CN114202645A

CN114202645A - Plastic near infrared spectrum classification and identification precision verification method

Info

Publication number: CN114202645A
Application number: CN202111553369.1A
Authority: CN
Inventors: 王斌; 谭成章; 徐晓轩
Original assignee: Nankai University
Current assignee: Nankai University
Priority date: 2021-06-29
Filing date: 2021-12-17
Publication date: 2022-03-18

Abstract

A method for verifying the identification accuracy of plastic classes by using near infrared spectrum. The method comprises respectively collecting near infrared spectra of a plurality of PP (polypropylene) reclaimed materials, PE (polyethylene) reclaimed materials, PP reclaimed materials and PE reclaimed material samples; preprocessing the near infrared spectrum data and establishing a support vector machine model; and selecting the most appropriate data preprocessing and support vector machine model method. Establishing a one-dimensional convolution neural network model: and respectively comparing the experimental precision of the near infrared spectrum data of the model adopting the preprocessing-support vector machine and the selected neural network model, thereby verifying the identification precision of different models aiming at different types of plastics.

Description

Plastic near infrared spectrum classification and identification precision verification method

Technical Field

The invention relates to the field of spectrum detection, in particular to application of a neural network and a support vector machine to near infrared spectrum analysis of plastics.

Background

The plastic is convenient for production and life of human beings since the invention, and is widely applied. However, most plastic products are chemically stable and non-degradable, and a large amount of plastic waste is generated. Data published by the environmental planning agency in 2018 indicate annual production of plastic waste of about 3 hundred million tons worldwide. However, China is one of ten major plastic product production and consumption countries in the world, and the classification, recycling and reutilization of plastic wastes is very important, otherwise, the environment is polluted and resources are wasted.

Improving the collection and sorting of plastic waste is the best way to improve the quality of recycled plastic. Traditional plastic classification methods such as manual classification, optical sorting, flotation and the like are time-consuming and labor-consuming, and in order to save labor cost and accurately and efficiently classify plastics, people research an intelligent plastic classification algorithm. For example, Laser Induced Breakdown Spectroscopy (LIBS) techniques and Principal Component Analysis (PCA) have been successfully used to identify polyethylene terephthalate (PET), high density Polyethylene (PE), polypropylene (PP) and Polystyrene (PS). The LIBS technique can also identify plastic/polymer samples that have the same polymer matrix but contain different additives. X-ray absorption spectroscopy (XAS) in combination with PCA and Back Propagation Neural Network (BPNN) was used to identify 15 different plastics. Raman spectroscopy and K-nearest neighbor algorithm (KNN), Cyclic Subspace Regression (CSR), library search are also used for plastic classification. The attenuated total reflection Fourier transform infrared spectrum is combined with principal component analysis and system clustering analysis (HCA) to classify and identify seven types of waste plastics, and cosine and average distance methods are selected as inter-sample and inter-class distance functions to cluster data, so that 100% classification accuracy is obtained finally. Classification and regression models (CART) can find direct and simple classification conditions from near infrared spectral data.

The invention patent CN110441254A discloses a method for identifying plastics by combining a neural network and near infrared spectrum, which uses a near infrared optical frequency comb spectrometer; the overlapped interference image signals are subjected to convolution demodulation to obtain separated interference image signals, and finally Fourier transformation is carried out to obtain a frequency domain spectrogram corresponding to the sample. In view of this, many models are currently involved in the identification of plastics, and a quantitative analysis and determination method for the identification accuracy of the models is required. Therefore, it is necessary to provide a method for determining the model identification accuracy of a specific test object, plastic, quickly and nondestructively.

Disclosure of Invention

The invention provides a near infrared spectrum plastic identification method using a convolutional neural network and a support vector machine, and the identification effect of different preprocessed models is verified.

