Method for rapid detection of quality of artificial leather based on spectral analysis

Abstract

The application discloses a rapid detection method for quality of artificial leather based on spectral analysis, and particularly relates to the field of quality detection, which comprises the following steps: through multi-dimensional original data acquisition, the component composition characteristics, the surface and micro characteristics, and the dynamic response characteristics of the artificial leather are synchronously acquired; time stamp alignment and space calibration are performed to form a space-time correlation data set; then, a component analysis model, a surface correlation model, and a dynamic response model are constructed, and time-space deviation is eliminated through synchronous optimization of the models; after multi-dimensional features are extracted, optimized features are generated through cross-direction tensor fusion, attention enhancement, and gate recurrent unit time sequence fusion; finally, quality indexes are quantified based on a multi-task model, and quality grades and traceable reports are output in combination with a dynamic threshold value through hierarchical encoding, graph neural network correlation enhancement, and evidence fusion; the single sample acquisition time is less than 3 minutes, and the response time is less than 1 second, so that full-dimensional, rapid and accurate detection and quality tracing are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of quality detection, more particularly, the present application relates to a rapid quality detection method for artificial leather based on spectral analysis. BACKGROUND

[0002] As a substitute material widely used in furniture, clothing, automotive interiors and other fields, the quality level of artificial leather directly affects the use experience and market acceptance of the end product, and the industry has formed the demand for three core dimensions of component safety, surface state and dynamic performance for its detection.

[0003] In the field of component detection, chemical analysis methods are still used, while spectral technology is used for component feature recognition, which can capture the spectral signals of core chemicals; in surface state detection, manual visual inspection and special equipment detection coexist, and some equipment can collect multi-angle reflection characteristic data, and a few schemes introduce polarization detection or microscopic scanning technology to obtain local details; dynamic performance evaluation is mostly carried out offline by special equipment to monitor the performance change of the material under the action of machinery or heat.

[0004] With the improvement of production line efficiency, the industry has higher requirements for the real-time and comprehensiveness of detection, and the existing spectral detection scheme focuses on single-dimensional data processing, and has not yet formed a complete system of multi-dimensional coordination and space-time correlation, but its technical foundation provides support for subsequent integrated detection schemes.

[0005] However, in actual use, it still has some disadvantages, such as:

[0006] 1. Component detection is destructive and lagging: existing component detection mostly uses laboratory methods such as chemical digestion, which requires cutting samples, causing damage, and the detection period is long, which cannot adapt to the continuous detection rhythm of the production line; at the same time, only the average content of the components can be obtained, it is difficult to reflect the uniformity of the internal distribution, and it is easy to miss the problem of local harmful substances exceeding the standard;

[0007] 2. Surface quality detection is single and fuzzy: surface detection often relies on manual visual inspection or single equipment, manual detection is highly subjective and low in efficiency, and single equipment can only obtain macroscopic reflection data, and cannot distinguish the source of layered defects and microscopic structure abnormalities; the detection result lacks spatial coordinate correlation, and the defect positioning is fuzzy;

[0008] 3. Dynamic performance evaluation cannot be online coordinated: dynamic performance such as folding resistance and flexibility needs to be tested offline by special equipment, and the detection process needs manual intervention, which cannot be synchronized with the production process. And the test result has no data correlation with the component and surface characteristics, it is difficult to analyze the root cause of performance degradation;

[0009] 4. Multi-dimensional data fragmentation and traceability difficulty: existing spectral detection schemes focus on single-dimensional modeling, independent processing of component, surface and dynamic data, lack of space-time synchronization optimization, leading to fragmented quality information and poor consistency of detection results; and there is no complete data traceability chain, so that when quality problems occur, the equipment state and data of the production link cannot be associated. SUMMARY

[0010] In order to overcome the above-mentioned defects of the prior art, the present application provides a rapid detection method for artificial leather quality based on spectral analysis, which solves the problems raised in the background art through the following scheme.

[0011] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a rapid detection method for artificial leather quality based on spectral analysis, comprising:

[0012] S1: Multi-dimensional original data acquisition: acquiring component composition characteristics, surface and microstructure characteristics and dynamic response characteristic data of artificial leather samples, and performing data global synchronization and coordinate unification on the data to obtain preprocessed data;

[0013] S2: Multi-dimensional data modeling and function construction: based on the preprocessed data, a component composition characteristic analysis model, a surface and microstructure characteristic correlation model, and a dynamic response characteristic model are constructed, and the three models are optimized using a data synchronization optimization model to eliminate space-time bias, obtaining the model output parameters after synchronization optimization;

[0014] S3: Multi-dimensional feature extraction and fusion: based on the parameters output by the three synchronized models, multi-dimensional feature extraction and standardization are performed, and the standardized multi-dimensional features are subjected to basic cross-direction three-order tensor fusion to obtain cross-direction fusion features, and feature fusion and compression are performed based on an attention mechanism to obtain basic fusion features and real-time compression features, and a gated recurrent unit is used to perform time sequence correlation fusion on the real-time compression fusion features to generate time sequence optimization features;

[0015] S4: Quality evaluation and grading system: based on the basic fusion features and the time sequence optimization features, the original quantitative values of the multi-task quality evaluation model output indicators are constructed, and hierarchical feature coding is performed on the original quantitative values to obtain low-dimensional core features; the correlation rules between the low-dimensional core features are mined based on a graph neural network to generate correlation enhanced features; the correlation enhanced features are comprehensively decided using an evidence fusion rule to obtain the trust degree of each quality grade; the confidence, likelihood and uncertainty of the trust degree are quantified, and combined with a dynamic threshold adjustment strategy, quality grading is performed to output the quality grade of the artificial leather sample and a traceable detection report.

[0016] Preferably, the multi-dimensional original data acquisition comprises:

[0017] Component composition characteristic collection: Collect full-waveband original spectral data of the artificial leather sample, including characteristic peak reflectivity and mean value data of plasticizer, resin, and phthalate ester;

[0018] Surface and microscopic characteristic collection: Collect multi-angle reflection characteristic data BRDF value, polarization spectrum, and Mueller matrix data of the sample, standard diffuse reflection plate calibration data, and spatial and spectral three-dimensional data and microscopic coating thickness of the microscopic region, and attach the world coordinate system label corresponding to each collection point;

[0019] Dynamic response characteristic collection: Collect characteristic peak reflectivity fluctuation data of the sample over time under dynamic excitation of mechanical pressure and thermal stress, including time and reflectivity two-dimensional change curve before and after excitation and dynamic repeatability verification data;

[0020] Data global synchronization and coordinate unification: Time stamp alignment and spatial coordinate calibration are performed on the above three types of data to finally form an associated data set with additional space-time coordinates and time stamps.

