Method for rapidly identifying adulterated substances in soybean meal and determining content of adulterated substances in soybean meal based on near infrared spectrum

Abstract

The invention discloses a near infrared spectrum-based soybean meal adulterated substance rapid identification and content determination method, which is characterized in that on the basis of near infrared spectrum data of substances such as soybean meal, an artificial intelligence algorithm is utilized to construct a classification model and a prediction model, and efficient detection of soybean meal adulteration is realized. In the single adulteration classification model, the detection success rate of the sand, the corn and the wheat is high, and the detection success rate is relatively low due to the fact that the cottonseed meal and the rapeseed meal are high in component similarity with the soybean meal. The mixed adulteration classification model adopts a plurality of algorithm models, and finally a neural network model is selected. The single adulteration prediction model and the mixed adulteration prediction model also achieve better effects, and meet the basic requirements of the prediction models. According to the method, rapid and nondestructive detection of soybean meal adulteration is realized, the efficiency of soybean meal quality detection can be improved, and a scientific basis can be provided for striking soybean meal adulteration behaviors.

Description

Technical Field

[0001] This invention relates to the field of feed quality testing technology, specifically to a rapid identification and content determination method for adulterants in soybean meal based on near-infrared spectroscopy. Background Technology

[0002] Traditional methods for detecting adulteration in soybean meal, such as combustion tests, sensory identification methods, and litmus paper methods, suffer from drawbacks such as being time-consuming, costly, damaging to samples, and having low detection accuracy. These methods fail to meet the demands of the modern feed industry for rapid, non-destructive, and efficient detection. Near-infrared spectroscopy (NIR), a green analytical technique integrating computer science, chemometrics, and informatics, can rapidly obtain information about the composition of substances by detecting the near-infrared light absorption characteristics of functional groups such as C, N, O, and S. It offers advantages such as being non-destructive, efficient, safe, and low-cost, demonstrating broad application prospects in the field of feed testing.

[0003] Currently, although there are studies on feed adulteration detection based on near-infrared spectroscopy technology both domestically and internationally, problems such as insufficient detection accuracy and poor model applicability still exist in complex adulteration scenarios (such as mixed adulteration of multiple substances or low-proportion adulteration). For example, cottonseed meal, rapeseed meal, and soybean meal have high component similarity and significant overlap in spectral characteristics, resulting in low success rates for single adulteration detection; when mixed adulteration occurs, the spectra of multiple substances interfere with each other, making it difficult to accurately identify the types and contents of adulterants. Therefore, there is an urgent need to develop a method that can rapidly identify and accurately determine the content of multiple adulterants in soybean meal. Summary of the Invention

[0004] The purpose of this invention is to overcome the defects of the prior art and provide a rapid identification and content determination method for adulterants in soybean meal based on near-infrared spectroscopy. By optimizing sample preparation, spectral preprocessing and model construction strategies, this invention achieves efficient identification and content determination of single and mixed adulterants such as sand, corn, wheat, cottonseed meal and rapeseed meal in soybean meal, meeting the needs of the feed industry for rapid, non-destructive and accurate detection.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A rapid identification and content determination method for adulterants in soybean meal based on near-infrared spectroscopy includes five core steps: sample preparation, spectral data acquisition, spectral data preprocessing, model construction, model validation, and detection, as detailed below: Step S1, Sample Preparation: Select 5 kg of soybean meal sample for initial screening to remove impurities. Then, use a high-speed grinder to pulverize the soybean meal and pass it through a 40-mesh sieve to ensure uniform particle size and reduce interference from particle size differences on spectral detection. Weigh the sieved material into 4 g portions to prepare a total of 1080 preliminary samples, divided into 6 groups. Five groups are adulterated samples, with sand, corn, wheat, cottonseed meal, and rapeseed meal added to each group respectively. Six adulteration ratios are set for each adulterant: 5%, 10%, 15%, 20%, 25%, and 30%. 30 samples are prepared for each ratio, resulting in 180 samples for each adulterant. One control group consists of 180 unadulterated soybean meal samples, serving as a blank control for model training and validation.

[0006] Step S2, spectral data acquisition: A Fourier transform near-infrared spectrometer is used as the detection instrument, in conjunction with SpectraSmart spectral data acquisition software. At room temperature, near-infrared spectral data are acquired for all samples prepared in step S1. The acquisition range covers the near-infrared bands corresponding to the characteristic functional groups of soybean meal and adulterants to ensure complete spectral information is obtained.

