Cornu cervi pantotrichum variety identification method based on multi-dimensional data fusion and artificial intelligence optimization algorithm

Abstract

The invention relates to a pilose antler variety rapid identification method and system based on multi-dimensional data fusion and an artificial intelligence optimization algorithm, and the method comprises the steps: collecting multi-dimensional feature data of colors, textures, component contents and the like of pilose antler pieces of different varieties by using computer vision (CV), high performance liquid chromatography (HPLC) and a rapid meteorological electronic nose technology (UF-GC-E-nose); comprising sika antler, hairy antler, domesticated antler and camel antler, efficient, accurate and automatic variety identification and traceability are realized, the characteristic that the whole is superior to the sum of parts in a complex system is reflected, a good traceability result is generated, and a constructed intelligent algorithm based on multi-dimensional data fusion provides reference and data support for food variety identification and traceability.

Description

Technical Field

[0001] The present invention belongs to the intersection of intelligent traceability technology, food quality control and artificial intelligence applications. Specifically, it relates to a method and system for identifying deer antler varieties based on multi-dimensional perception data fusion and intelligent classification algorithm. The method and system can be widely used in scenarios such as quality testing of traditional Chinese medicines, identification of functional foods, traceability of agricultural products, and source identification of high-value animal-derived products. Background Art

[0002] Deer antlers (velvet antlers) are the densely hairy, unossified antlers of male deer (Cervus nippon Temminck) or red deer (Cervus elaphus Linnaeus), members of the Cervidae family. They are known for strengthening kidney yang, nourishing essence and blood, and strengthening tendons and bones, making them a precious traditional Chinese medicinal ingredient. Since the Shennong Compendium of Materia Medica, deer antlers have been considered a top-grade medicinal herb. Medical texts throughout history document their ability to "produce essence and replenish marrow, nourish blood and invigorate yang." They are widely used clinically for symptoms such as consumptive emaciation, impotence, spermatorrhea, and soreness and coldness of the waist and knees. Modern research indicates that different antler varieties exhibit significant differences in their medicinal components and efficacy due to their origins. Sika deer antlers and red deer antlers are listed as legal medicinal herbs in the Chinese Pharmacopoeia, while reindeer and moose antlers, due to their lower content of active ingredients, are often counterfeited. However, deer antler slices often have highly similar morphologies after processing. Traditional identification relies primarily on empirical observation of characteristics such as velvet distribution, velvet skin texture, and color, which is highly subjective and inaccurate. It is common to see low-priced deer antlers being passed off as high-priced varieties in the market, which seriously restricts the quality control and clinical application of deer antler medicinal materials.

[0003] In recent years, research on velvet antler identification has largely focused on single techniques: High-performance liquid chromatography (HPLC) can detect differences in the content of characteristic amino acids such as glycine and proline in velvet antlers, but cannot distinguish between morphologically similar species; computer vision techniques analyze velvet texture and color for preliminary classification, but are sensitive to ambient lighting and camera conditions; and electronic nose technology can capture the fingerprint of volatile odor components in velvet antlers, but is susceptible to interference from sample moisture content and storage conditions. A single detection method struggles to fully characterize the multidimensional characteristics of velvet antlers, resulting in insufficient identification accuracy.

[0004] In view of this, it is urgent to build a multi-technique integrated rapid identification system for deer antler slices. This study proposes to integrate three types of feature data: HPLC detection of amino acid composition, computer vision extraction of texture and color parameters, and electronic nose analysis of odor fingerprints. It also innovatively introduces the Whale Optimization Algorithm (WOA) to adaptively optimize the hyperparameters of the Random Forest (RF) model. By integrating multi-source data fusion with machine learning algorithms, it breaks through the limitations of traditional methods. Preliminary experiments have shown that this method can accurately distinguish between sika deer antler, red deer antler, reindeer antler and moose antler, with an identification rate of 100%, providing efficient and reliable technical support for variety identification, quality evaluation and market supervision of deer antler medicinal materials. Summary of the Invention

[0005] The present invention aims to solve the problems of existing antler variety identification methods, such as complex operation, low efficiency, low accuracy, and difficulty in meeting the rapid traceability needs of the modern market. It provides a antler variety identification method and system based on multi-dimensional data fusion and artificial intelligence optimization algorithm. Through multi-source perception of image information, odor components and chemical composition, combined with an intelligent classification model, it can achieve rapid, accurate and efficient identification of different antler varieties.

[0006] Technical Solution

[0007] A method for identifying antler varieties based on multidimensional data fusion and artificial intelligence optimization algorithm, comprising the following steps:

[0008] (1) collecting multidimensional feature data of antler samples of different varieties, wherein the feature data includes color, texture, smell and chemical composition information;

[0009] (2) Computer vision technology was used to extract color and texture features, ultra-fast gas phase electronic nose technology was used to extract odor features, and high performance liquid chromatography technology was used to extract amino acid and other component information;

[0010] (3) Perform multivariate statistical analysis on the characteristic data to screen out key discriminant factors with variable importance in the projection (VIP) greater than 1 and P value less than 0.05;

[0011] (4) Using the key feature factors as input variables, a WOA-RF classification model that integrates the whale optimization algorithm and random forest is constructed;

[0012] (5) The WOA-RF classification model is used to automatically classify and identify the varieties of the antler samples to be tested, thereby achieving efficient traceability and identification of antler varieties.

[0013] The method described herein comprises four varieties of antler samples, namely, sika deer antler, red deer antler, reindeer antler and moose antler.

[0014] In the method, the color features include RGB (Red, Green, Blue), and HSV includes hue (Hue, H), saturation (Saturation, S), and value (Value, V); the texture features are extracted by a local binary pattern (LBP) algorithm.

[0015] The method described herein is to extract the odor characteristics by using a Heracles Neo ultra-fast gas phase electronic nose and analyze them using a low-polarity chromatographic column (MXT-5) and a polar chromatographic column (MXT-WAX).

[0016] In the method, the chemical components are amino acid components, including hydroxyproline, glycine, proline, leucine and lysine.

[0017] The method described above, wherein the WOA-RF classification model optimizes the feature subset and hyperparameters of RF through the WOA algorithm, and the model iteration number is no less than 200 times.