One method of the invention is to provide a method for verifying the identification precision of plastic types by using near infrared spectrum, which comprises the following steps:

step 1, respectively collecting near infrared spectrums of a plurality of PP (polypropylene) reclaimed materials, PE (polyethylene) reclaimed materials, PP virgin materials and PE virgin material samples;

step 2, dividing a sample set into a training set and a verification set according to the ratio of the training set to the verification set which is approximately 3: 1 by adopting a random selection method (RS) for each type of the four types of samples;

step 3, carrying out data preprocessing on the data in the step 2 and establishing a support vector machine model; wherein relaxation coefficients and penalty factors C are introduced for the linearly indivisible data set; inputting the near infrared spectrum data and corresponding class labels, and randomly selecting a verification set according to a proportion to perform an experiment; the model output comprises predicted labels, training set accuracy and verification set accuracy; selecting the most appropriate data preprocessing and support vector machine model method as MSC-SVM according to the experimental result;

step 4, establishing a one-dimensional convolution neural network model: constructing a 6-layer 1DCNN for plastic classification; the total area neural network model comprises an input layer, a convolutional layer C1, a pooling layer S2, a full connecting layer F3, a full connecting layer F4 and an output layer, wherein a label value '0, 1, 2, 3' of a plastic class represented by a training set is converted into one-hot vector input, and the input dimension of spectral data of each sample is 15011; regular terms and random inactivation (Dropout) are added in the neural network, the complexity of the model is reduced, only one convolution layer is used, 8 convolution kernels are used, the size is 3 x 1, and the step size is 1;

and 5, respectively comparing the experimental precision of the near infrared spectrum data adopting the MSC-SVM and the 1DCNN model, thereby verifying the identification precision of different models aiming at different types of plastics.

Preferably, the verification in the step 5 results in: predicting PP fresh feed by using a 1DCNN model; PE virgin materials can be correctly classified by using two models, namely an MSC-SVM model and a 1DCNN model.

Preferably, in step 3, for the linearly separable data set, the objective function is:

obeying the constraint (2):

y_i(ω^Tx_i+b)≥1，i＝1，2，...，n (2)

for linearly inseparable data sets, a relaxation coefficient xi is introduced_i≧ 0 and a penalty factor C, the objective function and constraint condition become equations (3) and (4).

y_i(ω^Tx_i+b)≥1-ξ_i，i＝1，2，...，n (4)

For equations (1) - (4), n is the number of samples. Omega and b are respectively hyperplane omega^Tx + b is 0 weight and bias parameter. x is the number of_iAnd y_iA vector representing the ith input and an ith dependent variable value.

Preferably, theAnd in the step 5, the classification accuracy of the training set and the test set is used as a model evaluation index. The accuracy is obtained from equation (5) and is the number of correctly classified samples N_cAccounts for the total number of samples N_rThe ratio of (a) to (b).

Preferably, the data preprocessing method in step 3 is a first derivative method, a second derivative method, centering, normalization, a Savitzky-Golay smoothing method, multivariate scattering processing, and a standard normal transformation method, respectively.

Another aspect of the present invention provides the use of the above method for near infrared spectroscopy identification of plastic species.

Another aspect of the invention provides a computer storage medium having stored thereon an executable program comprising a method of performing identification accuracy verification as described in any one of the above.

The inventive aspects of the present invention include the following lower limitations, but are not limited to:

(1) the detection introduces neural network modeling analysis for plastics. Technical solutions exist in the prior art that combine neural networks with spectroscopic techniques. But the identification of plastics is different from other types of sample identification. PP or PE as the two types of plastic sources with the largest usage amount, if the introduced neural network is not optimized as described in the present invention, the resolution accuracy is not high. Therefore, the method introduces the SVM model, and performs experimental comparison on the model after each kind of data preprocessing to obtain the optimized model. Through comparison of various pretreatments and data verification after model establishment, the prediction accuracy can be improved by selecting different models for different plastic types.

(2) Relaxation coefficient xi is introduced to linear inseparable data set in support vector machine_i≧ 0 and a penalty factor C, thereby defining the objective function and the constraint condition. And solving the extreme value of the constraint condition by using a Lagrange multiplier method. After modeling, 4-fold cross validation is carried out, and different data are compared to preprocess the methodAnd selecting the SVM model with the highest accuracy.

(3) A6-layer 1DCNN is constructed for plastic classification, and comprises an input layer, a convolutional layer C1, a pooling layer S2, a full-link layer F3 and a full-link layer F4, an output layer, wherein label values of plastic classes represented by a training set are converted into one-hot vector input, and the input dimension of each sample spectral data is 1501 multiplied by 1. In order to avoid the over-fitting phenomenon as much as possible, regularization terms and random inactivation (Dropout) are added in the neural network. And the complexity of the model is reduced, only one convolution layer is used, 8 convolution kernels are used, the size is 3 multiplied by 1, and the step length is 1.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the principles of the invention. In the drawings:

FIG. 1(a) is a near-infrared raw spectrum of a plastic sample according to an embodiment of the present invention;

FIG. 1(b) is a spectrum of a plastic sample after removing abnormal values in the near infrared spectrum according to an embodiment of the present invention;

fig. 2 is a schematic diagram of each layer structure of the CNN model according to the embodiment of the present invention.

