Multi-model fusion wheat berry mildew screening method and device and storage medium

Abstract

The invention discloses a multi-model fusion wheat berry mildew screening method, equipment and a storage medium, relates to the technical field of food safety and detection, and solves the problem of how to quickly and accurately detect aflatoxin in wheat berries. A fluorescence signal in a wheat berry sample is excited through a laser, and laser is introduced into the wheat berry sample through an optical fiber; a fluorescence signal generated after excitation is transmitted to a spectrograph to collect fluorescence spectrum data in real time; preprocessing the fluorescence spectrum data through a combined strategy of an SG smoothing method and a standard normal variable transformation method SNV to eliminate noise and light scattering interference; and inputting the preprocessed fluorescence spectrum data into a plurality of trained machine learning models for prediction, fusing prediction results of the plurality of models through an improved weighted average method, and outputting the aflatoxin pollution level of the sample to be analyzed according to a multi-model fusion result. The detection time of the aflatoxin in the wheat berry is greatly shortened, and the detection efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of food safety and testing technology, specifically to a multi-model fusion method, equipment, and storage medium for screening mold growth in wheat kernels. Background Technology

[0002] Wheat kernels, as an important food crop, are widely used in food processing, feed production, and other fields. However, during storage and transportation, wheat kernels are prone to mold growth and contamination with fungal toxins due to changes in environmental conditions. Aflatoxin B1 (AFB1) is a common fungal toxin produced by mold growth in agricultural crops; it is the most toxic and carcinogenic of all fungal toxins. AFB1 is mainly produced by fungi such as Aspergillus flavus and Aspergillus parasiticus and is widely found in grain crops and their products. These toxins pose a serious threat to human health, potentially causing liver damage and even inducing cancer. Therefore, rapid and accurate detection of mold and related toxins in wheat kernels is crucial for ensuring food safety.

[0003] As countries increasingly stringent limits for AFB1, the requirements for the accuracy and sensitivity of AFB1 detection technologies are becoming increasingly stringent. Traditional AFB1 detection methods include thin-layer chromatography (TLC), high-performance liquid chromatography (HPLC), and enzyme-linked immunosorbent assay (ELISA). TLC was the earliest method used for AFB1 detection, offering advantages such as low detection limits, high recovery rates, and low environmental requirements. However, it is susceptible to interference from background substances, leading to inaccurate results, and has poor experimental safety. HPLC offers high sensitivity, good specificity, and accurate and reliable results, but sample pretreatment is complex, and the equipment is expensive and complicated to operate, making it unsuitable for online detection. Compared to TLC and HPLC, ELISA is a more suitable method for AFB1 detection. While traditional instrumental analysis techniques are simple to operate, have fast reaction speeds, and low detection costs, the stability of the results is poor due to the numerous factors affecting enzyme activity, leading to a high likelihood of false positives. In summary, although traditional instrumental analysis techniques have achieved good sensitivity and reliability after years of development, the sample processing methods are relatively cumbersome, and they require expensive equipment and specialized personnel, making them unsuitable for rapid on-site detection.

[0004] With the continuous development of laser-induced fluorescence (LIF) technology, its application in food testing has gradually attracted attention. LIF technology has advantages such as high sensitivity, rapid response, and non-destructive nature, enabling fluorescence feature analysis of samples in a short time. However, the accuracy and reliability of LIF technology alone in detecting complex samples still need improvement. Summary of the Invention

[0005] The technical solution of this invention is used to solve the problem of how to quickly and accurately detect aflatoxin in wheat kernels.

[0006] The present invention solves the above-mentioned technical problems through the following technical means: This invention provides a multi-model fusion method for screening mold growth in wheat kernels, comprising the following steps: S1 Place the wheat kernel sample to be tested into the detection dark chamber, activate the laser to emit a laser beam, introduce the laser beam into the detection dark chamber through an optical fiber, use the laser to excite the fluorescence signal in the wheat kernel sample, transmit the generated fluorescence signal through the optical fiber to the filter, filter out the interference of the excitation light, and then send it into the spectrometer for real-time capture and analysis to generate a fluorescence spectrum dataset. S2 inputs the fluorescence spectrum dataset into the computer data processing unit, and preprocesses the fluorescence spectrum data using a combination of SG smoothing and standard normal variable transformation to eliminate noise and light scattering interference, and extracts feature parameters related to AFB1. S3 divides the preprocessed fluorescence spectrum dataset into training and testing sets according to the proportions. The training set is used to train the SVM, RF, and KNN models. The SVM, RF, and KNN models are then fused and tested. Each model outputs a prediction label. The prediction labels of each model are fused using a weighted average method. Based on the result of the multi-model fusion, the aflatoxin contamination level of the sample to be analyzed is output.