[0021] Preferably, the component composition characteristic analysis model comprises:

[0022] Based on the full-waveband spectral data, characteristic peak reflectivity, and mean value data in the preprocessed component composition characteristic collection data, baseline correction, smoothing and denoising are first performed on the spectral data, the effective spectral response at the wavelength is calculated by combining the Gaussian function, the noise interference is eliminated, and the standardized spectral data set is obtained; based on the data set, a characteristic peak positioning module is constructed, the characteristic peak center wavelength and the corresponding reflectivity interval of different components are located by comparing the known component standard spectrum library; the quantitative mapping relationship of reflectivity, component concentration, and purity is constructed by combining the Lambert-Beer law and using the standard sample with known component content to calibrate the molar absorption coefficient and the optical path parameter, the matrix effect is eliminated to improve the accuracy, and a quantitative analysis module is formed.

[0023] Preferably, the surface and microscopic characteristic correlation model comprises:

[0024] Surface reflection uniformity sub-model: based on the multi-angle BRDF value and standard diffuse reflection plate calibration data collected by the surface and microscopic characteristic collection as input, the mapping relationship between incident light irradiance and reflected radiation brightness is constructed based on the BRDF core function, the anisotropy coefficient is calculated to quantify the surface reflection uniformity at each reflection angle, and the full-angle reflection uniformity comprehensive score is output;

[0025] Layered defect positioning sub-model: based on the polarization spectrum and Mueller matrix data as input, the polarization characteristic comprehensive response of all layers of the artificial leather is described by multi-layer Mueller matrix product operation, the spatial position of the layered defect is located by extracting the depolarization ratio and dichroism ratio parameters of each layer from the decomposition matrix, and the defect layer position identifier and defect probability are output;

[0026] Micro-macro performance correlation sub-model, which takes micro-space and spectral three-dimensional data, micro-coating thickness and corresponding world coordinate system label as input, performs volume integral operation on micro-local performance based on weight function, builds cross-scale correlation between microstructure and macro performance, and outputs macro performance parameter prediction value.

[0027] Preferably, the dynamic response characteristic model comprises:

[0028] Relaxation parameter extraction submodule, based on time and reflectance dynamic data, preset initial relaxation amplitude, and using double exponential relaxation equation to fit the relaxation curve of reflectance from initial value to equilibrium value, calculate core index relaxation time constant, simultaneously extract attenuation coefficient, fast relaxation coefficient, and output relaxation characteristic parameter group;

[0029] Dynamic performance quantification submodule, which takes relaxation time constant and dynamic scene label as input, and builds parameter mapping relationship.

[0030] Preferably, the multi-dimensional feature extraction comprises:

[0031] Composition feature extraction, which takes the concentration parameters output by the composition characteristic analysis model as input, decomposes the 1735nm, 1720nm and 1240nm characteristic peak intensity curves by 5 layers through self-adaptive Morlet wavelet, selects the top 10% wavelet coefficients, extracts 6 key wavelength point features, and outputs a 90-dimensional composition feature vector;

[0032] Surface and micro-feature extraction, which takes the parameters output by the surface and micro-characteristic correlation model as input, extracts the anisotropy coefficients and mean values of 9 angles, a total of 10 dimensions, the layer depolarization ratio and defect probability, a total of 8 dimensions, the micro-coating thickness statistics, a total of 20 dimensions, and finally outputs a 38-dimensional surface and micro-feature vector;

[0033] Dynamic response feature extraction submodule, which takes the relaxation time constant and folding resistance index parameters output by the dynamic response characteristic model as input, decomposes the relaxation curve into 8 eigenmode functions, calculates the instantaneous frequency mean, amplitude decay rate and peak time through Hilbert transform, and outputs a 24-dimensional dynamic feature vector;

[0034] Single-direction feature standardization submodule, which adopts Z-Score standardization processing on the above three types of feature vectors, converts them into standard features with mean value 0 and standard deviation 1, eliminates dimension differences, and finally outputs the standardized composition, surface and dynamic feature vectors.

[0035] Preferably, the basic fusion feature and time sequence optimization feature comprise:

[0036] The basic fusion feature is based on the standardized multi-dimensional feature vector, and a query vector, a direction key vector and a value vector are constructed through a Q-K-V attention mechanism to calculate the attention weight of each direction feature and then to weight and sum the 128-dimensional value vector to generate a 128-dimensional basic fusion feature.

[0037] The time sequence optimization feature sub-module takes the 64-dimensional real-time compression feature obtained by Tucker decomposition of the basic fusion feature as input, acquires continuous feature frames at 100 ms / frame, inputs the filtered abnormal value into a GRU network, fuses historical and current frame information through update gate and reset gate, and outputs one 64-dimensional aggregated feature every 8 frames.

[0038] Preferably, the multi-task quality evaluation model comprises:

[0039] Input feature fusion is based on the basic fusion feature and the time sequence optimization feature, and 128-dimensional final input features are generated according to a static and dynamic weighted fusion rule;

[0040] Shared feature extraction is based on the 128-dimensional final input features, and 64-dimensional shared features are obtained through a 4-layer CNN: each layer adopts a 3*3 convolution kernel, a step of 1 and a ReLU activation function, and the hierarchical processing of local feature capture, middle layer aggregation, high layer abstraction and feature condensation is performed to output the 64-dimensional shared features;

[0041] Task-specific output is based on the 64-dimensional shared features, and three fully connected layers are provided to correspond to three major tasks: component safety task output component purity and harmful substance concentration; surface quality task output defect rate and gloss uniformity; and dynamic performance task output folding resistance and tensile strength, with a total of six original quantitative values of indexes;

[0042] The loss function adopts a weighted MSE loss combined with an L2 regularization term.

