A method for identifying signature pen writing traces based on LIBS combined with machine learning

Abstract

The invention discloses a signature pen writing trace recognition method based on LIBS combined with machine learning, including a spectrum data set of different ink handwritings collected by a LIBS method, the laser pulse energy of the LIBS method is 30-40 mJ, and 25-35 pulsed lasers are accumulated as a spectrum; a spectrum with a high signal-to-noise ratio is obtained while the spectrum is collected by line scanning to reduce sample loss, wherein the moving speed of the three-dimensional platform is 1.5-2.5 mm / s; the spectrum data set of different ink handwritings is normalized, and the spectrum data set after normalization is reduced by principal component analysis to obtain the spectrum feature vector of different ink handwritings; a variety of machine learning algorithms are trained using a 10-fold cross validation method based on the spectrum feature vector of different ink handwritings, and the machine learning algorithm trained with the highest evaluation index based on the evaluation criteria of the model classification effect is used as a handwriting recognition model. The method can quickly and accurately identify handwriting, and the amount of inspection material is small.

Description

Technical Field

[0001] The present application relates to the technical field of spectral analysis and identification of types of signature pen handwriting, and in particular to a signature pen writing trace identification method based on LIBS combined with machine learning. Background Art

[0002] Black neutral signature pens are the most popular writing tools at home and abroad. Due to the complexity and instability of their ink composition, they are easily affected by temperature, humidity and storage conditions, resulting in changes in physical and chemical properties. The interaction between the ink and the base paper, as well as the confidentiality of the ink formula as a trade secret, greatly increase the difficulty of accurately identifying the ink. At present, there is no unified standard for the standardized detection method of the composition and characteristics of black signature pen ink, which undoubtedly brings certain uncertainties to the investigation and trial process of the case. Therefore, how to overcome these technical obstacles and establish a unified signature pen handwriting identification system has become an important issue that needs to be urgently solved in the current field of forensic science research.

[0003] At present, researchers have proposed some solutions for the identification of black neutral pen writing traces, such as paper spray mass spectrometry (PS-MS), thin layer chromatography (TLC) and capillary electrophoresis (CE), but there are still some problems in the application of these technologies to handwriting inspection, as follows:

[0004] The document Mcbride EM, Mach PM, Dhummakupt ES, et al. Paper Spray Ionization: Applications and Perspectives [J]. TrAC Trends in Analytical Chemistry, 2019, 118. discloses the paper spray mass spectrometry method: (1) The implementation process is: by placing the analyte directly on the paper, then wetting it with a solvent (usually methanol or a solution of methanol and deionized water), and applying a high voltage to the wet paper to ionize the analyte, and using a mass spectrometer to analyze the generated ionized spray. (2) Its advantages are: this method has a fast detection speed, high sensitivity and accurate results. (3) The technical disadvantage is: since the sample needs to be cut into triangular slices suitable for mass spectrometry analysis, and the detection area needs to be sprayed with methanol solvent to wet it, it will cause large-scale damage to the sample.

[0005] The literature Kravchenko AV, Makeeva DV, Tumkin I, et al. Rapid analysis and classification of blue ballpoint pen inks using high-performance thin-layer chromatography and digital color analysis [J]. Talanta, 2024, 275. discloses thin layer chromatography: (1) The implementation process is: extracting ink by punching or scraping from paper, manually preparing the extract in a solvent, and then applying it to the stationary phase. (2) The technical disadvantages are: the pretreatment method is complicated and time-consuming, and the sample is severely damaged; at the same time, TLC's identification of components depends on the comparison of fingerprint chromatograms, and it may be difficult to distinguish samples with similar ink compositions.

[0006] Regarding the literature Copper C, Erbach T, Mcdonough M, et al. Analysis of Ballpoint Pen Inks Directly from Paper using Capillary Electrophoresis [J]. Forensic Chemistry, 2019, 13., the capillary electrophoresis method is disclosed: (1) Its advantages are: only a very small amount of material is required for analysis, which means that the destruction of evidence documents is minimal and the integrity of the questioned documents is maintained; in addition, the CE method has a high resolution. (2) The technical disadvantage is that CE has high requirements for sample processing, especially for the extraction and purification of handwriting ink, which is more complicated. This involves multiple steps such as cutting ink, dissolving, centrifuging, filtering, etc., which increases the difficulty of operation and the source of error.

