EVA-based laminated non-woven fabric product quality traceability method and system

Abstract

The present application relates to the technical field of non-woven fabric production quality control, in particular to a quality tracing method and system for EVA laminated non-woven fabric products, comprising: synchronously collecting time sequence process parameters of the whole production chain through a sensor network to generate original parameter flow bound with production volume, obtaining baseline, rhythm and disturbance components through multi-scale decomposition, and mapping them to product raw material layer, interface bonding layer and surface layer to generate layered quality parameter set. A defect mode sample library is constructed, potential defect evolution paths are identified through unsupervised clustering, dynamic matching degree between current parameter set and early stage of defect evolution path is calculated in real time, and if the dynamic matching degree exceeds the threshold, parameter tracing and root cause positioning of the corresponding process link are triggered. The present application realizes the hierarchical correlation between process parameters and product structure, can identify defect evolution characteristics in advance, and improves the accuracy and proactivity of EVA laminated non-woven fabric quality tracing.

Description

Technical Field

[0001] This invention relates to the field of quality control technology for nonwoven fabric production, and in particular to a method and system for tracing the quality of EVA-coated nonwoven fabric products. Background Technology

[0002] Existing EVA-coated nonwoven fabric production quality traceability technologies collect process parameters from raw material feeding, melt blending, casting and lamination, and winding stages using sensors at the production site. These parameters are simply bound and stored to individual production rolls, providing only a general record and post-production review of the raw parameter stream. They fail to perform multi-scale decomposition of the time-series parameter stream, making it impossible to distinguish different types of parameter variations. Furthermore, existing technologies do not establish a correspondence between process parameters and the physical structure layers of EVA-coated nonwoven fabrics. They cannot correlate parameter changes with the quality formation processes of the raw material layer, interface bonding layer, and surface layer. Quality traceability relies solely on passive analysis of finished product inspection results, failing to achieve a precise correspondence between parameter characteristics and product structural quality.

[0003] Existing traceability solutions can only check parameters of a single process step for existing defects, failing to identify common defect evolution patterns through historical data mining. They lack a mechanism for quantitative identification and dynamic matching of early defect characteristics, making it difficult to trigger traceability processes in the early stages of defect formation. The tracing process suffers from problems such as ambiguous feature differentiation, broad process location, and inaccurate root cause determination. This invention aims to complete multi-scale decomposition of the original production parameter flow and hierarchical mapping with the product structure. Simultaneously, it uses unsupervised clustering to identify potential defect evolution paths, achieving dynamic matching between current production parameters and early defect characteristics, thus enabling precise backtracking and location of defect root causes. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a product quality traceability method and system based on EVA-coated nonwoven fabric.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for quality traceability of EVA-coated nonwoven fabric products, comprising:

[0006] By deploying a sensor network on the production site, the timing process parameters of the raw material feeding, melt blending, casting and laminating and winding stages that constitute the EVA laminated nonwoven fabric production process chain are collected synchronously, and a unique original parameter stream is generated for each independent production roll.

[0007] The original parameter flow is decomposed into a multi-scale component to separate the baseline component representing long-term process drift, the rhythm component representing periodic fluctuations, and the disturbance component representing instantaneous anomalies.

[0008] The baseline component, rhythm component, and disturbance component are mapped according to the physical structure hierarchy of the laminated nonwoven fabric and associated with the raw material layer, interface bonding layer, and surface layer, respectively, to generate a layered quality parameter set.

[0009] Based on the hierarchical quality parameter set of historical production volumes and the final product inspection records, a defect pattern sample library is established. Unsupervised clustering method is used to identify common potential defect evolution paths from the defect pattern sample library.

[0010] For the current production roll, the dynamic matching degree between its hierarchical quality parameter set and the early stages of each potential defect evolution path is calculated in real time. When the dynamic matching degree exceeds a preset sensitivity threshold, a quality traceability process for the current production roll is triggered. The quality traceability process includes parameter backtracking and root cause localization of the process links involved in the matched potential defect evolution path.

[0011] As a further aspect of the present invention, a sensor network deployed on the production site is used to synchronously collect the timing process parameters of the raw material feeding stage, melt blending stage, casting and laminating stage, and winding stage constituting the EVA laminated nonwoven fabric production process chain, generating an original parameter stream uniquely bound to each independent production roll, specifically including:

[0012] During the raw material feeding stage, batch codes, feeding times, and initial temperature parameters of different batches of EVA particles and non-woven fabric substrates were collected.

[0013] In the melt blending process, the continuous variation curves of temperature, melt pressure and melt viscosity in each zone of the screw of the blending equipment are collected;

[0014] In the casting and lamination process, data on die temperature distribution, cooling roller temperature, traction speed, and online thickness spectrum are collected.

[0015] During the winding process, the results of visual inspection of winding tension, roll diameter changes, and surface defects are collected.

[0016] Each independent production roll is assigned a globally unique identifier, and the time-series process parameters of all stages are spliced ​​and packaged using timestamps as indexes to form the original parameter stream of the production roll.

[0017] As a further aspect of the present invention, the original parameter flow is decomposed into multiple scales to separate the baseline component characterizing long-term process drift, the rhythm component characterizing periodic fluctuations, and the disturbance component characterizing instantaneous anomalies, specifically including:

[0018] An adaptive sliding window mean filtering method is used to process the original parameter stream, filter out high-frequency noise, and the resulting smooth trend line is used as the baseline component characterizing long-term process drift.

[0019] The baseline component is subtracted from the original parameter stream to obtain the residual sequence. The residual sequence is subjected to spectral analysis to identify the frequency components related to the equipment operation cycle and feeding cycle. The periodic components are extracted and reconstructed as rhythmic components characterizing periodic fluctuations.