[0043] Preferably, the evidence fusion rule comprises:

[0044] Evidence source preprocessing takes the results of the three major tasks output by the multi-task quality evaluation model as input, and defines the component safety, surface quality and dynamic performance as three independent evidence sources; according to the quality level framework of excellent [90, 100], good [80, 89], medium [60, 79] and poor [0, 59], the task indexes are converted into trust value of the corresponding level through a membership function to output an evidence source trust matrix;

[0045] Conflict coefficient calculation is based on the above-mentioned trust matrix to calculate the conflict coefficient between the evidence sources to measure the consistency, wherein a coefficient of 0 represents complete consistency of the evidence, a coefficient of 1 represents complete conflict of the evidence, and 0.5 is used as a boundary to distinguish low conflict and high conflict scenarios, and the conflict coefficient and the consistency judgment result are output.

[0046] Hierarchical fusion decision, according to the conflict judgment result, the adaptive strategy is executed: in the low conflict scene, the trust degree fusion is carried out by D-S evidence theory; in the high conflict scene, the trust degree of each evidence source is adjusted through the weighted average method, and then the fusion operation is executed, and the fused each level basic probability distribution is output;

[0047] The fusion result is output, the normativity of the probability distribution is verified, and the final trust degree data of each quality level is output.

[0048] Preferably, the quality grading comprises:

[0049] Based on the confidence, likelihood and uncertainty of each quality level, the comprehensive score Q is calculated according to the weights of confidence 0.6, likelihood 0.3 and uncertainty 0.1, the level with the maximum Q value is selected as the candidate quality level, and the candidate level and the corresponding comprehensive score are output; and the threshold size is dynamically adjusted according to the accuracy deviation;

[0050] The comprehensive score is compared and verified with the final threshold, the matching quality level is confirmed, the sample identification, detection time and world coordinate system range traceability information are associated, the threshold adjustment basis is recorded synchronously, and the determined quality level result is output.

[0051] The technical effects and advantages of the present application are:

[0052] 1. Non-destructive real-time and accurate analysis of component detection: the non-contact detection unit composed of a high-resolution fiber spectrometer and an integrating sphere is adopted in the scheme, the single sample component acquisition time is less than 20 seconds, and the detection can be completed without cutting the sample; through feature peak positioning and uniformity analysis, not only the content of components such as plasticizer and phthalate can be quantified, but also the internal distribution consistency can be reflected, effectively solving the problems of destructiveness and hysteresis of traditional chemical methods;

[0053] 2. Multi-dimensional accurate positioning of surface quality detection: multi-angle BRDF acquisition, polarization detection and micro-scan technology are integrated to synchronously acquire macroscopic reflection characteristics, layered polarization data and microstructure information. Through world coordinate system calibration of the production line, the detection data and the spatial coordinates are bound, the sources of layered defects can be accurately distinguished and the micro-abnormalities can be positioned, and the defects of insufficient detection dimension and fuzzy defect positioning of traditional single equipment are made up;

[0054] 3. Online collaborative evaluation of dynamic performance: with the aid of mechanical pressure and thermal stress dynamic excitation unit, combined with high-speed spectral detection and time sequence control module, dynamic response data of materials are captured in real time. Through the correlation of dynamic response model and component, surface model, the performance such as folding resistance and flexibility can be predicted online, and the production process can be synchronized without offline test, solving the problem of disconnection between dynamic performance evaluation and production;

[0055] 4. Multi-dimensional data collaboration and traceability: Through time and coordinate calibration, the spatial and temporal synchronization of composition, surface and dynamic data is realized, and a unified quality information system is formed through multi-model collaboration and feature fusion. The report can trace the sample identification, equipment status and detection data chain, and when quality problems occur, it can quickly associate the production link to completely solve the pain points of traditional detection data fragmentation and traceability difficulties. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 It is a schematic diagram of the overall structure of the present application.

[0057] Figure 2 It is a flowchart of S1 and S2 of the present application.

[0058] Figure 3 It is a flowchart of S3 of the present application.

[0059] Figure 4 It is a flowchart of S4 of the present application. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0061] Reference Figures 1-4 The artificial leather quality rapid detection method based on spectral analysis shown includes:

[0062] S1: Multi-dimensional raw data acquisition

[0063] Through the collaborative acquisition of composition characteristics, surface and micro characteristics, and dynamic response characteristics, combined with global space-time synchronization, full-dimensional raw data covering internal composition, surface state and dynamic performance are generated to adapt to the rapid detection needs of the production line, and the single sample acquisition time is less than 3 minutes, and the specific steps are as follows:

[0064] S101: Composition characteristic acquisition: quickly capture the spectral characteristics of the core chemical composition of artificial leather, establish a direct correlation between wavelength and composition content, provide original spectral data for subsequent composition purity calculation and harmful substance screening, and eliminate the thermodynamic dimension to shorten the acquisition period, and the single sample acquisition is less than 20 seconds, and the specific steps are as follows:

[0065] The spectral detection unit is composed of a preset high-resolution optical fiber spectrometer, a 50W halogen tungsten lamp and an integrating sphere. The ambient temperature is controlled at a constant 25±2°C by an environmental control unit to eliminate environmental temperature interference. A high-precision trigger associated only with the spectrometer is used to ensure a synchronization accuracy of 1 μs. The exit of the integrating sphere is fixed at a distance of 5 cm from the sample to ensure consistent optical path. The reflected light is directly connected to the spectrometer through an optical fiber to avoid spectral deviation caused by uneven illumination. The integration time is adjusted according to the color of the base material (50 ms for light-colored PU and 200 ms for dark-colored PVC), and is recorded as: Each sample continuously collects 3 spectra to reduce random noise, and the average value is taken.

[0066] Based on the collected full-waveband original spectral intensity data and component characteristic peak reflectivity data, including: continuous reflectivity curve in the wavelength range of 400-2500 nm and reflectivity mean value of plasticizer (1735 nm), resin (1720 nm) and phthalate (1240 nm);

[0067] S102: Surface and Microscopic Property Collection: Synchronously acquire multi-scale data of artificial leather macroscopic surface reflection, layered polarization characteristics and microscopic structure details, establish a traceability chain of spatial coordinates, surface state and microscopic defects, and provide multi-dimensional original data for subsequent surface uniformity evaluation, defect positioning and macroscopic performance prediction. The specific steps are as follows:

[0068] Pre-set multi-angle reflection unit: 360° rotating sample table, 0-90° rotating detector arm, standard diffuse reflection plate; polarization detection unit: 4-direction polarizer at light source end (0° / 45° / 90° / 135°), synchronous rotating polarization analyzer (4×4 polarization system) at detector end; microscopic scanning unit: 100× optical objective (NA=0.9, spatial resolution 5 μm), hyperspectral camera (400-2500 nm), automatic focusing module; linkage control unit: production line world coordinate system controller (X for width, Y for moving direction, Z for perpendicular to surface, conversion error ≤0.1 mm);