[0007] Although the above methods have achieved some results, they still have defects such as large amount of sample consumption, complicated processing procedures and long detection time.

[0008] Therefore, it is necessary to establish a fast and effective analysis method to identify the handwriting of black signing pens. Laser induced breakdown spectroscopy (LIBS) has the advantages of fast, real-time, in-situ, micro-destructive analysis, remote detection and simultaneous multi-element analysis. It is widely used in environmental science, deep space exploration, scientific archaeology, medical testing and other fields. LIBS technology uses high-energy pulsed lasers to excite trace substances on the surface of samples to produce plasma. During the cooling process of plasma, an emission spectrum is generated, and then the type of sample is identified by analyzing the spectrum.

[0009] However, there are still some challenges in the actual application of LIBS technology in handwriting recognition: on the one hand, due to the influence of writing habits, the ink of the signature pen and the properties of the paper, the thickness of the ink penetrating into the base paper is uneven, and how to optimize the working parameters to meet the analysis requirements of micro-damage is a problem. On the other hand, there are a large number of similar spectral lines in the handwriting spectra of different types of signature pens, which makes LIBS spectrum classification difficult. How to achieve fast and accurate handwriting recognition through machine learning methods is particularly important. Summary of the invention

[0010] The present invention provides a signing pen writing trace recognition method based on LIBS combined with machine learning. The method can be used to quickly and accurately recognize handwriting with a small amount of sample material.

[0011] The present invention provides a method for identifying writing traces of a signature pen based on LIBS combined with machine learning, comprising:

[0012] S1. Collecting spectrum data sets of different ink handwritings by LIBS method, wherein the laser pulse energy of the LIBS method is 30-40 mJ, and 20-35 pulse lasers are accumulated into one spectrum;

[0013] S2, normalizing the spectrum data set of different ink handwritings, and reducing the dimension of the normalized spectrum data set by principal component analysis to obtain the spectrum feature vectors of different ink handwritings;

[0014] S3, using 10-fold cross validation method to train multiple machine learning algorithms based on the spectral feature vectors of different ink handwritings, and taking the trained machine learning algorithm with the highest evaluation index as the handwriting recognition model based on the evaluation criteria of model classification effect;

[0015] When applied, the spectrum data set of the ink handwriting to be recognized is normalized and then the category of the ink handwriting to be recognized is obtained through principal component analysis and handwriting recognition model in turn.

[0016] The present invention controls the laser pulse energy and the number of pulse laser shots accumulated in each spectrum, so that LIBS can only burn the ink part and avoid burning the paper as much as possible. The generated plasma is as much as possible for the ink part, with less noise introduced by the paper, thereby providing a relatively pure training sample for subsequent machine learning. Even if the pulse energy is too high and the number of pulse laser shots accumulated in each spectrum is too much, it is easy to cause the paper to burn through, not only introducing a large amount of noise but also destroying the judicial documents. Through the regulation of the present invention, the laser ablation radius at the surface of the paper sample is 0.1-0.2mm. It can be seen that the method provided by the present invention consumes very little sample material. Through the method provided by the present invention, the LIBS method is applied to the spectrum collection of handwriting without ink for the first time.

[0017] Since there are a large number of similar spectral lines in the spectral lines of different ink handwritings collected by the LIBS method, the workload of manual identification is very large. The present invention uses a trained machine learning algorithm for the first time to identify the categories of different ink handwritings, thereby improving efficiency and accuracy.

[0018] Preferably, each laser pulse of the LIBS method acts on a different position of the corresponding ink mark. By limiting the position where each laser pulse acts, the noise introduced into the paper substrate and even the risk of overburning the paper are further reduced.

[0019] Preferably, the handwriting of different inks is placed on a three-dimensional platform, and when collecting the spectrum, the moving speed of the three-dimensional platform along the handwriting direction of each ink is 1.5-2.5 mm / s. By controlling the moving speed of the three-dimensional platform, it is ensured that each laser pulse acts on a different position of the corresponding ink handwriting.

[0020] Preferably, the spectrometer gate width of the LIBS method is 20-40 μs, and the delay time of the enhanced charge coupled device (ICCD) is 0.5-1 μs.

