A method is proposed for rubber identification based on terahertz time-domain spectroscopy (THz-TDS) and support vector machine (SVM). In order to improve the accuracy, the cuckoo search algorithm (CS) is used to optimize the penalty factor C and kernel function parameter g of SVM. The SVM model optimized by the cuckoo search algorithm is abbreviated as CS-SVM. Principal component analysis (PCA) is applied to decrease the dimension of the spectral data. The top ten principal component factors, whose accumulated variance contribution rate reaches 93.93%, are extracted from the original spectra data and then are applied to CS-SVM. The identification rate of testing sets for CS-SVM is 100%, which is significantly higher than 96.67% identification rate of testing sets for PSO-SVM and Grid search. Experimental results show that CS-SVM can accomplish nondestructive identification for different rubber. This method lays a theoretical foundation for the application of terahertz spectroscopy in rubber classification and identification.
Natural rubber (NR), isobutylene isoprene rubber (IIR), and styrene-butadiene rubber (SBR) are the main materials of tires. With the continuous development of the automotive industry, their consumption has increased rapidly. The quality of the tire is closely related to the type and quality of the rubber material, so it is important to quickly and accurately detect the type and composition of the rubber [
This paper focuses on the identification of natural rubber (NR), isobutylene isoprene rubber (IIR), and styrene-butadiene rubber (SBR). The spectra of three types of rubber have been investigated in time and frequency domain, where some obvious characteristic absorption peaks can be observed in 0.3 ~ 1.6THz.
The experimental apparatus consists of Z-3 terahertz time-domain spectroscopy (THz-TDS) system (ZOMEGA, USA) and ultrafast femtosecond fiber laser (TOPTICA Photonics AG, Germany). Ultrafast femtosecond laser generates laser pulses at around 800nm of 100fs duration at a central wavelength with a repetition frequency of 80 MHz. The laser beam is divided into a pump beam and a probe beam by a beam splitter (CBS) [
Schematic diagram of THz-TDS system.
In order to decrease the strong absorption of moisture to the THz wave, the experiment was conducted in a particular environment, where dry air was continuously injected to make the relative humidity below 3%, and the temperature was kept at room temperature.
In this paper, there are three types of rubber to be identified, which are IIR, SBR, and NR. These three types of rubber look similar and are difficult to distinguish. Five experimental samples were made for each type of rubber. The sample was in the form of a disk having a thickness of about 1 mm and a diameter of about 12 mm. The inside of the sample was uniform and the upper and lower surfaces are parallel to each other. Considering the strong absorption of moisture to the THz wave, all the experimental samples were dried in a vacuum oven at a constant temperature of 50°C for 2 to 3 hours to reduce the moisture content before the experiment. During the experiment, each sample was measured 12 times, and the front and back sides were measured 6 times each. Therefore, there are 60 sets of data for each type of rubber. In order to eliminate coarse errors, any two of the 60 sets of data were averaged to obtain 30 sets of data for each type of rubber, and 90 sets of data were obtained for the three types of rubber.
The reference signals
In the above formulas (
Support vector machine (SVM), an algorithm of learning machine based on statistical learning theory, is first proposed by Vapnik [
SVM classification principle is to construct the classification hyperplane in the feature space and use
After solving the dual problem, the optimal classification function is as follows:
Cuckoo search (CS) is a novel heuristic global optimization algorithm proposed by YANG Xinshe and DEB Suash [
Flow chart of the CS-SVM algorithm.
Experimental data of three types of rubber (IIR, SBR, and NR) were processed by MATLAB and imported into Origin software. Figure
Spectra of three types of rubber.
Time-domain waveforms
frequency-domain spectra
The frequency-domain spectra are transformed from the corresponding time-domain spectra by using FFT showed in Figure
Owing to the low signal-to-noise ratio over 1.4 THz for the spectral instrument, the effective spectral range is 0.3 ~ 1.4THz. Figure
The THz absorption peak of different rubber.