[0018] The method described above achieves a classification accuracy of the model of over 98% on both the training set and the test set.

[0019] The method also includes inputting the multidimensional feature data into multivariate analysis software such as SIMCA to perform principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA) to assist in modeling and visualization.

[0020] According to the method described, among the four species of antlers, namely sika deer antler, red deer antler, reindeer antler and moose antler, 15 odor components were detected in sika deer antler, and 3-methyl-2-butenal, 2,3-octanedione and limonene were only found in sika deer antler; there were 16 odor components in red deer antler, among which isopentanol, 1,3-dioctene and n-octanal were only detected in red deer antler; there were 10 odor components detected in reindeer, among which methyl mercaptan, acetaldehyde, acetone, isobutyraldehyde and isovaleraldehyde accounted for the least among the four antler species; there were 7 odor components detected in moose antler, n-butyraldehyde and nonanal were detected in the other three species but not in moose antler. There were a total of 7 odor components contained in all four antler species, namely methyl mercaptan, acetaldehyde, acetone, isobutyraldehyde, isovaleraldehyde, 2-methylbutyraldehyde and n-hexanal.

[0021] According to the method described, the content of hydroxyproline in four varieties of antlers, namely sika deer antler, red deer antler, reindeer antler and moose antler is compared. Among them, sika deer antler has the highest content and red deer antler has the lowest content; for glycine, moose antler has the highest content and red deer antler has the lowest content; sika deer antler has the highest proline content and red deer antler has the lowest; red deer antler has the highest leucine content and moose antler has the lowest content; and red deer antler has the highest lysine content and moose antler has the lowest content.

[0022] Specifically:

[0023] 1. Sample collection and data acquisition:

[0024] The color, texture, odor, and chemical composition data of four types of velvet antlers were collected, totaling 120 batches. These included sika deer velvet antler (SVA), red deer velvet antler (WVA), reindeer velvet antler (RVA), and moose velvet antler (MVA).

[0025] 2. Feature extraction:

[0026] (1) Using computer vision technology to collect images and extract color features in RGB and HSV color spaces (including hue H, saturation S, and lightness V);

[0027] (2) Use LBP algorithm to extract cross-sectional texture features of the sample;

[0028] (3) Using ultrafast gas phase electronic nose (UF-GC-E-nose) to extract the odor information of the sample;

[0029] (4) Quantitative analysis of the contents of five representative amino acids (hydroxyproline, glycine, proline, leucine, and lysine) was performed using high performance liquid chromatography (HPLC);

[0030] 3. Feature screening and modeling:

[0031] Multivariate statistical analysis methods (OPLS-DA and VIP analysis) were used to screen out 162 key characteristic factors with VIP values ​​greater than 1 and P values ​​less than 0.05;

[0032] 4. Intelligent model construction:

[0033] The filtered key features are input into the WOA-RF model that combines the Whale Optimization Algorithm (WOA) and the Random Forest (RF) for classification training and parameter optimization;

[0034] 5. Variety classification and traceability identification:

[0035] The WOA-RF model is used to identify the species of the antler samples to be tested, output specific species labels, and complete species classification and traceability.

[0036] Compared with the prior art, the present invention has the following significant advantages:

[0037] (1) High accuracy: The WOA-RF classification model was used to identify antler species, achieving 100% classification accuracy in both the training set and the test set, which is far superior to traditional linear discriminant or single feature recognition methods;

[0038] (2) Strong data fusion: integrating image, odor and component information, reflecting the emergent property of "the whole is better than the sum of its parts" in complex systems, effectively improving the expressiveness and adaptability of the model;

[0039] (3) Strong algorithm optimization capability: The whale optimization algorithm is used to perform dual optimization of feature selection and model parameters to enhance the robustness, generalization ability and computational efficiency of the model;

[0040] (4) High application value: The present invention provides an accurate, fast and intelligent solution for antler variety identification, quality control and market supervision, and can be widely applied to meet the intelligent traceability needs in various fields such as Chinese herbal medicines, functional foods, and animal-derived agricultural products;

[0041] (5) Cross-domain transferability: This method can be extended to complex environmental identification tasks with high noise and small samples, such as archaeological samples, ecological materials, food anti-counterfeiting, and agricultural quality control, and has good versatility and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The chromatograms are of mixed reference solution and test solution.

[0043] Figure 2 This is a quantitative analysis chart of the properties of deer antler slices from different sources.

[0044] Figure 3 This is an analysis of the odor differences of deer antler slices from different sources.

[0045] Figure 4 This is an analysis chart of the amino acid content of deer antler slices from different sources.

[0046] Figure 5 This is a multivariate statistical analysis of multidimensional data fusion of deer antler slices from different sources.

[0047] Figure 6 This is the result of the WOA-RF classification algorithm for deer antler slices from different sources. DETAILED DESCRIPTION

[0048] The present invention will be further described below with reference to specific embodiments. However, the present invention is not limited to the following embodiments. Without departing from the essential spirit of the present invention, technicians in the relevant technical field can make various equivalent deformations and modifications, which should all be included in the scope of protection of the present invention.

[0049] The biological materials in the present invention are derived from the following: unossified, densely hairy young horns of male deer of the Cervidae family, including sika deer Cervus nippon Temminck, red deer Cervus elaphus Linnaeus, elk Cervus alces Linnaeus, and reindeer Cervus tarandus Linnaeus, purchased from Jilin Aodong Pharmaceutical Group Co., Ltd.

[0050] The information of instruments and reagents used in the following examples are as follows:

[0051] The reference substance was a series of alkane mixtures (nC6-nC16, Lot: A0142930) purchased from RESTEK Co. Ltd. Amino acid reference substances were L-hydroxyproline (Shanghai Macklin Biochemical Co., Ltd., batch number C11478024, purity 99%); glycine (batch number S29A10I96410, purity ≥99%), L-proline (batch number Z12O11H127142, purity ≥99%), L-leucine (batch number J07GB154079, purity ≥98%), and L-lysine (batch number J19IB217736, purity ≥98%), all purchased from Shanghai Yuanye Bio-Technology Co., Ltd.; methanol and acetonitrile (chromatographic grade, ThermoFisher Scientific-CN); and phenyl isothiocyanate (batch number C12095842, Shanghai Macklin Biochemical Co., Ltd.). Biochemical Co., Ltd); triethylamine (batch number 20141217), n-hexane (batch number 20221124) and anhydrous sodium acetate (batch number 20231128) were all analytically pure and purchased from Sinopharm Chemical Reagent Co., Ltd.

