectra without real clean spectral labels or ideal reference spectra, and improve demodulation stability and accuracy in low SNR scenarios, remains a key technical problem that urgently needs to be solved in this field. Summary of the Invention
[0003] This invention proposes a low signal-to-noise ratio (SNR) interferometric spectral denoising and robust demodulation method based on a self-supervised diffusion model. Without requiring distance tags or an ideal reference spectrum, it achieves phase-preserving denoising and reconstruction of the interferometric spectral signal by learning the statistical distribution of real interferometric spectral noise. Combined with traditional frequency domain demodulation methods, it significantly improves ranging stability and accuracy under low SNR conditions. The method includes the following steps: Step 1: Acquisition and Sample Construction of Real Interference Spectra. Under the same measurement conditions, the same object or the same measurement location is repeatedly sampled to obtain multiple real interference spectral signals with consistent physical information but different noise components; Step 2: Construct an interferometric spectral denoising network based on a diffusion model. A diffusion model network, SIDNet, for interferometric spectral denoising and reconstruction is constructed, using U-Net as the baseline structure. Step 3: Self-supervised training based on the conditional diffusion model. The diffusion model network constructed in Step 2 is trained using the interference spectrum sample set built in Step 1 under self-supervised conditions. Step 4: Phase-preserving denoising and reconstruction of low signal-to-noise ratio (SNR) interferometric spectra. After completing self-supervised training, the low SNR interferometric spectrum to be processed is input into the trained diffusion model network; Step 5: Robust frequency domain demodulation based on the SPF method. The denoised interferometric spectrum output from Step 4 is input into the traditional frequency domain demodulation method SPF (spectral peak fitting method) for distance inversion processing.
[0004] In step 1 above, a real interference spectrum sample set is obtained for self-supervised training. The process includes repeated acquisition of real interference spectra, preprocessing, and construction of spectral sample pairs.
[0005] Specifically, under the same measurement conditions, multiple repeated acquisitions are performed on the same object under test or the same measurement location to obtain multiple true interference spectral signals. The same measurement conditions include keeping at least the optical system structural parameters, light source operating state, detector configuration, and the position state of the object under test unchanged, thereby ensuring that the multiple interference spectra obtained from repeated acquisitions correspond to the same interference information in a physical sense. m A true interference spectrum can be represented as: in, I (m) ( k ) indicates the first m The true interference spectrum obtained from the second acquisition I 0( k () represents the ideal interference spectrum signal under the same measurement conditions. n (m) ( k ) indicates the first m Random noise components introduced during the acquisition process. k This represents the wavenumber variable. Because the noise source is random, the noise components at different acquisition times... n (m) ( k They are independent or approximately independent of each other, while the ideal interference spectrum I 0( k The data collection process remained consistent across multiple data collection sessions.
[0006] The repeatedly acquired interferometric spectral signals undergo uniform preprocessing, including but not limited to wavelength axis or wavenumber axis alignment, sampling point rearrangement, intensity normalization, and outlier removal, to eliminate the impact of system drift and sampling inconsistencies on subsequent training. After preprocessing, spectral sample pairs for self-supervised training are constructed from the repeatedly acquired interferometric spectra. Each spectral sample pair consists of two real interferometric spectra corresponding to the same interferometric information but containing different noise components, and their relationship can be expressed as: in, I (1) ( k )and I(2) ( k This constitutes a set of self-supervised training sample pairs. n (1) ( k )and n (2) ( k These are mutually independent or approximately independent noise components. The spectral sample pairs do not rely on ideal reference spectra or clean spectral labels, nor do they contain corresponding distance parameter labels; they are constructed solely using the actual measurement data itself.
[0007] In step 2 above, a diffusion model network, SIDNet, is constructed for low signal-to-noise ratio interferometric spectral denoising and reconstruction. The SIDNet network adopts a U-Net structure consisting of an encoding network, a bottleneck network, and a decoding network, and is used to perform multi-scale feature modeling and step-by-step reconstruction of the interferometric spectral signal during diffusion denoising.
[0008] Specifically, the encoding network consists of multi-level residual convolutional modules, used to extract features from the input interference spectral signal step by step. This reduces the feature resolution while expanding the receptive field to obtain local structural information and mid-to-low frequency feature representations of the interference fringes. The encoded features at each level are preserved through skip connections for feature fusion in the subsequent decoding stage, thereby preventing the loss of crucial information during downsampling.