[0021] 0.5-1 μs after the burning of ink handwriting to generate plasma, the influence of bremsstrahlung and recombination radiation on the spectrum can be avoided. After collecting the radiation photons of the energy level transition of material particles in 20-40 μs through ICCD, a spectrum with a high signal-to-noise ratio can be obtained, thereby obtaining high-quality training samples to train the machine learning algorithm and obtain a handwriting recognition model with higher ink handwriting recognition ability.

[0022] Preferably, before collecting handwriting with different inks by LIBS method, the handwriting with different inks should be placed in a drying oven and kept away from light for at least 4 hours. Since different types of signature pen inks have different fluidity when in contact with paper fibers, resulting in different drying times, in order to avoid this factor affecting the spectrum collection, the paper with the handwriting of the signature pen is placed in a drying oven and kept away from light for more than 4 hours to make it completely dry.

[0023] Preferably, the spectrum data sets of different ink marks are normalized, and the normalization process is a maximum and minimum value normalization process. The normalization process reduces the influence of laser pulse energy fluctuation and air flow on the generated plasma.

[0024] Preferably, a plurality of machine learning algorithms are trained using a 10-fold cross validation method based on spectral feature vectors of different ink handwritings, wherein the plurality of machine learning algorithms include a quadratic support vector machine (Quadratic SVM), a linear discriminant analysis (LDA), a k-nearest neighbor (KNN) or a bagged decision tree (Bagged Trees).

[0025] Preferably, the trained machine learning algorithm with the highest evaluation index based on the evaluation criteria of the model classification effect is used as the handwriting recognition model, and the evaluation index includes sensitivity, accuracy and specificity.

[0026] Preferably, before training multiple machine learning algorithms, the spectral feature vectors of different ink handwritings are divided into a training sample set and a verification sample set, the multiple machine learning algorithms are trained with the training sample set, and the trained machine learning algorithms are randomly verified with the verification sample set.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] The present invention achieves minimal damage to physical evidence by controlling the laser pulse energy of LIBS and the accumulated amount of pulsed laser in each spectrum, and minimizes the introduction of paper noise during the plasma generation process, so as to obtain ink plasma as much as possible, thereby providing relatively pure training samples for subsequent machine algorithm training, and further being able to train a model with strong ink handwriting recognition capability.

[0029] The amount of spectral data of handwritings in different inks collected by the LIBS method is large, and there are a large number of similar spectral lines in the spectral lines of various ink handwritings. The present invention combines machine algorithms for the first time to quickly and accurately identify the type of handwriting from a large number of similar spectral lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present invention will be further described below in conjunction with the accompanying drawings.

[0031] Figure 1 A flowchart of a method for identifying signature pen writing traces based on LIBS combined with machine learning provided in a specific embodiment of the present invention;

[0032] Figure 2 Etching effect diagrams provided for Example 1, Example 2 and Comparative Example 1 of the present invention;

[0033] Figure 3 Spectrum diagram of 200 to 473 nm of 20 kinds of black neutral signature pen handwriting provided in a specific embodiment of the present invention;

[0034] Figure 4 LIBS spectrum of the handwriting of the 2# signature pen provided in the specific embodiment of the present invention;

[0035] Figure 5 The contribution rate and cumulative contribution rate diagram of the first 300 principal components after PCA dimensionality reduction provided by the specific embodiment of the present invention;

[0036] FIG6( a ) is a confusion matrix of the signature pen handwriting identification result of the PCA-Quadratic SVM model provided in a specific embodiment of the present invention;

[0037] FIG6( b ) is a confusion matrix of the signature pen handwriting identification result of the PCA-LDA model provided in a specific embodiment of the present invention;

[0038] FIG6( c ) is a confusion matrix of the signature pen handwriting identification result of the PCA-KNN model provided in a specific embodiment of the present invention;

[0039] FIG6( d ) is a confusion matrix of the signature pen handwriting identification result of the PCA-Bagged Trees model provided in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to enable those skilled in the art to better understand the scheme of the present invention, an embodiment of the present invention is specifically described below with reference to the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work should belong to the protection scope of the present invention.

[0041] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below in conjunction with the embodiments.