[0069] Macroscopic reflection and polarization linkage collection: the rotating sample table is initially at 0°, the incident light is fixed at 45°, and the detector is from 0° to 80°, with a step of 10°, a total of 9 steps to collect BRDF values, and the spectra of 4 polarization directions are recorded synchronously;

[0070] Microscopic scanning and macroscopic positioning association: first, 10 μm step conventional scanning of macroscopic area, then 2 μm step encryption scanning according to BRDF abnormal points;

[0071] Coordinate calibration: all data are attached with production line world coordinate system (x, y, z) labels to ensure spatial alignment of macroscopic and microscopic data;

[0072] Based on the collected BRDF values at 9 angles from 0° to 80°, each angle corresponding to 1 reflectivity curve from 400-2500 nm; Mueller matrix data at each reflection angle with 4 incident polarizations and 4 detection polarizations; conventional scanning: 400-2500 nm spectra corresponding to (x, y) coordinates under a 200 μm x 200 μm field of view, and high-resolution spectra corresponding to (x, y) coordinates in the defect area for encrypted scanning, to obtain microscopic local performance, denoted as: ; (x, y, z) coordinates corresponding to all macroscopic and microscopic data; BRDF values of a standard diffuse reflection plate, i.e., BRDF calibration data; micro-layer thickness d;

[0073] S103: Dynamic response characteristic acquisition: capturing the spectral dynamic changes of artificial leather under mechanical extrusion and local thermal stress, recording the transient response process of time and reflectivity, providing original dynamic data for subsequent quantification of molecular chain movement speed and evaluation of folding resistance or flexibility, with single scene acquisition time less than 30 seconds;

[0074] Pre-set dynamic excitation unit: mechanical pressure device (0.1-0.5 MPa instantaneous pressure, response ≤10 ms), 10 mW, 650 nm pulsed laser (thermal stress, trigger delay ≤5 μs); high-speed detection unit: high-speed photodetector (response ≤10 μs), high-resolution fiber spectrometer (shared with S101); timing control unit: high-frequency trigger (sampling frequency 1 kHz, 1 data point per millisecond).

[0075] Mechanical stress scenario: start collecting 50 ms before pressure application and continue for 200 ms after pressure, tracking the fluctuation of 1735 nm plasticizer peak reflectivity;

[0076] Thermal stress scenario: start collecting at 0 ms when the laser is triggered, continue for 500 ms, and record the relaxation curve of reflectivity from the initial value to the equilibrium value ;

[0077] Repeatability verification: repeat each scenario 2 times to ensure that the standard deviation of dynamic data is less than 5%;

[0078] Based on the collected two-dimensional data sets of time and reflectivity, characteristic peak dynamic fluctuation data, and dynamic repeatability data, and labeling each dynamic data with mechanical stress and thermal stress;

[0079] S104: Data global synchronization and coordinate unification: eliminate the spatiotemporal bias of S101-S103, correlate the component data, surface microscopic data, and dynamic data to the same spatial point of the same detection sample, avoid information misplacement during subsequent modeling, and ensure the synergy of multi-dimensional data;

[0080] Time synchronization: Connect S101 spectrometer, S102 coordinate calibration, S103 time sequence control unit with GPS time trigger (accuracy 1 μs) to unify the start and stop time of data acquisition;

[0081] Space synchronization: Record the position of Y-axis of the material (accuracy 0.1 mm) through the production line encoder, convert the timestamp of S103 into Y-axis coordinate, and align it with the world coordinate system (x, y, z) of S102;

[0082] Periodic calibration: Calibrate the conversion matrix of the local coordinate system of each device to the world coordinate system (error ≤0.1 mm) with a 0.01 mm precision checkerboard calibration plate every hour;

[0083] Synchronized output: Space-time related data set: each data point is attached with world coordinates (x, y, z), timestamp, and excitation label; Coordinate system conversion matrix: conversion matrix of S101 stage, S102 scanning table, S103 pressure device to the world coordinate system of the production line;

[0084] S2: Multi-dimensional data modeling and function construction

[0085] The original spectrum and space-time data collected in S1 are converted into parameterized results that can be directly related to the quality of artificial leather through four models: component composition analysis, surface micro-correlation, dynamic response quantification, and global synchronization optimization. The specific steps are as follows:

[0086] S201: Component composition characteristic analysis model: From the wavelength and reflectivity of the original spectrum of S101, remove noise interference, extract characteristic signals directly related to component content, realize quantitative analysis of harmful substance concentration, quantify chemical composition and harmful substances, and provide parameter support for component safety evaluation. The specific mathematical function is as follows:

[0087]

[0088] Where, is the effective spectral response at wavelength , is the spectral wavelength, is the original spectral intensity at wavelength at time t, is the reflectivity reference of the component characteristic peak center wavelength in the component characteristic peak reflectivity data, is the spectral half-width coefficient, which is [2 nm, 3 nm], and is fine-tuned according to the specific characteristic peak, is the effective window time of spectral acquisition; and based on this model, the concentration of each harmful substance is calculated by eliminating the matrix effect according to the Lambert-Beer law or standard curve method;

[0089] S202: Surface and microcosmic property correlation model: analyze the surface reflection uniformity, layered defect position, microcosmic and macroscopic performance correlation, cover the artificial leather appearance grade and structure integrity evaluation requirements;

[0090] Surface reflection uniformity model: The derived parameters for quantifying uniformity: Wherein The reflection radiation luminance generated by a unit incident light irradiance in a certain reflection direction, The polar and azimuth angles of the incident light, The polar and azimuth angles of the reflected light, the polar angle The angle between the incident light and the sample surface normal, S102 is fixed at 45°, the polar angle The angle between the reflected light and the sample surface normal, S102 collects from 0° to 80°, the azimuth angle is fixed at 0°, The reflection radiation luminance, directly collected by the detector of S102, The incident irradiance, calibrated by the standard diffuse reflection plate of S102, A is the anisotropy coefficient, the comprehensive score of surface reflection uniformity;