[0009] The bottleneck network, located between the encoding and decoding networks, is used for global semantic modeling of interference spectral features at the lowest resolution level. A Softmax-based bottleneck gating module is incorporated into the bottleneck network to selectively modulate the high-level features output by the encoding network. This highlights effective features related to the main structure of the interference fringes and suppresses noise-dominated invalid responses, thereby enhancing the network's stability under low signal-to-noise ratio conditions.
[0010] The decoding network structure is symmetrical to the encoding network and is used to progressively recover a high-resolution representation of the interference spectrum. During the decoding process, features from corresponding levels in the encoding network are introduced into the decoding network through multi-level skip connections, achieving multi-scale feature fusion. To further enhance the matching between encoded and decoded features, a diffusion conditional attention module is set on the skip connection path to adaptively weight and align features from the encoding network, enabling the decoding network to more effectively utilize conditional features for reconstruction during diffusion denoising.
[0011] In step 3 above, the SIDNet interference spectrum denoising network constructed in step 2 is trained in a self-supervised manner using the real interference spectrum sample set constructed in step 1, so as to learn the noise statistical characteristics of low signal-to-noise ratio interference spectra under given conditions.
[0012] Specifically, spectral sample pairs obtained from repeated acquisitions are selected from the spectral sample set. One real interference spectrum is used as the conditional input, and the other real interference spectrum is used as the target output. During training, noise perturbation is gradually introduced into the target spectrum through a forward diffusion process, and the SIDNet network is guided to predict and suppress noise through a reverse diffusion process. This allows the network to learn the noise distribution characteristics of the real interference spectrum without relying on an ideal reference spectrum or distance parameter labels. The above training process can be understood as modeling the noise distribution of the real interference spectrum under conditional spectral constraints. Its objective can be expressed as learning the following conditional probability distribution relationship: in, I (1) This represents the actual interference spectrum used as a conditional input. I (2) This represents the corresponding real interference spectrum containing different noise components. By learning the above conditional distribution, the SIDNet network can recover its corresponding denoised spectral representation given any real interference spectrum as a conditional input.
[0013] In step 4 above, the interference spectrum denoising network SIDNet, which was trained in step 3, is used to denoise and reconstruct the low signal-to-noise ratio interference spectrum in order to obtain an interference spectrum signal with higher physical demodulation consistency.
[0014] Specifically, the low signal-to-noise ratio (SNR) interference spectrum to be processed is input into the trained SIDNet network. During diffusion denoising, the SIDNet network extracts multi-scale features from the input spectrum through the encoding network and selectively enhances high-level semantic features using a global feature modulation mechanism in the bottleneck network, thereby suppressing noise-dominated invalid responses. Simultaneously, during the stepwise restoration of spectral resolution, the decoding network adaptively weights and aligns the encoded features through a diffusion conditional attention module, enabling the reconstruction process to fully utilize effective information related to the interference fringe structure under conditional constraints.
[0015] In step 5 above, the denoised reconstructed interference spectrum output from step 4 is input into the traditional frequency domain demodulation method SPF for distance inversion processing to achieve robust demodulation under low signal-to-noise ratio conditions.
[0016] Specifically, the denoised interferometric spectrum undergoes frequency domain transformation to extract the main peak features related to the distance parameter. The position of the main peak is then determined through peak fitting, thereby completing the inversion calculation of the distance parameter. Since the interferometric spectrum processed in step 4 shows significant improvements in signal-to-noise ratio, peak stability, and spectral structure consistency, the SPF method can more stably identify the main peak and suppress spurious peak interference during demodulation, thus reducing the risk of misjudgment.
[0017] The beneficial effects of this invention are as follows: This invention proposes a low signal-to-noise ratio (SNR) interferometric spectral denoising and robust demodulation method based on a self-supervised diffusion model. By constructing a self-supervised training mechanism using repeatedly acquired data from real interferometric spectra, it achieves effective denoising and reconstruction of low SNR interferometric spectra without requiring ideal reference spectra or distance parameter labels. This significantly reduces the dependence on high-quality labeled data and improves the applicability of the method in practical engineering environments. This invention introduces a diffusion model and a denoising network, SIDNet, to model and suppress adverse factors such as random noise, spurious peak interference, multipath interference, and temperature drift in real interferometric spectra. During the denoising process, it maintains the phase continuity and frequency domain structure consistency of the interference fringes, thereby outputting an interferometric spectrum with higher physical demodulation consistency. Furthermore, by combining the denoised reconstruction results with the traditional peak fitting method SPF, the stability of main peak identification under low signal-to-noise ratio conditions is significantly improved without changing the existing demodulation process. This reduces the probability of false peak misjudgment and improves the stability and demodulation accuracy of interferometric spectral ranging. This method combines the adaptive capability of self-supervised learning with the interpretability of traditional physical demodulation methods, and has good robustness and engineering promotion value. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the method described in this application.