[0042] In order to solve the problems in the prior art that the inspection loss of samples is large, and due to the complex detection process, the time-consuming handwriting recognition is long and the accuracy is poor, the specific embodiment of the present invention can obtain the spectral information of handwritings with different inks under the condition of low inspection loss by adjusting the parameters of the LIBS method, and can quickly and accurately identify the categories of different handwritings in combination with machine learning, so as to achieve rapid and accurate identification of the types of handwritings of signature pens and meet the needs of judicial organs in adjudicating such cases. The specific embodiment of the present invention provides a method for identifying handwriting traces of signature pens based on LIBS combined with machine learning, such as Figure 1 As shown, the specific description is as follows:

[0043] S1. Collect spectral data sets of different ink handwritings by LIBS method: select 20 black neutral signing pens of different brands or grades: Deli 6600ES, Deli S18-B, Chenguang AGP02323, Chenguang ARPM2301, Chenguang AGPJ9001, Chenguang AGPK2301, Zebra JJZ58-BK, Zebra JJ29-R1-BK1, Chenguang AGPH5603, News NS570, Pilot BL-P50-B, Pilot LJU-10EF-B, Deli ARMOUS19, Chenguang AGPJ9101, Chenguang GP-1008, Chenguang ARPM2401, Chenguang AGPV3401, Chenguang AGPK3705, Mitsubishi UM-100, and number the handwritings written by the above black neutral signing pens as 1 # -20 # . The method provided in the specific embodiment of the present invention is used to # -20 # The base paper is ordinary office paper (Gold Flagship, copy paper, white). Use different black signature pens to mark 1×1cm on the paper at the same time. 2 Draw similar parallel lines in the area and deposit ink to form the same handwriting.

[0044] In a specific embodiment, different types of signature pen inks have different fluidity when in contact with paper fibers, resulting in different drying times. To prevent this factor from affecting spectral acquisition, the paper with signature pen marks is placed in a drying oven away from light for 4 hours to allow it to dry completely.

[0045] In a specific embodiment, the configuration and measurement principle of the LIBS device used in this embodiment are as follows: the ablation light source is a Q-switched Nd:YAG solid laser (TINY-D100L, Beijing Grace Company, pulse width of 6ns, repetition frequency of 10Hz), wavelength of 1064nm, maximum output energy of 140mJ, duration of about 6ns, repetition frequency of 10Hz. The handwriting sample is placed on an electrically controlled XYZ three-dimensional platform in a sample chamber (LIBSpector, LTB, Germany). The laser beam is coupled through the optical path and focused on the three-dimensional platform in the sample chamber, and vertically incident on the sample surface, ablating a small amount of sample and generating plasma. The strong light generated by the energy level transition when the plasma is cooled is coupled to the optical fiber through the collection optical path, and then transmitted to the fiber optic spectrometer (Aryelle Butterfly, LTB, Germany, spectral range: 190~430nm, resolution 5 pm) equipped with an ICCD detector (iStar 340t, Andor Technology, UK) to measure the LIBS spectrum.

[0046] In order to obtain the spectrum data set of different ink handwritings in the specific embodiment of the present invention, it is necessary to optimize the working parameters of the LIBS method, and the optimized working parameters include the optimization of the laser pulse energy, the spectrometer gate width and the delay time.

[0047] In judicial evidence inspection, it is necessary to consider the protection of documents. Therefore, in order to ensure that the characteristic elements have a higher signal intensity under the minimum amount of ablation and reduce the interference of the substrate paper, in Example 1, the pulse energy provided is optimized to 39.7mJ, and the laser ablation radius at the sample surface is 0.2mm. The optimized spectrometer gate width is 20 μs, and the delay time of ICCD is 0.5μs, which avoids the influence of bremsstrahlung and composite radiation on the spectrum to a certain extent and improves the signal-to-noise ratio. Before the experiment begins, the wavelength is first calibrated by Hg lamp. Under the same experimental conditions, LIBS spectral data of 20 handwriting samples are obtained. During spectrum acquisition, in order to prevent the surface of the handwriting paper from being excessively ablated and causing the paper to penetrate, and to reduce the influence of laser energy fluctuations, the sample is fixed on a three-dimensional platform, and the platform moves at a speed of 2mm / s along the handwriting writing direction, so that each laser pulse acts on a new position, and 30 pulses are accumulated as a spectrum. 60 spectra are collected for each handwriting sample as subsequent training samples, and the total number of spectra of 20 samples is 1200. The experiments were carried out under standard atmospheric pressure, indoor temperature of 25°C and relative humidity of 37%.

