Method for rapidly identifying luxury fiber components

Abstract

The invention discloses a method for rapidly identifying luxury fiber components. The method comprises the following steps: (A) pretreating an actual sample; (B) detecting the sample by adopting a screen mesh collision-microtube plasma ionization mass spectrometry to obtain mass spectrum data; and (C) comparing the obtained mass spectrum data with a random forest classification model to obtain an identification result. The invention further discloses an establishment process of the random forest classification model. The method has high customizability and instrument compatibility, experimental operation is simple, the ionization process is stable, data consistency is good, the requirement of customs for on-site rapid detection of textiles can be met, and a feasible technical means is provided for data transverse comparison, cross-laboratory standardized large database establishment and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of analytical chemistry, in particular to a method for rapid identification of luxury product fiber components by micro-tube plasma ionization mass spectrometry combined with machine learning. BACKGROUND

[0002] At present, the domestic related industry identifies the textile fiber components usually uses the method specified in GB / T 2910 series national standards and FZ / T01057 series, SN / T 3236-2012 and other industry standards. Among them, FZ / T 01057 parts 2, 4, 5 specifies the chemical qualitative analysis methods such as combustion method, dissolution method, color reaction method. These methods are simple in operation, short in analysis time, but cannot accurately quantify, and the determination of test results depends on the experience of detection personnel, which is highly subjective. FZ / T 01057 parts 3, 6, 7, 9 specifies the physical methods such as microscope method, melting point method, density gradient method, birefringence method, etc. These methods can get relatively accurate analysis results by comparing with standard atlas or data, but the sample pretreatment process is relatively complex, which is not suitable for on-site detection. GB / T 2910 series standards adopt ISO 1833-1:2006 standard, which specifies the quantitative chemical analysis methods for various two-component fiber mixtures. These methods are accurate in quantification, but the sample pretreatment process is more time-consuming, and a large amount of auxiliary reagents are consumed, which is difficult to meet the basic requirements of green analytical chemistry, and there is a lot of room for improvement. In addition to the above methods, a series of textile identification methods based on modern instrumental analysis technology have been developed and applied. For example, the textile fiber identification test method based on infrared and Raman spectroscopy is specified as a standard method in FZ / T 01057.8 and SN / T 3236-2012. This kind of method has wide application range and is convenient to operate, but is easily disturbed by water, carbon dioxide, dye molecules, etc., and increasing the spatial resolution will significantly increase the risk of damaging the sample. Thermal analysis method can also identify fiber types by comparing the thermal characteristics of different fibers, which has the advantages of small sample size, simple operation, high accuracy, etc., but often relies on special precision instruments and stable laboratory environment, and the single analysis takes a long time. Traditional chromatography-mass spectrometry has high sensitivity and specificity, and is suitable for the analysis of trace samples and complex mixtures, but usually requires a complex sample pretreatment process, which is time-consuming and environmentally polluting. SUMMARY

[0003] In order to solve the problems existing in the prior art, the present application establishes a new method suitable for rapid identification of fiber types of worn luxury product samples. The method uses electric iron burning-sieve collision plasma ionization mass spectrometry to collect textile fiber component mass spectrum data, and combines random forest algorithm to realize rapid and accurate identification of sample fiber types.

[0004] A method for rapidly identifying the fiber components of luxury goods, comprising the following steps:

[0005] (A) Pretreatment of the actual sample;

[0006] (B) Detecting the sample by using the screen mesh collision-micro tube plasma ionization mass spectrometry to obtain mass spectrum data;

[0007] (C) Obtaining the identification result after processing the mass spectrum data of the actual sample by using the random forest classification model.

[0008] The method for rapidly identifying the fiber components of luxury goods, wherein the establishment process of the random forest classification model comprises the following steps:

[0009] (a) Pretreatment of the standard sample;

[0010] (b) Detecting the standard sample by using the screen mesh collision-micro tube plasma ionization mass spectrometry to obtain original mass spectrum data;

[0011] (c) Preprocessing the original mass spectrum data to obtain a mass spectrum data set, which is a list file for saving the mass number and corresponding signal intensity data;

[0012] (d) Evaluating the quality of the mass spectrum data set by using an unsupervised classification method;

[0013] (e) Establishing the random forest classification model.

[0014] The method for rapidly identifying the fiber components of luxury goods, wherein the preprocessing method of the original mass spectrum data in step (c) specifically comprises the following steps:

[0015] (I) Extracting the mass spectrum information in the mass axis range of 100.0-550.0 Da, including the mass number and corresponding signal intensity;

[0016] (II) Using local peak detection combined with signal-to-noise ratio screening to reduce noise, setting the noise estimation window width to 50 Da, the peak detection minimum intensity threshold to 10 5 cps, and the noise threshold to 3 times the signal-to-noise ratio;

[0017] (III) Normalizing the maximum value of the retained signal intensity to the range of 0 to 1;

[0018] (IV) Taking 100.0 and 550.0 Da as the starting point and the ending point of binning respectively, 0.1 Da as the binning step, and ±0.05 Da as the binning radius, establishing continuous non-overlapping global binning;

[0019] (VI) Assigning the retained ion signals to each bin according to the mass number, and retaining the maximum value of multiple signal intensities in the same bin.

[0020] (VII) merging all sample mass spectrum information to form a mass spectrum dataset; the obtained mass spectrum dataset needs to be subjected to quality evaluation in step (d) and meet the requirements before being used for subsequent random forest model training.

[0021] The method for rapidly identifying the fiber components of luxury goods according to the present application, wherein the unsupervised classification method is used to evaluate the quality of the mass spectrum dataset in step (d) specifically includes the following steps: first, the mass spectrum dataset is reduced to 25 dimensions by PCA to compress redundant information, retain the core data structure, and reduce the computational complexity; second, the K-means clustering algorithm is used to perform unsupervised classification on the reduced dataset to evaluate its data clustering property, and the number of clusters used is optimized according to the silhouette coefficient.

[0022] The quality of the dataset is evaluated by the silhouette coefficient, adjusted Rand index, Davies-Bouldin index, and Calinski-Harabasz index, and the value ranges of these indexes are as follows:

[0023] The silhouette coefficient (Silhouette Coefficient) has a value range of [-1, 1]. When the silhouette coefficient is greater than 0.5, the sample clustering is clear; when the silhouette coefficient is between 0.25 and 0.5, there is a clustering structure but it is relatively vague; and when the silhouette coefficient is less than 0.25, the clustering effect is poor or the data has no clear structure.

[0024] The adjusted Rand index (ARI) has a value range of [-1, 1]. When the ARI is greater than 0.5, the clustering result is highly consistent with the standard sample category label; when the ARI is between 0.3 and 0.5, the clustering result has moderate consistency with the label; and when the ARI is less than 0.3, the clustering result has poor consistency with the label.

[0025] The Davies-Bouldin index (DBI) has a value range of [0, +∞). When the DBI is less than 0.5, the separation degree between the clustering clusters is high and the clusters are tight; when the DBI is between 0.5 and 1.0, the clustering effect is moderate; and when the DBI is greater than 1.0, the clusters are seriously overlapped or the clusters are loose.