[0019] Figure 2 This is a schematic diagram of the SIDNet interferometric spectral denoising network based on a diffusion model.
[0020] Figure 3 This is a schematic diagram of the DCA Block structure for diffusion-conditional attention modules.
[0021] Figure 4 This is a schematic diagram of the bottleneck gating module SBGM Block structure. Detailed Implementation
[0022] To make the technical solution of this application clearer, the following will describe in detail the implementation method, technical features and advantages of this application with reference to illustrated examples, and explain them through specific embodiments and corresponding drawings.
[0023] like Figure 1The diagram shown is a flowchart of the method proposed in this application. This application proposes a low signal-to-noise ratio (SNR) interferometric spectral denoising and robust demodulation method based on a self-supervised diffusion model. This method is used to suppress the adverse effects of noise, spurious peak interference, multipath interference, and temperature drift under low SNR conditions, and to achieve stable demodulation of the interferometric spectral signal. The method includes the following steps: Step 1: Acquisition and Sample Construction of Real Interference Spectra. Under the same measurement conditions, multiple acquisitions are performed on the same object or the same measurement location to obtain multiple real interference spectrum signals with consistent physical information but different noise components. The acquired interference spectra are preprocessed, including wavelength or wavenumber axis alignment and intensity normalization. Spectral sample pairs are constructed from the repeatedly acquired spectra, ensuring that the two spectra in each pair correspond to the same interference information but contain different random noise components, thus forming a real interference spectrum sample set for self-supervised training.
[0024] Step 2: Construct the SIDNet interferometric spectral denoising network based on the diffusion model, the structure of which is as follows: Figure 2 As shown, SIDNet adopts a U-Net structure consisting of an encoding network, a bottleneck network, and a decoding network. The encoding network is used to extract multi-scale features from the input interference spectrum, the bottleneck network is used to perform global feature modulation and semantic modeling in the low-resolution feature space, and the decoding network is used to fuse multi-scale features and recover a high-resolution representation of the interference spectrum step by step.
[0025] Furthermore, such as Figure 2 As shown, a diffusion conditional attention module (DCA) block is set on the skip connection path of the corresponding layer of the encoding network and the decoding network, and its structure is as follows. Figure 3 As shown, this is used for adaptive weighting and alignment of encoded and decoded features, thereby enhancing the network's ability to model the statistical characteristics of real interferometric spectral noise. Simultaneously, a bottleneck gating module (SBGM Block) is set in the bottleneck network of SIDNet, with the structure shown below. Figure 4 As shown, this is used to selectively modulate high-level features, highlighting effective information related to the main structure of the interference fringes.
[0026] Step 3: Self-supervised training based on the conditional diffusion model. The SIDNet network constructed in Step 2 is trained in a self-supervised manner using the real interferometric spectrum sample set built in Step 1. During training, one real interferometric spectrum from each sample pair is used as the conditional input, and the other real interferometric spectrum is used as the target output. The network is guided to learn the noise distribution characteristics of the real interferometric spectrum under given conditions through forward and reverse diffusion processes. This allows the network to acquire the ability to denoise and model low signal-to-noise ratio interferometric spectra without requiring an ideal reference spectrum or distance parameter labels.
[0027] Step 4: Phase-preserving denoising and reconstruction of low signal-to-noise ratio (SNR) interferometric spectra. After completing self-supervised training, the low SNR interferometric spectrum to be processed is input into the trained SIDNet network. During the diffusion denoising process, SIDNet gradually suppresses random noise, spurious peak interference, multipath interference, and temperature drift in the spectrum through the synergistic effect of multi-scale feature extraction of the encoding network, global modulation of the bottleneck gating module, and diffusion conditional attention module, while maintaining the phase continuity and frequency domain structure consistency of the interference fringes, and outputs the denoised and reconstructed interferometric spectrum signal.