[0048] In Example 2, the spectrum acquisition parameters were changed, and the pulse energy provided was set to 30.3 mJ. The spectrometer gate width was set to 40 μs, the ICCD delay time was 1 μs, the platform movement speed was 2 mm / s, and 30 pulses were accumulated into one spectrum. The other experimental conditions were the same as in Example 1 (standard atmospheric pressure, room temperature 25°C, relative air humidity 37%).

[0049] In Comparative Example 1, the spectrum acquisition parameters were changed, and the pulse energy provided was set to 82.7 mJ. The spectrometer gate width was set to 10 μs, the ICCD delay time was set to 0.2 μs, the platform moving speed was set to 2 mm / s, and 60 pulses were accumulated into one spectrum so that each laser pulse acted on a new position. The other experimental conditions were the same as those in Example 1 (standard atmospheric pressure, indoor temperature 25°C, relative air humidity 37%).

[0050] like Figure 2As shown, in Comparative Example 1, when the pulse energy is 82.7 mJ, which exceeds the range of 30-40 mJ, the ablation radius at the sample surface is 1000 μm, which is much larger than the ablation radius of 0.1-0.2 mm in Example 1; at the same time, different from Example 1, when 60 pulse lasers are accumulated into a spectrum, when it exceeds 25-35, the length of the connection line formed by the point-shaped ablation pits during the movement, that is, the area of ​​the ablated sample, increases significantly, which cannot meet the application requirements of judicial handwriting inspection, while the parameters provided in Examples 1 and 2 can avoid ablation to the paper substrate when ablation is performed.

[0051] S2. Normalize the spectrum data set of different ink handwritings, and reduce the dimension of the normalized spectrum data set by principal component analysis (PCA) to obtain the spectrum feature vectors of different ink handwritings.

[0052] like Figure 3 As shown, 20 types of ink handwriting obtained by parameter collection using the LIBS method provided in Example 1 were analyzed, and the elemental composition of the samples was determined with reference to relevant data in the database of the National Institute of Standards and Technology (NIST). During the writing process, due to the properties of ink, paper, and the different ways of using pens on paper, the penetration depth of ink deposited on paper is irregular, resulting in uneven layers. At the same time, due to the influence of laser pulse energy fluctuations and air flow on the generated plasma, the absolute value of the LIBS spectrum intensity of the sample will fluctuate to a certain extent. Therefore, in order to eliminate these influences to a certain extent and facilitate the comparison of spectra of different types of handwriting samples, the data is normalized to the maximum and minimum values, and the spectral data is mapped to the interval [0, 1]. By Figure 3 The comparison shows that there are a lot of similar characteristic spectra between the handwritings of the 20 black signature pens, indicating that there are a lot of similar chemical components in different inks. Therefore, elemental analysis was performed on the spectrum of one of the handwritings. Figure 4 For 2 #In the LIBS spectrum of the handwriting sample, characteristic spectral lines of elements such as C, Mg, Cu, Al, Ca, Si, and Fe can be observed in the spectrum. A CN molecular band with Δv = 0 is also observed, which is formed by CO2 and N2 in the air. At the same time, there are also slight differences in the emission line intensities of some elements in different handwritings, such as Fe, Cu, Al, etc. This may be because commercial black inks contain purple, blue, and even red pigments in addition to black pigments, considering different writing effects such as adhesion, storage time, and drying time; at the same time, some black inks also contain synthetic black dye mixtures made by heating a mixture of nitrobenzene, aniline, and hydrochloric acid in the presence of Cu or Fe. The incorporation of these different components leads to differences in the content of these elements. From the above analysis, it can be seen that there are indeed differences in the LIBS spectra of different handwriting samples, but it is difficult to distinguish them by observation alone. At the same time, the workload of labeling the characteristic spectral lines of each element in the spectrum is heavy, so it is necessary to conduct data mining and analysis with the assistance of machine learning.