[0020] The rhythmic components are further removed from the residual sequence, and the remaining signal is the disturbance component that represents the instantaneous anomaly. A dynamic threshold is set for the disturbance component to mark the abnormal pulse.

[0021] As a further aspect of the present invention, the baseline component, rhythm component, and disturbance component are mapped according to the physical structure hierarchy of the laminated nonwoven fabric, and respectively associated with the raw material layer, interface bonding layer, and surface layer to generate a layered quality parameter set, specifically including:

[0022] The baseline components of the melt blending process, especially the long-term trend of melt viscosity, are mapped to raw material layer quality parameters that affect overall uniformity.

[0023] The rhythmic components of the casting and laminating process, especially the periodic fluctuations in temperature and speed, are mapped as interfacial bonding layer quality parameters that affect the bonding strength between the EVA film and the nonwoven substrate.

[0024] The disturbance components of each stage, especially abnormal pulses exceeding the dynamic threshold, are mapped to surface quality parameters that lead to surface defects or localized weak points.

[0025] The mapped raw material layer quality parameters, interface bonding layer quality parameters, and surface layer quality parameters are organized according to the same unique identification code for production volume to form a structured set of layered quality parameters.

[0026] As a further aspect of this invention, a defect pattern sample library is established based on the hierarchical quality parameter set of historical production volumes and the final product inspection records. Unsupervised clustering methods are then used to identify common potential defect evolution paths from the defect pattern sample library, specifically including:

[0027] From historical production records, select production rolls with final finished product inspection records showing non-conforming products, extract the stratified quality parameter set corresponding to the production rolls, and form an initial defect sample set;

[0028] For each sample in the initial defect sample set, the hierarchical quality parameters are sliced ​​at equal time intervals from the start of production to the time when the defect is detected, forming a series of parameter state snapshots arranged in chronological order.

[0029] All parameter state snapshots are input into an unsupervised clustering model, which assigns the snapshots to different clusters based on their proximity in the multidimensional feature space.

[0030] The transition sequences of a series of snapshots belonging to the same historical defect sample are analyzed among different clusters. Frequently occurring transition sequences are extracted and defined as a potential defect evolution path from an early process state to a final defect state. A path feature template is then established for each path.

[0031] As a further aspect of the present invention, for the current production roll, the dynamic matching degree between its layered quality parameter set and the early stages of each potential defect evolution path is calculated in real time, specifically including:

[0032] The hierarchical quality parameter set of the current production volume from the start to the current moment is acquired in real time, and it is divided into a parameter state snapshot sequence with the same interval as the potential defect evolution path definition in time order;

[0033] The current parameter status snapshot sequence is compared stage by stage with the path feature template of each potential defect evolution path;

[0034] The comparison process calculates the multidimensional feature distance between the current sequence and the corresponding path template at the center of the stage cluster at each stage, and performs a weighted summation according to the time weight to obtain the overall dynamic matching score between the current production state and the potential defect evolution path.

[0035] Maintain a real-time dynamic matching score for each potential defect evolution path.

[0036] As a further aspect of the present invention, when the dynamic matching degree exceeds a preset sensitivity threshold, a quality traceability process for the production roll is triggered. This quality traceability process includes parameter backtracking and root cause localization of the process steps involved in the potential defect evolution path of the matched roll, specifically including:

[0037] Preset sensitive thresholds for the evolution paths of potential defects of different severity levels;

[0038] The system monitors the dynamic matching score of all potential defect evolution paths in real time. When any score exceeds its corresponding sensitivity threshold, the system automatically triggers the quality traceability process.

[0039] The tracing process first identifies the potential defect evolution path that exceeds the matching limit, analyzes the feature template of the potential defect evolution path, and determines the defect type and key impact stage it represents.

[0040] Based on the key impact stage, trace back to the raw material layer, interface bonding layer, or surface layer parameters corresponding to the stratified quality parameter set.

[0041] Further linking back to the original parameter stream, the system pinpoints the specific production process, equipment, or even sensor that generated the abnormal parameter data, and generates a preliminary root cause analysis report that includes the time range, location, abnormal parameter values, and associated defect types.

[0042] As a further aspect of the present invention, the step of using an adaptive sliding window mean filtering method to process the original parameter stream, filtering out high-frequency noise, and obtaining a smooth trend line as a baseline component characterizing long-term process drift specifically includes:

[0043] First, set an initial sliding window length for each process parameter timing data. The sliding window length is predefined based on the physical change inertia of the parameter and the sampling frequency.

[0044] During data processing, the standard deviation of the data within the sliding window is calculated in real time to measure the degree of fluctuation of the data within the window;

[0045] The standard deviation within the current window is compared with the standard deviation sequence of historical windows, and the length of the next sliding window is dynamically adjusted. If the fluctuations intensify, the window length is appropriately reduced to improve the response speed to trend changes. If the fluctuations are stable, the window length is maintained or appropriately increased to enhance the smoothing effect.

[0046] Within a defined sliding window, calculate the arithmetic mean of all time-series data points covered by the window, and assign the arithmetic mean to the baseline component value at the center of the window.

[0047] The sliding window is moved point by point along the time axis, and the mean value within the window is repeatedly calculated and assigned to the center point. Finally, a smooth trend line that runs through the entire production roll time axis and filters out high-frequency fluctuations is obtained. The smooth trend line is the baseline component that characterizes long-term process drift.

[0048] As a further aspect of the present invention, the step of inputting all parameter state snapshots into an unsupervised clustering model, and the unsupervised clustering model assigning the snapshots to different clusters based on their proximity in the multidimensional feature space, specifically includes:

[0049] Construct a multi-dimensional feature space whose dimensions consist of all key feature indicators of the raw material layer quality parameters, interface bonding layer quality parameters, and surface layer quality parameters contained in the hierarchical quality parameter set;

[0050] For all parameter state snapshots extracted from the historical defect sample set, their corresponding feature index values ​​are standardized to eliminate the influence of different parameter units and numerical ranges, and mapped to a coordinate point in the multidimensional feature space.

