Research on non-destructive identification about vehicle paints by DT-KNN-FDA

YAN Wenjie; CHEN Junming; SONG Yajun; KONG Hao; JIA Zhenjun

doi:10.7510/jgjs.issn.1001-3806.2021.02.009

An identification method based on fingerprint spectroscopy combined with decision tree, k-nearest neighbor, and Fisher discriminant analysis (DT-KNN-FDA) model was proposed to achieve the rapid and non-destructive identification of the vehicle paints and performed by theoretical analysis and experimental verification. The infrared absorption spectroscopy for a total of 60 samples of car paint were collected and obtained as the experimental data. Through the selection of characteristic wave numbers, a multi-classification model based on the DT, KNN analysis, and FDA was established and compared. 58 sets of adjustment data were extracted through correlation analysis, and a classification model was constructed based on this. The results show that the overall discrimination accuracy of DT classification model, KNN classification model and FDA classification model for each sample is 77.80%, 72.31%, and 85.00%, respectively; infrared spectroscopy combined with DT-KNN-FDA analysis can realize the distinction between products of different brands is ideal for classification. This method is fast, accurate, and effective, and has certain universality and significance.

HTML

引言

在交通肇事案件中，执法人员经常会在肇事现场、受害人衣物上发现并提取到车漆碎片。通过对车漆进行分析与鉴定，进一步确定其品牌、生产厂家等信息，进而追溯肇事车辆，从而为确认或排除嫌疑人和嫌疑车辆提供一定的线索，为案件的诉讼和判决提供一定的证据。因此，车漆的检验鉴定对侦破交通肇事案件具有十分重要的意义。

不同品牌和生产厂家的车漆有不同的配方和工艺，即在成分和其含量上均存在一定差异。即不同品牌的车漆样本间存在一定差异，对这一差异的挖掘将有助于执法人员推断并确定现成提取的碎片检材的品牌和生产厂家。目前，车漆检验主要有光学显微镜法^[1]、扫描电镜法^[2]和光谱成像技术^[3]等。光学显微镜法只能对车漆碎片的形态学特征进行初步解读，这易受主观因素影响，且耗时耗力；扫描电镜法在确定车漆中元素含量上有一定优势，但对其品牌和生产厂家信息的解读不够全面。高发的交通肇事案件和提取到的大量车漆碎片物证给执法人员的工作带来了极大的挑战。如何降低鉴定所需的时间精力等成本，提高鉴定效率，实现对车漆碎片的快速无损鉴定，是当下执法人员关注的重点之一。

鉴于此，实验中借助红外光谱分析技术，通过对特征波数的选择，建立基于决策树分析(decision tree, DT)、k近邻分析(k-nearest neighbor, KNN)、Fisher判别分析(Fisher discriminant analysis, FDA)的车漆样本光谱分类鉴别模型，从而实现对车身油漆品牌较为准确区分与归类，为法庭科学中车漆无损、准确地检验鉴定提供一定的参考和借鉴。

1. 实验

1.1. 实验样本

从市场上收集了常见的诚得利等4种品牌共计60个不同品牌和生产厂家的车漆样本。采集车身前部、两侧、后部共计4处位置的车漆碎片，为避免采集过程中人为因素带来的误差，每处随机采集3份样本。首先，将采集的样本用酒精棉擦拭样品，从而除去样本表面残留的灰尘等污物；而后将样本放入盛有去离子水的烧杯中，并超声清洗2次，每次10min；最后用酒精棉将样本擦拭干净，进样检测。

1.2. 实验设备

采用Nicolet 5700型傅里叶变换红外光谱仪(Thermo Fisher Scientific公司)，配有衰减全反射附件(Thermo Fisher Scientific公司)^[4-5]。光谱数据处理软件OMNIC8.2，光谱采集范围为4000cm^-1~400cm^-1，每个样本均采集3次，取其平均值作为实验数据^[4-5]。

3. 结论

本文中采用红外吸收光谱与DT-KNN-FDA方法，实现了对车漆样本较为准确地分类与识别。通过相关性分析筛选出58组的特征数据，以此为基础构建分类模型。DT分类模型、KNN分类模型和FDA分类模型对各样本的总体区分准确率分别为77.80%, 72.31%和85.00%。综上所述，红外吸收光谱结合相关性分析及FDA模型可较好地实现对车漆不同品牌间较为准确地区分，且分类结果较为理想。本实验中在一定程度上消除了传统鉴别方法中因主观判断造成误差、人工鉴别效率较低以及对检材损耗较大的缺点，为车漆的分类鉴别提供了一种新的参考思路，同时，本方法也为其它鉴别手段提供了一定的借鉴。值得注意的是，车漆是多组分样本，对多组分分析是一个挑战，因为不同的分子可能导致相似的光谱形状，使它很难从一个复杂的系统中分离出某些分子信息。因此，如何改进红外光谱技术以满足日益增长的物证分析需求，是今后研究的热点之一。

Reference (15)

