Method for identifying true and false anoectochilus formosanus based on near infrared spectrum

Abstract

The invention discloses a method for identifying true and false anoectochilus formosanus based on near infrared spectroscopy, which comprises the following steps: firstly, collecting samples of true and false anoectochilus formosanus, making an anoectochilus formosanus spectral data set, and dividing the anoectochilus formosanus spectral data set into a training set and a test set; and secondly, constructing a convolutional neural network model based on a one-dimensional convolutional network 1D-CNN, inputting the spectral data of the training set into the convolutional neural network model for training, predicting the anoectochilus formosanus sample, and outputting an anoectochilus formosanus authenticity identification result. According to the method, the rapid, accurate and nondestructive authenticity identification of the anoectochilus formosanus sample is realized through the spectral data.

Description

Technical Field

[0001] The present invention belongs to the fields of visible-near infrared spectroscopy technology and deep learning technology, and specifically relates to a method for identifying the authenticity of golden thread vine based on near infrared spectroscopy. Background Art

[0002] Anoectochilus roxburghii contains polysaccharides, flavonoids and steroids. Due to its excellent medicinal value, it has been widely used in medicine, health care and food in recent years. With the increase in demand for Anoectochilus roxburghii and the shortage of market supply, counterfeits are becoming increasingly prevalent in the market, such as Goodyera schlechtendaliana and Ludisia discolor. Therefore, how to quickly, accurately and non-destructively identify the authenticity of Anoectochilus roxburghii samples has become an urgent problem to be solved. At present, there are few studies on methods for identifying the authenticity of Anoectochilus roxburghii, and there are the following shortcomings: (1) Manual screening methods have high accuracy, but consume a lot of labor and have low efficiency, which is not conducive to the intelligent development of the traditional Chinese medicine industry; (2) There are methods for identifying the authenticity of Anoectochilus roxburghii based on chemical markers and DNA markers. These methods require expensive experimental equipment and require the experimenters to have certain prior knowledge. At the same time, the detection speed is slow, which is not conducive to large-scale promotion; (3) In terms of visible-near infrared spectroscopy technology, most of the current analysis is based on chemometrics and machine learning methods. This type of method requires preprocessing of the spectrum and extraction of characteristic bands, which requires a lot of prior knowledge and the modeling process is relatively complicated.

[0003] At present, some scholars have done some research on the identification of the authenticity of Anoectochilus roxburghii using visible-near infrared spectroscopy technology. Lin Yu et al. collected the near-infrared spectrum of Anoectochilus roxburghii from 1000nm to 2500nm, and used the first-order derivative method and the standard canonical variation method to pre-process the obtained spectrum, and finally used the PCA-Cluster clustering analysis method to identify the authenticity of Anoectochilus roxburghii samples. However, this method uses Anoectochilus roxburghii powder as the sample, which will damage the sample. Summary of the invention

[0004] In view of the shortcomings and defects in the prior art, the purpose of the present invention is to provide a method for identifying the authenticity of Anoectochilus roxburghii based on near infrared spectroscopy. The technical solution of the present invention is as follows:

[0005] A method for identifying the authenticity of golden thread lily based on visible-near infrared spectroscopy technology and deep learning technology, comprising the following steps:

[0006] Collect samples of authentic and counterfeit Anoectochilus roxburghii and create a Anoectochilus roxburghii dataset: Build a spectral data collection platform to collect spectral data of authentic and counterfeit Anoectochilus roxburghii in the Anoectochilus roxburghii dataset.

[0007] The spectral acquisition mode is transmission, and the band response range is 400nm-1170nm. The collected spectral data are then divided into training set and test set.

[0008] A convolutional neural network model based on the one-dimensional convolutional network 1D-CNN was constructed, and the spectral data of the training set was input into the convolutional neural network model for training. The golden thread vine samples were predicted and the authenticity identification results of the golden thread vine were output.