[0026] The Calinski-Harabasz index (CHI) has a value range of [0, +∞). The larger the CHI value is, the better; and when the CHI value is less than 50, the data structure is not obvious.

[0027] The method for rapidly identifying the fiber components of luxury goods according to the present application, wherein the method for establishing the random forest classification model in step (e) includes the following steps:

[0028] A random forest model based on the Scikit-learn machine learning library is used to identify the category of a textile product. In the model training process, the data set is standardized and dimensionality reduction is performed using principal component analysis. The initial hyperparameters of the random forest model for actual sample identification are as follows: the number of decision trees is 100; the maximum depth of a single decision tree is limited to 10; the maximum number of features algorithm is log2, that is, log2 features are randomly selected when splitting a node; the minimum number of samples for node splitting is set to 2; the minimum number of samples for leaf nodes is 4; the training subset is generated using the bootstrap sampling method; and the out-of-bag data is retained.

[0029] The quality of the random forest model is evaluated by the out-of-bag score, 5-fold cross-validation average accuracy, precision, recall, and F1 score. The value range of these indicators is as follows:

[0030] The out-of-bag score has a value range of [0, 1] (or in percentage form). When the out-of-bag score is greater than 0.85, the model has strong generalization ability and the data fitting and generalization balance is good; when the out-of-bag score is between 0.7 and 0.85, the model performs moderately and may have slight overfitting or underfitting; when the out-of-bag score is less than 0.7, the model performance is poor.

[0031] The 5-fold cross-validation average accuracy has a value range of [0, 1] (or in percentage form). When the average accuracy is greater than 0.85, the model has strong classification ability; when the average accuracy is between 0.7 and 0.85, the model performs moderately; and when the average accuracy is less than 0.7, the model fails to capture effective patterns.

[0032] The precision (Precision) has a value range of [0, 1] (or in percentage form). When the precision is greater than 0.8, the model's prediction reliability for positive classes is high; when the precision is between 0.6 and 0.8, the model's credibility is moderate; and when the precision is less than 0.6, the model has more misjudgments.

[0033] The recall (Recall) has a value range of [0, 1] (or in percentage form). When the recall is greater than 0.8, the model can effectively capture positive samples; when the recall is between 0.6 and 0.8, the model's coverage is moderate; and when the recall is less than 0.6, the model has serious omissions.

[0034] The F1 score (F1-Score) has a value range of [0, 1] (or in percentage form). When the F1 score is greater than 0.8, the precision and recall are balanced and excellent; when the F1 score is between 0.6 and 0.8, the model's balance is moderate; and when the F1 score is less than 0.6, the model's overall performance is poor.

[0035] To ensure the repeatability of the experiment, the random number seed is fixed throughout the modeling process, including hyperparameter optimization, model training and validation stage.

[0036] The method for rapidly identifying the fiber components of luxury goods according to the present application, wherein in step (C), the mass spectrum data of the actual sample is processed by using the trained random forest model to obtain a classification result (the fiber type to which the sample belongs), the confidence is calculated by counting the parallel identification results, and finally the conclusion is obtained.

[0037] The result statistical method is as follows, n times of parallel identification are performed:

[0038] If the results of n times of parallel identification are all the same, the actual sample type is consistent with the identification result, and the confidence is 100%;

[0039] If m times of the results of n times of parallel identification are the same, m < n, the actual sample type is consistent with the m times of the same identification result, the confidence is m / n x 100%, and when the confidence is greater than 60%, it is considered that the identification result has a relatively high reliability;

[0040] If the results of n times of parallel identification are all different, the actual sample type cannot be determined, and the mass spectrum data needs to be collected again for identification.

[0041] The method for rapidly identifying the fiber components of luxury goods according to the present application, wherein the sample in step (A) and step (a) needs to ensure that the sample taken is representative and the weight is not less than 5 mg, and before sampling, the sample surface is cleaned by using a dust-free paper dipped in methanol.

[0042] The method for rapidly identifying the fiber components of luxury goods according to the present application, wherein the screen mesh collision-microtube plasma ionization mass spectrometry (MC-μTP-MS) in step (B) and step (b) adopts a screen mesh collision-microtube plasma ionization device, when detecting, a handheld electric iron is used to burn the textile sample at 350-450 DEG C to generate smoke for MC-μTP-MS analysis, each standard sample is sampled at least 6 times in parallel, and the actual sample is sampled at least 3 times in parallel; when the total ion current signal intensity of the mass spectrum decreases, or ions from the sample smoke appear in the mass spectrum background, the capillary, screen mesh and electric iron head should be cleaned with methanol before sampling.

[0043] The method for rapidly identifying the fiber components of luxury goods according to the present application, wherein the working conditions of the screen mesh collision-microtube plasma ionization device are: an alternating voltage of 2.0 kV, a frequency of 10 kHz, a carrier gas flow rate of 0.5 mL / min, a negative pressure of -76 MPa between the device and the mass spectrometer, and a room temperature; and the working conditions of the electrostatic field orbitrap high-resolution mass spectrometer are: an ion transmission tube temperature of 320 DEG C, an S-lens voltage of 55 V, a negative ion mode, a full scan mode for data acquisition, a scanning range of m / z 100-750, an automatic gain control target value of 10 6 , a maximum injection time of 50 ms, and a microscan number of 1 per second.

[0044] The method for rapidly identifying the fiber components of luxury goods according to the present application, wherein the screen mesh collision-microtube plasma ionization device comprises a microtube plasma generator, a molecular ion reactor and a mass spectrometer interface connected in sequence; the microtube plasma generator comprises a first T-shaped tee, the molecular ion reactor comprises a second T-shaped tee, and the straight pipes of the second T-shaped tee are connected to one end of the straight pipe of the first T-shaped tee and the mass spectrometer interface, respectively.

[0045] A first connecting pipe is used to connect between the first T-shaped tee and the second T-shaped tee, a second flexible sleeve and a fused quartz capillary are arranged in the interior of the connecting position, a first flexible sleeve is arranged in the interior of the other end of the straight pipe of the first T-shaped tee; the device further comprises a flat-end stainless steel hollow needle, one end of which is connected to high-frequency high-voltage alternating current, and the other end of which penetrates through the first flexible sleeve and the second flexible sleeve and is arranged in the fused quartz capillary; the vertical pipe of the first T-shaped tee is connected to a helium bottle through a helium pipe; the vertical pipe of the second T-shaped tee is connected to a gas collection hood; a copper screen mesh is arranged at the connecting position between the second T-shaped tee and the mass spectrometer interface and is fixed in the second connecting pipe; the shell is connected to the second T-shaped tee through the second connecting pipe.

[0046] The two ends of the second flexible sleeve are arranged in the first T-shaped tee and the first connecting pipe, respectively.

[0047] One end of the fused quartz capillary is arranged in the first T-shaped tee, and the other end thereof is arranged in the middle of the second T-shaped tee through the second flexible sleeve; in the interior of the second T-shaped tee, the distance between the tip of the flat-end stainless steel hollow needle and the end of the fused quartz capillary is 5 mm.

[0048] A gap is left between the flat-end stainless steel hollow needle and the fused quartz capillary for gas to pass through.