[0091] Layered defect positioning model: describe the polarization characteristic difference of each layer of artificial leather through Mueller matrix, locate the defect position through matrix decomposition, solve the problem that macroscopic detection cannot distinguish the defect source, and the specific mathematical function is: Wherein The Mueller matrix data of artificial leather, representing the polarization characteristic comprehensive response of all layers, including the depolarization ratio and dichroism ratio of each layer; The incident light polar angle, fixed at 45°, N is the number of layers of artificial leather, The Mueller matrix of the kth layer;

[0092] Microcosmic and macroscopic performance correlation model: realize the cross-scale correlation of microcosmic data and macroscopic quality by establishing the quantitative relationship between microcosmic local performance and macroscopic performance, and the specific mathematical function is as follows:

[0093]

[0094] Wherein, The macroscopic performance parameter represents the overall use performance of artificial leather, The local performance of microcosmic position (x, y, z) represents the local quality of microstructure, The weighting function represents the contribution weight of the micro-region to the macro-performance: based on the failure mechanism of artificial leather, the surface region z≤5μm has a weight of 1.0, because the surface is the primary area affected by wear and scratches, and its micro-quality directly determines the macro-wear resistance; the internal region z>5μm has a weight of 0.3, because the internal substrate only plays a major role in macro-stretching and has a small contribution to the surface-related performance. V is the volume of the test sample, representing the representative area of ​​the micro-analysis: fixed at 1mm×1mm×0.1mm.

[0095] S203: Dynamic response characteristic model: By fitting the time and reflectivity dynamic data collected by S103 through the double exponential relaxation equation, parameters such as relaxation time constant are extracted to quantify the molecular chain motion capability of artificial leather under mechanical stress and thermal stress.

[0096] Its specific mathematical function is: ,in The relaxation time constant (ms) is a core indicator representing the speed of molecular chain movement: it reflects the time it takes for the molecular chain of artificial leather to recover from the deformed state to the equilibrium state after being subjected to external force and heat. If the value is less than 200ms, it means that the molecular chain moves fast, the material is flexible, and the number of folding cycles is greater than 100,000, which indicates that it is a high-quality product.

[0097] If the value is greater than 500ms, it indicates that the molecular chain movement is slow, the material is rigid, and the folding resistance is less than 10,000 times, which is a defective product; otherwise, it is a normal product. Initial time The characteristic peak reflectance, For stable moments The characteristic peak reflectance, The attenuation coefficient is... The initial relaxation amplitude is preset. The fast relaxation coefficient, For dynamic response time;

[0098] S204: Data Synchronization Optimization Model: Eliminates the spatiotemporal offset of the three model parameters: composition, surface, and dynamics. Its specific mathematical functions are as follows:

[0099]

[0100] in, For serial number, For component parameters, For surface micro parameters, For dynamic response parameters, These are the parameters after the offset. , where is the time offset. ) represents the spatial offset. The data offset sum of squares is a quantitative index of multi-dimensional parameter synchronization, and the value less than 0.01 indicates that the synchronization is qualified and can be used for subsequent feature fusion; otherwise, the synchronization fails and the equipment needs to be recalibrated;

[0101] S3: Multi-dimensional feature extraction and fusion

[0102] Taking the interpretable quality parameters output by S2 as input, through single-dimensional feature extraction, standardization, cross-direction fusion, attention enhancement, and time sequence optimization, the high-dimensional and dispersed quality parameters are converted into low-dimensional, correlated, and high-information fusion features, which not only retain the core quality information of components, surfaces, and dynamics, but also adapt to the real-time detection needs of the production line, with a response time of less than 1s. The specific analysis is as follows:

[0103] S301: Component composition feature extraction: From the component purity and harmful substance concentration parameters output by S201, sensitive features to component safety are selected, and redundant noise is removed to form a quantifiable component feature vector;

[0104] Through adaptive Morlet wavelet transform, the feature peak intensity data at room temperature output by S201 is input, containing 3 feature peaks: 1735nm, 1720nm, and 1240nm;

[0105] Wavelet decomposition: Each feature peak intensity curve is decomposed by 5 layers of wavelet, the 1st layer is 400-800nm, the 5th layer is 2000-2500nm, and each layer is 400nm;

[0106] Feature selection: Calculate the energy proportion of each layer of wavelet coefficients, and retain the top 10% coefficients;

[0107] Feature dimension: A total of 90 features are extracted: 5 layers x 3 feature peaks x 6 key wavelength points;

[0108] S302: Surface and microscopic feature extraction: From the surface reflection uniformity, layered defects, microscopic and macroscopic correlation parameters output by S202, sensitive features to appearance grade and structural integrity are extracted to form a surface microscopic feature vector;

[0109] Input data: Surface reflection parameters: A values at each angle , polarization defect parameters: each layer of depolarization ratio, dichroism ratio, microscopic correlation parameters: microscopic coating thickness d(x, y, z) output by S202;

[0110] Extracted features: surface reflection features: 9 angle A values and one full-angle mean, a total of 10 dimensions; polarization defect features: 3-layer depolarization ratio, 3-layer dichroism ratio, and 2-layer defect probability, a total of 8 dimensions; microscopic correlation features: statistical quantities of microscopic coating thickness: mean, standard deviation, maximum, minimum, and defect area ratio, a total of four groups of data, a total of 20 dimensions;

[0111] S303: Dynamic response feature extraction: from the dynamic relaxation curve and relaxation time output from S203, extract time-frequency features sensitive to flexibility and folding resistance to form a dynamic response feature vector;

[0112] Input data: dynamic relaxation curve and relaxation time constant ; decompose into 8 intrinsic modal functions: , do Hilbert transform for each IMF to calculate instantaneous frequency , take the mean ;

[0113] Calculate the amplitude of the IMF , calculate the decay rate: , calculate the time when the IMF amplitude reaches the maximum , extract three parameters of the eight IMFs , a total of 24 dimensions;

[0114] S304: Single-direction feature standardization preprocessing: eliminate the dimension and numerical range differences of the three types of features of composition, surface, and dynamic to avoid certain features from masking other features due to large values;

[0115] Use Z-Score standardization: convert the original features to standard normal distribution features with a mean of 0 and a standard deviation of 1;

[0116] S305: Cross-direction third-order tensor fusion: construct the internal correlation of composition, surface, and dynamic through a third-order tensor, extract correlation features that can reflect multi-dimensional collaborative quality, and avoid single-dimensional misjudgment;

[0117] Through CP decomposition, decompose the third-order tensor of composition, surface, and dynamic into the sum of factor vectors outer product, retain the correlation information of the three directions, and realize dimension compression at the same time;