[0007] Furthermore, the method for preprocessing fluorescence spectral data using a combination of SG smoothing and standard normal variable transformation is as follows: The SG smoothing method is used to perform polynomial fitting on the data points within a sliding window to eliminate high-frequency noise in the fluorescence spectral data. The calculation formula is as follows:

[0008] in, For the data points after sliding, The average value of the sample spectral sequence. For the original data points, The radius of the sliding window, H is the smoothing coefficient, H is the normalization coefficient, and i is the counting factor; The standard normal variable transformation method is used to eliminate light scattering interference. The calculation formula is as follows:

[0009] in, These are the data points after processing using the standard normal variable transformation method. denoted as the original data points, and s as the standard deviation of the sample spectral sequence.

[0010] Furthermore, the method of fusing the predicted labels of each model using a weighted average method is as follows:

[0011]

[0012] in, To integrate the prediction results, The prediction results output by the SVM model. The prediction results output by the PF model. The prediction results output by the KNN model. The weights of the predicted labels output by the SVM model. The weights of the predicted labels output by the PF model. The weights of the predicted labels output by the KNN model. Let be the accuracy of the j-th model. This represents the recall rate of the j-th model, where j can be 1, 2, or 3; This is the sum of the products of model accuracy and model recall, where m can be 1, 2, or 3.

[0013] Furthermore, 5-fold cross-validation was used to perform a fusion test on the SVM model, RF model, and KNN model.

[0014] Furthermore, the characteristic parameters include: fluorescence peak position and intensity.

[0015] Furthermore, the wavelength of the laser is in the ultraviolet wavelength range, specifically the 365nm band.

[0016] The present invention also provides an apparatus comprising at least one processor and at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the above-described method by invoking the program instructions.

[0017] The present invention also provides a computer-readable storage medium storing computer instructions that cause the computer to perform the above-described method.

[0018] The advantages of this invention are: This invention preprocesses fluorescence spectral data using a combination of SG smoothing and Standard Normal Variable Transform (SNV) to eliminate noise and light scattering interference. By combining a multi-model fusion algorithm and leveraging the advantages of multiple machine learning models, it significantly improves the accuracy of identifying AFB1 contamination levels and reduces the false positive rate. It enables real-time online detection, greatly shortening the detection time for aflatoxin in wheat kernels and improving detection efficiency. Attached Figure Description

[0019] Figure 1 This is a structural diagram of the wheat kernel mold screening system of the present invention; Figure 2 This is a flowchart of the multi-model fusion wheat kernel mold screening method of the present invention; Figure 3 The fluorescence spectra of the present invention are shown under different AFB1 contamination levels. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 like Figure 1 The diagram shown is a structural diagram of a wheat kernel mold screening system, which includes: a laser, an optical fiber transmission system, a spectrometer, a data processing unit, filters, and a detection darkroom; The system works as follows: The wheat kernel sample to be tested is placed in the detection dark chamber, and the laser is activated to emit a laser beam. The laser beam is introduced into the detection dark chamber through an optical fiber. The laser excites the fluorescence signal in the wheat kernel sample. The generated fluorescence signal is transmitted to a filter through the optical fiber. The filter removes the interference of the excitation light and then sends it to the spectrometer for real-time capture and analysis to generate fluorescence spectrum data. Finally, the fluorescence spectrum data is input into the computer data processing unit for analysis.

[0022] The system in this embodiment uses laser-induced fluorescence technology without chemical pretreatment, and combined with a darkroom design to avoid ambient light interference, significantly improving detection efficiency. It overcomes the shortcomings of traditional biochemical methods, which are time-consuming and complex. The system also effectively filters background noise through filters, improving the signal-to-noise ratio.

[0023] like Figure 2 The diagram shows a flowchart of a multi-model fusion method for screening moldy wheat kernels based on a wheat kernel mold screening system, including the following steps: Step 1: Acquisition of fluorescence spectroscopy dataset.

[0024] The wheat kernel sample to be tested is placed in the detection dark chamber, and the laser is activated to emit a laser beam. The laser beam is introduced into the detection dark chamber through an optical fiber. The laser is used to excite the fluorescence signal in the wheat kernel sample. The generated fluorescence signal is transmitted to a filter through an optical fiber. The filter removes the interference of the excitation light and then sends it to a spectrometer for real-time capture and analysis to generate a fluorescence spectrum dataset.