[0049] The mass spectrometer interface comprises a housing connected with the second T-shaped tee and an ion transmission tube of a mass spectrometer, and a vacuum pump connected with the housing;

[0050] The high-frequency high-voltage alternating current is provided by an adjustable frequency high-voltage alternating current source; the first T-shaped tee is a polyether ether ketone material threaded tee, and the threaded interface size is 1 / 16 inch; the helium pipeline is a polyethylene hose with an inner diameter of 1 / 16 inch; the flat head stainless steel hollow needle has a size of 30G; the first flexible sleeve and the second flexible sleeve are polyether ether ketone materials with an outer diameter of 1 / 16 inch and an inner diameter of 310-650 microns; the fused quartz capillary has an outer diameter of 650 microns and an inner diameter of 530 microns; the first connecting pipe and the second connecting pipe are polyethylene material hard pipes with an outer diameter of 6 mm; the second T-shaped tee is a polybutylene terephthalate material quick connector tee with an interface inner diameter of 6 mm; the copper screen has a diameter of 4-6 mm and a mesh number of 200 meshes; the gas collection cover is a hemispherical plastic material with a sampling port diameter of 5 cm; the housing of the mass spectrometer interface is a stainless steel material, the interface inner diameter connected with the second T-shaped tee is 6 mm, and the pagoda interface specification connected with the vacuum pump is 8 mm; the ion transmission tube is a stainless steel material with an inner diameter of 40 microns.

[0051] The method for rapidly identifying the fiber components of luxury goods of the present application is different from the prior art in that:

[0052] The method for rapidly identifying the fiber components of luxury goods of the present application establishes a textile fiber component rapid detection method based on electric iron burning-screen collision plasma ionization mass spectrometry, and realizes rapid identification of the fiber types of imported wearing class luxury goods samples in combination with a random forest algorithm. The method uses a self-made ionization device, has strong customizability and instrument compatibility, is simple in experimental operation, stable in ionization process, and good in data consistency, can meet the needs of the customs for on-site rapid detection of textiles, and provides a feasible technical means for developing data horizontal comparison, establishing a standardized large database across laboratories, etc.

[0053] The application establishes a rapid detection technology based on electric iron burning-screen collision-micro tube plasma ionization mass spectrometry (MC-μTP-MS) and random forest algorithm. The method does not need sample pretreatment, is simple to operate, and is suitable for rapid acquisition of textile mass spectrum characteristic data. The mass spectrum data of 27 identified textile samples are collected under the optimized conditions, to form a data set containing 181 groups of data and 4501 characteristic dimensions, which are used for training the random forest prediction model. The obtained model has an out-of-bag score of 0.9762, a 5-fold cross-validation accuracy of 0.9683 (±0.0299), and an F1 score of cotton, cashmere / wool, polyester fiber and acrylic fiber all higher than 0.889, and has good classification performance. The model is used to analyze 30 customs inspection samples of wearing class luxury goods, and it is found that the identification components of 28 samples are consistent with the identified components; the identification results of two samples are consistent with the identified components but have low confidence, which may be false positive samples or model overfitting. The results show that the method has great application potential in the quality and safety supervision of imported luxury goods, and provides a feasible idea for the comprehensive evaluation of textile quality and safety.

[0054] The method for rapidly identifying fiber components of luxury goods of the application will be further described below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 Figure 1 is a structural schematic diagram of the screen collision-micro tube plasma ionization device in the application;

[0056] Figure 2 Figure 3 is an optimization diagram of the MC-μTP ionization device alternating voltage (A), carrier gas flow rate (B) and VAPUR interface negative pressure (C) in the application;

[0057] Figure 3 Figure 5 is a representative mass spectrum diagram of the textile standard sample with different fiber components in the application;

[0058] Figure 4 Figure 7 is a PCA gravel diagram (A) and a K-means clustering three-dimensional visualization diagram (B) based on the textile mass spectrum data set in the application;

[0059] Figure 5 Figure 9 is a classification report heat map (A) and a test set confusion matrix (B) of the trained random forest model in the application;

[0060] All English translations appearing in the drawings of the application are as follows:

[0061] Voltage: voltage;

[0062] carrier gas flow rate: carrier gas flow rate;

[0063] negative pressure: negative pressure

[0064] intensity: intensity

[0065] Relative intensity: relative intensity

[0066] Mass: mass

[0067] Explained Variance Ratio: explained variance ratio

[0068] Principal Component: principal component

[0069] Variance Explained: explained variance

[0070] Culmulative Explained Variance: cumulative explained variance

[0071] Acrylic: acrylic

[0072] Wool / Cashmere: wool / cashmere

[0073] Cotton: cotton

[0074] Polyester: polyester

[0075] Cluster centers: cluster centers

[0076] Precision: precision

[0077] Recall: recall

[0078] F1-score: f1-score DETAILED DESCRIPTION

[0079] I. Instruments and reagents

[0080] CTP-2000K low temperature plasma experimental power (Nanjing Sumen Plasma Technology Co., Ltd.); QExactive electrostatic field orbitrap mass spectrometer, equipped with Exactive Tune mass spectrometer control software and Xcalibur QualBrowser data processing software (Thermo Scientific, USA); VAPUR interface (Ionsense, USA); 2652C-50 dry vacuum pump (Welch, USA); MF5712-N-200 gas mass flow meter (Siargo, USA); WSD 71 handheld electric iron (Weller, Germany).

[0081] The data preprocessing script, the compilation and running environment of the random forest model are Python 3.10.4, the operating system is Windows 10 Professional (64-bit), the computer processor is Intel Core TM i5-13600KF, and the machine band RAM is 32.0 GB.

[0082] T-shaped PEEK tee (1 each of threaded and pneumatic quick connector), 30G flat head stainless steel hollow needle, fused silica capillary, various specifications of PEEK threaded fasteners and sleeves, and various specifications of polyethylene tubes (purchased online); 200-mesh copper screen (Ionsense Company, USA).

[0083] 27 textile standard samples were provided and identified by the Inner Mongolia Autonomous Region Market Supervision Administration Evaluation and Inspection Center, and the specific information is listed in Table 1. 30 actual samples of wearing luxury goods were provided by the Shanghai Customs Industrial Products and Raw Materials Detection Technology Center, and the specific information is listed in Table 3. Pierce ESI anion calibration solution (Thermo Scientific Company, USA); High-purity helium (99.999%, Beijing Beixing Joint Gas Co., Ltd.); Methanol (chromatographically pure, Shanghai Anpu Experimental Technology Co., Ltd.); Dust-free paper (Kimberly-Clark Company, USA).

[0084] Table 1 Textile standard samples for establishing mass spectrometry data set

[0085]

[0086]

[0087] II. Experimental methods

[0088] The screen collision-microtube plasma ionization device was constructed as follows:

[0089] As shown in Figure 1 , the screen collision-microtube plasma ionization device comprises a microtube plasma generator, a molecular ion reactor and a mass spectrometer interface connected in sequence; the microtube plasma generator comprises a first T-shaped tee 1, the molecular ion reactor comprises a second T-shaped tee 2, and the straight pipes at both ends of the second T-shaped tee 2 are connected to one end of the straight pipe of the first T-shaped tee 1 and the mass spectrometer interface, respectively.

