CN-121994781-A - Soybean origin tracing detection method based on laser-induced breakdown spectroscopy technology

CN121994781ACN 121994781 ACN121994781 ACN 121994781ACN-121994781-A

Abstract

The invention is applicable to the technical field of agricultural product traceability detection, and discloses a soybean production area traceability detection method based on a laser-induced breakdown spectroscopy technology, which comprises the following steps: firstly, soybean samples are collected, a multi-producing-area mixed sample model is built, soybeans are pressed into tablets according to a certain pretreatment method, LIBS spectrums are scanned, after spectrum pretreatment such as baseline subtraction, smoothing and normalization, data sets are divided, a 1D-ResNet soybean producing area judging model is built and trained, and model performance is evaluated through a series of indexes. The detection method effectively improves the accuracy and stability of the soybean classification model, and meanwhile, compared with other machine learning algorithms, the 1D-ResNet also improves the judgment accuracy of the small sample amount samples and the mixed samples of the multiple production areas, and has important significance for quick evaluation and deployment of soybean production area tracing.

Inventors

Kang Chuanbin
YAN WENQIAN
LU SHAN
LI KEFEI
WANG KAIQIANG
JU LEI

Assignees

中国海洋大学

Dates

Publication Date: 20260508
Application Date: 20260320

Claims (9)

1. A soybean origin tracing detection method based on a laser-induced breakdown spectroscopy technology is characterized by comprising the following steps: S1, collecting soybean samples of different producing areas, and screening out impurities and defective soybean grains; s2, weighing soybean samples of different producing areas in the step S1, placing the soybean samples of different producing areas in the same container to enable the soybean samples to be uniformly distributed, and constructing different mixed sample models by adjusting source of the producing areas or sample batches; s3, randomly taking out the soybean samples in S1 and S2 to obtain a set number of soybean grains, grinding the soybean grains into powder, and sieving the powder; S4, taking out part of soybean powder in the step S3, uniformly mixing the soybean powder with microcrystalline cellulose according to a set weight ratio, and pressing the mixture into tablets with consistent specification for spectrum test; S5, scanning the tabletting sample of the S4 by using a laser-induced breakdown spectroscopy LIBS device, and collecting LIBS spectrums of different types of soybean samples; S6, preprocessing the LIBS spectrum acquired in the S5; S7, dividing the data set into a training set and a testing set according to a proportion after preprocessing in the S6; And S8, constructing a 1D-ResNet soybean origin discrimination model by using the training set in the S7, and evaluating the performance of the model by using the testing set.
2. The method for detecting soybean origin tracing based on laser induced breakdown spectroscopy according to claim 1, wherein in the step S1, soybean samples from different origins are collected, at least 5 soybean samples from different origins are collected, and each origin needs to cover different batches of soybean samples.
3. The method for detecting soybean origin tracing based on the laser-induced breakdown spectroscopy according to claim 1, wherein the construction of each mixed sample model in the step S2 at least comprises soybean samples from two different origins, and the weight ratio of any two soybean samples from different origins is controlled to be 0.25 ‒ 1 when the soybean samples from different origins are weighed.
4. The soybean origin tracing detection method based on the laser-induced breakdown spectroscopy technology of claim 1, wherein the weight of the soybean sample taken in the step S3 is greater than 80 g.
5. The method for detecting soybean origin tracing based on the laser-induced breakdown spectroscopy technology according to claim 1, wherein the weight of the soybean powder used in the step S4 is not less than 0.5 g, the mixing weight ratio of microcrystalline cellulose and the soybean powder is not more than 1:2, and the weight of each tablet is not more than 0.5: 0.5 g and the diameter is not more than 15mm when the tablets are pressed.
6. The soybean origin tracing detection method based on the laser-induced breakdown spectroscopy technology according to claim 1, wherein in the step S5, random point positions on a tablet plane are tested in parallel, and elements contained in LIBS spectra include one or more elements of calcium Ca, potassium K, magnesium Mg, sodium Na, carbon C, hydrogen H and oxygen O.
7. The soybean origin tracing detection method based on the laser induced breakdown spectroscopy technology according to claim 1, wherein in the step S6, the pretreatment method comprises baseline removal, spectrum smoothing and normalization.
8. The soybean origin tracing detection method based on the laser induced breakdown spectroscopy technology according to claim 1, wherein 1D-ResNet in the step S8 is a residual network ResNet variant of adjusting an input structure for one-dimensional spectral data characteristics, and a residual module is introduced, and the module adds an input identity map and a nonlinear transformed output through skip connection, so as to change a learning target of the network from a direct fitting expected map H (x) to a fitting residual map F (x), and the specific calculation is expressed as: , where x is the input of the residual block, F (x) represents the residual map learned by the stacked convolutional layers, and H (x) is the map that the block expects to output.
9. The soybean origin tracing detection method based on the laser-induced breakdown spectroscopy technology according to claim 1, wherein the indexes for evaluating the model performance in the step S8 include Accuracy, precision, recall and F1 Score as indexes for evaluating the model performance, and the calculation method is as follows: , , , 。