[0052] Example 1: Establishment of a method for identifying antler varieties based on multidimensional data fusion and artificial intelligence optimization algorithm 1. Sample preparation

[0053] 120 batches of dried velvet antler samples from various sources were collected from a deer farm under Jilin Aodong Pharmaceutical Group Co., Ltd. in Jilin, China. The samples were stored in a refrigerator at -4°C until use. Detailed information about the samples is provided in Table 1. Thirty batches of sika deer antler, 30 batches of red deer antler, 30 batches of reindeer antler, and 30 batches of moose antler were collected from Jilin (JL) and Liaoning (LN) provinces and cities. These 120 batches of velvet antler samples were authenticated as authentic by Professor Chen Jianwei of Nanjing University of Chinese Medicine. As a renewable biological resource, velvet antler has long enjoyed widespread application worldwide. Velvet antler is primarily harvested during the antler's juvenile stage, when the hairs have not yet ossified and the antlers grow rapidly. Spring and summer are the optimal harvesting times. It is noteworthy that velvet antler does not pose long-term health risks to deer, as it is a renewable tissue and the deer are not harmed during the velvet growth process. Therefore, harvesting velvet antler meets basic animal welfare requirements and does not involve harming or sacrificing animals. Furthermore, modern deer farming has adopted sustainable development and scientific management practices, such as artificial breeding and antler management, to ensure that antler harvesting does not harm deer populations and contributes to the conservation of natural resources. This ensures that antler harvesting complies with environmental protection requirements while ensuring animal welfare during the farming process. Therefore, for these reasons, our study used antler as the research subject, which not only adheres to animal ethics but also meets the standards of modern farming and animal protection. The unique physiological structure of antler antler results in significant variations in the properties and quality of antler slices cut from different locations on the same antler. Considering market demand for antler antler, we divided the entire antler into five equal sections along its growth direction from bottom to top. The samples used in this study were all from the 3 / 5 section (i.e., the middle section). Because antler antler is primarily used in slice form for both edible and medicinal purposes, the samples used in this study were uniformly sliced ​​antler slices. Finally, each batch of samples was numbered, such as Sika Deer Velvet Antler (SVA) as SVA-1, Wapiti Velvet Antler (WVA) as WVA-1, Reindeer Velvet Antler (RVA) as RVA-1, and Moose Velvet Antler (MVA) as MVA-1. Prior to analysis, the samples were pulverized using a grinder, sieved through a 50-mesh sieve, and stored in a dry, airtight environment.

[0054] Table 1 Detailed information of the sample

[0055]

[0056]

[0057] 2. Antler trait extraction based on computer vision

[0058] 2.1 Extracting the appearance color of antler samples from different sources based on Python language

[0059] A Python algorithm was developed to extract the color features of antler antler samples. This algorithm primarily utilizes the summation of RGB value vectors, incorporating additional metrics such as hue, saturation, and value to achieve a more comprehensive color analysis. The RGB color space linearly represents color using red, green, and blue channels, while the HSV color space provides a more intuitive description of color features using hue, saturation, and value. Therefore, the HSV color space is more consistent with human subjective color perception. During the experiment, images of antler samples were acquired using automatic white balance and exposure to avoid interference caused by color gain. All samples were placed on a platform below the camera under uniform background and environmental conditions. One image was captured for each batch of samples, and all images were saved in JPG format with a size of 4000 × 3000 pixels. Using color extraction techniques, we extracted chromaticity values ​​from the collected images. Using computer vision techniques, we generated a dataset of 120 samples covering four categories. This approach ensured data accuracy and consistency, providing a solid foundation for subsequent color analysis.

[0060] 2.2 Establish a Python algorithm for extracting texture features of DVA samples

[0061] A Python algorithm was established to extract the texture features of PCR samples. The method is the same as 2.1.

[0062] We applied the Local Binary Pattern (LBP) algorithm to quantify the texture parameters of DVA samples. LBP is a widely used method for extracting image texture features, recognized for its rotational and grayscale invariance. The LBP algorithm primarily extracts local texture information from an image: within a 3x3 pixel window, the grayscale values ​​of eight adjacent pixels are compared with the center pixel, using the center pixel as the threshold. If the surrounding pixel values ​​are higher than the center pixel value, the location is marked as 1; otherwise, it is marked as 0. Through this comparison, the eight points within the 3x3 neighborhood generate an 8-bit binary number (typically converted to a decimal LBP code, which has 256 possible values) that reflects the texture characteristics of the region. Ultimately, a total of 256 texture features were extracted from each image, resulting in a dataset of 256 for subsequent analysis. The detailed texture extraction algorithm steps are detailed in Table S3. This method effectively captures and analyzes the texture information of samples, laying the foundation for further scientific research.

[0063] 3. UF-GC-E-nose analysis

[0064] This experiment mainly uses the Heracles Neo ultra-fast gas chromatography electronic nose and a PAL RSI fully automatic headspace sampler (Alpha MOS, France). It has two chromatographic columns with different polarities, namely the MXT-5 (low polarity) column and the MXT-WAX (polar) column. The sample is injected once and the two columns are analyzed simultaneously. The two chromatographic columns output the results simultaneously. After deducting the blank reference, the odor component information of deer antler can be obtained.

[0065] 3.1 Sample incubation

[0066] In this analysis, automated headspace sampling was used. The sample stirring rate, incubation temperature, and incubation time were determined. During the incubation process, the odor components of the sample evaporated and concentrated in the upper portion of the sealed gas phase bottle, eventually reaching equilibrium. A 1.0 g sample was placed in a 20.0 mL gas phase bottle and transported by a robotic arm to an incubator for incubation. A stirring rate of 500 rpm and a stirring time of 7 s per cycle (5 s stirring, 2 s rest) were selected. The temperature was set to 80°C, and the incubation time was 25 minutes.