[0090] The first connecting pipe 10 is connected between the first T-shaped tee 1 and the second T-shaped tee 2, and the inside of the connecting position is provided with the second flexible sleeve 3 and the fused quartz capillary tube 4; the inside of the other end of the straight pipe of the first T-shaped tee 1 is provided with the first flexible sleeve 5; the device further comprises a flat-end stainless steel hollow needle 6, one end of which is connected with high-frequency high-voltage alternating current, and the other end of which penetrates through the first flexible sleeve 5 and the second flexible sleeve 3 and is arranged in the fused quartz capillary tube 4; the vertical pipe of the first T-shaped tee 1 is connected with a helium bottle through a helium pipe 7; the vertical pipe of the second T-shaped tee 2 is connected with a gas collecting cover 8; a copper screen 9 is arranged at the connecting position of the second T-shaped tee 2 and a mass spectrometer interface, and is fixed in the second connecting pipe 11; an outer shell 12 is connected with the second T-shaped tee 2 through the second connecting pipe 11.

[0091] The two ends of the second flexible sleeve 3 are arranged in the first T-shaped tee 1 and the first connecting pipe 10 respectively.

[0092] One end of the fused quartz capillary tube 4 is arranged in the first T-shaped tee 1, and the other end is arranged in the middle of the second T-shaped tee 2 through the second flexible sleeve 3; in the inside of the second T-shaped tee 2, the distance between the tip of the flat-end stainless steel hollow needle 6 and the end of the fused quartz capillary tube 4 is 5 mm.

[0093] A gap is left between the flat-end stainless steel hollow needle 6 and the fused quartz capillary tube 4 for gas to pass through.

[0094] The mass spectrometer interface comprises the outer shell 12 and a vacuum pump 14 connected with the outer shell 12, and the outer shell 12 is connected with the second T-shaped tee 2 and an ion transmission pipe 13 of a mass spectrometer.

[0095] The high-frequency high-voltage alternating current is provided by an adjustable frequency high-voltage alternating current source; the first T-shaped tee 1 is a polyether ether ketone material threaded tee, and the sizes of the threaded interfaces are all 1 / 16 inch; the helium pipe 7 is a polyethylene hose with an inner diameter of 1 / 16 inch; the size of the flat-end stainless steel hollow needle 6 is 30G; the first flexible sleeve 5 and the second flexible sleeve 3 are polyether ether ketone materials with an outer diameter of 1 / 16 inch and an inner diameter of 310-650 microns; the fused quartz capillary tube 4 has an outer diameter of 650 microns and an inner diameter of 530 microns; the first connecting pipe 10 and the second connecting pipe 11 are polyethylene hard pipes with an outer diameter of 6 mm; the second T-shaped tee 2 is a polybutylene terephthalate material quick connector tee with an interface inner diameter of 6 mm; the copper screen 9 has a diameter of 4 mm and a mesh number of 200; the gas collecting cover 8 is a hemispherical plastic material with a sampling port diameter of 5 cm; the outer shell 12 of the mass spectrometer interface is a stainless steel material, the interface inner diameter connected with the second T-shaped tee 2 is 6 mm, and the pagoda interface specification connected with the vacuum pump 14 is 8 mm; the ion transmission pipe 13 is a stainless steel material with an inner diameter of 40 microns. In the assembly process, the stainless steel needle and the capillary tube axis need to be aligned to avoid plasma deviation.

[0096] The device can realize real-time collection and direct ionization of aerosol, transmit sample ions to a mass spectrometer for analysis, and obtain chemical component information of the aerosol. The device solves problems in a traditional aerosol mass spectrum analysis process, such as complex sample processing, difficulty in direct sampling, and low ionization efficiency. A porous screen is embedded in a transmission channel to diffuse fine beam-shaped plasma into a planar shape, greatly improving the contact probability of neutral molecules and active plasma, thereby improving the ionization efficiency and the detection sensitivity, which is 5-10 times higher than that of a conventional microtube plasma mode.

[0097] The device is composed of a microtube plasma generator, a molecular ion reactor, and a mass spectrometer interface. In the plasma generator, helium gas is broken down to generate plasma under the action of high-frequency high-voltage alternating current. A flat-end stainless steel hollow needle is used as an internal electrode to significantly reduce background noise signals. The plasma enters the molecular ion reactor and is diffused into a planar shape by the porous screen, greatly improving the reaction probability of neutral molecules and active ions, thereby improving the ionization efficiency and detection sensitivity. The chemical components in the aerosol are ionized into charged ions, which enter the mass spectrometer for analysis under the action of vacuum negative pressure extraction. The un-ionized aerosol waste is separated out by a diaphragm pump.

[0098] The establishment process of the random forest classification model includes the following steps:

[0099] 1. Standard sample pretreatment: Direct analysis of textiles by microtube plasma ionization mass spectrometry using an electric iron burning method. In general, no sample pretreatment process is required. This method has no requirements for sample morphology and size, and only needs to ensure that the sample taken is representative and weighs no less than 5 mg. Before sampling, it is recommended to clean the sample surface with dust-free paper dipped in methanol to remove dust and potential interferents.

[0100] 2. Detecting the standard sample by the screen collision-microtube plasma ionization mass spectrometry method to obtain mass spectrum data;

[0101] The working conditions of the screen collision-microtube plasma ionization device are as follows: alternating voltage (equivalent value) 2.0 kV, frequency 10 kHz; carrier gas (He) flow rate 0.5 mL / min; negative pressure of the device and mass spectrometer interface-76 MPa; working temperature is room temperature. The working conditions of the electrostatic field orbitrap high-resolution mass spectrometer are as follows: ion transmission tube temperature 320℃, S-lens voltage 55 V, negative ion mode, data acquisition uses full scan mode, scan range is m / z 100-750, automatic gain control target value is 10 6 , maximum sampling time is 50 ms, and microscanning number is 1 per second.

[0102] A handheld soldering iron (as an example, 350 °C) was used to burn the textile samples, generating smoke for MC-μTP-MS analysis. Each standard sample was sampled in parallel for at least 6 times, and the actual sample was sampled in parallel for at least 3 times. When the total ion current signal intensity of the mass spectrum decreased, or ions from the sample smoke remained in the mass spectrum background, the capillary, screen and soldering iron head should be cleaned with methanol before sampling.

[0103] 3. Preprocessing of raw mass spectrometry data (saving a list file of mass numbers and corresponding signal intensity data);

[0104] The raw mass spectrometry data was preprocessed using a script compiled in the Python 3.10.4 environment, with the following specific steps:

[0105] (I) Extracting mass spectrometry information within the mass axis range of 100.0-550.0 Da, including mass numbers and corresponding signal intensities;

[0106] (II) Using local peak detection combined with signal-to-noise ratio screening to denoise the imported data, setting the noise estimation window width to 50 Da, the peak detection minimum intensity threshold to 10 5 cps, and the noise threshold to 3 times the signal-to-noise ratio;

[0107] (III) Maximum value normalization of the retained signal intensities, scaling them to the range of 0 to 1;

[0108] (IV) Establishing continuous non-overlapping global bins with 100.0 and 550.0 Da as the starting and ending points of binning, 0.1 Da as the binning step, and ±0.05 Da as the binning radius;

[0109] (VI) Assigning the retained ion signals to each bin according to the mass number, and retaining the maximum value of multiple signal intensities in the same bin;

[0110] (VII) Merging all sample mass spectrometry information to form a mass spectrometry data set.