DETAILED DESCRIPTION OF EMBODIMENT (S) OF INVENTION

The plastic samples were provided by the Ningbo City inspection and quarantine office, and the total number of the plastic samples included PP (polypropylene) and PE (polyethylene) was 100. The number of PE recycled materials (PE recycled, abbreviated as PEr) is 32, the number of PE new raw materials (PE new raw material, abbreviated as PEN) is 36, the number of PP recycled materials (PP recycled, abbreviated as PPr) is 15, the number of PP new raw materials (PP new raw material, abbreviated as PPn) is 17, and the categories are respectively marked as 0, 1, 2 and 3.

Near infrared spectroscopy data of the plastic is collected by a near infrared spectrometer using a prism technique S450. The wavelength range is set to 900-.

The spectra of the 100 samples collected are shown in FIG. 1 (a). And (3) according to the comparison data of the spectrogram, eliminating data of 2 abnormal samples in the PP reclaimed material to obtain the spectrogram shown as a graph (b) in FIG. 1.

The final total number of valid samples is 98. The four types of plastic samples are divided into sample sets according to the proportion of approximately 3: 1 of a training set and a verification set by adopting a random selection method (RS) in each type. The total number of training set samples is 72, and the number of verification set samples is 26. Specific information is shown in table 1.

TABLE 1 statistical table of plastic sample information

Unit：PCS

Data preprocessing:

in order to reduce the influence of background noise and sample scattering on the model in the spectrum, the spectral data is preprocessed by the first derivative method (1 st Der), the second derivative method (2 nd Der), centralization (centralization), normalization (standardization), Savitzky-Golay smoothing method (SG), multiple scattering process (MSC), and standard normal transform (SNV), respectively, before modeling. After the support vector machine model is established, the most appropriate data preprocessing method is selected according to the experimental result.

Support vector machine model:

support vector machines are commonly used to solve the two-class problem, and now can also handle the multiple-class problem. One-to-many (OVA) or one-to-one (one-to-one, OVO) approaches may be used to convert a multi-classification problem into a two-classification problem. The basic principle is to find a hyperplane omega^Tx + b is 0, so that different classes of points in the training set fall on both sides of the hyperplane, while maximizing the blank area on both sides of the hyperplane. The samples can be mapped to a high-dimensional space by using different kernel functions to find a hyperplane, so that linear classification and nonlinear classification are supported.

For a linearly separable data set, the objective function is:

obeying the constraint (2):

y_i(ω^Tx_i+b)≥1，i＝1，2，...，n (2)

y_i(ω^Tx_i+b)≥1-ξ_i，i＝1，2，...，n (4)

For equations (1) - (4), n is the number of samples. Omega and b are respectively hyperplane omega^Tx + b is 0 weight and bias parameter. x is the number of_iAnd y_iA vector representing the ith input and an ith dependent variable value. The extrema can be solved using the lagrange multiplier method. And after modeling, performing 4-fold cross validation, comparing the accuracy rates of different data preprocessing methods, and selecting the model with the highest accuracy rate. And the parameters that make the model work best are selected: using OVO method, i.e. constructing a two-classification SVM model between every two classes; the penalty factor C is set to 256 and the kernel function is a linear kernel function. Near infrared spectrum data of the plastic and corresponding class labels (0, 1, 2, 3) are input, and 10 verification sets are randomly selected according to the proportion of table 1 to carry out 10 experiments. The model output comprises predicted labels, training set accuracy and verification set accuracy.

One-dimensional convolutional neural network model:

the convolutional neural network is used as a nonlinear model, can effectively extract local information in the spectrum, and is strong in learning ability.

