CN-121982327-A - Raman spectrum characteristic extraction method for physical constraint countermeasure enhancement and cooperative sensing

CN121982327ACN 121982327 ACN121982327 ACN 121982327ACN-121982327-A

Abstract

The invention provides a Raman spectrum characteristic extraction method of physical constraint contrast enhancement and cooperative perception, which is characterized in that one-dimensional Raman spectrum data is converted into a two-dimensional structure characteristic diagram through a gram angle field algorithm, on the basis of the two-dimensional structure characteristic diagram, a physical prior constraint generation contrast network is introduced through Lorentz distribution, an enhanced two-dimensional structure characteristic diagram with physical consistency is generated, one-dimensional physical form characteristics, two-dimensional space structure characteristics and frequency domain significant characteristics of Raman spectrum are aimed at, multi-mode characteristic extraction is carried out through heterogeneous neural network branches, an adaptive attention mask is constructed by utilizing the space structure characteristics and the frequency domain significant characteristics, self-adaptive weighting calibration is carried out on the one-dimensional physical form characteristics through an attention mechanism, and cooperative purification characteristics are constructed. The invention ensures that the finally generated collaborative purification features have the spatial structure stability and the frequency domain distribution significance at the same time, and greatly improves the classification capacity and the robustness of the classification decision layer.