[0111] 4. Evaluation of textile mass spectrometry data set quality using unsupervised classification method;

[0112] To evaluate the quality of the textile mass spectrometry data set and reveal its potential structure, unsupervised classification method was used for data analysis, outlier detection and feature effectiveness verification. First, the textile mass spectrometry data set was reduced to 25 dimensions by PCA to compress redundant information, retain core data structure, and reduce computational complexity. Next, the K-means clustering algorithm was used for unsupervised classification of the reduced data set to evaluate its data clustering property, and the number of clusters (n_clusters) was optimized according to the silhouette coefficient.

[0113] The data set quality is evaluated by the silhouette coefficient, adjusted Rand index, Davies-Bouldin index, and Calinski-Harabasz index, and the value ranges of these indexes are as follows:

[0114] The silhouette coefficient has a value range of [-1, 1]. When the silhouette coefficient is greater than 0.5, the sample clustering is clear; when the silhouette coefficient is between 0.25 and 0.5, there is a clustering structure but it is relatively vague; and when the silhouette coefficient is less than 0.25, the clustering effect is poor or the data has no clear structure.

[0115] The adjusted Rand index (ARI) has a value range of [-1, 1]. When the ARI is greater than 0.5, the clustering result is highly consistent with the standard sample category label; when the ARI is between 0.3 and 0.5, the clustering result has moderate consistency with the label; and when the ARI is less than 0.3, the clustering result has poor consistency with the label.

[0116] The Davies-Bouldin index (DBI) has a value range of [0, +∞). When the DBI is less than 0.5, the separation degree between clustering clusters is high and the cluster is tight; when the DBI is between 0.5 and 1.0, the clustering effect is moderate; and when the DBI is greater than 1.0, the clusters are seriously overlapped or the cluster is loose.

[0117] The Calinski-Harabasz index (CHI) has a value range of [0, +∞). The larger the CHI value is, the better; and when the CHI value is less than 50, the data structure is not obvious.

[0118] 5. Establishment of a random forest classification model;

[0119] A random forest algorithm-based textile category identification model is constructed using the Scikit-learn machine learning library in the Python environment, which is used to identify actual samples of wearing luxury goods. In the model training process, the data set used is subjected to Z-score standardization, and principal component analysis (PCA) is used for dimension reduction (n_component=25). The initial hyperparameter settings of the random forest model used for actual sample identification are as follows: the number of decision trees (n_estimators) is 100; the maximum depth of a single decision tree (max_depth) is limited to 10; the maximum number of features (max_features) algorithm is log2, that is, log2 (maximum number of features) features are randomly selected at the split node; the minimum number of samples for node splitting (min_samples_split) is set to 2; and the minimum number of samples for leaf node (min_samples_leaf) is 4. Bootstrap sampling method (Bootstrap=True) is used to generate training subsets, and out-of-bag data (OOB samples) is retained.

[0120] The quality of the random forest model is evaluated by out-of-bag score, 5-fold cross-validation average accuracy, precision, recall, and F1 score, which have the following value ranges:

[0121] Out-of-bag score, with a value range of [0, 1] (or in percentage form). When the out-of-bag score is greater than 0.85, the model has strong generalization ability, and the data fitting and generalization are well balanced; when the out-of-bag score is between 0.7 and 0.85, the model performs moderately, and there may be slight overfitting or underfitting; when the out-of-bag score is less than 0.7, the model performance is poor.

[0122] 5-fold cross-validation average accuracy, with a value range of [0, 1] (or in percentage form). When the average accuracy is greater than 0.85, the model has strong classification ability; when the average accuracy is between 0.7 and 0.85, the model performs moderately; when the average accuracy is less than 0.7, the model fails to capture effective patterns.

[0123] Precision, with a value range of [0, 1] (or in percentage form). When the precision is greater than 0.8, the model's prediction of positive classes is highly reliable; when the precision is between 0.6 and 0.8, the model's credibility is moderate; when the precision is less than 0.6, the model has more misjudgments.

[0124] Recall, with a value range of [0, 1] (or in percentage form). When the recall is greater than 0.8, the model can effectively capture positive class samples; when the recall is between 0.6 and 0.8, the model's coverage is moderate; when the recall is less than 0.6, the model has serious omissions.

[0125] F1 score, with a value range of [0, 1] (or in percentage form). When the F1 score is greater than 0.8, the precision and recall are balanced and excellent; when the F1 score is between 0.6 and 0.8, the model's balance is moderate; when the F1 score is less than 0.6, the model's overall performance is poor.

[0126] To ensure experimental repeatability, fix the random number seed throughout the modeling process, including hyperparameter optimization, model training, and validation stages.

[0127] The detection method for the actual sample includes the following steps:

[0128] (A) Pretreatment of the actual sample (the specific steps are the same as the treatment method for the standard sample);

[0129] (B) Detecting the sample by screen collision-microtube plasma ionization mass spectrometry to obtain mass spectrometry data (the specific steps are the same as the treatment method for the standard sample);

[0130] (C) The identification result is obtained after processing the actual sample mass spectrum data using the random forest classification model.

[0131] The identification result (the fiber type to which the sample belongs) is obtained by processing the actual sample mass spectrum data using the trained random forest model. The confidence rate is calculated by counting the parallel identification results, and the final conclusion is obtained.

[0132] The statistical method is as follows, n parallel identifications are performed:

[0133] If the n parallel identification results are the same, the actual sample type is consistent with the identification result, and the confidence is 100%;

[0134] If m of the n parallel identification results are the same, m < n, the actual sample type is consistent with the m same identification results, and the confidence is m / n x 100%. When the confidence is greater than 60%, it is considered that the identification result has high reliability.

[0135] If the n parallel identification results are all different, the actual sample type cannot be determined, and the mass spectrum data needs to be collected again for identification.

[0136] The above statistical method is taken as an example of 3 parallel classification identifications:

[0137] The 3 parallel identification results are the same, the actual sample type is consistent with the identification result, and the confidence is 100%.

[0138] 2 of the 3 parallel identification results are the same, the actual sample type is consistent with the 2 same identification results, and the confidence is 66.7%.

[0139] The 3 parallel identification results are all different, the actual sample type cannot be determined, and the mass spectrum data needs to be collected again for identification.