A typical convolutional neural network model comprises an input layer, a convolutional layer, a pooling layer, a fully-connected layer, and an output layer. For spectral data, the input layer inputs one-dimensional spectral data more efficiently than a two-dimensional spectral matrix. After the data and the labels are input, the convolutional layer uses a plurality of one-dimensional convolutional kernels with set size and step length to obtain a characteristic diagram after convolution operation. The number of convolutional layers may be one, two, or more, but too many convolutional layers and convolutional kernels may result in overfitting. Pooling layers are typically used after convolutional layers to extract local features of the data. The pooling method comprises a maximum pooling method and an average pooling method, and the maximum pooling method and the average pooling method mainly extract characteristic information. Features are mapped to sample space for classification through one or more fully connected layers. To build a non-linear model, an activation function needs to be added. While the gradient vanishing problem can be avoided by using the ReLU function, in the classification problem, the last layer of the neural network usually uses the Softmax function to map the input to be between 0 and 1, and the probabilities of different classes are obtained for classification.

When the model is trained, firstly, the weight is initialized, the near infrared spectrum data and the class labels of the plastic sample training set are input, and the final output result is obtained through each layer of the neural network. And calculating a model loss function value, transmitting the loss function value from the last layer to each layer of the network through back propagation, updating the weight according to the direction of minimizing the loss function value, and continuing training.

This experiment constructed a 6-layer 1DCNN for plastics classification, including the input layer-convolutional layer C1-pooling layer S2-fully-connected layer F3-fully-connected layer F4-output layer, as shown in fig. 2. The label value "0, 1, 2, 3" of the training set representing the plastic class is converted into one-hot vector input, and the input dimension of each sample spectral data is 1501 × 1. In order to avoid the over-fitting phenomenon as much as possible, regularization terms and random inactivation (Dropout) are added in the neural network. And the complexity of the model is reduced, only one convolution layer is used, 8 convolution kernels are used, the size is 3 multiplied by 1, and the step length is 1. The pooling layer uses the maximum pooling method, the size of the kernel is 2 × 1, and the step size is 2. The last layer of the full-junction layer of the model uses a Softmax activation function, the optimizer is AdamaOptizer, the learning rate is 0.0001, the number of convolution kernels is 3, the number of neurons of the full-junction layer is 60, and the iteration number is 5000. Model training is based on the Tensorflow framework GPU version.

And (3) carrying out experiments by using 10 groups of verification sets which are the same as the MSC-SVM model to obtain the prediction category and the prediction accuracy.

Evaluation indexes are as follows:

the classification accuracy of the training set and the test set is used as a model evaluation index in the experiment. The accuracy P is obtained from the formula (5) and is the number of correctly classified samples N_cAccounts for the total number of samples N_rThe ratio of (a) to (b).

When the SVM model is established, the results of cross validation experiments after modeling without data preprocessing and different data preprocessing methods are shown in table 2.

TABLE 2 SVM model accuracy for different data preprocessing

It can be seen that the accuracy of the model is highest after MSC. And selecting the MSC-SVM model with the best classification performance, and performing 10 random experiments on 98 samples to obtain a result of 980 times in total, wherein the training set is 720 times, and the verification set is 260 times. And (3) counting the 980 results, recording the classification results of different plastic types, and embodying the true values and the predicted values of the classification results in a table to obtain a confusion matrix as shown in table 3. Similarly, the confusion matrix recording the results of 10 random experiments and 980 classification of 1DCNN model is shown in table 4, and the verification set of each random experiment is the same as the corresponding verification set used by the MSC-SVM model. Diagonal elements in the confusion matrix represent correctly classified samples, and if the experimental results of the training set and the verification set are distributed in a form in a diagonal manner, the classification accuracy of the model is high.

TABLE 3 MSC-SVM model 980 times classification result confusion matrix

Unit:label(s)

TABLE 41 DCNN model 980 classification result confusion matrix

Unit:label(s)

As can be seen from Table 3, the training set results of the MSC-SVM model are distributed on the diagonal, which shows that the classification in the experiment is completely accurate; and the verification set has the result of classification error except the PE fresh material. The data in table 4 are not all distributed on the diagonal line, which indicates that there are few error results in the classification results of the 1DCNN model training set and the verification set. In a comprehensive view, the 1DCNN model for predicting the category of the PP new raw material has better effect, and the MSC-SVM model can misjudge the PP regenerated material with higher probability. The PE regenerated material has a high probability of being misjudged as a new PE raw material. PE virgin material can be classified almost correctly using both models.

The results of 10 random experiments on the MSC-SVM model and the 1DCNN model were compared together, and Table 5 is obtained.