[0067] 3.2 Sample testing

[0068] The specific parameters for sample testing are as follows: injection volume, 5000 μL; injection rate, 125 μL / s; inlet temperature, 200°C; detector capture temperature, 80°C; acquisition time, 50 s; column chamber initial temperature, 50°C; detector temperature, 260°C. Data acquisition time, 110 s, acquisition cycle, 0.01 s. Carrier gas: hydrogen, flow rate 30 mL / min. Each batch of antler samples was measured once.

[0069] 3.3 Analysis of volatile compounds

[0070] The Kovats retention index of the sample peaks was used to characterize the odor components in the sample. The Kovats retention index of the peaks collected by the two columns was used to search for the corresponding chemical components in the ArochemBase database.

[0071] 4. Determination of amino acid composition

[0072] Protein is an important component of deer antler, and amino acids are the basic building blocks of protein. Therefore, amino acids are the most abundant organic nutrient in deer antler, and their type and content directly affect the quality of the protein. Deer antler contains more than 20 types of amino acids, with a total content exceeding 50%. Previous studies have found that glycine has the highest content, followed by proline. In addition, hydroxyproline is considered to be a specific amino acid for collagen. It does not participate in the synthesis of other proteins, but is converted by hydroxylase during collagen synthesis. Gong Ruize et al. regard hydroxyproline as one of the quality markers (Q-Markers) of deer antler. Deer antler also contains a variety of essential amino acids that cannot be synthesized by the human body, among which lysine and leucine are present in relatively high levels. Therefore, hydroxyproline, glycine, proline, lysine, and leucine were selected as detection indicators.

[0073] Method optimization was conducted based on the "Second Public Notice on the Guiding Principles for Amino Acid Analysis" (https: / / www.chp.org.cn / ), issued by the Chinese Pharmacopoeia Commission. The amino acid content in deer antler was primarily determined using the internal standard method. Standard curves were constructed using reference substances for hydroxyproline (Hyp), glycine (Gly), proline (Pro), lysine (Lys), and leucine (Leu) as the benchmarks, with peak area as the ordinate and concentration as the abscissa to calculate amino acid content.

[0074] 4.1 Preparation of mixed reference solution

[0075] Accurately weigh 5 kinds of amino acid standard powders and use 0.1 mol·L -1 Dissolve in dilute hydrochloric acid to make Hyp: 1.242 7mgmL -1 , Gly:2.844 2mg·mL -1 , Pro:2.000 6mg mL -1 、Leu:1.008 9mg mL -1 Lys: 0.814 5mg mL -1 ) mixed solution. Accurately pipette 300 μL of the mixed reference solution into a 5 mL volumetric flask, and add 1 mol·L -1 triethylamine acetonitrile solution and 0.1 mol·L -1 Prepare 0.5 mL of each PITC acetonitrile solution, mix thoroughly, and let stand at room temperature for 1 hour. Then, add 50% acetonitrile to the mark and shake thoroughly. Take 1 mL of the solution, add 1 mL of n-hexane, shake, and let stand for 10 minutes. Remove the lower layer and filter through a 0.22 μm microporous filter. The filtrate is the mixed reference stock solution. 4.2 Preparation of Test Solution and Blank Solution

[0076] Accurately weigh 0.1 g of antler powder and place it in an amino acid hydrolysis tube. Add 6 mol·L-1 7 ml of hydrochloric acid solution was placed in a 150 ° C oven for hydrolysis for 3 h, cooled, transferred to an evaporating dish, and washed with water several times. The washing liquid was combined into the evaporating dish and evaporated to dryness in a water bath. The residue was added with 0.1 mol·L -1 Dissolve in hydrochloric acid, transfer to a 10 mL volumetric flask, and add 0.1 mol·L -1 Add hydrochloric acid to the mark and shake well. Accurately pipette 300 μL of the above solution and follow the procedure in “2.2.1” to prepare the test solution.

[0077] 4.3 Chromatographic conditions

[0078] Chromatographic column: Merck Star LP RP-18 endcapped (250 mm × 4.6 mm, 5 μm); 0.1 mol·L -1 Sodium acetate (adjusted to pH 6.5 with acetic acid) was used as mobile phase A, and acetonitrile was used as mobile phase B. Gradient elution was performed (0-10 min, 95% A; 10-24 min: 95% → 92% A; 24-35 min, 92% → 85% A; 35-40 min, 85% → 80% A; 40-50 min, 80% → 75% A; 50-65 min, 75% → 70% A); detection wavelength was 254 nm; column temperature was 35°C; flow rate was 1 mol / min; injection volume was 2 μL. The chromatograms of the mixed reference solution and test solution are shown in the figure below. Figure 1 A. Figure 1 B.

[0079] 4.4 Validation of the assay method

[0080] 4.4.1 Investigation of linear relationships

[0081] Accurately measure the mixed reference substance stock solution under item "4.1" and dilute it with 0.1 mol / L hydrochloric acid to prepare reference substance solutions of different concentrations. Inject the samples separately and record the peak area of ​​each index component. Draw the standard curve with the chromatographic peak area as the ordinate (Y) and the mass concentration of each reference substance as the abscissa (X). Perform linear regression and obtain the regression equation: hydroxyproline: Y = 3 910X - 0.352 1, r = 0.999 9, linear range: 4.66-74.56 μg / mL; glycine: Y = 6 289.1X - 1.664 6, r = 0.999 7, linear range: 10.67-170.65 μg / mL; proline: Y = 4 549.8X - 0.464 6, r = 0.999 4, the linear range was 7.50-120.04 μg / mL; leucine: Y=4002.5X+2.477 1, r=0.999 5, the linear range was 3.78-60.54 μg / mL; lysine: Y=5869X-1.364 6, r=0.999 8, the linear range was 3.05-48.87 μg / mL.