[0025] The generation of fluorescence spectral data includes: collecting fluorescence data from 200 groups of wheat kernel samples, corresponding to samples with different AFB1 contamination levels; dividing the fluorescence data into 70% training and 30% testing components; using 5-fold cross-validation; and conducting multi-model training and fusion testing. A laser beam of a specific wavelength is generated using the laser described above to excite the fluorescence signal in the wheat kernel samples; the wavelength of the laser is in the ultraviolet range, specifically the 365nm band.

[0026] Step 2: Preprocessing and feature parameter extraction of fluorescence spectroscopy dataset.

[0027] Fluorescence spectral data were input into a computer data processing unit for analysis. The fluorescence spectral data were preprocessed using a combination of SG smoothing and standard normal variable transformation (SNV) to eliminate noise and light scattering interference. A spectrometer was used to extract characteristic parameters related to AFB1, including fluorescence peak position and intensity.

[0028] The method for preprocessing fluorescence spectral data using a combination of SG smoothing and standard normal variable transformation is as follows: The SG smoothing method is used to perform polynomial fitting on the data points within a sliding window to eliminate high-frequency noise in the fluorescence spectral data. The calculation formula is as follows:

[0029] in, For the data points after sliding, The average value of the sample spectral sequence. For the original data points, The radius of the sliding window, H is the smoothing coefficient, H is the normalization coefficient, and i is the counting factor; The standard normal variable transformation method is used to eliminate light scattering interference. The calculation formula is as follows:

[0030] in, These are the data points after processing using the standard normal variable transformation method. denoted as the original data points, and s as the standard deviation of the sample spectral sequence.

[0031] Step 3: Multi-model fusion prediction.

[0032] The preprocessed fluorescence spectral data is input into multiple trained machine learning models for prediction. The prediction results of multiple models are fused using an improved weighted average method. The aflatoxin contamination level of the sample to be analyzed is output based on the result of the multi-model fusion.

[0033] The step of inputting the preprocessed fluorescence spectral data into multiple trained machine learning models for prediction includes: The preprocessed fluorescence spectral data is input into multiple pre-trained models, each outputting a predicted label, i.e., the contamination level of the sample. These models include: SVM, RF, and KNN models; the output predicted label values ​​are... , where m is the type of the model.

[0034] The methods for determining pollution levels are shown in Table 1 below.

[0035] Table 1. Training data labels are manually graded based on AFB1 concentration levels:

[0036] The method of fusing the prediction results of multiple models by weighted averaging includes: The preprocessed spectral features of the sample to be tested are input into SVM, RF, and KNN models. Each model outputs a predicted label, and the predicted labels from each model are fused using a weighted average method. The formula for the weighted average method is as follows:

[0037]

[0038] in, To integrate the prediction results, The prediction results output by the SVM model. The prediction results output by the PF model. The prediction results output by the KNN model. The weights of the predicted labels output by the SVM model. The weights of the predicted labels output by the PF model. The weights of the predicted labels output by the KNN model. Let be the accuracy of the j-th model. This represents the recall rate of the j-th model, where j can be 1, 2, or 3; This is the sum of the products of model accuracy and model recall, where m can be 1, 2, or 3.

[0039] Experimental verification and result analysis Fluorescence data of 200 wheat kernel samples were collected using a laser-induced fluorescence system, corresponding to samples with different AFB1 contamination levels, such as... Figure 3 The figure shows a comparison of fluorescence spectra under different AFB1 contamination levels, where the AFB1 contamination level labels are known. The model was divided into 70% training and 30% testing, and 5-fold cross-validation was used for multi-model training and fusion testing. As shown in Table 2, the experimental data shows that the fusion model achieved an accuracy of 99.83% in the AFB1 contamination classification task and significantly improved the recall rate of low-concentration samples, outperforming the single model and verifying the effectiveness and practicality of the method of this invention.