TABLE 5 MSC-SVM model vs. 1DCNN model accuracy

As can be seen from Table 5, the MSC-SVM model performed well on the training set with an accuracy of 100%. On the verification set, the accuracy of the 1DCNN model is 91.5 percent and is slightly better than that of the MSC-SVM model. For different types of plastics, the prediction effects of two models of PE reclaimed materials and PP reclaimed materials are similar, and the accuracy of judging the types of the PP reclaimed materials is not high; the PE new raw material category judgment accuracy rate reaches 100% on the verification set; the PP fresh feed is classified by using a 1DCNN model with the accuracy rate of 100%. And recording the average program execution time required by single experiment training of the MSC-SVM model and the 1DCNN model, wherein the MSC-SVM model is 2.84 seconds, and the 1DCNN model is 24.55 seconds. It can be seen that the MSC-SVM model is faster than the 1DCNN model under the condition of smaller experimental data quantity. Generally, the CNN has obvious advantages under the condition of large data volume, and an overfitting phenomenon is easy to occur on a small data set, but the experiment proves that as long as the number of convolution layers is reasonable, a one-dimensional convolution kernel suitable for spectral data is used, parameters are set properly, and the accuracy can also achieve better effect by using some methods for avoiding overfitting. And the CNN has lower requirements on data preprocessing, even does not need data preprocessing in some cases, and does not need to consider the characteristics of sample data, so the CNN has great application potential in near infrared spectrum data analysis as a universal method. The method is one of the invention points, and meets the detection accuracy requirements of different samples by reasonably selecting parameters such as the number of convolution layers and the like.

The above embodiments are only used for illustrating the technical solutions of the present application, and not for limiting the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions in the embodiments of the present application.

It should be understood that in the present application, "at least one" means one or more, "a plurality" means two or more. "and/or" for describing an association relationship of associated objects, indicating that there may be three relationships, e.g., "a and/or B" may indicate: only A, only B and both A and B are present, wherein A and B may be singular or plural. The character "/" generally indicates that the former and latter associated objects are in an "or" relationship. "at least one of the following" or similar expressions refer to any combination of these items, including any combination of single item(s) or plural items. For example, at least one (one) of a, b, or c, may represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", wherein a, b, c may be single or plural.

Claims

1. A method for verifying the identification precision of plastic classes by using near infrared spectrum comprises the following steps:

step 2, dividing a sample set into a training set and a verification set according to the ratio of approximately 3: 1 of the training set to the verification set by adopting a random selection method (RS) for each type of sample in the four types of samples;

step 3, carrying out data preprocessing on the sample centralized data in the step 2 and establishing a support vector machine model (SVM); wherein relaxation coefficients and penalty factors C are introduced for linearly indivisible data sets in the support vector machine model; inputting the near infrared spectrum data and corresponding class labels, and randomly selecting a verification set according to a proportion to perform an experiment; the model output comprises predicted labels, training set accuracy and verification set accuracy; selecting the most appropriate data preprocessing and support vector machine model method according to the experimental result;

and 5, comparing the experimental precision of the near infrared spectrum data by adopting the most appropriate data preprocessing and support vector machine model method in the step 3 and the 1DCNN model in the step 4 respectively, and verifying the identification precision of different models aiming at different types of plastics.

2. The method of claim 1, wherein the most suitable data preprocessing and support vector machine model method in step 3 is MSC-SVM;

the verification in step 5 yields: predicting PP fresh feed by using the 1DCNN model; PE fresh feed was correctly classified using both the MSC-SVM and the 1DCNN models.

3. The method of claim 1, wherein in step 3, for a linearly separable data set, the objective function is:

obeying the constraint (2):

y_i(ω^Tx_i+b)≥1，i＝1，2，...，n (2)

y_i(ω^Tx_i+b)≥1-ξ_i，i＝1，2，...，n (4)

4. The method according to claim 1, wherein the classification accuracy of the training set and the test set is used as a model evaluation index in the step 5. The accuracy is obtained from equation (5) and is the number of correctly classified samples N_cAccounts for the total number of samples N_rThe ratio of (a) to (b).

5. The method of claim 1, wherein the data preprocessing method in step 3 is a first derivative method, a second derivative method, centralization, normalization, Savitzky-Golay smoothing method, multivariate scattering processing, and standard normal transformation method, respectively.

6. Use of the method according to any one of claims 1 to 5 for near infrared spectroscopy in the identification of plastic species.

7. A computer storage medium having stored thereon an executable program comprising a method of performing identification accuracy verification according to any one of claims 1-5.