[0009] The trained 1D-CNN model was exported, and the test set was used to conduct classification experiments on authentic and fake samples of Anoectochilus roxburghii.

[0010] Furthermore, the preparation of the Anoectochilus roxburghii dataset is as follows:

[0011] For each roxburghii sample, ten areas were divided, and a transmission spectrum was collected from each area. A total of ten transmission spectra were collected for one roxburghii sample.

[0012] The ten spectra were averaged to obtain the average spectrum corresponding to the Anoectochilus roxburghii sample.

[0013] The average spectra corresponding to all the golden thread vine samples were obtained, randomly shuffled, and then divided into training sets and test sets in proportion. The data set division process was repeated ten times to obtain ten training sets and ten test sets.

[0014] Furthermore, the testing process is as follows:

[0015] The data of the training set is input into the established 1D-CNN model for training. After the training is completed, the weights of the model are exported, and then the data of the test set is predicted, and the accuracy, recall rate, precision and F1 score of the test set are calculated.

[0016] The model training and testing process is repeated ten times using different training sets and test sets, and the final performance index of the model is the average of these ten times.

[0017] Furthermore, the spectral data acquisition platform includes a spectrometer, an infrared quartz fiber probe, and a halogen lamp light source with a light source band range of 360nm-2500nm; the golden thread vine sample is placed statically on a bracket, the halogen lamp light source is placed 10 cm below the bracket, and the fiber optic probe is placed 3 cm above the bracket to collect transmission spectrum signals. The collected spectrum signals are refracted by the optical fiber to reach the spectrometer, and the spectral data acquisition process is completed.

[0018] Compared with the prior art, the present invention has the following effects:

[0019] The present invention builds a spectrum acquisition system for Anoectochilus roxburghii, uses a spectrometer to collect visible-near infrared spectra of Anoectochilus roxburghii samples, makes Anoectochilus roxburghii data sets, and then establishes a 1D-CNN-based deep learning model to analyze the spectrum. Through the spectral data, the rapid, accurate and non-destructive identification of the authenticity of Anoectochilus roxburghii samples is achieved. This method for distinguishing the authenticity of Anoectochilus roxburghii based on visible-near infrared spectroscopy technology and deep learning technology effectively overcomes the problem that manual sorting is greatly affected by subjective factors and consumes a lot of time and labor. Compared with other methods, the present invention overcomes the difficulty that the traditional method of detecting the authenticity of Anoectochilus roxburghii will damage the sample, and is more suitable for fast, accurate and non-destructive identification of Anoectochilus roxburghii in the actual production process. At the same time, the 1D-CNN deep learning model established by the present invention uses full-band modeling, does not need to perform complex and cumbersome pre-processing and feature band selection on the spectrum, and further improves the efficiency and universality of the authenticity identification of Anoectochilus roxburghii. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of the method for detecting the authenticity of Anoectochilus roxburghii based on visible-near infrared spectroscopy technology and deep learning technology proposed in the present invention;

[0021] Figure 2 This is a flow chart of step S3 of a specific embodiment of the present invention;

[0022] Figure 3 The spectral data acquisition device of Anoectochilus roxburghii designed in the present invention;

[0023] Figure 4 The spectra of the authentic and fake roxburghii roxburghii were collected in the present invention;

[0024] Figure 5 The accuracy of the convolutional neural network established by the present invention on the training set. DETAILED DESCRIPTION

[0025] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, without departing from the concept of the present invention, corresponding improvements and modifications may be made thereto. These all belong to the protection scope of the present invention.

[0026] like Figure 1 , the present invention further describes a method for detecting the authenticity of golden thread lotus based on visible-near infrared spectroscopy technology and deep learning technology:

[0027] S1: Anoectochilus samples without obvious external defects were purchased for data set preparation. All samples were provided by the Anoectochilus base of Jinhua Academy of Agricultural Sciences and authenticated by researchers from Jinhua Academy of Agricultural Sciences. The authentic Anoectochilus samples include three varieties: Hongxia, Andijiangkang and Zhejiang Jianye, and the fake Anoectochilus samples include two varieties: Banyelan and Xueyelan. All samples were stored in a refrigerator until the experiment was completed.