[0053] The specific embodiment of the present invention uses PCA to perform feature dimension reduction on spectral data. PCA is used to perform dimension reduction on LIBS spectral data of 20 different handwriting samples, where there are 60 groups of LIBS spectral data of each handwriting, totaling 1200 groups. Figure 5 As shown in the figure, a represents the change curve of the individual contribution rate of different principal components, and b represents the change of the cumulative contribution rate of the principal components with the number of principal components. After PCA dimension reduction, the cumulative contribution rate of the first 300 principal components (PC1, PC2, PC3, ..., PC300) reached 72.69%. They were used as the characteristic spectral lines of handwriting LIBS to construct the characteristic vectors of LIBS spectra of different handwriting samples, forming a 1200×300 data matrix. These characteristic vectors were used as the input of four machine learning algorithms: Quadratic SVM, LDA, KNN, and Bagged Trees.

[0054] S3. Based on the spectral feature vectors of handwritings with different inks, a 10-fold cross-validation method is used to train multiple machine learning algorithms. Based on the evaluation criteria of the model classification effect, the trained machine learning algorithm with the highest evaluation index is used as the handwriting recognition model.

[0055] The specific embodiment of the present invention adopts a 10-fold cross validation method to randomly divide the sample set into a training sample set and a test sample set at a ratio of 2:1. That is, the sample is randomly divided into 3 equal parts, and each time 1 part is used as a validation sample set, and the remaining 2 parts are used as training sample sets for training, and the average accuracy of the 10 results is used as an estimate of the accuracy of the training set. Finally, the results of handwriting sample recognition are shown by the confusion matrix, and these results are shown in Figures 6(a) to 6(d) respectively.

[0056] Based on four machine learning algorithms, PCA-Quadratic SVM, PCA-LDA, PCA-KNN, and PCA-Bagged Trees, 20 different types of signature pen handwriting were recognized. The cross-validation recognition results are shown in Table 1. The sensitivity, precision, and specificity are also marked in the table. Sensitivity indicates the proportion of handwriting predicted correctly by the model to the real handwriting category; precision indicates the proportion of handwriting categories predicted correctly by the model to the category predicted by the model; specificity indicates the proportion of other categories predicted correctly by the model to other handwriting categories predicted by the model. These three are all discriminant indicators of the classification effect of the machine learning model. The higher the value, the better the classification effect of the model.

[0057] The data in Table 1 show that the accuracy of these four recognition models exceeds 93.91%, among which the PCA-LDA model has the best recognition result, with an accuracy of 99.91% and a sensitivity of 5. # Handwriting samples, precision and specificity except 8 # The results showed that PCA effectively extracted the most representative features in the data, enabling the machine learning algorithm to better distinguish different categories in low-dimensional space. Compared with complex nonlinear algorithms (such as SVM), the calculation process of the LDA machine learning model is relatively simple. It only needs to calculate the class mean, intra-class variance and inter-class variance matrix, and has higher calculation efficiency when processing large-scale data. At the same time, LDA can construct linear discriminant boundaries and can effectively classify data with linearly separable features. The results show that the PCA-LDA model has good identification performance for different types of black signature pen handwriting, and the combination of LIBS technology and machine learning algorithms can achieve high-precision handwriting recognition.

[0058] When applied, the spectrum data set of the ink handwriting to be recognized is normalized, and then the category of the ink handwriting to be recognized is obtained through principal component analysis and handwriting recognition model.

[0059] The specific embodiment of the present invention proposes a model for identifying the handwriting of a black signing pen using LIBS technology combined with a machine learning algorithm, which realizes efficient and accurate qualitative identification of handwriting. The experiment collects handwriting spectra of different types of black neutral signing pens through a LIBS device, and identifies the characteristic elements of the spectra. Since there are a large number of similar spectral lines in the spectrum, the necessity of using a machine learning algorithm is proposed. The machine learning model is applied to identify the handwriting samples after reducing the data dimension using the PCA method. The results show that the machine learning model has a high recognition accuracy rate, as well as high sensitivity, precision and specificity. LIBS technology can perform micro-destructive analysis on the ink components of black signing pens, and the machine learning algorithm provides a way to efficiently identify these handwritings. The application of LIBS combined with machine learning methods in handwriting identification has good feasibility and has high practical value in the field of forensic science.