Inventors

PAN ZHENG
JIN XIU

Assignees

安徽农业大学

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. The Raman spectrum characteristic extraction method for physical constraint antagonism enhancement and cooperative sensing is characterized by comprising the following steps of: Converting the preprocessed one-dimensional Raman spectrum data into a two-dimensional structural feature map through a Grahm angle field algorithm, and synchronously extracting one-dimensional energy distribution features through discrete wavelet transformation; on the basis of the two-dimensional structure feature map, generating an enhanced two-dimensional structure feature map with physical consistency by introducing Lorentz distribution as a generation countermeasure network of physical priori constraint; Aiming at the one-dimensional physical morphological characteristics, the two-dimensional spatial structural characteristics and the frequency domain significant characteristics of the Raman spectrum, carrying out multi-mode characteristic extraction through heterogeneous neural network branches, constructing a self-adaptive attention mask by utilizing the spatial structural characteristics and the frequency domain significant characteristics, carrying out self-adaptive weighting calibration on the one-dimensional physical morphological characteristics through an attention mechanism, and constructing a collaborative purification characteristic; And constructing a classification model based on a collaborative purification feature mode set, introducing a global joint loss function containing center loss to perform multi-model collaborative optimization, and realizing the high clustering of Raman fingerprint feature peaks by contracting the intra-class distribution radius of different class features in a high-dimensional space to finish the collaborative optimization of the feature extraction model and the classification decision model.
2. The method for extracting features of raman spectra for physical constraint antagonism enhancement and co-sensing according to claim 1, wherein the step of converting the features into a two-dimensional structural feature map by a glamer angle field algorithm and synchronously extracting the energy distribution features of the one-dimensional raman spectra comprises: normalizing the pretreated Raman spectrum sequence X, and converting the pretreated one-dimensional spectrum data into two-dimensional Raman spectrum data through a gram angle field algorithm; Discrete wavelet transformation is selected to extract the energy distribution characteristics of one-dimensional Raman spectrum, and L layers of multi-scale decomposition is carried out on the sequence X to obtain wavelet coefficients of all levels; constructing a one-dimensional multi-scale energy distribution vector E as one-dimensional energy distribution characteristics according to detail coefficients of all scales, wherein energy values are calculated respectively by detail coefficients of all L decomposition scales And arranging the energy distribution characteristic vectors according to a scale sequence to obtain a one-dimensional energy distribution characteristic vector E.
3. The method for extracting features of raman spectra for physical constraint challenge-enhancement and co-sensing according to claim 2, wherein in said step of generating a challenge network by introducing lorentz distribution as physical prior constraint, two-dimensional raman spectrum structural data obtained by a garland angle field transformation is obtained Sampling N random vectors from standard normal distribution N (0, 1), and collecting random vectors The random vector set Z is input to a generator, and the generator outputs data with the same input size , wherein, For the nth two-dimensional raman spectral structure data sample, For the nth random vector noise, Output data of the generator to the nth noise vector; constructing a physical constraint module P, and outputting data from a generator through a gram angle field inverse transformation operator Reduction to one-dimensional pseudo-spectrum sequence ; The lorentz distribution was introduced as a physical a priori constraint of the raman spectrum generation countermeasure network, expressed as: ; In the formula, Representing the lorentz distribution, v being the wavenumber of the raman spectrum, Is the intensity coefficient of the i-th raman peak, Is the center wavenumber of the ith raman feature peak, Is the half-width of the ith raman characteristic peak.
4. A method of extracting features of raman spectra for physical constrained challenge enhancement and co-sensing as claimed in claim 3 wherein the step of generating an enhanced two-dimensional structural feature map with physical consistency comprises: loss function of model generated in generator by combining Raman intrinsic physical prior constraint Expressed as: ; where z is from a priori distribution Is used to generate the noise vector that is randomly generated, Is a priori distribution The expected noise vector to be generated is then, Is the data generated by the generator and, Is the decision of the arbiter on the generated data, outputs a value between 0 and 1 for representing the probability that the data is real data, As a physical balance factor, the balance factor is, Is a physical consistency loss term; through the physical consistency loss function, the generator parameters reach the optimal balance point after updating the weights for a plurality of times, and the generator outputs data And inputting the true sample Y into a discriminator, wherein the loss function of the discriminating model in the discriminator Expressed as: ; where y is the distribution from the real data The real data of the mid-sample is sampled, Is distributed in real data The following expectations for all possible real sample data y, Is a priori distribution The expected noise vector to be generated is then, Is the judgment of the discriminator on the real data, outputs a value between 0 and 1, represents the probability that the data is the real data, Is a characteristic pattern generated by the generator and, Is the decision of the discriminator on the generated feature pattern.
5. The method for extracting features of raman spectra for physical constraint challenge enhancement and co-sensing according to claim 4, wherein the step of extracting multi-modal features through heterogeneous neural network branches comprises: constructing a 1D-CNN convolutional neural network, and capturing physical morphological characteristics of a spectrum from a one-dimensional Raman sequence by alternately stacking one-dimensional convolutional layers and maximum pooling layers ; One-dimensional Raman spectrum characteristics of preprocessing original one-dimensional Raman spectrum data Inputting the features into a 1D-CNN model to obtain features mapped by k layers of features 。
6. The method for extracting features of raman spectra for physical constraint challenge enhancement and co-sensing according to claim 5, wherein the step of extracting multi-modal features by heterogeneous neural network branches further comprises: Constructing a 2D-CNN convolutional neural network, inputting a two-dimensional Raman spectrogram subjected to physical constraint countermeasure enhancement by alternately stacking two-dimensional convolutional layers and average pooling layers, and extracting spatial structural features ; Two-dimensional Raman spectrum after physical constraint antagonism enhancement Inputting the features into a 2D-CNN model to obtain features mapped by m layers of features 。
7. The method for extracting features of raman spectra for physical constraint challenge enhancement and co-sensing according to claim 6, wherein the step of extracting multi-modal features through heterogeneous neural network branches further comprises: constructing a multi-layer perceptron MLP comprising a full-connection layer, inputting a one-dimensional energy distribution characteristic vector E obtained by wavelet transformation decomposition into the multi-layer perceptron, and extracting frequency domain significance characteristics reflecting vibration energy levels 。
8. The method for extracting features of raman spectrum for physical constraint antagonism enhancement and co-perception according to claim 7, wherein the step of constructing an adaptive attention mask using spatial structural features and frequency domain salient features comprises: to two-dimensional space structure characteristics And frequency domain saliency features Performing dimension stitching to generate an adaptive attention mask M, which is expressed as: ; wherein each element in the adaptive attention mask M corresponds to a weight value for representing the importance degree of the corresponding feature position, For Sigmoid activation functions, the output values are mapped in the 0-1 interval, W 1 and W 2 represent a learnable weight matrix, The feature concatenation operator is represented as a function of the feature concatenation operator, The characteristics of the spatial structure are represented, Representing the frequency domain saliency features.
9. The method for extracting raman spectral features of physical constraint antagonism enhancement and cooperative sensing according to claim 8, wherein the step of adaptively weighted calibrating the one-dimensional physical morphological feature by an attention mechanism comprises: Physical morphological characteristics of one-dimensional Raman spectrum by utilizing cooperative sensing mask M Performing self-adaptive weighting calibration, and performing dynamic gain adjustment to obtain an optimal feature set after recalibration 。
10. The method for extracting raman spectral features of physical constraint contrast enhancement and co-perception according to claim 9, wherein in the step of introducing a global joint loss function containing a center loss for multi-model co-optimization, the global joint loss function And center loss Expressed as: ; ; Where m represents the number of batch samples, Indicating the co-purification vector of the i-th sample after the co-recalibration , Representing the feature center of the class to which the i-th sample belongs, As a weight balance factor, the weight of the object is calculated, Is a cross entropy loss function for quantifying the discrimination loss of the classification layer based on the collaborative purification feature set, drives the classification layer to continuously optimize parameters through back propagation, Is the center loss function by minimizing And category feature center The Euclidean distance of the Raman spectrum characteristic is promoted to be gathered to the class center height of the Raman spectrum characteristic in the ultra-high dimensional characteristic space.