[0118] Third-order tensor T: dimension 90*38*24, each element is:

[0119]

[0120] where, is the normalized single-direction feature;

[0121] ​Decomposition rank R: R = 40; Factor vectors: component factor vectors (90-dimensions, r = 1-40); surface factor vectors (38-dimensions, r = 1-40); dynamic factor vectors (24-dimensions, r = 1-40), representing the core characteristics of the r-th association pattern in each direction;

[0122] Outer product operation: each element is: , generating a 90x38x24 third-order tensor slice, denoted as

[0123] Fusion features: the core strength values of the 40 outer product tensor slices, denoted as the strength of the 40 association patterns;

[0124] S306: Attention enhancement and feature compression: according to the detection scene, automatically enhance the weight of the corresponding direction feature; and compress the fusion features from the original 40+152-dimensional standardized features to 64-dimensional;

[0125] First, adjust the weight through attention, and then compress the dimension through Tucker decomposition;

[0126] Q-K-V attention mechanism: global query vector Q: 128-dimensional global query vector generated by linear transformation and nonlinear activation of 40-dimensional fusion features through fully connected layer on 40-dimensional fusion features, the mathematical formula is: , wherein is the weight matrix, is the bias vector, which is denoted as reflecting the current detection scene demand;

[0127] Direction key vector K: 3 128-dimensional vectors: component key : generated from 90-dimensional component features; surface key : generated from 38-dimensional surface features; dynamic key : generated from 24-dimensional dynamic features; its representation is the identity of each direction feature;

[0128] Direction value vector : 3 128-dimensional vectors, one-to-one corresponding to K: , , , representing the core information value of each direction feature;

[0129] Attention weight: calculated by softmax: , representing the importance proportion of each direction feature under the current scene;

[0130] 128-dimensional basic fusion features: ;

[0131] Tucker decomposition: Decomposes the 128-dimensional features into a core tensor and factor matrices, outputting 64-dimensional compressed features. ;

[0132] S307: Temporal Fusion: For scenarios where roll materials move continuously on a production line, it fuses multi-frame compression features, smooths random noise in single-frame features, and simultaneously achieves quality assessment of continuous areas, avoiding misjudgment in single frames.

[0133] Gated Circular Unit (GRU): Controls the fusion ratio of historical features and current features by updating and resetting gates;

[0134] One frame of 64-dimensional compressed features is acquired every 100ms from the input feature frame: , For the first Frame features;

[0135] Initial hidden state 64-dimensional, with values ​​representing the features of the first frame. ;

[0136] Update Gate : , It is the sigmoid function, with values ​​[0, 1].

[0137] Reset door : The value can be [0, 1].

[0138] Candidate hidden state : , For element-wise multiplication, the values ​​are: [-1, 1];

[0139] Final hidden state : ;

[0140] Timing optimization feature: Output 1 every 8 frames 64 dimensions in total;

[0141] S4: Quality Assessment and Grading System

[0142] Through index quantification, feature concentration, correlation enhancement, evidence fusion, reliability assessment, and graded output, the system achieves simultaneous assessment of component safety, surface quality, and dynamic performance, and outputs traceable reports.

[0143] S401: Multi-task quality assessment model: Transforms fusion features into specific quality indicators that can be benchmarked against industry standards;

[0144] Basic input: 128-dimensional basic fusion features of the S306 output after standardization;

[0145] Dynamic input: 64-dimensional time-series optimized features output by S307, filtered by 3σ outliers;

[0146] Input strategy: static and dynamic weighted fusion: time-series optimized feature weight 0.3, basic fusion feature weight 0.7, generate 128-dimensional final input features;

[0147] Model structure: shared feature layer: 4-layer CNN, each layer contains convolution operation and ReLU activation function, where the convolution and uniformity are 3x3, step 1, output 64-dimensional shared features;

[0148] Output layer: 3 task-specific fully connected layers:

[0149] Component safety task 1: component purity, harmful substance concentration, directly related to the concentration of each harmful substance in S201;

[0150] Surface quality task 2: defect rate, gloss uniformity, the former is based on the hierarchical defect position and micro-defect proportion calculated by S202, and the latter is directly quantified by the anisotropy coefficient A of S202;

[0151] Dynamic performance task 3: folding resistance, tensile strength, the former is based on the relaxation time constant conversion, and the latter is based on the macro-performance prediction value of S202.

[0152] Use loss function to ensure accuracy: , where , , MSE loss of component, surface, and dynamic tasks, respectively, L2 regularization term to prevent overfitting, output six original quantitative values of the indicators.

[0153] S402: Hierarchical quality feature encoding: encode the six indicators output by S401 through an encoding network to condense them into low-dimensional features;

[0154] Input data: component purity, harmful substance concentration, defect rate, gloss uniformity, folding resistance, and tensile strength, and normalize them;

[0155] Encoding network structure: the encoder passes through a 5-layer fully connected network with dimension changes of 6, 16, 32, 64, and 6, and is based on the Sigmoid activation function, outputting 6-dimensional hierarchical quality features, where the first dimension is component safety, dimensions 2-3 are surface quality, and dimensions 4-6 are dynamic performance;

[0156] Decoder reconstructs the original indicators through a symmetric 5-layer fully connected network with dimensions varying as: 6, 64, 32, 16, 6, and reconstruction error;

[0157] Output 6-dimensional hierarchical quality features;

[0158] S403: Feature correlation enhancement: mining the internal correlation rules among 6-dimensional quality features, correcting isolated indicator bias, and improving evaluation consistency;

[0159] Input data and correlation graph construction:

[0160] Input data: 6-dimensional hierarchical quality features output by S402 and pre-computed Pearson correlation coefficient matrix between features.

