Search

CN-121982201-A - Continuous spectrum space reconstruction method based on depth generation model

CN121982201ACN 121982201 ACN121982201 ACN 121982201ACN-121982201-A

Abstract

The invention discloses a continuous spectrum space reconstruction method based on a depth generation model, which comprises the following steps of 1, data organization and problem formalization, 2, robust normalization and scale alignment, 3, end point anchoring and continuous condition parameterization, 4, sample pairing and track construction, 5, end point condition variation self-encoder structure modeling, 6, physical consistency constraint and joint loss optimization, 7, missing interval generation and structured complementation, and 8, unified verification of generation quality and task availability, wherein a generation model which has explicit response to continuous physical parameters and can maintain spectral peak morphology and space statistics consistency can be constructed, so that the structured complementation and continuous reconstruction of the missing interval are realized.

Inventors

  • LOU JUNGANG
  • Shen Haoqi
  • LIU ZHENFANG
  • SHEN QING

Assignees

  • 湖州师范学院

Dates

Publication Date
20260505
Application Date
20260119

Claims (9)

  1. 1. A continuous spectrum space reconstruction method based on a depth generation model is characterized by comprising the following steps: step 1, data organization and problem formalization; step 2, robust normalization and scale alignment; Step3, end point anchoring and continuous condition parameterization; step 4, sample pairing and track construction; step 5, modeling an endpoint condition variation self-encoder structure; step 6, optimizing physical consistency constraint and joint loss; step 7, generating a missing interval and performing structural complementation; And 8, generating unified verification of quality and task availability.
  2. 2. The method for continuous spectrum space reconstruction based on depth generation model as set forth in claim 1, wherein said step 1 comprises the steps of: 1-1 representing each spatially displaced Raman spectrum sample as a two-dimensional spectrum image xE RH x W, wherein For the raman shift/wave Duan Wei, For the spatial displacement dimension, corresponding physical parameter labels are associated for each sample Recording according to discrete level or continuous value; 1-2 specify two endpoints on parameter axes And (3) with Define the missing interval as The goal is to generate intermediate state samples meeting the physical laws in the interval 。
  3. 3. The method for continuous spectrum space reconstruction based on depth generation model as set forth in claim 1, wherein said step 2 comprises the steps of: 2-1, performing unified dimension specification on an original matrix; 2-2 robust normalization with quantile truncation by computing the upper and lower bounds of the truncation on the training set of visible samples Truncating and linearly mapping data to ; 2-3, In the missing interval evaluation scene, the missing interval samples are only calculated based on training visible samples, so that the missing interval samples cannot indirectly participate in model fitting in a global statistical form.
  4. 4. The method for continuous spectrum space reconstruction based on depth generation model as set forth in claim 1, wherein said step 3 comprises the steps of: 3-1 for any target parameter Constructing normalized continuous condition variables ; Enabling the model to perceive the positional relationship of 'from end point to intermediate state' in a continuous manner; 3-2 during training and reasoning, endpoint samples As a global up-down Wen Maodian, As continuous condition coordinates for controlling intermediate state generation.
  5. 5. The method for continuous spectrum space reconstruction based on depth generation model as set forth in claim 1, wherein said step 4 establishes an endpoint-aligned trajectory structure for the samples under different parameters, comprising the steps of: 4-1 determining a fixed for each track based on the left endpoint sample index ; 4-2 Right endpoint samples And other parameters and the sample Performing one-to-one matching, and enhancing consistency of endpoint contexts by adopting a matching strategy based on sequence alignment or spectrum characteristics; 4-3 random sampling and sampling from the trajectory during each training iteration Corresponding target parameters Forming training tuples 。
  6. 6. The method for continuous spectrum space reconstruction based on depth generation model as in claim 1, wherein said step 5 uses endpoint condition variation to perform generation learning from encoder, and the key modules include: Endpoint context encoder to As input, extracting multi-scale context features for determining global boundary conditions for intermediate state generation; Posterior encoder to Combined input, estimation of posterior distribution parameters of latent variables The system is used for variation inference and uncertainty modeling of a training stage; decoder with endpoint context features, latent variables And conditions Output intermediate state prediction for input ; FRaman modulating the conditions Affine modulation parameters mapped to multi-layer feature channels To Modulating the multi-scale characteristics at the decoding end.
  7. 7. The method for continuous spectrum space reconstruction based on depth generation model as set forth in claim 1, wherein said step 6 of constructing joint loss for training optimization comprises the steps of: 6-1, a robust reconstruction term, namely, adopting Huber/SmoothL1 to restrict the whole spectrum shape and the space structure; 6-2, reconstructing peak sensitivity weighting, namely constructing a wave band weight graph based on the mean spectrum change rate of the training visible sample, so that a peak area obtains higher weight in optimization and is used for improving the fidelity of a characteristic peak; 6-3, spectrum gradient consistency, namely constraining first-order difference consistency of the wave band direction, and keeping local spectrum deformation rate and peak-shaped edge structure; 6-4, carrying out statistical form constraint, namely constraining a statistical consistency term, wherein the statistical consistency term is used for keeping the strength form of the space displacement dimension stable; 6-5, overshoot penalty, namely suppressing the non-physical enhancement of the systematic high prediction in the peak area; And 6-6, KL regularization and annealing, namely applying KL regularization to the latent variable distribution, and gradually enhancing KL weight by using Warmup strategy to improve the stability of potential spatial structuring and training.
  8. 8. The method for continuous spectrum space reconstruction based on depth generation model as set forth in claim 1, wherein said step 7 comprises the steps of: 7-1 given target deletion parameters Calculate the correspondence ; 7-2 Selection of endpoint context Latent variable Setting to zero vector to obtain deterministic output, or sampling from standard normal to obtain diversity output; 7-3 Generation by decoder Inversely transforming the normalization result to an original scale to obtain a complement sample of the missing interval; And 7-4, carrying out structural preservation on the generated samples according to the parameter sequence and the number, and using the samples for subsequent database amplification and downstream quantitative modeling.
  9. 9. The method for continuous spectrum space reconstruction based on depth generation model as set forth in claim 1, wherein said step 8 simultaneously verifies at output level the following aspects: Spectrum shape consistency, which includes mean spectrum consistency, peak area error, gradient consistency and space statistics consistency; Task driving consistency, namely taking 'missing training-missing+generating complement-full real training' as a unified comparison framework, checking whether the generating complement can approach to the performance upper bound of full real data in a quantitative regression task, and verifying the effectiveness and availability of the generating data in downstream modeling.