[0082] 4.4.2 Precision test

[0083] Accurately weigh 0.1 g of antler antler powder (S12) and prepare the test solution according to the method under "4.2". Inject 2 μL of the sample six times continuously and record the peak area. The RSDs of the peak areas of hydroxyproline, glycine, proline, leucine, and lysine were calculated to be 0.37%, 0.93%, 0.34%, 0.53%, and 1.18%, respectively.

[0084] 4.4.3 Repeatability test

[0085] Accurately weigh 0.1 g of velvet antler powder (S12) and prepare six parallel sample solutions according to the method in "4.2." Record the peak areas and calculate the relative standard deviations (RSDs) of the components hydroxyproline, glycine, proline, leucine, and lysine. The RSDs for the mass fractions of hydroxyproline, glycine, proline, leucine, and lysine were 0.71%, 0.59%, 0.64%, 0.42%, and 0.58%, respectively.

[0086] 4.4.4 Stability test

[0087] Accurately weigh 0.1 g of pilose antler slice powder (S12) and prepare the test solution according to the method under "4.2". Samples were injected and analyzed at 0, 2, 4, 8, 12, and 24 h, and the peak areas of the five components were recorded. The RSDs of the peak areas of hydroxyproline, glycine, proline, leucine, and lysine were 0.63%, 2.36%, 0.78%, 1.96%, and 3.31%, respectively.

[0088] 4.4.5 Sample recovery test

[0089] Accurately weigh 0.05 g of velvet antler powder (SVA-1) in 6 portions. Add a reference substance equal to the concentration of each component to each portion of the test sample. Prepare the test solution according to the method in "4.2". Inject the sample, record the peak area, calculate the content of the five amino acids in each sample, and calculate the recovery rate. The average recovery rate of the five amino acid components was 95.63% to 100.56%, with an RSD of 0.74% to 2.65%, which meets the requirements and demonstrates the good accuracy and reliability of the determination results of this method.

[0090] 5. Analysis of Random Forest Classification Algorithm Based on Whale Optimization Algorithm (WOA)

[0091] The Random Forest Algorithm (RF), a popular and powerful machine learning algorithm, belongs to the bagging category of ensemble learning methods. It improves model accuracy and robustness by building multiple decision trees and aggregating their predictions. However, RF also faces challenges, including difficult-to-interpret predictions, high computational overhead, the risk of overfitting, and high memory consumption. Therefore, optimization is necessary to improve performance. The Whale Optimization Algorithm (WOA) is an innovative swarm intelligence optimization method that mimics the cooperative and competitive behavior of whales to find optimal solutions. Across approximately 30 test functions and six structural design problems, WOA demonstrated its outstanding competitiveness and uniqueness, particularly its low parameter count and computational simplicity. The key to optimizing random forests with WOA lies in its ability to enhance model prediction accuracy and reduce the risk of overfitting by optimizing feature selection and hyperparameter settings. Furthermore, WOA's global search capability improves computational efficiency and resource utilization, and it also automates hyperparameter tuning, reducing the tedious manual adjustment process. In the combination of WOA and random forest (WOA-RF), WOA is used to optimize the decision tree structure and feature selection of random forest. The specific process includes initializing random forest and WOA parameters, optimizing feature selection and decision tree structure, and performing iterative optimization. This systematic combination fully demonstrates the global optimization capability and applicability of WOA in optimizing random forest. In this study, the appearance characteristics, odor components, and amino acid content of quantified antler samples were processed through multivariate statistical analysis and used as the input layer to perform the classification task. The final output involves four nominal output variables: 1 represents sika deer antler, 2 represents red deer antler, 3 represents reindeer antler, and 4 represents moose antler.

[0092] 6. Data Analysis

[0093] Data are presented as mean ± SD and analyzed using orthogonal partial least squares discriminant analysis (OPLS-DA) and variable importance projection (VIP) analysis in SIMCA 14.1. Data visualization was performed using GraphPad Prism (Version 9, GraphPad Software, San Diego, CA, USA) and Origin (Version 2021, OriginLab Corporation).

[0094] 7. Results and Discussion

[0095] 7.1 Quantitative Analysis of Characteristics of Deer Antler Samples from Different Sources Based on Computer Vision

[0096] The color of food is an important indicator of its sensory properties and a key factor in measuring its quality. This paper uses the basic three primary colors RGB value vector summation method, combined with hue (H) = arctan (B / G), saturation (S) = (1-(3*min(R,G,B)) / (R+G+B))*100 and lightness (V) = (R+G+B) / 3 to construct a radar chart of the chromaticity value of deer antler slices ( Figure 2 B). Studies have shown that different varieties of deer antler exhibit significant differences in the chromaticity parameters R, G, B, H, S, and V. Among them, the order of saturation (H) is RVA>WVA>MVA>SVA, with a difference of 92.00 between the maximum and minimum values; the order of S is SVA>WVA>RVA>MVA, with a difference of 8.31; and the order of V is SVA>RVA>MVA>WVA, with a difference of 7.63. Specifically for color values, the order and difference of R, G, and B are: R value SVA>RVA>MVA>WVA (difference 8.50), G value SVA>RVA>MVA>WVA (difference 6.30), and B value RVA>SVA>WVA>MVA (difference 2.90). The differences in saturation (H) among the four varieties are particularly obvious, indicating that it is a key factor affecting the color differences of deer antler slices. Sika deer antlers exhibit high brightness and warm tones; red deer antlers exhibit lower saturation and brightness; reindeer antlers display low saturation and neutral tones; and moose antlers exhibit neutral tones. These color characteristics are not only a visual reflection but may also be the result of the dynamic adaptation of organisms within the ecosystem.