[0028] S2: Build a spectrum acquisition platform for Anoectochilus roxburghii. The main experimental instruments used in the spectrum acquisition system are as follows: a portable high-speed fiber optic spectrometer (FX2000, Shanghai Fuxiang Optics Co., Ltd., Shanghai, China), an infrared quartz fiber probe with a core diameter of 600 μm and a length of 2.0 m (FIB-600-NIR, Shanghai Fuxiang Optics Co., Ltd., Shanghai, China), and a halogen lamp with a wavelength range of 360-2500 nm (HL2000, Shanghai Fuxiang Optics Co., Ltd., Shanghai, China). Figure 2 The golden thread vine sample was placed on a bracket, and a halogen lamp light source was placed 10 cm below the bracket. A fiber optic probe was placed 3 cm above the bracket to collect the transmission spectrum signal. The collected light signal was finally refracted by the optical fiber to reach the spectrometer, and the data collection process was completed.

[0029] The flowchart of the specific step S3 is as follows Figure 3 As shown:

[0030] S31: When collecting spectra, fix the sample leaves on a horizontal bracket, the ambient temperature is about 26℃, and the humidity is about 75%RH. First, number each sample and divide each sample into ten areas. Each area collects a transmission spectrum, and a total of ten transmission spectra are collected for one sample. The spectrum collection mode is transmission, and the band response range is 400-1170nm.

[0031] S32: The ten spectra are averaged to serve as the average spectrum of each sample. Finally, the feature data length of each average spectrum is 1608. Figure 4 Shown are images of one of the authentic average spectra and one of the counterfeit average spectra.

[0032] S33: Randomly shuffle the data set, and use 90% of the data as the training set and 10% of the data as the validation set. Repeat the data set division process ten times, each division is random, and a total of ten training sets and ten test sets are obtained for ten-fold cross validation.

[0033] A deep learning model network based on 1D-CNN is constructed. The convolutional neural network model based on the one-dimensional convolutional network 1D-CNN is composed of a cascade of several convolutional units and a final classifier. Each convolutional unit is composed of a one-dimensional convolutional neural network, a normalization layer, and a pooling layer cascade. The classifier is composed of a fully connected layer and a softmax function. The structure is shown in Table 1.

[0034] Table 1 1D-CNN model network structure

[0035]

[0036]

[0037] S41: Visible-near infrared spectrum is a one-dimensional sequence data, so the convolutional neural network model based on the 1D-CNN structure is used. 1D-CNN consists of a convolutional layer (Conv1d), a batch normalization layer (BN), a maximum pooling layer (MaxPooling), a flattening layer (Flatten), a fully connected layer (Linear), and an input and output layer.

[0038] S42: The model input is the collected spectral data, i.e., 1×1608 spectral sequence data, which passes through three Conv blocks consisting of Conv1d, BN, and MaxPooling. Each block starts with a Conv1d layer, which performs convolution operations on the spectral sequence through a sliding window to extract local features of the data. The mathematical model of the convolution layer can be described as:

[0039]

[0040] Among them, M j is the input feature vector, l is the lth layer network, k is the convolution kernel, b is the network bias, is the output of layer l, is the input of the lth layer. The activation function is ReLu (rectified linear unit). ReLu introduces nonlinear units and enhances the expression ability of the network. The formula of ReLu is as follows:

[0041]

[0042] The convolutional layer is followed by the BN layer, which is used to accelerate the training speed of the model, strengthen the robustness of the model, and prevent gradient explosion. Finally, there is the MaxPooling layer, which divides the input one-dimensional data according to the set window size, and then takes the maximum value in each window as the output, which can play a role in feature extraction and dimensionality reduction. The formula for maximum pooling is as follows:

[0043]

[0044] is the eigenvalue of the nth eigenvector of the lth layer, n∈[(j-1)V+1,jV], V is the width of the pooling layer, is the value of the neuron in the l+1th layer.