Description

Continuous spectrum space reconstruction method based on depth generation model Technical Field The invention relates to the technical field of spectrum modeling, in particular to the technical field of a continuous spectrum space reconstruction method based on a depth generation model. Background The space shift Raman spectrum (SPATIALLY OFFSET RAMAN SPECTROSCOPY, SORS) can acquire deep Raman response under the condition that surface shielding or a multilayer structure exists, so that the space shift Raman spectrum has important value in quantitative analysis and composition inversion of a complex system, the quantitative modeling of the space shift Raman spectrum is highly dependent on a standard sample database with sufficient coverage in engineering practice, and the construction of the database often faces a plurality of bottleneck problems. The Chinese patent with the application number of CN201910446804.7 discloses a Raman spectrum quantitative analysis technology based on a half-peak high-distance method, which takes the ratio of the strongest two characteristic peaks in a Raman spectrum of a sample to be detected as the basis for judging the quantitative analysis of substances, maximally reduces the interference of other substances, fully considers the peak type and the peak intensity and the errors caused by the change of the spectral peaks along with the change of the concentration of the substances, maximally eliminates the errors caused by the artificial and objective environments in the detection process, but still has the defects that the method is only suitable for the detection of the sample with single component or the concentration of the target object far higher than that of the other components and is not suitable for the Raman spectrum analysis detection of the multicomponent sample. The Chinese patent with the application number of CN202111546557.1 discloses a spectral data enhancement method based on conditional variation self-coding, which generates a virtual spectrum with the same distribution as the concentration of components to enhance a training set, is convenient for developing a calibration model, and adopts a regression learning model based on a semi-supervised ladder network for modeling by using the generated virtual spectrum. In the prior art, the construction of a standard sample database often faces the following common bottlenecks: The contradiction between the sampling discreteness and the physical continuity is prominent that the concentration gradient or the thickness parameter is a continuous physical quantity in nature, but is constrained by the sample preparation cost, the acquisition time and the stability, the actual data can be sampled discretely only at a limited concentration point, and particularly common continuous intervals are missing, namely, the lack of intermediate state samples in the parameter intervals leads to the model training to only cover end points or partial sparse nodes, so that the prediction of the missing intervals is closer to extrapolation rather than interpolation; The traditional interpolation/mixing method is difficult to keep physical forms of spectrum peaks, such as linear interpolation, weighted mixing and the like, pixel-level averaging is carried out in an observation space, nonlinear evolution process of Raman spectrum along with concentration change is difficult to be described, peak shape blurring, weak peak inundation, even 'ghost/double peak' and other non-physical phenomena are caused, and a generated result lacks scientific effectiveness; Typical condition generation countermeasure network, standard variation self-encoder and the like often follow the paradigm of dense supervision and coverage learning, when a missing interval sample is insufficient or completely invisible, the model is easy to degrade to end point copy, detail collapse or condition insensitivity, a stable end point-intermediate state continuous evolution track is difficult to form, and the problem is more remarkable for a space displacement Raman spectrum, because data is highly structured in a wave Duan Weishang, statistical form constraint exists in a space displacement dimension, due to high sample preparation and acquisition cost, concentration gradient or thickness parameter is often only discretely sampled, actual data presents remarkable sparsity in a physical parameter axis, and particularly common continuous interval missing is that a modeling task has the problems of incomplete database, learning bias, extrapolation degradation and the like due to the fact that the intermediate state sample is absent in the parameter interval. Disclosure of Invention The invention aims to solve the problems in the prior art, and provides a continuous spectrum space reconstruction method based on a depth generation model, which can construct a generation model which has explicit response to continuous physical parameters and can keep consistency of spectrum pea