[0097] The texture of cross-sectional deer antler slices is one of their most important characteristics. Local Binary Pattern (LBP), a commonly used texture analysis method that calculates the frequency of patterns and textures in an image, was used to characterize its texture features. LBP is widely used in computer vision and medical imaging due to its rotational and grayscale invariance (Lan, Liao, Fan, Hu, & Pan, 2023; Qiao et al., 2022). This study used high-resolution deer antler samples as the primary data. Using a Python platform, sliding window calculations were performed to extract LBP texture features from DVA samples. A 3×3 window with a step size of 1 was used to ensure accurate and reliable feature extraction. The LBP operator, acting as a local feature descriptor, was used to generate LBP images that describe the local texture features of the image. These images revealed the texture distribution of different deer antler species, showing significant regional consistency. Figure 2 Figure A shows the LBP (local binary pattern) texture features of four different varieties of antlers (SVA, WVA, RVA, MVA). The antler slices all have irregular circular edges and are attached with cortex. The cross-sections are densely covered with tiny honeycomb pores, which make them contrast sharply with the background, thereby improving the recognition accuracy. Specifically, the texture characteristics of the SVA variety are characterized by relatively uniform honeycomb eyes and a circular shape; WVA is unique, with larger central honeycomb eyes and smaller and more delicate surrounding honeycomb eyes, which may be related to the growth shape of the antler. RVA is similar to WVA, with the central honeycomb eyes significantly larger than the edge parts; while the texture of MVA appears rougher and the honeycomb eyes are larger. In the study, special attention was paid to the consistency of the acquisition environment to ensure the consistency of the background pattern. In order to comprehensively extract the image texture features, we performed histogram statistical analysis on the LBP image, extracted 256 local texture features from the images of the four antler varieties, and constructed an LBP local texture heat map (see Figure 2 C) to demonstrate the overall texture characteristics of antlers, thus laying a solid foundation for in-depth texture analysis.

[0098] OPLS-DA is a supervised discriminant analysis method that combines PLS-DA and orthogonal signal filtering technology. It can separate information that is irrelevant to the pre-defined and classification from the original matrix to the greatest extent possible, thereby concentrating the most relevant factors on the first principal component and then finding the orthogonal correction axis direction of the principal component. This improves the separation of samples between groups, weakens the differences within the groups, and maximizes the differences between groups. It is suitable for separating samples (Yan et al., 2024). OPLS-DA (orthogonal partial least squares discriminant analysis) was performed on the color and texture characteristics of deer antler slices. Figure 2 D and Figure 2E. The analysis results show that the antler samples are not clearly distributed across the different regions shown in the figure, and there is a lack of clear separation between the regions. This indicates that external traits alone cannot effectively distinguish the antler samples of these four varieties. To achieve more accurate classification, a comprehensive analysis combining data from multiple dimensions is necessary.

[0099] 7.2 Odor Difference Analysis of Deer Antler Samples from Different Sources Based on UF-GC-E Nose

[0100] The Heracles Neo ultrafast gas chromatography electronic nose has two chromatographic columns with different polarities. With one injection, the two columns are analyzed simultaneously, and the spectrum shows the separation results of the two columns at the same time. After deducting the blank reference, the gas chromatographic overlay of different varieties of deer antler is shown in Figure 3 A. Comparing the odor information obtained using two chromatographic columns of varying polarity revealed that a low-polarity column was more effective in separating the odor components in sika deer antler, so the peak area analysis using the MXT-5 column was selected. Component peaks with peak areas greater than 5000 were screened, and their molecular formulas were identified using their CAS numbers. The components were then identified based on relevant literature and compound properties. Compounds present in at least half of the antler batches were identified as the primary odor components.

[0101] A total of 19 odor components were detected in 4 different varieties of deer antlers, including 10 aldehydes, 3 ketones, 2 alcohols, 2 hydrocarbons, 1 furan, and 1 sulfur compound, all of which showed varying degrees of fishy odor ( Figure 3B) Fifteen odor components were detected in sika deer antler, with 3-methyl-2-butenal, 2,3-octanedione, and limonene being unique to sika deer antler. Elk antler contained 16 odor components, of which isopentanol, 1,3-dioctene, and n-octanal were detected only in elk antler. Ten odor components were detected in reindeer antler, with methyl mercaptan, acetaldehyde, acetone, isobutyraldehyde, and isovaleraldehyde being the least abundant among the four antler species. Only seven odor components were detected in elk antler, with n-butyraldehyde and nonanal being detected in the other three species but not in elk antler. Seven odor components were common to all four antler species, including methyl mercaptan, acetaldehyde, acetone, isobutyraldehyde, isovaleraldehyde, 2-methylbutyraldehyde, and n-hexanal. Among them, sulfur-containing compounds often have the smell of rotten cabbage. Aldehyde compounds occupy an important position among odor substances due to their small relative molecular mass, high volatility, and low olfactory threshold, and have a pungent odor. Among the five aldehydes, acetaldehyde is relatively high in moose, while the other four are predominant in sika deer. Ketone compounds may be derived from fatty acid oxidation and amino acid degradation, and have a fatty or burnt smell, which can enhance the fishy smell. Their threshold is high, but their content is high in all four species, making a certain contribution to the overall odor. In addition, for the cutting of antler slices, the dried antler is generally softened before cutting. Different vendors use different reagents for softening, some use water directly, while others use white wine. MVA is softened with white wine, so the ethanol content in the detected MVA odor components is very high. Then, to make the differences in the odor components of the four antler varieties more intuitive, we conducted an OPLS-DA analysis. In the scatter plot, the four antler varieties are clustered together, indicating that the established OPLS-DA model can achieve the distinction between raw antler varieties ( Figure 3 C).

[0102] Overall, the four antler varieties exhibit significant differences in odor characteristics, which are closely related to their biological characteristics, husbandry methods, and environmental factors. Understanding and controlling odor components will help improve the market competitiveness and consumer acceptance of antler velvet, providing theoretical support for antler product development. Furthermore, based on these studies, further investigations into the impact of odor components on pharmacological effects can be conducted to promote the widespread application and further development of antler velvet.