[1]	KRUGLAK K J, DUBNICKA M, KAMMRATH B. The evidentiary significance of automotive paint from the northeast: A study of red paint[J]. Journal of Forensic Sciences, 2019, 64(5): 1345-1358. doi: 10.1111/1556-4029.14007
[2]	MALEK M, NAKAZAWA T, KANG H W. Multi-modal compositional analysis of layered paint chips of automobiles by the combined application of ATR-FTIR imaging, Raman microspectrometry, and SEM/EDX[J]. Molecules, 2019, 24(7): 1381-. doi: 10.3390/molecules24071381
[3]	ISHIKAWA A, HARA S, TANAKA T. Cross-polarized surface-enhanced infrared spectroscopy by fano-resonant asymmetric metamaterials[J]. Scientific Reports, 2017, 7(1): 3205-. doi: 10.1038/s41598-017-03545-8
[4]	HE X L, WANG J F. The identification about the automotive bumper based on Newton interpolation polynomial-infrared derivative spectroscopy[J]. Laser Technology, 2020, 44(3): 333-337.
[5]	HOU W, WANG J F. Rapid identification the black marker ink based on infrared fingerprint spectros copy combined with multilayer perceptron[J]. Laser Technology, 2020, 44(4): 436-440.
[6]	JI J H, WANG J F, WANG G X. Raman spectrum identification of waterborne wood coating based on radial basis function[J]. Laser Technology, 2020, 44(6): 762-767.
[7]	ZHOU X, SUN J, WU X H. Research on moldy tea feature classification based on WKNN algorithm and NIR hyperspectral imaging[J]. Spectrochimica Acta, 2019, A206(14): 378-383.
[8]	ZHENG X, LV G, ZHANG Y. Rapid and non-invasive screening of high renin hypertension using Raman spectroscopy and different classification algorithms[J]. Spectrochimica Acta, 2019, A215(5): 244-248.
[9]	WANG G J, XIE C, STANLEY H E. Correlation structure and evolution of world stock markets: Evidence from pearson and partial correlation-based networks[J]. Computational Economics, 2016, 51(3): 607-635. doi: 10.1007/s10614-016-9627-7
[10]	ZHOU H, DENG Z, XIA Y. A new sampling method in particle filter based on Pearson correlation coefficient[J]. Neurocomputing, 2016, 216(12): 208-215.
[11]	FREDL M A, BRODLEY C E. Decision tree classification of land cover from remotely sensed data[J]. Remote Sensing of Environment, 1997, 61(3): 399-409. doi: 10.1016/S0034-4257(97)00049-7
[12]	HE Y, WANG J F. Rapid nondestructive identification of wood lacquer using Raman spectroscopy based on characteristic-band-Fisher-K nearest neighbor[J]. Laser & Optoelectronics Progress, 2020, 57(1): 13001-.
[13]	HE X L, CHEN L B, WANG J F. Raman spectral analysis of plastic steel Windows based on k-nearest neighbor algorithm[J]. Advances in Laser and Optoelectronics, 2018, 55(5): 053001-. doi: 10.3788/LOP55.053001
[14]	HE X L, WANG J F, LI Q Sh. Multilayer perceptron-Fisher discriminant analysis based on infra red spectrum identification of vehicle bumper[J]. China Test, 2019, 45(5): 74-78.
[15]	HE X L, MA Y, WANG J F. Rapid qualitative and quantitative detection of vehicle bumpers by mid-infrared spectroscopy[J]. Engineering Plastics Applications, 2019, 47(5): 122-126.

characteristic wavenumber/cm^-1	spectral data	characteristic wavenumber/cm^-1	spectral data	characteristic wavenumber/cm^-1	spectral data	characteristic wavenumber/cm^-1	spectral data
501	68.30139	729	70.59264	802	71.68636	922	70.73677
679	68.22303	733	69.40947	806	71.35540	926	70.43494
683	67.50737	737	68.54868	810	71.42767	930	70.13481
687	66.34003	741	67.97828	814	71.63278	1003	70.44617
690	64.28992	744	67.37168	818	71.80663	1057	67.02615
694	60.26677	748	66.37009	822	71.85703	1092	66.28056
698	56.67580	771	70.23232	891	71.66743	1095	65.56648
702	58.64999	775	71.62344	895	71.50935	1099	64.66272
706	65.09207	779	72.36163	899	71.30006	1103	63.62498
710	70.65044	783	72.76534	903	71.05427	1107	62.63717
714	72.93565	787	72.91515	906	70.95933	1146	59.33903
717	73.04305	791	72.90823	910	70.95816	1176	63.28623
721	72.39362	795	72.68170	914	70.95504	1250	71.51285
725	71.59163	798	72.25454	918	70.89352	1277	73.97288

brands	Chengdeli	Munchsett	Sanhe	Sangmei
classification accuracy/%	0.00	100.00	94.30	0.00

brands	Chengdeli	Munchsett	Sanhe	Sanmei
classification accuracy/%	0.00	0.00	96.80	25.00

function	variance contribution rate/%	correlation	function test	Wilks’lambda	significance
f₁	63.7	0.810	1~3	0.153	0.000
f₂	30.0	0.688	2~3	0.444	0.001
f₃	6.3	0.398	3	0.842	0.006

Research on non-destructive identification about vehicle paints by DT-KNN-FDA

Abstract

References

Proportional views

通讯作者: 陈斌, bchen63@163.com

Proportional views