[0060] Compared with traditional TLC, MS and CE methods, it does not require long sample processing time and complex experimental steps, which reduces the impact of human factors and sample processing process to a certain extent. It can complete the spectral acquisition of samples in a short time, and combined with the automated data processing and analysis of machine learning algorithms, the detection process is more efficient and can achieve highly accurate handwriting identification.

[0061] The specific embodiment of the present invention optimizes parameters such as laser energy, has low destructiveness to samples, and only requires a trace amount of material for analysis, thereby ensuring the integrity of physical evidence. Considering the need to preserve original physical evidence in forensic identification, the present invention has obvious advantages over other chemical analysis methods that require the consumption of large amounts of samples.

[0062] The method provided by the specific embodiment of the present invention has strong robustness and can effectively deal with the noise and irregularity in the spectrum of the signature pen trace. Due to the properties of ink and paper and the inhomogeneity of laser pulses, the spectral data has noise and fluctuations. By adopting a machine learning algorithm, the classification model can be dynamically optimized according to the sample data, and cross-validation is used to improve the model stability, so that it is suitable for handwriting detection under different conditions, and shows good identification effect in practical applications.

[0063] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0064] Table 1. Classification accuracy, sensitivity, precision and specificity of different models

[0065]

Claims

1. A method for identifying writing traces of a signature pen based on LIBS combined with machine learning, characterized in that: include: S1. Collecting spectrum data sets of different ink handwritings by LIBS method, wherein the laser pulse energy of the LIBS method is 30-40 mJ, and 20-35 pulsed lasers are accumulated into one spectrum; S2, normalizing the spectrum data set of different ink handwritings, and reducing the dimension of the normalized spectrum data set by principal component analysis to obtain spectrum feature vectors of different ink handwritings; S3, using a 10-fold cross validation method to train multiple machine learning algorithms based on the spectral feature vectors of different ink handwritings, and using the trained machine learning algorithm with the highest evaluation index as the handwriting recognition model based on the evaluation criteria of the model classification effect; When applied, the spectrum data set of the ink handwriting to be identified is normalized and then the category of the ink handwriting to be identified is obtained by principal component analysis and handwriting recognition model in turn; The handwritings of different inks are placed on a three-dimensional platform, and when collecting spectra, the moving speed of the three-dimensional platform along the handwriting direction of each ink is 1.5-2.5 mm / s; The spectrometer gate width of the LIBS method is 20-40 μs, and the delay time of the ICCD is 0.5-1 μs.

2. The method for identifying signature pen writing traces based on LIBS combined with machine learning according to claim 1, characterized in that: Each laser pulse of the LIBS method acts on a different position of the corresponding ink mark.

3. The method for identifying signature pen writing traces based on LIBS combined with machine learning according to claim 1, characterized in that: Before collecting handwriting with different inks by the LIBS method, the handwriting with different inks was placed in a drying oven away from light for no less than 4 hours.

4. The method for identifying signature pen writing traces based on LIBS combined with machine learning according to claim 1, characterized in that: The spectrum data sets of different ink handwritings are normalized, and the normalization process is a maximum and minimum value normalization process.

5. The method for identifying signature pen writing traces based on LIBS combined with machine learning according to claim 1, characterized in that: A 10-fold cross validation method is used to train multiple machine learning algorithms based on spectral feature vectors of different ink handwritings. The multiple machine learning algorithms include quadratic support vector machine, linear discriminant analysis, k-nearest neighbor or bagged decision tree.

6. The method for identifying signature pen writing traces based on LIBS combined with machine learning according to claim 1, characterized in that: The trained machine learning algorithm with the highest evaluation index is used as the handwriting recognition model based on the evaluation criteria of the model classification effect, and the evaluation indexes include sensitivity, accuracy and specificity.

7. The method for identifying signature pen writing traces based on LIBS combined with machine learning according to claim 1, characterized in that: Before training multiple machine learning algorithms, the spectral feature vectors of different ink handwritings are divided into a training sample set and a verification sample set. The multiple machine learning algorithms are trained with the training sample set, and the trained machine learning algorithms are randomly verified with the verification sample set.

Images