[0051] Unsupervised clustering models employ density-based clustering algorithms to scan the entire feature space and automatically find regions where data points are densely distributed.

[0052] The clustering algorithm sets a neighborhood radius and a minimum point threshold, and groups the core points in the feature space that contain no less than the minimum point threshold within the neighborhood radius and the points whose density is reachable into the same cluster;

[0053] Through this process, all input parameter state snapshots are automatically divided into several different clusters based on their spatial clustering in the multidimensional feature space, with each cluster representing a type of process state that is similar in quality parameter characteristics.

[0054] As a further aspect of the present invention, the present invention also includes a quality traceability system for EVA-coated nonwoven fabric products. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the above-described quality traceability method for EVA-coated nonwoven fabric products.

[0055] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0056] Multi-scale decomposition was performed on the original parameter flow that was uniquely bound to the independent production roll of EVA laminated nonwoven fabric. The baseline component representing long-term process drift, the rhythm component representing periodic fluctuations, and the disturbance component representing instantaneous anomalies were separated. The three types of components were mapped to the raw material layer, the interface bonding layer, and the surface layer according to the physical structure of the laminated nonwoven fabric, respectively, and a layered quality parameter set was generated. Different changes in process parameters can directly correspond to the quality influence dimensions of different structural layers of the product. The objects of long-term drift, periodic fluctuations, and instantaneous anomalies can be clearly distinguished. The relationship between process parameters and product structural quality presents a concrete layered correspondence. The structural level of defect association can be directly and clearly defined.

[0057] A defect pattern sample library is built based on the hierarchical quality parameter set of historical production volumes and the final product inspection records. Unsupervised clustering method is used to identify common potential defect evolution paths from the sample library. The dynamic matching degree between the hierarchical quality parameter set of the current production volume and the early stage of each potential defect evolution path is calculated in real time. When the matching degree exceeds the preset sensitivity threshold, the parameter backtracking and root cause location of the corresponding process links involved in the evolution path are triggered. The common evolution law of defect formation can be extracted autonomously. The matching relationship between the current production status and the early characteristics of defects can be quantified in real time. The traceability process can be triggered in the early stage of defect formation. The traceability scope is accurately locked to the process links associated with the corresponding evolution path. The location of the defect root cause can be directly focused on the corresponding production link. Attached Figure Description

[0058] Figure 1 This is a flowchart of a product quality traceability method based on EVA-coated nonwoven fabric as described in this invention;

[0059] Figure 2 A flowchart for generating a hierarchical quality parameter set for mapping;

[0060] Figure 3 Box-line comparison chart of comprehensive quality parameter scores for three-layer EVA laminated nonwoven fabric;

[0061] Figure 4 This is the real-time defect matching degree monitoring curve for the current production roll;

[0062] Figure 5 Heat map for locating abnormalities in the process. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0064] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0065] See Figure 1 A quality traceability method for EVA-coated nonwoven fabric products, the overall implementation scheme of which includes the following steps:

[0066] By deploying a sensor network on the production site, the time-series process parameters of the raw material feeding, melt blending, casting and laminating, and winding stages of the EVA laminated nonwoven fabric production process chain are collected synchronously, generating a raw parameter stream uniquely bound to each independent production roll. The raw parameter stream is decomposed at multiple scales to separate the baseline component representing long-term process drift, the rhythm component representing periodic fluctuations, and the disturbance component representing instantaneous anomalies. The baseline, rhythm, and disturbance components are mapped according to the physical structure hierarchy of the laminated nonwoven fabric, respectively associating them with the raw material layer, interface bonding layer, and surface layer, generating a layered quality parameter set. Based on the layered quality parameter set of historical production rolls and the final product inspection records, a defect pattern sample library is established. Unsupervised clustering methods are used to identify common potential defect evolution paths from the defect pattern sample library. For the current production roll, the dynamic matching degree between its hierarchical quality parameter set and the early stages of each potential defect evolution path is calculated in real time. When the dynamic matching degree exceeds the preset sensitivity threshold, the quality traceability process for the current production roll is triggered. The quality traceability process includes parameter backtracking and root cause location of the process links involved in the matched potential defect evolution path.

[0067] In one embodiment of the present invention, during the raw material feeding stage, batch codes, feeding times, and initial temperature parameters of different batches of EVA particles and nonwoven fabric substrates are collected. During the melt blending stage, continuous variation curves of temperature, melt pressure, and melt viscosity in different zones of the blending equipment screw are collected. During the casting and laminating stage, die temperature distribution, cooling roller temperature, traction speed, and online thickness spectrum data are collected. During the winding stage, winding tension, roll diameter changes, and visual inspection results of surface defects are collected. A globally unique identifier is assigned to each independent production roll, and the time-aligned process parameters of all stages are spliced ​​and packaged using timestamps as indexes to form the original parameter stream for that production roll.

[0068] In practical implementation, sensors in the raw material feeding stage collect batch codes, feeding times, and initial temperature parameters of different batches of ethylene-vinyl acetate copolymer (EVA) particles, as well as batch codes, unwinding times, and surface temperature parameters of the nonwoven fabric substrate. Sensors in the melt blending stage collect temperature parameters of each section of the screw in the blending equipment, melt pressure parameters at the melt pump outlet, and melt viscosity parameters measured by an online melt viscometer; these parameters are recorded as continuously varying curves. Sensors in the casting and laminating stage collect temperature parameters distributed laterally on the die head, surface temperature parameters of the cooling roller, speed parameters of the traction motor, and thickness spectral data measured by an online infrared thickness gauge. Sensors in the winding stage collect winding tension parameters fed back by the tension roller, roll diameter variation parameters measured by the encoder, and surface defect visual inspection results images and coordinate data output by the industrial camera system.