[0140] 3 Results and Discussion

[0141] 3.1 Electric iron burning-sieve collision-microtube plasma ionization mass spectrometry suitable for textile mass spectrum data collection

[0142] In order to reduce the signal residue after each mass spectrum analysis and reduce the dimension of the obtained mass spectrum data, the negative ion mode is used to collect the mass spectrum data of the textile sample. On this basis, the total ion current (TIC) signal intensity of a specific sample (Table 1, serial number 1) is used as an evaluation index, and the three working parameters of the MC-μTP ionization device, namely the alternating voltage (equivalent value), the carrier gas flow rate and the VAPUR interface negative pressure, are optimized, and the results are as follows Figure 2The results show that the plasma ionization efficiency increases with the increase of the AC voltage equivalent value; when the voltage is greater than 2.0 kV, the TIC signal intensity slowly decreases; when the voltage is greater than 2.5 kV, discharge occurs between the microelectrode of the ionization device and the inlet of the ion transmission tube of the mass spectrometer, and the mass spectrometry signal cannot be generated. With the increase of the carrier gas flow rate, the plasma ionization efficiency increases; when the flow rate is higher than 0.5 mL / min, the TIC tends to be stable. With the increase of the negative pressure at the atmospheric pressure interface of the mass spectrometer, the TIC first increases and then decreases; when the negative pressure is-76 MPa, the TIC reaches the maximum value. It is worth noting that when the negative pressure is lower than-86 MPa, although the TIC decreases, the residual mass spectrometry signal is obviously suppressed. Considering the TIC signal intensity, instrument stability and detection cost, the AC voltage (equivalent value) of 2.0 kV, the carrier gas flow rate of 0.5 mL / min and the negative pressure of-76 MPa at the interface between the device and the mass spectrometer are selected as the default working parameters for the subsequent experiments.

[0143] 3.1.2 Textile mass spectrometry data acquisition

[0144] The handheld electric iron is used to burn the textile sample to generate smoke, which is ionized by the MC-μTP device and then analyzed by the electrostatic field orbitrap high-resolution mass spectrometer. Considering that the thermal decomposition temperatures of cotton and wool / silk involved in this study are 353℃ and 227℃, respectively, and the thermal decomposition temperatures of polyester fiber and acrylic fiber are 350℃ and 200-250℃, respectively, the electric iron burning temperature is set to 350℃ to obtain the mass spectrometry information of the thermal decomposition characteristic products of different types of fibers while maximizing the suppression of the residual smoke signal. After preprocessing, the collected mass spectrometry data is combined into a mass spectrometry data set (containing 181 groups of data and 4501 variable dimensions) for subsequent machine learning. Figure 3 The characteristic mass spectra of cotton, cashmere / wool, polyester fiber and acrylic fiber after maximum value normalization are shown in FIG. 3. It can be seen that there are great differences in the types and quantities of fragment ions and the base peak mass-to-charge ratio of the characteristic mass spectra of the four types of fibers, indicating that it is feasible to identify the fiber types of unknown samples by training a prediction model based on the standard sample mass spectrometry data set.

[0145] 3.2 Evaluation of the quality of the textile mass spectrometry data set using unsupervised classification method

[0146] In order to evaluate the quality of the textile mass spectrometry data set and reveal its potential structure, unsupervised classification method is used for data analysis, outlier detection and feature effectiveness verification. First, the textile mass spectrometry data set is reduced to 25 dimensions by PCA to compress redundant information, retain the core data structure and reduce computational complexity. Figure 4A shows the PCA scatter plot, in which the first 10 principal components cumulatively explain 95% of the variance, indicating that the data after dimensionality reduction can still represent the main differences of the original features. Next, the K-means clustering algorithm is used to perform unsupervised classification on the dimensionally reduced dataset to evaluate its data clustering performance, and the number of clusters (n_clusters) is optimized to 4 according to the silhouette coefficient. As shown in Figure 4 B, the sample points in the dataset are clustered in the dimensionally reduced space of the first three principal components, and the distance between clusters is large, indicating that the textile mass spectrum data is highly separable. It is worth noting that the sample points of cashmere and wool are overlapped in space and cannot be independently clustered, which is due to the fact that the main chemical components of both are α-Kelatin, and only the amino acid composition and content have small differences, resulting in insufficient mass spectrum feature discrimination. In subsequent experiments, the classification labels of the two types of samples are unified as Wool / Cashmere. Finally, the Adjusted Rand Index (ARI), Davies-Bouldin Index (DBI), and Calinski-Harabasz Index (CHI) are used to evaluate the unsupervised classification results. The results show that the ARI is 0.833, indicating that the clustering results are highly consistent with the sample grouping labels, and the data has a clear class structure; the DBI is 0.518, indicating that all the samples in the cluster are tightly packed and there are no obvious overlapping cluster pairs; and the CHI is 299.347, indicating that the global clustering structure is clear and the class differences are significant. In summary, the quality of the obtained textile mass spectrum dataset can meet the requirements of supervised learning.

[0147] 3.3 Building a Random Forest Classification Model Based on the Textile Mass Spectrum Dataset

[0148] The textile mass spectrum dataset is randomly divided into a training set and a test set in a 7:3 ratio to build a random forest classification model and evaluate its performance. First, based on the training set after Z-score standardization and PCA dimensionality reduction (n_components = 25), the hyperparameters of the random forest model are optimized through grid search, and the value ranges of the 5 hyperparameters involved and the optimization results are listed in Table 2. Next, the performance of the optimized model is evaluated using out-of-bag (OOB) data and the test set. The results show that the OOB score of the model reaches 0.9762, indicating that the model performs well on the samples that are not used by individual trees during the training process (i.e., OOB samples), reflecting that the model has strong generalization ability and low risk of overfitting. The 5-fold cross-validation average accuracy is 0.9683 (±0.0299), with a small standard deviation, indicating that the model has small fluctuations in performance on different data subsets and robust adaptability. As shown in Figure 5As shown in A, the model's precision, recall, and F1 score for the four fiber classifications are all no less than 0.889, indicating that the model has a balanced ability to distinguish between different categories and no significant category bias occurs. The accuracy of the test set is 0.9636, which is similar to the performance of the training set. From the confusion matrix of the test set ( Figure 5 As can be seen in B), cashmere / wool and acrylic each have one misclassification, which may be caused by insufficient feature discrimination or sample annotation noise. Polyester and acrylic only have 8 and 9 test samples respectively, and their high accuracy may have the risk of overfitting.

[0149] Table 2 Optimizing the hyperparameters of the random forest model based on the training set

[0150]

[0151]

[0152] 3.4 Actual sample identification

[0153] Mass spectrometric data from samples of luxury apparel inspected by customs authorities were collected using soldering iron scorching-screen collision-microtube plasma ionization mass spectrometry. Each sample was collected three times in parallel. The training and test sets were then combined, and a random forest model was trained using optimized hyperparameters. The samples were then subjected to fiber composition identification, and the fiber composition was finally determined by combining the three test results. As shown in Table 3, the identified fiber compositions of 30 samples from eight luxury brands were largely consistent with the labeled compositions. Twenty samples achieved a prediction confidence level of 1.00, indicating that all three replicates were consistent with the labeled compositions. These samples were primarily cotton, cashmere, and wool, demonstrating that a larger sample size in a textile mass spectrometry dataset has a positive impact on accurate classification. Eight samples achieved a confidence level of 0.67, primarily in the wool, polyester, and acrylic fiber categories. This is likely due to the presence of a small number of outliers in the actual sample mass spectrometry data. Increasing the number of replicates can reduce the impact of outliers and improve confidence. The confidence level for one sample each of polyester fiber and acrylic fiber was only 0.33. The reason for the low confidence level may be that there is indeed a difference between the actual fiber composition of the sample and the labeled composition, or it may be a false positive caused by the small number of training samples. Other identification methods need to be introduced for further comparison and confirmation.