[0045] The one-dimensional feature obtained after three Conv blocks has a size of (1, 51, 64), indicating that the feature is one-dimensional, with a length of 51 and a number of channels of 64. The one-dimensional feature is then input into the second type of block consisting of two Conv1d and BN layers. The MaxPooling layer is removed from the second type of block. The function of the MaxPooling layer is to downsample the feature map and reduce the size of the feature map to extract higher-level features. However, due to the short length of the one-dimensional feature, in order to avoid over-compression of the feature size and retain more local details in the classification task, the MaxPooling layer is removed in the second block so that more local information can be retained in the network. Doing so helps improve the model's perception of fine-grained features, thereby better capturing the difference between the authenticity of the golden thread lotus samples.

[0046] Finally, after being flattened, the fully connected layer and the SoftMax activation function are connected, and the final output is the prediction result of the authenticity of the golden thread lotus. The formula of the SoftMax activation function is as follows:

[0047]

[0048] Among them i is the output of the ith node, C is the number of output nodes, that is, the number of classification categories, and C of the present invention is 2, namely, genuine and counterfeit.

[0049] S43: In order to further illustrate the complexity of the model and verify whether the model meets the conditions for rapid prediction, the parameters of the entire model are calculated using the following formula:

[0050] ParaM Conv =K×CH in ×CH out +CH out (5)

[0051] ParaM BN =2×CH in (6)

[0052] ParaM FC =L in ×L out +L out (7)

[0053] ParaMConv Represents the total number of parameters of the one-dimensional convolutional layer, ParaM BN Represents the total number of parameters of the BN layer, ParaM FC represents the total number of parameters of the fully connected layer. K represents the size of the convolution kernel, CH in and CH out Represents the number of channels of input and output features respectively. in and L out Represents the size of input and output respectively. As shown in the table, the number of parameters of the entire model is 241,922, and the memory occupied by the parameters is 0.92MB.

[0054] S44: Introduce the cross entropy loss function. The cross entropy loss is often used in the classification tasks of deep learning models. It can effectively represent the gap between the predicted value and the true value, and cooperate with the softmax activation function to guide the parameter update of the model. The relevant formula is as follows:

[0055]

[0056] where y i is the true value, is the predicted value. The parameters related to the training model are as follows: The model finally divides the input spectral data labels into two categories, and the training batch (epochs) is 200. The initial learning rate is 0.0001, and the learning rate dynamic adjustment strategy is adopted. The learning rate is reduced to half of the original every 100 epochs. The optimizer of the model is Adam Weight DecayOptimizer (AdamW). The weight decay introduced by AdamW can effectively alleviate the overfitting problem and enhance the generalization ability of the model. The weight decay of the optimizer is set to 0.01, and the batch size is 4.

[0057] S45: The present invention uses accuracy (AC), precision (PR), recall (RE) and F1 score to represent the performance of 1D-CNN model and other comparison models. TP (true positive) refers to the number of samples correctly identified in the current category, TN (true negative) refers to the number of samples correctly identified in other categories, FN (false negative) refers to the number of samples incorrectly predicted in the current category, and FP (false positive) refers to the number of samples incorrectly predicted in other categories. The relevant formula is as follows:

[0058]