[0007] Step S3, spectral data preprocessing: Since the spectral acquisition process is easily affected by factors such as ambient temperature, humidity, instrument noise, and sample scattering, the raw spectral data needs to be preprocessed. Through visual observation combined with statistical analysis, abnormal spectral curves that deviate significantly from the trend of most spectral curves or have obvious abnormal peaks or valleys are removed to reduce the impact of abnormal data on the model. Smoothing (SG) is used to eliminate high-frequency noise, Standard Normal Variable Correction (SNV) is used to correct sample scattering differences, First Derivative Transform (FD) is used to enhance spectral resolution, and Second Derivative Transform (SD) is used to separate overlapping peaks. The optimal preprocessing combination is selected based on the characteristics of different adulterants. For example, SNV and second derivative are used for organic adulterants (corn, wheat, cottonseed meal, rapeseed meal), while multivariate scattering correction (MSC) and first derivative are used for inorganic adulterants (sand).

[0008] Step S4, Model Building Step S4.1: Classification Model Construction. The classification model is used to identify whether soybean meal is adulterated and the types of adulterants. It is divided into a single adulteration classification model and a mixed adulteration classification model. For the single adulteration classification model, the spectral data of each single adulterated sample after preprocessing are merged with those of the unadulterated sample. After PCA principal component analysis, 216 spectral data points are retained and divided into a training set (144 samples) and a validation set (72 samples) at a 2:1 ratio. Partial Least Squares Discriminant Analysis (PLS-DA) algorithm is used to construct the model. The category label for unadulterated soybean meal samples is set to 0, and the category label for adulterated soybean meal samples is set to 1. The threshold is adjusted to 0.4-0.6 according to the characteristics of each adulterant (e.g., 0.42 for the sand adulteration model and 0.54 for the corn adulteration model). For the mixed adulteration classification model, 1080 spectral data points from all single adulterated samples and unadulterated samples are mixed and divided into a training set and a test set at a 7:3 ratio. Three algorithms—random forest, XGBoost, and neural network—were used to construct basic models. GridSearchCV was used to iterate through parameter combinations and select the optimal parameters for each model. Then, VotingClassifier was used for soft voting ensemble learning (random forest weight 40%, neural network weight 40%, XGBoost weight 20%). The detection success rates of each model were compared, and the neural network model with a test set accuracy of 99.69% was selected as the final hybrid classification model.

[0009] Step S4.2, prediction model construction. The prediction model is used to determine the content of adulterants in soybean meal and is divided into a single adulteration prediction model and a mixed adulteration prediction model. For the single adulteration prediction model, PCA principal component analysis is performed on the spectral data of each single adulterated sample and the unadulterated sample after pretreatment. Feature bands are extracted and 210 spectral data points are retained. These are divided into a training set (140 samples) and a prediction set (70 samples) at a 2:1 ratio. After normalization, a partial least squares regression (PLSR) algorithm is used to construct the model, with each model containing 360 spectral data points. For the mixed adulteration prediction model, the spectral data of all single adulterated samples and the unadulterated sample are mixed. After PCA principal component analysis (retaining 210 spectral data points) and normalization, a PLSR algorithm is used to construct the model to determine the total adulteration amount in the mixed adulterated sample.

[0010] Step S5, Model Validation and Testing Classification model validation: Model performance was evaluated using AUC (Area Under the ROC Curve) and confusion matrix. A higher AUC value (closer to 1) and a larger number of true positives and true negatives in the confusion matrix indicate higher model accuracy. Success rates for single adulteration classification models: sand 93%, corn 89%, wheat 89%, cottonseed meal 78%, rapeseed meal 84%; success rate for mixed adulteration classification models (neural network) 99.69%. Predictive Model Validation: The correlation coefficient (R²) and root mean square error (RMSE) were used to evaluate model performance. The closer R² is to 1 and the closer RMSE is to 0, the higher the model's predictive accuracy. In the single adulteration prediction models, the adulterated cottonseed meal model had R² = 0.967 and RMSE = 1.557%, the adulterated sand model had R² = 0.933 and RMSE = 2.193%, and the mixed adulteration prediction model had R² = 0.903 and RMSE = 2.609%. All models met the prediction requirements.