[0028] Step 5: Robust frequency-domain demodulation based on the SPF method. The denoised interferometric spectrum output from Step 4 is input into the traditional peak fitting method SPF for frequency-domain demodulation. The distance parameter is inverted by identifying and fitting the main peak of the spectrum. Since the denoised interferometric spectrum is significantly improved in terms of signal-to-noise ratio, main peak stability, and spectral structure consistency, the demodulation stability and ranging accuracy of the SPF method under low signal-to-noise ratio conditions are significantly improved, thus achieving robust demodulation of the interferometric spectral signal.
[0029] The specific embodiments of this patent application have been described in detail above. However, the above embodiments are merely illustrative and do not constitute a limitation on the scope of protection of this patent application. The low signal-to-noise ratio interferometric spectral denoising and robust demodulation method based on a self-supervised diffusion model proposed in this patent application is applicable to various interferometric spectral measurement and signal processing scenarios. For those skilled in the art, any equivalent substitutions, modifications, or improvements made to the method of this application without departing from the technical concept and core idea of this patent application should be covered within the scope of protection of this patent application.
Claims
1. A method for low signal-to-noise ratio interferometric spectral denoising and robust demodulation based on a self-supervised diffusion model, characterized in that... Includes the following steps: Step 1: Acquisition and Sample Construction of Real Interference Spectra. Under the same measurement conditions, the same object or the same measurement location is repeatedly sampled to obtain multiple real interference spectral signals with consistent physical information but different noise components; Step 2: Construct an interferometric spectral denoising network based on a diffusion model. Build A diffusion model network, SIDNet, for interference spectral denoising and reconstruction, uses U-Net as the baseline structure. Step 3: Self-supervised training based on the conditional diffusion model. The diffusion model network constructed in Step 2 is trained using the interference spectrum sample set built in Step 1 under self-supervised conditions. Step 4: Phase-preserving denoising and reconstruction of low signal-to-noise ratio (SNR) interferometric spectra. After completing self-supervised training, the low SNR interferometric spectrum to be processed is input into the trained diffusion model network; Step 5: Robust frequency domain demodulation based on the SPF method. The denoised interferometric spectrum output from Step 4 is input into the traditional frequency domain demodulation method SPF (spectral peak fitting method) for distance inversion processing.
2. The method for low signal-to-noise ratio interferometric spectral denoising and robust demodulation based on a self-supervised diffusion model according to claim 1, characterized in that, In step 1, a set of real interference spectra samples for self-supervised training is obtained. This process includes repeated acquisition of real interference spectra, preprocessing, and construction of spectral sample pairs. Under the same measurement conditions, multiple acquisitions are performed on the same object or the same measurement location to obtain multiple real interference spectra signals. The same measurement conditions include keeping at least the optical system structural parameters, light source operating state, detector configuration, and the position state of the object under test unchanged, thereby ensuring that the multiple interference spectra obtained from repeated acquisitions correspond to the same interference information in a physical sense. The repeatedly acquired... m A true interference spectrum can be represented as: in, I (m) ( k ) indicates the first m The true interference spectrum obtained from the second acquisition I 0( k () represents the ideal interference spectrum signal under the same measurement conditions. n (m) ( k ) indicates the first m Random noise components introduced during the acquisition process. k This represents the wavenumber variable. Because the noise source is random, the noise components at different acquisition times... n (m) ( k They are independent or approximately independent of each other, while the ideal interference spectrum I 0( k The data collection process remained consistent across multiple data collection sessions. The repeatedly acquired interferometric spectral signals undergo uniform preprocessing, including but not limited to wavelength axis or wavenumber axis alignment, sampling point rearrangement, intensity normalization, and outlier removal, to eliminate the impact of system drift and sampling inconsistencies on subsequent training. After preprocessing, spectral sample pairs for self-supervised training are constructed from the repeatedly acquired interferometric spectra. Each spectral sample pair consists of two real interferometric spectra corresponding to the same interferometric information but containing different noise components, and their relationship can be expressed as: in, I (1) ( k )and I (2) ( k This constitutes a set of self-supervised training sample pairs. n (1) ( k )and n (2) ( k These are mutually independent or approximately independent noise components. The spectral sample pairs do not rely on ideal reference spectra or clean spectral labels, nor do they contain corresponding distance parameter labels; they are constructed solely using the actual measurement data itself.