[0154] Table 3 Identification of fiber composition of 30 luxury clothing samples using the optimized random forest model

[0155]

[0156]

[0157] 4 Conclusion

[0158] The established electric iron burning-sieve collision-micro-tube plasma ionization mass spectrometry method has small sample requirement, no need for pretreatment and simple operation. The random forest prediction model trained based on the collected mass spectrometry data set of textiles has good comprehensive performance, and basically meets the needs of rapid and accurate identification of fiber components of wearing luxury samples. This method has application potential in customs sampling, product authenticity identification and other scenes, and has certain guiding significance for related industries. Next, the method will be optimized from the following aspects to further improve the sampling stability, data comparability and identification accuracy. First, reference is made to the conclusions of fiber thermal analysis related work to optimize the electric iron burning temperature, improve the compatibility of the method for different types of fibers (textiles, leather, etc.), and obtain higher thermal decomposition efficiency. Second, the fiber mass spectrometry database is further expanded to improve the performance of the random forest model and reduce the risk of overfitting. Third, other identification methods (spectroscopy, microscopic imaging, etc.) are combined to realize mutual confirmation of the results, and a comprehensive evaluation method based on different sources of identification results is established. Fourth, the feature importance analysis of the random forest model is combined with tandem mass spectrometry technology to realize the identification of different types of fiber thermal decomposition characteristic chemical products. Fifth, the fiber component identification method is combined with the non-target screening technology of chemical risk substances to comprehensively evaluate the quality of wearing luxury goods from the aspects of authenticity and safety.

[0159] The above-described embodiments are merely preferred embodiments of the present application, and are not intended to limit the scope of the present application. Various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art without departing from the design spirit of the present application shall fall within the scope of protection of the claims of the present application.

Claims

1. A method for rapidly identifying the fiber composition of luxury goods, characterized by: The steps include: (A) Pretreatment of actual samples; (B) The sample was detected using screen collision-microtube plasma ionization mass spectrometry to obtain mass spectral data; (C) Identification results were obtained after processing the actual sample mass spectrometry data using the random forest classification model.

2. The method for rapidly identifying the fiber composition of luxury goods according to claim 1, characterized in that: The process of establishing the random forest classification model includes the following steps: (a) Pretreatment of standard samples; (b) Screen collision-microtube plasma ionization mass spectrometry is used to detect the standard sample to obtain raw mass spectrum data; (c) preprocessing the raw mass spectrometry data to obtain a mass spectrometry data set, which is a list file that stores mass numbers and corresponding signal intensity data; (d) Use unsupervised classification methods to evaluate the quality of mass spectrometry datasets; (e) Build a random forest classification model.

3. The method for rapidly identifying the fiber composition of luxury goods according to claim 2, characterized in that: The method for preprocessing the raw mass spectrometry data in step (c) specifically comprises the following steps: (I) Extracting mass spectrum information within the mass axis range of 100.0 to 550.0 Da, including mass number and corresponding signal intensity; (II) Local peak detection combined with signal-to-noise ratio screening was used to reduce the noise of the imported data. The noise estimation window width was set to 50 Da, and the minimum intensity threshold for peak detection was set to 10 5 cps, the noise threshold is set to 3 times the signal-to-noise ratio; (III) performing maximum normalization on the retained signal intensities and scaling them to the range of 0 to 1; (IV) 100.0 and 550.0 Da were used as the binning start and end points, 0.1 Da as the binning step size, and ±0.05 Da as the binning radius to establish continuous non-overlapping global bins; (VI) assigning the retained ion signals to each bin according to mass number, and retaining the maximum value of multiple signal intensities in the same bin; (VII) Merging the mass spectrometry information of all samples to form a mass spectrometry dataset; the obtained mass spectrometry dataset must undergo quality evaluation in step (d) and meet the requirements before it can be used for subsequent random forest model training.

4. The method for rapidly identifying the fiber composition of luxury goods according to claim 3, characterized in that: The use of an unsupervised classification method to evaluate the quality of the mass spectrometry dataset in step (d) specifically comprises the following steps: first, reducing the mass spectrometry dataset to 25 dimensions using PCA to compress redundant information, preserve the core data structure, and reduce computational complexity; second, performing unsupervised classification on the reduced-dimensional dataset using a K-means clustering algorithm to evaluate its data clustering, with the number of clusters used being optimized based on the silhouette coefficient; The quality of the dataset is evaluated by the silhouette coefficient, adjusted Rand index, Davidson-Botting index, and Calinski-Harabasz index. The value ranges of these indicators are as follows: Silhouette coefficient, the value range is [-1, 1]; when the silhouette coefficient is greater than 0.5, the sample clustering is clear; when the silhouette coefficient is between 0.25 and 0.5, the clustering structure exists but is relatively vague; when the silhouette coefficient is less than 0.25, the clustering effect is poor or the data has no clear structure; The adjusted Rand index, or ARI, has a value range of [-1, 1]. When the ARI is greater than 0.5, the clustering results are highly consistent with the standard sample type labels. When the ARI is between 0.3 and 0.5, the clustering results are moderately consistent with the labels. When the ARI is less than 0.3, the clustering results are poorly consistent with the labels. The Davidson-Boulding Index (DBI) ranges from [0 to +∞). When DBI is less than 0.5, the clusters are highly separated and tightly packed. When DBI is between 0.5 and 1.0, the clustering effect is moderate. When DBI is greater than 1.0, the clusters overlap severely or are loosely packed. The Calinski-Harabasz index, or CHI, ranges from [0, +∞); the larger the CHI value, the better; when the CHI value is less than 50, the data structure is not obvious.

5. The method for rapidly identifying the fiber composition of luxury goods according to claim 4, characterized in that: The method for establishing the random forest classification model in step (e) comprises the following steps: Using the Scikit-learn machine learning library, a textile identification model based on the random forest algorithm was constructed for identifying actual samples of luxury wear. During model training, the dataset was standardized and dimensionality reduction was performed using principal component analysis. The initial hyperparameters of the random forest model used for sample identification were as follows: 100 decision trees; a maximum depth of 10 for each decision tree; and a log2 maximum feature count algorithm, which randomly selected log2 features when splitting a node. The minimum number of samples for node splitting is set to 2; the minimum number of samples for leaf nodes is 4; the training subset is generated using the bootstrap sampling method; out-of-bag data is retained; The quality of the random forest model is evaluated by out-of-bag score, 5-fold cross-validation average accuracy, precision, recall, and F1 score. The value ranges of these indicators are as follows: Out-of-bag score, the value range is [0, 1] or percentage form; when the out-of-bag score is greater than 0.85, the model generalization ability is strong, and the balance between data fitting and generalization is good; when the out-of-bag score is between 0.7 and 0.85, the model performance is medium, and there may be slight overfitting or underfitting; when the out-of-bag score is less than 0.7, the model performance is poor; The average accuracy of 5-fold cross-validation, with a value range of [0, 1] or percentage. When the average accuracy is greater than 0.85, the model has strong classification ability; when the average accuracy is between 0.7 and 0.85, the model has moderate performance; when the average accuracy is less than 0.7, the model has not captured the effective pattern. Precision, the value range is [0, 1] or percentage form; when the precision is greater than 0.8, the model has high prediction reliability for the positive class; When the precision is between 0.6 and 0.8, the model credibility is medium; when the precision is less than 0.6, the model has many misjudgments; Recall rate, the value range is [0, 1] or percentage form; when the recall rate is greater than 0.8, the model can effectively capture positive samples; when the recall rate is between 0.6 and 0.8, the model coverage is medium; when the recall rate is less than 0.6, the model has serious missed detections; F1 score, the value range is [0,1] or percentage; When the F1 score is greater than 0.8, the precision and recall are well balanced; when the F1 score is between 0.6 and 0.8, the model is moderately balanced; when the F1 score is less than 0.6, the overall performance of the model is poor; To ensure the reproducibility of the experiment, the random number seed is fixed throughout the modeling process, including hyperparameter optimization, model training, and validation phases.