[0011] For soybean meal samples to be tested, prepare the sample according to step S1 (crushing and sieving), collect near-infrared spectral data according to step S2, and perform preprocessing according to step S3. Then, input the pre-constructed classification model to identify whether the sample is adulterated and the type of adulterant. Then, input the corresponding prediction model to determine the content of adulterant substances, so as to achieve rapid, non-destructive and accurate detection of soybean meal adulteration. Beneficial effects

[0012] 1. This invention is based on near-infrared spectroscopy technology, which eliminates the need for complex sample pretreatment. The entire process of spectral acquisition and model analysis can be completed in a short time. Compared with traditional detection methods, it significantly shortens the detection time and meets the needs of rapid detection of batch samples in the feed industry. 2. The testing process does not require damage to the sample, allowing for sample reuse and reducing testing costs. It is especially suitable for testing scenarios involving precious samples or samples that need to be retained. 3. By optimizing spectral preprocessing methods (such as differential preprocessing), selecting appropriate algorithms (PLS-DA, PLSR, neural networks), and constructing an integrated model, the problems of spectral overlap and mixed adulteration interference of cottonseed meal, rapeseed meal, and soybean meal were effectively solved. The success rate of single adulteration identification reached 93%, the success rate of mixed adulteration identification reached 99.69%, and the prediction R² of adulteration content was greater than 0.9. 4. It can simultaneously identify single adulteration, mixed adulteration, and determine the content of five common adulterants: sand, corn, wheat, cottonseed meal, and rapeseed meal, meeting the needs of different testing scenarios; 5. The model construction and analysis are implemented using PyCharm software based on the Python language. The code is highly reusable, and operators can master it after simple training, making it easy to promote and apply in feed production enterprises and quality testing institutions.

[0013] 6. This invention provides a scientific and efficient technical means for detecting adulteration in soybean meal, which helps to improve the level of feed quality monitoring, combat adulteration, and protect the safety of livestock production and consumer health, and has important economic and social value. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2This is the spectrum of the present invention, which contains no adulterants. Figure 3 This is the original spectrum of the present invention after incorporation of sand; Figure 4 This is the confusion matrix after sand is incorporated into the present invention; Figure 5 This is the regression curve model of the present invention after the addition of sand; Figure 6 This is the regression curve model for all mixed adulterated samples of this invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to specific embodiments, but these embodiments do not limit the scope of protection of the present invention.

[0016] Example 1: Identification and content determination of a single adulterant (sand) Five kilograms of soybean meal were selected, pulverized, and sieved through a 40-mesh sieve to prepare 180 samples of adulterated sand (30 samples each of 5%, 10%, 15%, 20%, 25%, and 30%) and 180 unadulterated samples. Fourier transform near-infrared spectroscopy was used with SpectraSmart software to acquire spectral data for all samples. Abnormal spectra were removed, and preprocessing was performed using MSC+first derivative. A classification model was constructed using the PLS-DA algorithm with a threshold of 0.42. After PCA, 210 experimental samples were retained, and the training set (140 samples) and prediction set (70 samples) were split in a 2:1 ratio. A prediction model was constructed using the PLSR algorithm. The classification model had an AUC of 0.96 and a detection success rate of 93%; the prediction model had R² = 0.933 and RMSE = 2.193%. The preprocessed soybean meal samples containing adulterated sand were input into the model, successfully identifying and measuring the adulteration.

[0017] Example 2: Identification and content determination of a single adulterant (cottonseed meal) This embodiment demonstrates the effectiveness of the method of the present invention for cottonseed meal, which has a similar composition to soybean meal and exhibits significant spectral overlap. Similar to Example 1, 180 adulterated cottonseed meal samples (30 samples each of 5%, 10%, 15%, 20%, 25%, and 30%) and 30 unadulterated samples were prepared. After acquiring the spectra, a preprocessing combination of SNV and second derivative was used, which effectively enhances the spectral features related to chemical composition and separates overlapping peaks. The classification model was constructed using the PLS-DA algorithm. The final model AUC value was 0.92, and the detection success rate reached 78%. The prediction model was constructed using the PLSR algorithm. The model performed excellently, with R²=0.967 and RMSE=1.557%. Even for cottonseed meal, which is the most difficult to identify, this method still achieved high prediction accuracy, proving the effectiveness of the differentiated preprocessing strategy.

[0018] Example 3: Identification and Total Content Determination of Adulterants This embodiment simulates a more complex and realistic adulteration scenario. Five substances—sand, corn, wheat, cottonseed meal, and rapeseed meal—were mixed in different proportions and then adulterated with soybean meal, resulting in 900 mixed adulterated samples. 180 unadulterated samples were retained as a control. After collecting spectra, preprocessing was performed using a combination of methods, including smoothing, SNV, and derivative transformation. A neural network model was used for classification. Tested on 1080 mixed datasets, the model exhibited extremely high robustness, achieving a test set accuracy of 99.69%. The prediction model was constructed using the PLSR algorithm to predict the total adulterant content in the samples. The model achieved good results, with R²=0.903 and RMSE=2.609%. This invention can effectively handle complex mixed adulteration scenarios, successfully identifying adulterants while accurately estimating the total adulteration amount.