[0103] 7.3 Analysis of amino acid content in antler samples from different sources based on HPLC

[0104] The contents of five bioactive components in DVA were accurately determined using HPLC. The method validation results showed that the established HPLC analysis conditions met the analytical requirements and could accurately determine the contents of the five components. Representative chromatograms of DVA samples and mixed standards are shown in Figure 2. Figure 4As shown in A. Based on the content determination results (Table 2), we analyzed the differences in amino acid composition of antlers of different deer species. Amino acids are the basic building blocks of protein, and their content is closely related to the quality of antlers. Antler slices of different deer species contain these five amino acids, but the content varies. For example, the content of hydroxyproline is the highest in sika deer antler (SVA) and the lowest in red deer antler (WVA). Figure 4 B); for glycine, the content of moose velvet antler (MVA) is the highest, and that of red deer velvet antler (WVA) is the lowest ( Figure 4 C); the highest proline content is in sika deer antler (SVA), and the lowest is in red deer antler (WVA) ( Figure 4 D); the highest leucine content is in red deer velvet antler (WVA), and the lowest is in moose velvet antler (MVA) ( Figure 4 E); the highest lysine content is also in red deer antler (WVA), and the lowest is in moose antler (MVA) ( Figure 4 F). It can be seen that there are certain differences in the amino acid content of different varieties of deer antlers. In order to express the differences between sample groups more intuitively, OPLS-DA analysis was continued. Figure 4 G), different varieties of antlers are clustered together, indicating that the established OPLS-DA model can distinguish different varieties of antlers. In addition, the loading diagram is usually used to find the parameters that contribute most to the grouping, see Figure 4 H. In the loading plot, the farther from the origin, the greater the contribution to the grouping. Based on the current results, the amino acids Hyp, Gly, and Leu are the primary variables contributing to the differences in antler component content between different species, and may also contribute to the quality differences in antler antler slices. This phenomenon may be related to the living environment, dietary habits, physiological needs, and health status of different deer species.

[0105] Table 2 Content determination results

[0106]

[0107]

[0108]

[0109] 7.4 Multivariate Statistical Analysis of Multidimensional Data Fusion

[0110] Previous studies have found that multivariate statistical analysis of amino acid content and odor components can well distinguish the four varieties of deer antler. However, multivariate statistical analysis of color and texture did not provide sufficient differentiation. Therefore, efforts were made to integrate multidimensional data using SIMCA14.1 software to establish a Principal Component Analysis (PCA) model to reduce data complexity. PCA is currently very popular for distinguishing agricultural products based on their variety and geographic origin. Principal component analysis (PCA) was performed on the quantitative fused data of color, texture, odor, and amino acids from four DVA samples to describe the overall variability between and within sample groups. Figure 5 Figure A visually demonstrates the results of this analysis. The initial principal component contributed 36%, while the subsequent principal components contributed 15.3%. Overall, the cumulative contribution of these components was 51.3%. This observation suggests that these two components effectively capture most of the information in the sample. In addition, the DVA samples were divided into four discrete regions, consistent with their respective sources. It is worth noting that there are significant differences between samples from different regions, indicating that there are large differences in the comprehensive data of the DVA samples. In addition, an OPLS-DA model was established, and in the scatter plot, different varieties of antlers are clustered together, indicating that the established OPLS-DA model can achieve the distinction between different varieties of antlers. In order to prevent the model from overfitting, the model was verified by permutation test (200 times), and the results are shown in Figure 3. Figure 5 B. It is generally believed that R 2 The intercept of the fitting line on the Y coordinate axis is less than 0.3, indicating that the model is relatively reliable. 2 The intercept of the fitting line on the Y axis is less than 0.05, indicating that the model does not have overfitting. The results show that the model has a good fit and does not have overfitting. Figure 5 C), there are 162 components with VIP values ​​greater than 1. Among them, the amino acid components are Lys (VIP=1.1437), Leu (VIP=1.1914) and Hyp (VIP=1.0475); the color category is mainly contributed by the H value (VIP=1.0176); there are 27 texture components; the rest are contributed by odor components. The results show that the odor components of the four varieties of antlers are the most different, and the texture differences are the smallest. The reason may be that the cross-sections of the antlers of different varieties all show honeycomb-shaped pores, which may be due to the shortcomings of the existing texture extraction algorithm. Subsequent research can focus on optimizing related algorithms to find potential key differences.

[0111] 7.5 Constructing the WOA-RF classification algorithm to distinguish antler samples from different sources

[0112] During the training process of the WOA-RF classifier, 200 iterations are used for parameter optimization and feature selection. In each iteration, the WOA algorithm evaluates the quality of the feature subset based on the objective function and gradually adjusts the search strategy to select the optimal feature set. In addition, RF hyperparameters (such as the number of trees, maximum tree depth, and minimum number of sample splits) are also adjusted through WOA to ensure that the algorithm can achieve optimal performance on datasets of varying sizes. Compared with traditional single RF models, WOA-RF demonstrates greater robustness and adaptability when processing high-dimensional and noisy datasets.

[0113] First, in an experiment to distinguish DVA species, we used a random forest algorithm to classify our multidimensional data using 120 samples in a training to test ratio of 7:3. However, the recognition rate was only 53.333%. (See Table S3 for details.) WOA, a global optimization technique inspired by the hunting behavior of humpback whales, balances exploration and exploitation. It can be combined with RF to form a WOA-RF hybrid model, which enhances feature selection and predictive performance. WOA optimizes multidimensional data and is expected to improve the accuracy of DVA species classification.

[0114] Next, the study continued using the WOA-RF hybrid model. From the 120 batches of samples, 78 training samples and 30 test samples were randomly selected with a training to test ratio of 7.5:2.5. The training group consisted of 78 samples from four different sources, including 19 SVA, 23 WVA, 20 RVA, and 16 MVA. The test group consisted of 30 samples, including 11 SVA, 2 WVA, 10 RVA, and 7 MVA. After 200 optimization iterations, the WOA-RF model achieved 100% accuracy in both the training and test groups. This result demonstrates that the WOA-RF classification model can effectively distinguish between multiple varieties and exhibits high stability during the training process. The validation group also achieved 100% accuracy, further demonstrating the model's reliability and generalizability in practical applications. In the DVA source tracing experiment, the sample ratio of the training group and the test group was adjusted to 7:3, with a total of 90 training samples and 30 test samples. After 200 iterations of training, the recognition rate of the WOA-RF model still reached 100%. Figure 6(6A: Training set selection ratio; 6B: Training set results; 6C: Test set selection ratio; 6D: Test set results) These results demonstrate that WOA-RF provides stable performance in both variety differentiation and source tracing, even with incomplete or noisy data. Compared to traditional classification algorithms, WOA-RF not only improves classification accuracy but also effectively reduces model complexity and training time. WOA-RF demonstrates a particularly strong advantage in high-dimensional data and complex feature spaces.