6. The method for rapidly identifying the fiber composition of luxury goods according to claim 1, characterized in that: In step (C), the trained random forest model is used to process the actual sample mass spectrometry data to obtain the classification result, that is, the fiber type to which the sample belongs. The confidence level is calculated by statistically analyzing the parallel discrimination results, and finally a conclusion is obtained. The result statistical method is as follows. n parallel discriminations are carried out: If the results of n parallel discriminations are all the same, then the actual sample type is consistent with the discrimination result, and the confidence level is 100%. If m out of n parallel discrimination results are the same (m < n), then the actual sample type is consistent with the m same discrimination results, and the confidence level is m / n × 100%. When the confidence level is greater than 60%, the discrimination result is considered to be highly credible. If the results of n parallel discriminations are all inconsistent, then the actual sample type cannot be determined, and it is necessary to re-collect the mass spectrometry data for discrimination.

7. The method for rapidly identifying the fiber composition of luxury goods according to claim 2, characterized in that: The samples in step (A) and step (a) need to ensure that the samples taken are representative and weigh no less than 5 mg. Before sampling, use lint-free paper dipped in methanol to clean the sample surface.

8. The method for rapidly identifying the fiber composition of luxury goods according to claim 7, characterized in that: In the mesh collision-microtube plasma ionization mass spectrometry method described in step (B) and step (b), a mesh collision-microtube plasma ionization device is used. During detection, a handheld soldering iron is used to burn the textile sample at 350 - 450 °C to generate flue gas for MC-μTP-MS analysis. Each standard sample is sampled in parallel at least 6 times, and each actual sample is sampled in parallel at least 3 times. When the total ion current signal intensity of the mass spectrometry decreases, or ion residues from the sample flue gas appear in the mass spectrometry background, the capillary, mesh, and soldering iron tip should be cleaned with methanol before sampling again.

9. The method for rapidly identifying the fiber composition of luxury goods according to claim 8, characterized in that: The operating conditions of the screen collision-microtube plasma ionization device were: AC voltage 2.0 kV, frequency 10 kHz; carrier gas flow rate 0.5 mL / min; negative pressure at the interface between the device and the mass spectrometer -76 MPa; operating temperature was room temperature; the operating conditions of the electrostatic field orbital trap high-resolution mass spectrometer were: ion transfer tube temperature 320°C, S-lens voltage 55 V, negative ion mode, data acquisition using full scan mode, scanning range m / z 100-750, automatic gain control target value 10 6 , maximum injection time 50ms, micro-scan number 1 per second.

10. The method for rapidly identifying fiber composition of luxury goods according to claim 9, characterized in that: The mesh collision-microtube plasma ionization device includes a microtube plasma generator, a molecular ion reactor, and a mass spectrometer interface connected in sequence. The microtube plasma generator includes a first T-shaped three-way joint (1), the molecular ion reactor includes a second T-shaped three-way joint (2), and the two ends of the straight pipe of the second T-shaped three-way joint (2) are respectively connected to one end of the straight pipe of the first T-shaped three-way joint (1) and the mass spectrometer interface. A first connecting pipe (10) is used to connect between the first T-shaped three-way joint (1) and the second T-shaped three-way joint (2). Inside the connection, a second flexible sleeve (3) and a fused silica capillary (4) are provided. Inside the other end of the straight pipe of the first T-shaped three-way joint (1), a first flexible sleeve (5) is provided. The device also includes a flat-headed stainless steel hollow needle (6), one end of which is connected to high-frequency high-voltage alternating current, and the other end passes through the first flexible sleeve (5) and the second flexible sleeve (3) and is arranged inside the fused silica capillary (4). The vertical pipeline of the first T-shaped three-way joint (1) is connected to a helium gas cylinder through a helium gas pipeline (7). The vertical pipeline of the second T-shaped three-way joint (2) is connected to a gas collection hood (8). A copper mesh (9) is provided at the connection between the second T-shaped three-way joint (2) and the mass spectrometer interface and is fixed in a second connecting pipe (11). The outer shell (12) is connected to the second T-shaped three-way joint (2) through the second connecting pipe (11). The two ends of the second flexible sleeve (3) are respectively arranged in the first T-shaped tee (1) and the first connecting pipe (10); One end of the fused silica capillary (4) is arranged in the first T-shaped tee (1), and the other end passes through the second flexible sleeve (3) and is arranged in the middle of the second T-shaped tee (2); inside the second T-shaped tee (2), the distance between the tip of the flat-head stainless steel hollow needle (6) and the end of the fused silica capillary (4) is 5 mm; A gap is left between the flat-head stainless steel hollow needle (6) and the fused silica capillary (4) for gas to pass through; The mass spectrometer interface comprises a housing (12) and a vacuum pump (14) connected thereto, wherein the housing (12) is connected to the second T-shaped three-way connection (2) and an ion transfer tube (13) of the mass spectrometer; The high-frequency high-voltage alternating current is provided by a frequency-adjustable high-voltage alternating current power supply; the first T-type tee (1) is a threaded tee made of polyetheretherketone, and the thread interface size is 1 / 16 inch; the helium pipeline (7) is a polyethylene hose with an inner diameter of 1 / 16 inch; the size of the flat-head stainless steel hollow needle (6) is 30G; the first flexible sleeve (5) and the second flexible sleeve (3) are made of polyetheretherketone, with an outer diameter of 1 / 16 inch and an inner diameter of 310-650 microns; the fused silica capillary (4) has an outer diameter of 650 microns and an inner diameter of 530 microns; the first connecting pipe (10) and the second The connecting pipe (11) is a polyethylene hard pipe with an outer diameter of 6 mm; the second T-shaped tee (2) is a quick-connect tee made of polybutylene terephthalate, with an interface inner diameter of 6 mm; the copper screen (9) has a diameter of 4 mm and a mesh size of 200; the gas collecting cover (8) is a hemispherical plastic material with a sampling port diameter of 5 cm; the shell (12) of the mass spectrometer interface is made of stainless steel, the interface connected to the second T-shaped tee (2) has an inner diameter of 6 mm, and the pagoda interface connected to the vacuum pump (14) has a specification of 8 mm; the ion transfer tube (13) is made of stainless steel with an inner diameter of 40 microns.

Images