Claims

1. A method for rapid identification and determination of adulterants in soybean meal based on near infrared spectroscopy, characterized in that, Comprising the following steps: Step S1, sample preparation, select soybean meal samples for preliminary screening, crush and pass through a 40 mesh sieve, divide 4g of undersize into each portion to obtain preliminary samples; add sand, corn, wheat, cottonseed meal, rapeseed meal to part of the preliminary samples respectively, set 5%, 10%, 15%, 20%, 25%, and 30% six adulteration proportions for each adulterant, prepare 30 adulterated samples for each proportion, and reserve 180 non-adulterated soybean meal samples as controls; Step S2, spectral data acquisition, use a Fourier transform near-infrared spectrometer, and cooperate with SpectraSmart spectral data acquisition software to collect the near-infrared spectral data of all non-adulterated samples and adulterated samples in step S1 respectively; Step S3, spectral data preprocessing, at least one of smoothing, standard normal variate correction (SNV), first derivative transformation (FD), and second derivative transformation (SD) is used to preprocess the collected spectral data to eliminate spectral errors caused by environmental, instrumental, and sample self interference factors; Step S4, model construction; Step S4.1, classification model construction, divide the preprocessed spectral data into a training set and a validation set according to a 2:1 ratio, use partial least squares discriminant analysis (PLS-DA) algorithm to construct a single adulterant classification model, set the non-adulterated soybean meal sample category identifier as 0 and the adulterated soybean meal sample category identifier as 1, and set the threshold value as 0.4-0.6; for mixed adulterated samples, mix the spectral data of all adulterated samples and non-adulterated samples, divide the training set and the test set according to a 7:3 ratio, use random forest, XGBoost, and neural network algorithms to construct models respectively, optimize the parameters through GridSearchCV, and then use VotingClassifier for soft voting ensemble learning to select the neural network model with the highest detection success rate as the mixed adulteration classification model; Step S4.2, prediction model construction, perform principal component analysis (PCA) on the preprocessed spectral data to extract characteristic bands, divide the training set and the prediction set according to a 2:1 ratio and perform normalization processing, and use partial least squares regression (PLSR) algorithm to construct single adulterant prediction models and mixed adulterant prediction models respectively; Step S5, model validation and detection, use AUC value and confusion matrix to validate the classification model, use correlation coefficient (R²) and root mean square error (RMSE) to validate the prediction model; input the near-infrared spectral data of the soybean meal samples to be detected after preprocessing in step S3 into the constructed classification model to identify whether they are adulterated and the type of adulterant, and input the prediction model to determine the content of the adulterant.

2. The method of claim 1, wherein, In step S1, 5kg of soybean meal is selected for preliminary screening, a high-speed pulverizer is used for crushing to ensure uniform sample particle size.

3. The method of claim 1, wherein, In step S3, the spectral data preprocessing also includes the elimination of abnormal spectral curves, which are curves with large deviation from the trend of most spectral curves, obvious abnormal peaks or valleys.

4. The method of claim 1, wherein, In step S4.1, 210 experimental samples were processed for single adulterant classification model construction, and 1080 experimental samples were processed for mixed adulterant classification model construction. The test set accuracy of the neural network model reached 99.69%.

5. The method of claim 1, wherein, In step S4.2, 210 experimental samples were processed for prediction model construction, including 140 experimental samples for training set and 70 experimental samples for validation set. In the single adulterant prediction model, the R² of the adulterated cottonseed oil model was 0.967, and the RMSE value was 1.557%. The R² of the adulterated sand model was 0.933, and the RMSE value was 2.193%. The R² of the mixed adulterant prediction model was 0.903, and the RMSE value was 2.609%.

6. The method of claim 1, wherein, In step S4, PyCharm2021 software was used for model construction, and Python language was used to write code to realize PLS-DA, PLSR, random forest, XGBoost, neural network algorithm, and model training and verification.

7. The method of claim 1, wherein, In step S4.1, the ensemble model of the mixed adulterant classification model was composed of random forest model (weight 40%), neural network model (weight 40%), and XGBoost model (weight 20%). The test set accuracy of the ensemble model was 99.38%.

8. The method of claim 1, wherein, In step S4.2, SNV and second derivative preprocessing were used for organic adulterant substances (corn, wheat, cottonseed oil, and rapeseed meal), and multiple scattering correction (MSC) and first derivative preprocessing were used for inorganic adulterant substances (sand).

9. The method of claim 1, wherein, In step S5, the detection success rate of the classification model for identifying single adulterant substances was 93% for sand, 89% for corn, 89% for wheat, 78% for cottonseed oil, and 84% for rapeseed meal, which met the demand for rapid detection of soybean meal adulteration.

Images