[0040] Table 2: Performance Comparison of Different Models

[0041] To ensure the accuracy and usability of spectral data, this invention employs SG smoothing and SNV methods for preprocessing the original spectra. SG smoothing suppresses signal noise through multinomial fitting, while SNV normalization eliminates scattering differences between samples. The spectra are input into a model that integrates Support Vector Machine (SVM), Random Forest (RF), and K Nearest Neighbors (KNN). The SVM model uses a radial basis function kernel to construct the optimal hyperplane for classification, while the RF model uses multiple decision trees to discriminate samples. Each tree is trained using a subset of features and a subset of samples. The prediction results of the trees determine the final classification through a voting mechanism; the output is the probability or label of a sample belonging to a certain pollution level. K Nearest Neighbors (KNN) sets the number of neighbors and distance metrics, assigning weights based on the correlation between spectral features and AFB1 concentration. A KD tree is constructed from the preprocessed spectral data to accelerate nearest neighbor lookups, and weighted voting is performed on the pollution levels of the k nearest neighbors. After training, the model is saved for the analysis phase. In the analysis phase, the spectral data of unknown samples, after preprocessing, are input into multiple training models. Each model outputs its prediction results, which are then fused using a weighted average method to obtain the predicted contamination level of the sample. This invention combines the complementary advantages of multiple models, reducing the risk of misjudgment by a single model, and can quickly and accurately detect the contamination level of aflatoxin in wheat kernels.

[0042] Example 2 This embodiment also provides a processing device, including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the steps of the multi-model fusion wheat kernel mold screening method based on the wheat kernel mold screening system described in Embodiment 1 by calling the program instructions.

[0043] Example 3 This embodiment also provides a computer-readable storage medium storing computer instructions that cause the computer to execute the steps of the multi-model fusion wheat kernel mold screening method based on the wheat kernel mold screening system described in Embodiment 1.

[0044] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-model fusion method for screening mold growth in wheat kernels, characterized in that, Includes the following steps: S1 Place the wheat kernel sample to be tested into the detection dark chamber, activate the laser to emit a laser beam, introduce the laser beam into the detection dark chamber through an optical fiber, use the laser to excite the fluorescence signal in the wheat kernel sample, transmit the generated fluorescence signal through the optical fiber to the filter, filter out the interference of the excitation light, and then send it into the spectrometer for real-time capture and analysis to generate a fluorescence spectrum dataset. S2 inputs the fluorescence spectrum dataset into the computer data processing unit, and preprocesses the fluorescence spectrum data using a combination of SG smoothing and standard normal variable transformation to eliminate noise and light scattering interference, and extracts feature parameters related to AFB1. S3 divides the preprocessed fluorescence spectrum dataset into training and testing sets according to the proportions. The training set is used to train the SVM, RF, and KNN models. The SVM, RF, and KNN models are then fused and tested. Each model outputs a prediction label. The prediction labels of each model are fused using a weighted average method. Based on the result of the multi-model fusion, the aflatoxin contamination level of the sample to be analyzed is output.

2. The multi-model fusion method for screening moldy wheat kernels according to claim 1, characterized in that, The method for preprocessing fluorescence spectral data using a combination of SG smoothing and standard normal variable transformation is as follows: The SG smoothing method is used to perform polynomial fitting on the data points within a sliding window to eliminate high-frequency noise in the fluorescence spectral data. The calculation formula is as follows: in, For the data points after sliding, The average value of the sample spectral sequence. For the original data points, The radius of the sliding window, H is the smoothing coefficient, H is the normalization coefficient, and i is the counting factor; The standard normal variable transformation method is used to eliminate light scattering interference. The calculation formula is as follows: in, These are the data points after processing using the standard normal variable transformation method. denoted as the original data points, and s as the standard deviation of the sample spectral sequence.

3. The multi-model fusion method for screening moldy wheat kernels according to claim 1, characterized in that, The method for fusing the predicted labels of various models using a weighted average is as follows: in, To integrate the prediction results, The prediction results output by the SVM model. The prediction results output by the PF model. The prediction results output by the KNN model. The weights of the predicted labels output by the SVM model. The weights of the predicted labels output by the PF model. The weights for the predicted labels output by the KNN model. Let the accuracy of the j-th model be denoted as . This represents the recall rate of the j-th model, where j can be 1, 2, or 3; This is the sum of the products of model accuracy and model recall, where m can be 1, 2, or 3.

4. The multi-model fusion method for screening moldy wheat kernels according to claim 1, characterized in that, Five-fold cross-validation was used to perform fusion testing on the SVM model, RF model, and KNN model.

5. The multi-model fusion method for screening moldy wheat kernels according to claim 1, characterized in that, The characteristic parameters include: fluorescence peak position and intensity.

6. The multi-model fusion method for screening moldy wheat kernels according to claim 1, characterized in that, The laser has a wavelength in the ultraviolet range, specifically in the 365nm band.

7. A device, characterized in that, It includes at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the method as described in any one of claims 1 to 6 by invoking the program instructions.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause the computer to perform the method as described in any one of claims 1 to 6.

Images