[0059] Final experimental results: In order to verify the performance of the method proposed in this invention, a ten-fold cross validation method was used to evaluate the performance of the model. The model training accuracy of each of the ten folds is as follows: Figure 5As shown in Figure 2, by observing the model training accuracy graph of each fold, it can be found that as the number of training rounds increases, the accuracy of each fold gradually increases. When the number of rounds reaches 200, the training accuracy of eight folds exceeds 95%, and the accuracy of the remaining two folds is also above 90%. The evaluation indicators of model performance include accuracy (AC), precision (PR), recall (RE) and F1 score (F1). The evaluation indicator results of each fold on the test set are shown in Table 2. It can be seen from Table 2 that except for the accuracy of the tenth fold of 83.33%, the accuracy of the other nine folds is 91.67% or 100%. At the same time, by calculating the average value, the average accuracy, precision, recall and F1 value of the model are 94.17%, 94.05%, 95.00% and 94.21% respectively. The above results show that the 1D-CNN model established in this study can achieve high classification accuracy and model performance without using spectral preprocessing methods and wavelength selection methods, and can effectively identify adulterated samples of A. roxburghii.

[0060] Table 2. Ten-fold model training accuracy

[0061]

[0062]

Claims

1. A method for distinguishing the authenticity of Anoectochilus roxburghii based on near infrared spectroscopy, characterized in that: The following steps are involved: Collect samples of authentic and fake Anoectochilus roxburghii, create a Anoectochilus roxburghii spectral dataset, and divide it into a training set and a test set; A convolutional neural network model based on the one-dimensional convolutional network 1D-CNN was constructed, and the spectral data of the training set was input into the convolutional neural network model for training. The golden thread vine samples were predicted and the authenticity identification results of the golden thread vine were output.

2. The method for distinguishing the authenticity of Anoectochilus roxburghii based on near infrared spectroscopy according to claim 1, characterized in that: The collection of samples of authentic and counterfeit Anoectochilus roxburghii is specifically implemented by building a spectral data acquisition platform to collect spectral data of authentic and counterfeit Anoectochilus roxburghii from a Anoectochilus roxburghii data set. The spectral acquisition mode of the spectral data acquisition platform is transmission, and the band response range is 400nm-1170nm.

3. The method for distinguishing the authenticity of Anoectochilus roxburghii based on near infrared spectroscopy according to claim 2, characterized in that: The spectral data acquisition platform includes a spectrometer, an infrared quartz fiber probe, and a halogen lamp light source with a light source band range of 360nm-2500nm; the golden thread vine sample is placed on a bracket, a halogen lamp light source is placed 10cm below the bracket, and a fiber optic probe is placed 3cm above the bracket to collect transmission spectrum signals. The collected spectrum signals are refracted by the optical fiber to reach the spectrometer, and the spectral data acquisition process is completed.

4. The method for distinguishing the authenticity of Anoectochilus roxburghii based on near infrared spectroscopy according to claim 3, characterized in that: The specific implementation process of preparing the chromatographic data set of Anoectochilus roxburghii and dividing it into a training set and a test set is as follows: For each Anoectochilus roxburghii sample, ten regions were divided, and a transmission spectrum was collected from each region. A total of ten transmission spectra were collected from one Anoectochilus roxburghii sample. The ten spectra were averaged to obtain the average spectrum corresponding to the Anoectochilus roxburghii sample; The average spectra corresponding to all the golden thread vine samples were obtained, randomly shuffled, and then divided into training sets and test sets in proportion. The data set division process was repeated ten times to obtain ten training sets and ten test sets.

5. The method for distinguishing the authenticity of Anoectochilus roxburghii based on near infrared spectroscopy according to claim 4, characterized in that: The convolutional neural network model based on the one-dimensional convolutional network 1D-CNN is composed of a cascade of several convolutional units and a final classifier, each convolutional unit is composed of a cascade of a one-dimensional convolutional neural network, a batch normalization layer and a pooling layer, and the classifier is composed of a fully connected layer and a softmax function.

6. The method for distinguishing the authenticity of Anoectochilus roxburghii based on near infrared spectroscopy according to claim 5, characterized in that: It also includes exporting the trained 1D-CNN model and using the test set to conduct classification experiments on genuine and fake samples of Anoectochilus roxburghii.

Images