[0115] 8. Conclusion

[0116] This study established a method for detecting the color, texture, aroma, and component content of different DVA varieties. The WOA-RF classification algorithm was then constructed using the emergent principle to fuse multidimensional feature data. The results demonstrated that the interaction of multidimensional data and the algorithm optimization process can produce a holistic effect that surpasses that of a single technique, enabling accurate identification of DVA varieties and demonstrating the value of the emergent principle in complex systems. Significant differences in appearance, shape, and internal composition were observed between different DVA varieties, indicating a strong correlation between variety and detection parameters. Furthermore, the results showed that the combination of WOA and RF optimized the classification capabilities of the random forest algorithm, achieving high recognition rates during both training and validation. Significant differences were observed in color, aroma, and amino acid content between different DVA varieties. Compared with traditional analytical methods, this method offers advantages such as faster detection speed and reduced sample requirements. However, it also has some drawbacks, including the difficulty in absolute quantification using the rapid gas phase electronic nose, the inability to determine inorganic flavor components, and the need for optimization of the LBP texture extraction algorithm. Overall, the WOA-RF classification model trained on multidimensional feature data is successful, easy to implement, computationally inexpensive, and highly accurate. Future research may consider adding more varieties of DVA samples and further exploring the correlation between appearance traits and internal components and pharmacological effects to promote the widespread application and in-depth development of DVA.

[0117] Furthermore, the established multi-dimensional traceability framework of "intrinsic components-image-odor-intelligent optimization" demonstrates potential for cross-disciplinary technology transfer, beyond its specific application in the DVA industry. By integrating multi-source, heterogeneous data, the framework can be extended to fields such as global biological resource regulation, intelligent quality inspection of the food chain, and precision agricultural management. Its core algorithm's robust adaptability to high-noise, small-sample scenarios provides a novel analytical paradigm for solving interdisciplinary challenges such as archaeological material identification and environmental pollutant analysis. More uniquely, this technical system reduces the reliance on biological samples for traditional testing through non-invasive detection, offering a technical alternative for applications such as field ecological monitoring. The key technology combination of "dynamic feature screening-intelligent parameter optimization" revealed in the study can be rapidly adapted to scenarios such as medical-assisted diagnosis and industrial quality control through transfer learning. To enhance the universality of the technology, it is recommended that a cross-industry multimodal database and a lightweight model deployment system be developed in the future, enabling the framework to evolve from a specific industry solution to a universal technical foundation for intelligent traceability. This research approach, moving from methodological innovation to paradigm shift, effectively balances technical expertise with scientific universality.

[0118] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying antler varieties based on multidimensional data fusion and artificial intelligence optimization algorithm, characterized in that: The steps include: (1) collecting multidimensional feature data of antler samples of different varieties, wherein the feature data includes color, texture, smell and chemical composition information; (2) Computer vision technology was used to extract color and texture features, ultra-fast gas phase electronic nose technology was used to extract odor features, and high performance liquid chromatography technology was used to extract chemical composition information such as amino acids; (3) Perform multivariate statistical analysis on the characteristic data to screen out key discriminant factors with variable importance projection value VIP greater than 1 and P value less than 0.05; (4) Using the key feature factors as input variables, a Whale Optimization Algorithm (WOA) and a Random Forest (RF) fusion WOA-RF classification model are constructed; (5) The WOA-RF classification model is used to automatically classify and identify the varieties of the antler samples to be tested, thereby achieving efficient traceability and identification of antler varieties.

2. The method according to claim 1, characterized in that The antler samples include four varieties: sika deer antler, red deer antler, reindeer antler and moose antler.

3. The method according to claim 1, characterized in that The color features include RGB, namely Red, Green, Blue; HSV is Hue, Saturation, Value, namely hue H, saturation S and lightness V in the color space parameters; the texture features are extracted by Local Binary Pattern, namely LBP algorithm.

4. The method according to claim 1, wherein The odor characteristics were extracted using a Heracles Neo ultra-fast gas phase electronic nose and analyzed using a low-polarity chromatographic column MXT-5 and a polar chromatographic column MXT-WAX.

5. The method according to claim 1, wherein The chemical components are amino acid components, including hydroxyproline, glycine, proline, leucine and lysine.

6. The method according to claim 1, characterized in that The WOA-RF classification model optimizes the feature subset and hyperparameters of RF through the WOA algorithm, and the model iteration number is no less than 200 times.

7. The method according to claim 1, characterized in that The classification accuracy of the model reached 100% on both the training set and the test set.

8. The method according to claim 1, characterized in that The method also includes inputting the multidimensional feature data into multivariate analysis software such as SIMCA to perform principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA) to assist in modeling and visualization.

9. The method according to claim 2, characterized in that Among the four species of antlers, namely sika deer antler, red deer antler, reindeer antler and moose antler, 15 odor components were detected in sika deer antler, of which 3-methyl-2-butenal, 2,3-octanedione and limonene were only found in sika deer antler; there were 16 odor components in red deer antler, of which isopentanol, 1,3-dioctene and n-octanal were only detected in red deer antler; there were 10 odor components detected in reindeer antler, of which methyl mercaptan, acetaldehyde, acetone, isobutyraldehyde and isovaleraldehyde accounted for the least among the four antler species; there were 7 odor components detected in moose antler, of which n-butyraldehyde and nonanal were detected in the other three species but not in moose antler. There were a total of 7 odor components contained in all four antler species, namely methyl mercaptan, acetaldehyde, acetone, isobutyraldehyde, isovaleraldehyde, 2-methylbutyraldehyde and n-hexanal.

10. The method according to claim 2, characterized in that Comparing the content of hydroxyproline among the four species of sika deer antler, red deer antler, reindeer antler and moose antler, the sika deer antler has the highest content and the red deer antler has the lowest content; for glycine, the moose antler has the highest content and the red deer antler has the lowest content; the sika deer antler has the highest proline content and the red deer antler has the lowest; the red deer antler has the highest leucine content and the moose antler has the lowest content; and the red deer antler has the highest lysine content and the moose antler has the lowest content.

Images