[0069] In some embodiments, a globally unique identifier is assigned to each independent production roll. The generation rules for the globally unique identifier follow the GS1 standard, combining the enterprise manufacturer identification code, production date, production line number, and daily serial number. In specific implementation, all time-series process parameters collected from all production stages are indexed with timestamps with nanosecond precision. The batch input events in the raw material feeding stage, the melt output events in the melt blending stage, the film forming events in the casting and laminating stage, and the winding start events in the winding stage are time-aligned, with the die extrusion signal in the casting and laminating stage serving as the reference zero point for global time synchronization.

[0070] Optionally, the time-aligned timing process parameters are concatenated and encapsulated to form the raw parameter stream of the production volume. The encapsulation format uses a self-describing data structure, including a header and a data body. The header of the data structure stores a globally unique identifier, a production volume start timestamp, a data version number, and a parameter index table. The parameter index table records the starting byte offset, data length, sampling frequency, and unit of measurement for each type of process parameter within the data body. The data body stores the aligned sequence of raw process parameter values ​​from different sensors in chronological order. The formula for generating the globally unique identifier is expressed as:

[0071]

[0072] Where: characters This represents a globally unique identifier generated for each production volume; the characters are... Represents a fixed enterprise / manufacturer identification code, characters Represents the production date, in the format of year-month-day, characters Represents the production line number, character This indicates that the production line was on the date Production volume serial number, symbol This represents a string concatenation operation. The globally unique identifier, timestamp index, and encapsulated raw parameter stream are persistently stored in a time-series database for use in subsequent processing.

[0073] In one embodiment of the present invention, an adaptive sliding window mean filtering method is used to process the raw parameter stream, filtering out high-frequency noise, and the resulting smooth trend line serves as the baseline component characterizing long-term process drift. An initial sliding window length is set for each process parameter time series data. The sliding window length is predefined based on the physical change inertia of the parameter and the sampling frequency. During data processing, the standard deviation of the data within the sliding window is calculated in real time. The standard deviation within the current window is compared with the standard deviation sequence of historical windows, and the length of the next sliding window is dynamically adjusted. If fluctuations intensify, the window length is appropriately reduced to improve the response speed to trend changes; if fluctuations are stable, the window length is maintained or appropriately increased to enhance the smoothing effect. Within the determined sliding window, the arithmetic mean of all time series data points covered by the window is calculated, and the arithmetic mean is assigned as the baseline component value at the center of the window. The sliding window moves point by point along the time axis, and the mean within the window is repeatedly calculated and assigned to the center point. Finally, a smooth trend line that runs through the entire production roll time axis and filters out high-frequency fluctuations is obtained. This smooth trend line is the baseline component characterizing long-term process drift. The baseline component is subtracted from the original parameter stream to obtain the residual sequence. Spectral analysis is performed on the residual sequence to identify frequency components related to the equipment operation cycle and feeding cycle. Periodic components are extracted and reconstructed as rhythmic components characterizing periodic fluctuations. The rhythmic components are further removed from the residual sequence, and the remaining signal portion represents the disturbance component characterizing instantaneous anomalies. A dynamic threshold is set for the disturbance component to mark abnormal pulses.

[0074] In practice, an adaptive sliding window mean filtering method is used to process the raw parameter stream. The raw parameter stream contains melt pressure time-series data from the melt blending process. The initial sliding window length for the melt pressure time-series data is predefined as 100 data points based on the physical inertia of melt pressure parameter changes and a sampling frequency of 10 times per second. During data processing, the standard deviation of the 100 melt pressure data points within the current sliding window is calculated in real time. The standard deviation within the current sliding window is compared with the standard deviation sequence of the previous nine historical sliding windows. If the standard deviation of the current sliding window is greater than 1.5 times the average of the historical standard deviation sequences, the length of the next sliding window is dynamically adjusted to 50. To improve the response speed to trend changes, if the standard deviation of the current sliding window is less than 0.8 times the average of the historical standard deviation series, the length of the next sliding window is dynamically adjusted to 150 data points to enhance the smoothing effect. Within the determined sliding window, the arithmetic mean of all time series data points covered by the window is calculated, and the arithmetic mean is assigned as the baseline component value at the center of the window. The sliding window moves point by point along the time axis, and the mean within the window is repeatedly calculated and assigned to the center point. Finally, a melt pressure smoothing trend line that runs through the entire production roll time axis and filters out high-frequency fluctuations is obtained. This melt pressure smoothing trend line is the baseline component that characterizes long-term process drift.

[0075] In some embodiments, the melt pressure baseline component is subtracted from the original parameter stream to obtain the melt pressure residual sequence. A fast Fourier transform spectral analysis is performed on the melt pressure residual sequence to identify frequency components corresponding to the extruder screw rotation cycle and the frequency components corresponding to the periodic addition of EVA masterbatch. The amplitude and phase information of these specific frequency components are extracted, and the periodic components are reconstructed using sine wave superposition as melt pressure rhythm components characterizing periodic fluctuations. The melt pressure rhythm components are further removed from the melt pressure residual sequence, and the remaining signal portion represents the melt pressure disturbance components characterizing instantaneous anomalies. Dynamic thresholds based on the moving average and moving standard deviation are set for the melt pressure disturbance components to mark abnormal pulses. The calculation window length for the moving average and moving standard deviation is 500 data points.