[0161] Correlation graph construction:

[0162] Nodes: 6 quality features, node features are 6-dimensional hierarchical features;

[0163] Edge weight: Pearson correlation coefficient between nodes

[0164] GCN network structure and correlation enhancement logic:

[0165] Network structure: First layer: from 6-dimensional to 16-dimensional, activated by ReLU function, graph convolution formula: where ReLU activation function, is the output feature vector of node after the first layer of graph convolution, is the degree of node , and is the degree of node j;

[0166] Second layer: from 16-dimensional to 32-dimensional, activated by ReLU function, and aggregated high-order correlation;

[0167] Feature aggregation: 32-dimensional node feature average pooling to generate 32-dimensional correlation enhanced quality features; output 32-dimensional correlation enhanced quality features;

[0168] S404: Multi-evidence fusion decision: considering the outputs of the three tasks in S401 as independent evidence sources, and making a comprehensive judgment on the quality level;

[0169] Input: defining the three tasks in S401 as evidence sources, i.e., evidence source 1 component security, evidence source 2 surface quality, and evidence source 3 dynamic performance;

[0170] Define quality level framework: ;

[0171] And through the membership function to convert the index into each level of trust, that is, m value, the sum of all m values is 1;

[0172] Calculate the conflict coefficient k between evidences, which measures the consistency of evidence, and the calculation method is as follows:

[0173]

[0174] k=0 represents complete consistency, and k=1 represents complete conflict. When k>0.5, the weighted average method is used to adjust the m value;

[0175] For the evidence that meets the no conflict or low conflict k<0.5, the fusion method is as follows:

[0176]

[0177] Among them, The denominator is the conflict normalization factor, which avoids the probability exceeding 1;

[0178] Output the fused basic probability assignment m value;

[0179] S405: Confidence and uncertainty quantification: evaluate the reliability of the fusion decision to avoid misjudgment of the boundary level sample;

[0180] Input the fused m value output by S404, generate all subsets of the level framework Θ; Then calculate its confidence: Likelihood: Uncertainty:

[0181] Output the Bel, Pl and Unc values of each level;

[0182] S406: Quality grading: calculate the comprehensive score to determine the level;

[0183] Comprehensive score calculation: Select The maximum one is the candidate level;

[0184] And set the reference threshold: {excellent greater than or equal to 0.8, good [0.6, 0.79], medium [0.4, 0.59], poor less than 0.4};

[0185] Dynamic optimization: every 1000 samples are detected, if the accuracy is <95%, the threshold size is dynamically adjusted: L=1-accuracy, The basic threshold is The adjusted threshold size is

[0186] Final output: traceable quality inspection report, including: 6 original indicators, comprehensive score, quality level, reliability evaluation, threshold adjustment record, associated features.

[0187] Secondly: the drawings of the disclosed embodiments only involve the structures involved in the disclosed embodiments, other structures can refer to the usual design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other;

[0188] Finally: the above only describes the preferred embodiments of the present application and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for rapid detection of the quality of artificial leather based on spectral analysis, characterized in that, Comprise: S1: multi-dimensional original data acquisition: collect the composition characteristics, surface and microstructure characteristics and dynamic response characteristics data of the artificial leather sample, and perform data global synchronization and coordinate unification on the data to obtain preprocessed data; The multi-dimensional original data acquisition specifically comprises: Composition characteristic acquisition: collect full-waveband original spectrum data of the artificial leather sample, including characteristic peak reflectivity and mean value data of plasticizer, resin and phthalate; Surface and microstructure characteristic acquisition: collect multi-angle reflection characteristic data BRDF value, polarization spectrum and Mueller matrix data of the sample, standard diffuse reflection plate calibration data, synchronously collect spatial and spectral three-dimensional data and micro-coating thickness of the micro-region, and add world coordinate system labels corresponding to each collection point; Dynamic response characteristic acquisition: collect characteristic peak reflectivity fluctuation data of the sample over time under mechanical pressure and thermal stress dynamic excitation, including time and reflectivity two-dimensional change curve before and after excitation and dynamic repeatability verification data; Data global synchronization and coordinate unification: time stamp alignment and spatial coordinate calibration are performed on the composition characteristics, surface and microstructure characteristics and dynamic response characteristics data, and finally an associated data set with additional space-time coordinates and time stamps is formed; S2: multi-dimensional data modeling and function construction: based on the preprocessed data, a composition characteristic analysis model, a surface and microstructure characteristic correlation model and a dynamic response characteristic model are constructed, and the three models are optimized using a data synchronization optimization model to eliminate space-time deviation, and finally the model output parameters after synchronization optimization are obtained; The composition characteristic analysis model specifically comprises: Based on the full-waveband spectrum data, characteristic peak reflectivity and mean value data in the preprocessed composition characteristic acquisition data, baseline correction and smoothing denoising are first performed on the spectrum data, the effective spectral response at the wavelength is calculated by combining the Gaussian function, the noise interference is removed, and the standardized spectrum data set is obtained; A feature peak positioning module is constructed based on the data set, and the characteristic peak center wavelength and corresponding reflectivity interval of different components are located by comparing the known component standard spectrum library; The quantitative mapping relationship of reflectivity, component concentration and purity is constructed by combining the Lambert-Beer law and using the standard sample with known component content to calibrate the molar absorption coefficient and the optical path parameter, the matrix effect is eliminated to improve the accuracy, and a quantitative analysis module is formed; The surface and microstructure characteristic correlation model specifically comprises: Surface reflection uniformity sub-model: based on the multi-angle BRDF value and standard diffuse reflection plate calibration data collected in the surface and microstructure characteristic acquisition as input, the mapping relationship between incident light irradiance and reflected radiation brightness is constructed based on the BRDF core function, the anisotropy coefficient is calculated to quantify the surface reflection uniformity at each reflection angle, and the full-angle reflection uniformity comprehensive score is outputted; Layered defect positioning sub-model: based on the polarization spectrum and Mueller matrix data as input, the polarization characteristic comprehensive response of the artificial leather is described by multi-layer Mueller matrix multiplication operation, the spatial position of the layered defect is located by extracting the depolarization ratio and dichroism ratio parameters of each layer from the decomposition matrix, and the defect layer position identifier and defect probability are outputted; Microscopic and macroscopic performance correlation sub-model, which takes the microscopic space and spectral three-dimensional data, the microscopic coating thickness and the corresponding world coordinate system label as the input, performs volume integral operation on the microscopic local performance based on the weight function, constructs the cross-scale correlation between the microscopic structure and the macroscopic performance, and outputs the macroscopic performance parameter prediction value; The dynamic response characteristic model comprises: A relaxation parameter extraction submodule extracts a relaxation time constant, a decay coefficient, and a fast relaxation coefficient based on time and reflectivity dynamic data, and outputs a relaxation characteristic parameter group by presetting an initial relaxation amplitude and fitting a reflectivity relaxation curve from an initial value to an equilibrium value using a double exponential relaxation equation; and a dynamic performance quantification submodule constructs a parameter mapping relationship using the relaxation time constant and a dynamic scene label as inputs. S3: Multi-dimensional feature extraction and fusion: based on the parameters output by the three models after data synchronization, multi-dimensional features are extracted and standardized, and the standardized multi-dimensional features are subjected to basic cross-direction third-order tensor fusion to obtain cross-direction fusion features, and feature fusion and compression are performed based on an attention mechanism to obtain basic fusion features and real-time compression features, and a gated recurrent unit is used to perform time sequence correlation fusion on the real-time compression features to generate time sequence optimization features. S4: Quality evaluation and grading system: based on the basic fusion features and the time sequence optimization features, an original quantitative value of an index output by a multi-task quality evaluation model is constructed, and hierarchical feature encoding is performed on the original quantitative value to obtain low-dimensional core features; an associated enhanced feature is generated based on the correlation between the low-dimensional core features by a graph neural network; a comprehensive decision is made on the associated enhanced feature using evidence fusion rules to obtain the trust degree of each quality grade; the confidence, likelihood, and uncertainty of the trust degree are quantified, and quality grading is performed in combination with a dynamic threshold adjustment strategy to output the quality grade of the artificial leather sample and a traceable detection report.