3. The method for low signal-to-noise ratio interferometric spectral denoising and robust demodulation based on a self-supervised diffusion model according to claim 1, characterized in that, In step 2, a diffusion model network, SIDNet, is constructed for low signal-to-noise ratio interferometric spectral denoising and reconstruction. The SIDNet network adopts a U-Net structure consisting of an encoding network, a bottleneck network, and a decoding network, and is used to perform multi-scale feature modeling and step-by-step reconstruction of the interferometric spectral signal during diffusion denoising. Specifically, the encoding network consists of multi-level residual convolutional modules, used to extract features from the input interference spectral signal step by step. This reduces the feature resolution while expanding the receptive field to obtain local structural information and mid-to-low frequency feature representations of the interference fringes. The encoded features at each level are preserved through skip connections for feature fusion in the subsequent decoding stage, thereby preventing the loss of crucial information during downsampling. The bottleneck network, located between the encoding and decoding networks, is used for global semantic modeling of interference spectral features at the lowest resolution level. A Softmax-based bottleneck gating module is incorporated into the bottleneck network to selectively modulate the high-level features output by the encoding network. This highlights effective features related to the main structure of the interference fringes and suppresses noise-dominated invalid responses, thereby enhancing the network's stability under low signal-to-noise ratio conditions. The decoding network structure is symmetrical to the encoding network and is used to progressively recover a high-resolution representation of the interference spectrum. During the decoding process, features from corresponding levels in the encoding network are introduced into the decoding network through multi-level skip connections, achieving multi-scale feature fusion. To further enhance the matching between encoded and decoded features, a diffusion conditional attention module is set on the skip connection path to adaptively weight and align features from the encoding network, enabling the decoding network to more effectively utilize conditional features for reconstruction during diffusion denoising.
4. The method for low signal-to-noise ratio interferometric spectral denoising and robust demodulation based on a self-supervised diffusion model according to claim 1, characterized in that, In step 3, the SIDNet interference spectral denoising network constructed in step 2 is trained in a self-supervised manner using the real interference spectrum sample set constructed in step 1, so as to learn the noise statistical characteristics of low signal-to-noise ratio interference spectra under given conditions. Specifically, spectral sample pairs obtained through repeated acquisitions are selected from the spectral sample set. One real interference spectrum is used as the conditional input, and the other real interference spectrum is used as the target output. During training, noise perturbation is gradually introduced into the target spectrum through a forward diffusion process, and the SIDNet network is guided to predict and suppress the noise through a reverse diffusion process. This allows the network to learn the noise distribution characteristics of the real interference spectrum without relying on an ideal reference spectrum or distance parameter labels. The above training process can be understood as modeling the noise distribution of the real interferometric spectrum under conditional spectral constraints. Its objective can be expressed as learning the following conditional probability distribution relationship: in, I (1) This represents the actual interference spectrum used as a conditional input. I (2) This represents the corresponding real interference spectrum containing different noise components. By learning the above conditional distribution, the SIDNet network can recover its corresponding denoised spectral representation given any real interference spectrum as a conditional input.
5. The method for low signal-to-noise ratio interferometric spectral denoising and robust demodulation based on a self-supervised diffusion model according to claim 1, characterized in that, In step 4, the interference spectrum denoising network SIDNet, which was trained in step 3, is used to denoise and reconstruct the low signal-to-noise ratio interference spectrum in order to obtain an interference spectrum signal with higher physical demodulation consistency. Specifically, the low signal-to-noise ratio (SNR) interference spectrum to be processed is input into the trained SIDNet network. During diffusion denoising, the SIDNet network extracts multi-scale features from the input spectrum through the encoding network and selectively enhances high-level semantic features using a global feature modulation mechanism in the bottleneck network, thereby suppressing noise-dominated invalid responses. Simultaneously, during the stepwise restoration of spectral resolution, the decoding network adaptively weights and aligns the encoded features through a diffusion conditional attention module, enabling the reconstruction process to fully utilize effective information related to the interference fringe structure under conditional constraints.
6. The method for low signal-to-noise ratio interferometric spectral denoising and robust demodulation based on a self-supervised diffusion model according to claim 1, characterized in that, In step 5, the denoised reconstructed interference spectrum output from step 4 is input into the traditional frequency domain demodulation method SPF for distance inversion processing to achieve robust demodulation under low signal-to-noise ratio conditions. Specifically, the denoised interferometric spectrum undergoes frequency domain transformation to extract the main peak features related to the distance parameter. The position of the main peak is then determined through peak fitting, thereby completing the inversion calculation of the distance parameter. Since the interferometric spectrum processed in step 4 shows significant improvements in signal-to-noise ratio, peak stability, and spectral structure consistency, the SPF method can more stably identify the main peak and suppress spurious peak interference during demodulation, thus reducing the risk of misjudgment.