[0076] Optionally, the adjustment logic for the adaptive sliding window length is quantified by a formula:

[0077]

[0078] Where: characters Represents the calculated length of the next sliding window, character Represents the length of the current sliding window, characters Represents the standard deviation of the data within the current sliding window, character The average of the standard deviations of a preset number of historical sliding windows, character Represents the window reduction factor, character The character represents the judgment coefficient indicating increased volatility. Represents the window enlargement factor, character The coefficient representing the judgment of stable fluctuations, with the sign... Indicates rounding down, sign This indicates rounding up. It's understandable that the traction speed parameter in the casting and laminating process has relatively small physical inertia; therefore, the initial sliding window length is predefined as 50 data points, and its fluctuation amplification judgment coefficient... Set to 1.3, window reduction factor Setting it to 0.7 gives it the coefficient for judging stable fluctuations. Set to 0.7, window enlargement factor Set it to 1.5.

[0079] In practical implementation, after multi-scale decomposition of the raw traction speed parameter stream, the separated traction speed baseline component represents the long-term slow drift of tension during winding, the traction speed rhythm component characterizes the speed fluctuation caused by the periodic eccentricity of the guide roller, and the traction speed disturbance component may correspond to instantaneous slippage or sudden load changes. The visual inspection results of surface defects in the winding stage are binarized to form an event sequence. This event sequence, as a special type of raw parameter stream, has a baseline component reflecting the background noise level, a rhythm component that may be related to the flicker of the lighting source, and a disturbance component that directly corresponds to sudden defect events. For these parameters, the dynamic threshold is set according to the Poisson distribution model.

[0080] In one embodiment of the present invention, see [reference] Figure 2 The baseline components of the melt blending process, particularly the long-term trend of melt viscosity, are mapped to raw material layer quality parameters affecting overall uniformity. The rhythmic components of the casting and laminating process, especially the periodic fluctuations in temperature and speed, are mapped to interfacial bonding layer quality parameters affecting the bonding strength between the EVA film and the nonwoven substrate. The disturbance components of each process, especially abnormal pulses exceeding dynamic thresholds, are mapped to surface layer quality parameters leading to surface defects or localized weaknesses. These mapped raw material layer quality parameters, interfacial bonding layer quality parameters, and surface layer quality parameters are organized using the same unique production roll identifier to form a structured, layered quality parameter set.

[0081] In practice, the baseline component of the melt blending process is mapped to the raw material layer quality parameter. The baseline component of the melt blending process is derived from the smooth trend line obtained by multi-scale decomposition of the melt viscosity time series data in the original parameter stream. The long-term trend of melt viscosity, such as the monotonically increasing baseline that slowly rises from 850 Pa·s to 920 Pa·s in eight hours of continuous production, is extracted and recorded as the core raw material layer quality parameter affecting the overall uniformity of EVA laminated nonwoven fabric. This parameter is named "raw material layer-melt viscosity trend coefficient". In practical implementation, the rhythmic component of the casting and laminating process is mapped to the interface bonding layer quality parameter. The rhythmic component of the casting and laminating process is derived from the extraction of the periodic fluctuation components of the die temperature and traction speed parameters. The sinusoidal fluctuation of the die temperature in degrees Celsius with a period of 45 seconds and an amplitude of ±1.5 degrees Celsius, and the cosine fluctuation of the traction speed in meters per minute with a period of 45 seconds and an amplitude of ±0.2 meters per minute, are identified as the same equipment cycle and reconstructed as the rhythmic component. This component is mapped to the key interface bonding layer quality parameter that affects the bonding strength between the ethylene-vinyl acetate copolymer (EVA) film and the nonwoven substrate. This parameter is named the "interface layer-temperature-speed synergistic fluctuation index".

[0082] In some embodiments, the disturbance components of each stage are mapped to surface quality parameters. In the tension sensor signal disturbance component of the winding stage, abnormal pulses exceeding a dynamic threshold set based on the moving average are marked. A tension spike with an amplitude of 150% of the standard tension value, appearing at the timestamp "2026-02-17 14:23:11.345", is identified as an abnormal event. This event is mapped to a surface quality parameter that may cause stretch marks or localized thinning on the surface of the laminated nonwoven fabric, and is recorded as a "surface-tension impact event". In the casting and laminating stage, a transient thickness trough is identified in the disturbance component of the online thickness spectral data after decomposition. This trough is mapped to a potential surface weak point parameter. A specific mapping relationship between multi-scale components and layer quality parameters is shown in Table 1.

[0083] Table 1: Example of mapping relationship between process parameter components and layered quality parameters

[0084] Source process Parameter component type Mapped target quality parameter layer Specific quality parameter names and exemplary content melt blending Baseline components raw material layer Raw material layer-melt viscosity trend coefficient: describes the slope of viscosity change over several hours. Casting and lamination Rhythmic components Interface Integration Layer Interface layer-temperature-velocity co-fluctuation index: describes the phase difference and amplitude ratio of periodic fluctuations in temperature and velocity. Casting and lamination Disturbance components surface layer Surface-thickness perturbation spectrum intensity: describes the frequency domain energy of thickness anomaly events. Collect Disturbance components surface layer Surface-Tension Impact Events: Records the timestamps and amplitudes of abnormal tension pulses.

[0085] Optionally, the calculation of the interface bonding layer quality parameters may involve integrating multiple rhythmic components, such as combining the die head temperature rhythmic component and the traction speed rhythmic component, and generating a composite index through a weighted formula. This composite index can be understood as follows:

[0086]

[0087] Where: characters This represents the calculated overall quality index of the interface layer, in characters. Represents the number of major periodic harmonics being analyzed, character and Representing the first Weighting coefficients for temperature and velocity components under each harmonic, character and Representing the first The amplitude of temperature and velocity fluctuations under each harmonic, character and Representing the first Phase angle of temperature and velocity fluctuations under each harmonic. Weighting coefficient. The influence of this frequency fluctuation on the peel strength was determined through regression analysis of historical data.