Rubber name | Peak 1/THz | Peak 2/THz | Peak 3/THz | Peak 4/THz |
---|---|---|---|---|
IIR | 1.05 | 1.17 | 1.30 | none |
SBR | 0.96 | 1.09 | 1.26 | none |
NR | 1.04 | 1.16 | 1.25 | 1.36 |
The absorbance curves of three types of rubber.
In order to eliminate data redundancy caused by excessive data dimension, the PCA is utilized to reduce the dimension of absorption spectra. The absorption spectra matrix is reduced from 90 rows and 99 columns to 90 rows and 10 columns, and the top ten principal component factors are selected. Their cumulative variance contribution rate reaches 93.93%, and it means that the main information of original data is retained. The variance contribution rate and cumulative variance contribution rate of principal component are shown in Table
Variance contribution rate and cumulative variance contribution rate of principal component.
Component | Variance/% | Cumulative rate/% |
---|---|---|
PC1 | 32.08 | 32.08 |
PC2 | 24.45 | 56.53 |
PC3 | 17.90 | 74.43 |
PC4 | 6.84 | 81.27 |
PC5 | 4.45 | 85.72 |
PC6 | 2.58 | 88.30 |
PC7 | 1.87 | 90.17 |
PC8 | 1.50 | 91.67 |
PC9 | 1.14 | 92.81 |
PC10 | 1.12 | 93.93 |
The two-dimensional score of the first three principal components of three types of rubber is shown in Figure
Scattered scores plot PC1 versus PC2 (a) and PC1 versus PC3 (b) based on three types of rubber.
Scattered scores plot PC1 versus PC2
Scattered scores plot PC1 versus PC3
3D scattered scores plots PC1, PC2, and PC3.
The spectral data of three types of rubber, dimensionality of which is reduced by the principal component, are divided into two data sets: a training set and a testing set. There are 30 sets of data for each type of rubber, 20 of which are used as training sets and the remaining 10 sets of data are used as testing sets. The labels for IIR, SBR, and NR are defined as 1, 2, and 3, as showed in Table
Data classification and sample label of three types of rubber.
Rubber name | Number of train data | Number of test data | Sample label |
---|---|---|---|
IIR | 20 | 10 | 1 |
SBR | 20 | 10 | 2 |
NR | 20 | 10 | 3 |
The experimental data with the label is added into the CS algorithm prediction model to extract the top 10 principal component data sets with the highest contribution rate as the new feature identification data. The identification rates of the training sets and the testing sets are obtained, respectively, and compared with the identification rate of PSO-SVM and Grid search model [
Identification rate and time of three types of models.
Identification model | Identification rate of training sets /% | Identification rate | Identification time /s |
---|---|---|---|
CS-SVM | 100 | 100 | 0.81 |
Grid search | 100 | 96.67 | 0.91 |
PSO-SVM | 100 | 96.67 | 3.26 |
The identification results of CS-SVM, PSO-SVM, and Grid search for 30 testing sets of three types of rubber are shown in Figure
Classification and identification for 30 testing sets of three types of rubber.
CS-SVM method
Grid search method
PSO-SVM method
Time-domain spectra, frequency-domain spectra, and absorbance of three types of rubber are obtained by THz-TDS. PCA is used to reduce the dimension of the spectral feature data, and the top ten principal components with cumulative variance contribution rate of 93.93% are extracted to establish three classification and identification models. The identification rates of the training sets and the testing sets for CS-SVM are 100%, and the identification time is 0.81s. CS-SVM is better than PSO-SVM and Grid search in testing sets identification rate and time. Experimental results show that the method proposed in this paper can identify three types of rubber quickly and nondestructively. It lays a theoretical foundation for the application of terahertz spectroscopy in rubber classification and identification and provides a new approach for the nondestructive identification of other rubber.
The data used to support the findings of this study are available from the corresponding author upon request.
The authors declare that they have no conflicts of interest.
This work was supported in part by National Natural Science Foundation of China (11574059) and in part by Guangxi Key Laboratory of Automatic Detecting Technology and Instruments (YQ14113).