Description

Soybean origin tracing detection method based on laser-induced breakdown spectroscopy technology Technical Field The invention relates to the technical field of agricultural product traceability detection, in particular to a soybean origin traceability detection method based on a laser-induced breakdown spectroscopy technology. Background Soybeans are globally important grain and oil crops, and the soybeans in different producing areas have certain differences in nutritional ingredients and element compositions due to the differences in natural conditions such as soil, climate and the like. Accurate soybean origin tracing is of great significance for guaranteeing food safety, maintaining consumer rights and interests and normalizing trade order. Currently, common origin tracing techniques cover methods such as stable isotope ratio analysis, mineral element fingerprint analysis, molecular biology, metabonomics, and the like. These methods have advantages, but often have the disadvantages of complex pretreatment of the sample, long detection period, or the need of treatment with chemical reagents. Laser Induced Breakdown Spectroscopy (LIBS) is an emerging elemental analysis technique that uses high-energy pulsed laser to irradiate the sample surface, ablate, ionize and form a high temperature plasma at a transient high temperature. Along with the rapid cooling of the plasma, the element atoms and ions in the excited state relax and emit characteristic spectra, and the element composition information reflecting the characteristics of the production place of the sample is obtained. The method has the advantages of rapidness, no damage and simultaneous detection of multiple elements, and is widely focused in the field of agricultural product origin tracing in recent years. However, the spectrum signal obtained by LIBS detection often has problems of background noise interference, overlapping characteristic spectral lines and the like, and is difficult to accurately mine the characteristic information of the production place in the spectrum data by means of the traditional spectrum analysis method, so that the accurate judgment of the production place cannot be realized. Under the background, the machine learning provides an effective way for solving the problems by virtue of the strong data mining, feature extraction and pattern recognition capability, and becomes an important support for realizing accurate application of the LIBS technology in tracing the origin of agricultural products. Class algorithms such as Linear Discriminant Analysis (LDA) and Support Vector Machines (SVM) have been applied to classification tasks of LIBS spectra, but the class algorithms have difficulty in effectively dealing with problems such as high dimensionality, nonlinearity, spectral line overlapping and the like of spectrum data, and the tracing accuracy still has room for improvement. The deep learning algorithm such as Convolutional Neural Network (CNN) has obvious advantages in the aspects of processing high-dimensional and nonlinear LIBS spectrum data and solving the problem of spectral line overlapping by virtue of the unique convolutional structure and strong feature extraction capability. When the CNN is applied to LIBS spectrum data processing, a one-dimensional convolutional neural network (1D-CNN) can be constructed by adjusting the sliding dimension of the convolutional kernel so as to adapt to the high-dimensional tensor characteristic of the light spectrum data, and further deep characteristic information in the spectrum can be effectively mined. However, researches show that as the number of network layers increases, the input features of the neural network are gradually abstracted along with the deepening of the network depth, and although the deeper features can be extracted, shallow feature information which is beneficial to the source tracing task of the production area is possibly lost, in addition, the too deep number of network layers can obviously increase the complexity of a model, so that the training difficulty is increased, the fitting phenomenon is easily caused, and the accuracy and the stability of the source tracing of the soybean production area are further affected. Particularly, in practical application, the soybean trade circulation link is complex, the situation that soybeans in different producing areas are mixed often occurs, and soybean samples in a specific producing area are limited in number and are difficult to obtain in a large amount. However, most of existing traceability models based on machine learning are constructed aiming at ideal conditions of single production place and sufficient sample size, attention on discrimination capability of mixed samples is insufficient, and when the production place category with sparse sample size is processed, the models are extremely easy to be subjected to over fitting or under fitting due to insufficient training data, so that discrimination pr