[0088] In some embodiments, the mapped raw material layer quality parameters, interface bonding layer quality parameters, and surface layer quality parameters are organized according to the same globally unique identifier for each production roll, forming a structured hierarchical quality parameter set. This hierarchical quality parameter set exists in the database as a data table, with each row corresponding to a unique globally unique identifier for each production roll. The data columns are divided into raw material layer parameter groups, interface bonding layer parameter groups, and surface layer parameter groups. Each parameter group contains several specific quality parameter fields and their values, calculated through mapping.

[0089] See Figure 3 The box plot visually presents the quality score distribution characteristics of the raw material layer, interface bonding layer, and surface layer. From the overall distribution, the raw material layer shows the best quality index performance, with its box concentrated in the 84–91 score range, a median of approximately 85 points, and upper and lower limits extending to 74 and 100 points respectively. This indicates high quality stability and excellent overall quality in this layer, with a small number of high-scoring extreme samples. The interface bonding layer's quality index is second best, with its box distributed in the 65–75 score range, a median of approximately 71 points, and upper and lower limits ranging from 51 to 82 points. This reflects moderate quality fluctuation in this layer, with its core quality level in the upper-middle range. The surface layer's quality index is relatively the weakest, with its box distributed in the 55–71 score range, a median of approximately 63 points, and upper and lower limits ranging from 43 to 88 points. This indicates the greatest quality dispersion in this layer, with many low-scoring samples, making it a weak link in overall quality control. From the perspective of layered quality logic, the high stability and high score of the raw material layer correspond to the long-term uniformity of the baseline component in the melt blending process; the moderate score and fluctuation of the interface bonding layer reflect the periodic fluctuation of the rhythm component in the casting and coating process; while the low score and wide distribution of the surface layer quality are directly related to the inducing effect of abnormal pulses in the disturbance components of each process on surface defects, providing an intuitive basis for the layered quality differences for subsequent defect evolution path identification and quality traceability.

[0090] In one embodiment of the present invention, production rolls with final product inspection records indicating non-conforming products are selected from historical production records. The stratified quality parameter sets corresponding to these production rolls are extracted to form an initial defect sample set. For each sample in the initial defect sample set, the stratified quality parameters are sliced ​​at equal time intervals from the start of production to the time the defect is detected, forming a series of parameter state snapshots arranged in chronological order. A multidimensional feature space is constructed, whose dimensions consist of all key feature indicators of the raw material layer quality parameters, interface bonding layer quality parameters, and surface layer quality parameters contained in the stratified quality parameter set. The feature indicator values ​​corresponding to all parameter state snapshots extracted from the historical defect sample set are standardized to eliminate the influence of different parameter dimensions and numerical ranges, and mapped to a coordinate point in this multidimensional feature space. All parameter state snapshots are input into an unsupervised clustering model. This model employs a density-based clustering algorithm, scanning the entire feature space and automatically identifying densely distributed data points. The algorithm sets a neighborhood radius and a minimum point count threshold, grouping core points within the neighborhood radius that contain at least the minimum point count threshold, along with their density-reachable points, into the same cluster. Through this process, all input parameter state snapshots are automatically divided into several different clusters based on their spatial clustering in the multidimensional feature space. Each cluster represents a type of process state similar in quality parameter characteristics. The transition sequences of snapshots belonging to the same historical defect sample across different clusters are analyzed. Frequently occurring transition sequences are extracted and defined as a potential defect evolution path from an early process state to the final defect state, and a path feature template is established for each path. For the current production volume, the hierarchical quality parameter set from the start to the current moment is acquired in real time and segmented chronologically into parameter state snapshot sequences with the same intervals as defined in the potential defect evolution path. The current parameter state snapshot sequence is compared stage by stage with the path feature template of each potential defect evolution path. The comparison process calculates the multidimensional feature distance between the current sequence and the corresponding path template at each stage cluster center, and then performs a weighted sum based on time weights to obtain the overall dynamic matching score between the current production state and the potential defect evolution path. A real-time dynamic matching score is maintained for each potential defect evolution path.

[0091] In practice, production rolls with final product inspection records indicating non-conforming products are selected from historical production records. For example, two production rolls with globally unique identifiers "P20260217S03N0125" and "P20260218S01N0047" are selected. The finished product inspection records of these two production rolls are both marked as "insufficient peel strength". The layered quality parameter set corresponding to production rolls "P20260217S03N0125" and "P20260218S01N0047" is extracted to form the initial defect sample set. For each sample in the initial defect sample set, the layered quality parameters are sliced ​​at equal time intervals from the start of production to the time when the defect is detected. For example, for the production process of production roll "P20260217S03N0125" lasting 120 minutes, slices are made at one minute intervals to generate 120 parameter status snapshots arranged in chronological order. Each parameter status snapshot contains the instantaneous or statistical values ​​of the raw material layer quality parameters, interface bonding layer quality parameters, and surface layer quality parameters within that minute.

[0092] In some embodiments, a multidimensional feature space is constructed. The dimensions of the multidimensional feature space consist of all key feature indicators of the raw material layer quality parameters, interface bonding layer quality parameters, and surface layer quality parameters contained in the layered quality parameter set. The key feature indicators include the raw material layer-melt viscosity trend coefficient, the interface layer-temperature-rate co-fluctuation index, the surface layer-tension impact event count, and the surface layer-thickness perturbation spectrum intensity. For all parameter state snapshots extracted from the historical defect sample set, the corresponding key feature indicator values ​​are standardized using the Z-score method to eliminate the influence of different parameter dimensions and numerical ranges, and mapped to a coordinate point in the multidimensional feature space. All parameter state snapshots are input into an unsupervised clustering model. The unsupervised clustering model uses a density-based clustering algorithm to scan the entire feature space and automatically find densely distributed regions of data points. The density-based clustering algorithm sets the neighborhood radius ε to 0.5 and the minimum number of points MinPts to 5. It groups the core points in the feature space that contain no less than MinPts points within the neighborhood radius ε and the points that are density-reachable into the same cluster. Through this process, all input parameter state snapshots are automatically divided into several different clusters based on their spatial clustering in the multidimensional feature space. Each cluster represents a type of process state that is similar in quality parameter characteristics.