2. The rapid detection method for quality of artificial leather based on spectral analysis according to claim 1, characterized in that: The multi-dimensional feature extraction comprises: Component composition feature extraction, which takes the concentration parameters output by the component composition characteristic analysis model as input, decomposes the 1735 nm, 1720 nm, and 1240 nm characteristic peak intensity curves by 5 layers through an adaptive Morlet wavelet, selects the top 10% wavelet coefficients in terms of energy, extracts 6 key wavelength point features, and outputs a 90-dimensional component feature vector; Surface and microscopic feature extraction, which takes the parameters output by the surface and microscopic characteristic correlation model as input, extracts the anisotropy coefficients and mean values at 9 angles, a total of 10 dimensions, the layer-by-layer depolarization ratio and defect probability, a total of 8 dimensions, the microscopic coating thickness statistics, a total of 20 dimensions, and finally outputs a 38-dimensional surface and microscopic feature vector; A dynamic response feature extraction submodule takes the relaxation time constant and the folding resistance index parameter output by the dynamic response characteristic model as input, decomposes the relaxation curve into 8 eigenmode functions, calculates the instantaneous frequency mean, amplitude decay rate, and peak time through Hilbert transform, and outputs a 24-dimensional dynamic feature vector. A unidirectional feature standardization submodule, which adopts Z-Score standardization processing on the three types of feature vectors mentioned above, converts them into standard features with a mean of 0 and a standard deviation of 1, eliminates dimensional differences, and finally outputs the standardized component, surface, and dynamic feature vectors. 3.The method for rapidly detecting quality of artificial leather based on spectral analysis according to claim 1, characterized in that: The basic fusion feature and the timing optimization feature include: The basic fusion feature is based on the standardized multi-dimensional feature vector, and a query vector, a direction key vector, and a value vector that are adapted to the scene are constructed through a Q-K-V attention mechanism. After calculating the attention weights of each direction feature, the 128-dimensional value vector is weighted and summed to generate a 128-dimensional basic fusion feature. The timing optimization feature submodule takes the 64-dimensional real-time compressed feature obtained by Tucker decomposition of the basic fusion feature as input, collects continuous feature frames at a rate of 100 ms per frame, filters out abnormal values, and inputs them into a GRU network. Through the update gate and the reset gate, the historical and current frame information is fused, and one 64-dimensional aggregated feature is output every 8 frames.

4. The rapid detection method for quality of artificial leather based on spectral analysis according to claim 1, characterized in that: The multi-task quality evaluation model includes: Input feature fusion is based on the basic fusion feature and the timing optimization feature, and the 128-dimensional final input feature is generated according to the static and dynamic weighted fusion rules. Shared feature extraction is based on the 128-dimensional final input feature, and a 4-layer CNN is used to realize common feature abstraction: each layer uses a 3×3 convolution kernel, a step of 1, and a ReLU activation function. After hierarchical processing such as local feature capture, middle-level aggregation, high-level abstraction, and feature condensation, a 64-dimensional shared feature is output. Task-specific output is based on the 64-dimensional shared feature, and three fully connected layers are used to output the results of the three main tasks: component safety task outputs component purity and harmful substance concentration; surface quality task outputs defect rate and gloss uniformity; dynamic performance task outputs folding resistance and tensile strength, with a total of 6 original quantitative values. The loss function uses a weighted MSE loss combined with an L2 regularization term.

5. The rapid detection method for quality of artificial leather based on spectral analysis according to claim 1, characterized in that: The evidence fusion rule includes: Evidence source preprocessing takes the results of the three main tasks output by the multi-task quality evaluation model as input, and defines component safety, surface quality, and dynamic performance as three independent evidence sources. According to the quality level framework of excellent [90, 100], good [80, 89], medium [60, 79], and poor [0, 59], the task indicators are converted into trust value values corresponding to the level by using the membership function, and the evidence source trust value matrix is output. Conflict coefficient calculation is based on the above trust value matrix to calculate the conflict coefficient between each evidence source to measure consistency. A coefficient of 0 indicates complete agreement, a coefficient of 1 indicates complete conflict, and a coefficient of 0.5 is used as a boundary to distinguish between low conflict and high conflict scenarios. The conflict coefficient and consistency judgment result are output. Hierarchical fusion decision-making executes adaptive strategies according to the conflict judgment result: in a low conflict scenario, the D-S evidence theory is used for trust value fusion; in a high conflict scenario, the trust values of each evidence source are adjusted by using the weighted average method, and then fusion is performed, outputting the fused basic probability distribution of each level. The fusion result output verifies the standardization of the probability distribution and outputs the final trust value data of each quality level. 6.The method for rapidly detecting quality of artificial leather based on spectral analysis according to claim 1, characterized in that: The quality grading includes: Based on the confidence, likelihood and uncertainty of each quality level, the comprehensive score Q is calculated according to the weights of confidence 0.6, likelihood 0.3 and uncertainty 0.1, the level with the maximum Q value is selected as the candidate quality level, and the candidate level and the corresponding comprehensive score are output; and the threshold size is dynamically adjusted according to the accuracy deviation; The comprehensive score is compared with the final threshold, the matching quality level is confirmed, the sample identification, detection time and world coordinate system range traceability information are associated, the threshold adjustment basis is recorded synchronously, and the determined quality level result is output.

Images