Description

Raman spectrum characteristic extraction method for physical constraint countermeasure enhancement and cooperative sensing Technical Field The invention belongs to the technical field of artificial intelligence, in particular to the field of Raman spectrum feature extraction and modeling analysis, and particularly relates to a Raman spectrum feature extraction method for physical constraint countermeasure enhancement and cooperative sensing. Background As a molecular fingerprint spectrum based on the inelastic scattering principle, the Raman spectrum technology has the advantages of no damage, rapidness, high specificity and the like, and has irreplaceable effects in substance component identification and content analysis. Compared with the traditional analysis means, the Raman spectrum technology has the remarkable advantages of simple sample preparation, high detection speed, no damage and the like. At present, the technology is widely applied to industries such as biological medicine, chemistry and chemical engineering, food safety, environmental monitoring and the like. Because raman spectroscopy can provide very specific molecular fingerprint information, it has become an important tool for analyzing the relationship between the substance components and the molecular structure-activity in complex systems. However, in practical applications, the prior art still faces the following challenges: firstly, raman spectrum data has extremely high dimensionality and contains complex redundant information, and when the traditional statistical method or machine learning algorithm processes the data, the problem of dimension curse is extremely easy to face, so that the generalization performance of a model is reduced, and deep nonlinear spatial structural features and frequency significant features are difficult to capture. Secondly, in the prior art, when data are enhanced, only statistical similarity of data distribution is usually focused, intrinsic physical constraint behind a Raman spectrum is ignored, physical distortion of a generated sample in key dimensions such as spectrum peak line type, energy broadening and the like is easy to occur, and real energy level response characteristics of a substance under a specific physical environment are difficult to characterize. In addition, the existing raman spectrum feature extraction and classification modeling schemes lack multi-mode feature collaborative extraction. In raman spectroscopy, key fingerprint information is often submerged in a strong low-frequency fluorescent background, and it is difficult for a single-dimensional sensing mode to accurately focus weak fingerprint peaks while suppressing background noise, resulting in limited accuracy of classification layer models when dealing with high-sensitivity detection requirements in complex environments. Disclosure of Invention The embodiment of the invention aims to provide a Raman spectrum characteristic extraction method for physical constraint antagonism enhancement and cooperative sensing, which aims to solve the technical problems in the background technology. In order to achieve the above purpose, the present invention provides the following technical solutions. According to one embodiment of the present invention, there is provided a raman spectral feature extraction method of physical constraint challenge enhancement and cooperative sensing, comprising the steps of: Converting the preprocessed one-dimensional Raman spectrum data into a two-dimensional structural feature map through a Grahm angle field algorithm, and synchronously extracting one-dimensional energy distribution features through discrete wavelet transformation; on the basis of the two-dimensional structure feature map, generating an enhanced two-dimensional structure feature map with physical consistency by introducing Lorentz distribution as a generation countermeasure network of physical priori constraint; Aiming at the one-dimensional physical morphological characteristics, the two-dimensional spatial structural characteristics and the frequency domain significant characteristics of the Raman spectrum, carrying out multi-mode characteristic extraction through heterogeneous neural network branches, constructing a self-adaptive attention mask by utilizing the spatial structural characteristics and the frequency domain significant characteristics, carrying out self-adaptive weighting calibration on the one-dimensional physical morphological characteristics through an attention mechanism, and constructing a collaborative purification characteristic; And constructing a classification model based on a collaborative purification feature mode set, introducing a global joint loss function containing center loss to perform multi-model collaborative optimization, and realizing the high clustering of Raman fingerprint feature peaks by contracting the intra-class distribution radius of different class features in a high-dimensional space to fin