[0093] In practical implementation, the transition sequence of a series of snapshots belonging to the same historical defect sample between different clusters is analyzed. For example, the 120 snapshots of the historical defect sample "P20260217S03N0125" have the transition sequence "cluster A->cluster A->cluster B->cluster B->cluster C->cluster C". Similarly, the 90 snapshots of the historical defect sample "P20260218S01N0047" have the same transition sequence "cluster A->cluster A->cluster B->cluster B->cluster C->cluster C". This frequently occurring transition sequence "cluster A->cluster A->cluster B->cluster B->cluster C->cluster C" is extracted and defined as a potential defect evolution path from the early process state to the final defect state. A path feature template is established for this path. The path feature template records the cluster center feature vector corresponding to each stage in the path sequence (e.g., stage 1-2 corresponds to cluster A, stage 3-4 corresponds to cluster B, and stage 5-6 corresponds to cluster C). See Table 2 for an example of a potential defect evolution path identified from historical data.

[0094] Table 2: Feature Template Table of Potential Defect Evolution Path "Path-01"

[0095] Path phase Corresponding cluster identifier Phase duration (example) Example summary of cluster center feature vectors at this stage. 1-2 Cluster A Approximately 20-40 minutes The viscosity of the raw material layer is stable, and the fluctuation index of the interface layer is low. 3-4 Cluster B Approximately 20-40 minutes The viscosity of the raw material layer begins to rise slowly, and the fluctuation index of the interface layer increases. 5-6 Cluster C Approximately 20-40 minutes The viscosity of the raw material layer remains high, the interfacial layer fluctuates dramatically, and surface events increase.

[0096] Optionally, for the current production volume, the hierarchical quality parameter set from the start to the current moment is acquired in real time. For example, if the current production volume "P20260220S02N0081" has been produced for 50 minutes, its hierarchical quality parameter set is divided into 50 parameter status snapshot sequences at minute intervals. The current parameter status snapshot sequence is compared stage by stage with the path feature template of each potential defect evolution path. During the comparison process, the multidimensional feature distance between the current sequence and the corresponding path template at each stage is calculated at the center of the stage cluster. The multidimensional feature distance is calculated using Euclidean distance and weighted according to time weights to obtain the overall dynamic matching degree score between the current production state and the potential defect evolution path. The formula for calculating the dynamic matching degree score can be understood as:

[0097]

[0098] Where: characters Represents the calculated dynamic matching score, characters Represents the number of stages currently produced, character Representing the Time weight of each stage, character Represents the number of dimensions of the feature vector, characters This represents the current production roll number in the [number]. Phase 1 Values ​​on dimensional features, characters The template representing the potential defect evolution path is in the first Phase 1 Values ​​on the dimensional features (i.e., the centers of the clusters corresponding to this stage). Time weights. along with The weighting increases incrementally, assigning higher weights to recent stages. A real-time dynamic matching score is maintained for each potential defect evolution path.

[0099] See Figure 4 In the current real-time defect matching monitoring of production rolls, the linkage between the dynamic matching score and the early warning threshold enables early warning of potential defect evolution paths. Specifically, the real-time matching curve is obtained by comparing the hierarchical quality parameter set of the current production roll with the potential defect evolution path template stage by stage. The similarity between the current process state and the early stage of the defect is quantified through weighted multidimensional feature distance. The early warning threshold is set at 0.6, representing the sensitive critical value for triggering the quality traceability process: when the real-time matching score first exceeds this threshold within the 35-40 minute interval, the system will automatically lock the corresponding potential defect evolution path, trace back to the abnormal parameters in the raw material layer, interface bonding layer, or surface layer, and ultimately locate the specific production link and equipment. As can be seen from the curve trend, the dynamic matching score of the current production roll continuously increases with production time, reaching 0.78 at 50 minutes, significantly higher than the early warning threshold. This indicates that its process state is accelerating along a certain defect evolution path, requiring immediate initiation of the quality traceability process to locate the root cause.

[0100] In one embodiment of the present invention, preset sensitivity thresholds are established for the evolution paths of potential defects of different severity levels. The dynamic matching scores of all potential defect evolution paths are monitored in real time. When any score exceeds its corresponding sensitivity threshold, the system automatically triggers a quality traceability process. The traceability process first identifies the potential defect evolution path with the exceeding matching score, analyzes its path feature template, and determines the defect type and key impact stage it represents. Based on the key impact stage, it traces back to the corresponding raw material layer, interface bonding layer, or surface layer parameters in the hierarchical quality parameter set. Further, it correlates back to the original parameter stream, locating the specific production link, equipment, or even sensor that generated the abnormal parameter data, and generates a preliminary root cause analysis report including the time range, location, abnormal parameter values, and associated defect types.

[0101] In practical implementation, a sensitivity threshold is preset for the evolution path of potential defects of different severity levels. The system maintains a lookup table containing path identifiers, defect severity levels and corresponding sensitivity thresholds. For example, the potential defect evolution path "Path-01" is defined as an "interlayer peeling" defect with a severity level of "high" and its preset sensitivity threshold is 0.85. The potential defect evolution path "Path-02" is defined as an "uneven thickness" defect with a severity level of "medium" and its preset sensitivity threshold is 0.90. The potential defect evolution path "Path-03" is defined as a "surface scratch" defect with a severity level of "low" and its preset sensitivity threshold is 0.95. The system monitors the dynamic matching score of all potential defect evolution paths in real time. The dynamic matching score is continuously updated by an independent computing service and written to the monitoring queue. A monitoring process polls and checks the dynamic matching score of all paths in the queue once per second. When any dynamic matching score exceeds the sensitivity threshold corresponding to its path identifier in the lookup table, for example, the dynamic matching score of the current production volume "P20260220S02N0081" and the potential defect evolution path "Path-01" reaches 0.88, which exceeds the sensitivity threshold of 0.85 for "Path-01", the system automatically triggers the quality traceability process for production volume "P20260220S02N0081".

[0102] In some embodiments, the tracing process first identifies the potential defect evolution path with an excessive matching degree. The matching event log records "Time: 2026-02-20 10:15:30, Volume: P20260220S02N0081, Exceeding Path: Path-01, Matching Degree: 0.88". The system retrieves and loads the path feature template based on the path identifier "Path-01". The path feature template of the potential defect evolution path "Path-01" is parsed. The metadata segment of the path feature template records that the defect type it represents is "interlayer peeling". The stage definition segment of the path feature template indicates that its key impact stages are the third and fourth stages. The cluster center feature vectors corresponding to these two stages have significantly higher values ​​than the normal range in the dimension of "interface layer-temperature-rate co-fluctuation index". Based on the key impact stage, the system traces back to the corresponding parameters in the stratified quality parameter set. The system queries the stratified quality parameter set of production volume "P20260220S02N0081" and extracts snapshots of all parameter statuses in the third and fourth stages of the production timeline (e.g., from the 21st minute to the 40th minute). The analysis shows that in these snapshots, the value of "interface layer - temperature-rate synergistic fluctuation index" continuously exceeds the normal threshold upper limit of 2.5, while the parameters of "raw material layer - melt viscosity trend coefficient" and "surface layer - tension impact event" are within the normal range.

[0103] Optionally, the system further correlates back to the original parameter stream to pinpoint the specific production process, equipment, and even sensors that generated the abnormal parameter data. Based on the time period of the layered quality parameter anomalies, "2026-02-2009:56:00 to 2026-02-2010:15:00," the system retrieves the original parameter stream for production roll "P20260220S02N0081" within this time period. The index of the original parameter stream points to the die head temperature sensor T_zone5 and the traction speed sensor V_main in the casting and laminating process. Querying the time-series data of these two sensors reveals that, within the corresponding time period, the measured value of the die head temperature sensor T_zone5 exhibits fluctuations with a period of 45 seconds and an amplitude of ±3 degrees Celsius, while the measured value of the traction speed sensor V_main exhibits synchronous fluctuations with a period of 45 seconds and an amplitude of ±0.3 meters per minute. It can be understood that the preset logic of the sensitivity threshold can be adjusted based on the severity level of the path and the historical false alarm rate. A formula for dynamically adjusting the sensitivity threshold is expressed as:

[0104]

[0105] Where: characters This represents the newly calculated sensitivity threshold, for the character. Represents the base sensitivity threshold set based on the severity level of the defect, character Represents the adjustment rate coefficient, character Represents the target false alarm rate set by the system, characters This represents the false alarm rate actually observed over a period of time for this potential defect evolution path. Adjustment rate coefficient. It is a constant between 0 and 1, used to control the magnitude of threshold adjustment.

[0106] In some embodiments, the system generates a preliminary root cause analysis report containing time range, location, abnormal parameter values, and associated defect types. The report is output in a structured text format and includes the following: the globally unique identifier of the traceability volume "P20260220S02N0081", the trigger path "Path-01 (interlayer peeling)", the abnormal time period "2026-02-2009:56:00 to 2026-02-2010:15:00", the main abnormal link "casting and laminating process", the specific abnormal equipment "die head zone 5 temperature control unit, main traction motor", the key abnormal parameters "die head zone 5 temperature fluctuation exceeds limit, traction speed fluctuation exceeds limit", and the inferred root cause "periodic oscillation of the temperature control system leads to a decrease in interface bonding strength". In practice, when the system simultaneously monitors that the dynamic matching scores of multiple potential defect evolution paths exceed their respective sensitivity thresholds, the tracing process will start an independent tracing thread in parallel for each out-of-limit path. Each tracing thread independently executes the logic of locking the path, parsing the template, tracing parameters, locating the root cause, and generating a report. The final report will then present a combined statement, pointing out possible composite defect patterns.

[0107] See Figure 5 The heatmap visually presents the distribution of anomaly indices under different production stages and process parameter dimensions, providing a quantitative basis for tracing the root causes of defects. Specifically, the heatmap uses production stages (raw material feeding, melt blending, casting and laminating, and winding and forming) as the horizontal axis and process parameter dimensions (temperature, pressure, speed, and tension) as the vertical axis. Color gradients and numerical labels visually reflect the degree of anomalies under each dimension-stage combination: the numerical range covers 1.0~3.6, and the color from light to dark corresponds to anomaly indices from low to high. The color scale on the right clearly indicates the mapping relationship between values ​​and colors. Under the temperature, pressure, and tension dimensions, the anomaly indices of each production stage are all in the low range of 1.0~1.4, with stable fluctuations and no significant contribution to anomalies; speed... The temperature dimension is the core source of anomalies. The anomaly index in the casting and laminating stage reaches a peak of 3.6, corresponding to the darkest area, which is the key anomaly point for matching the "interlayer peeling" defect path in this quality traceability. The speed anomaly indices in the melt blending and winding stages are 2.9 and 2.8 respectively, also significantly higher than other parameter dimensions, indicating that speed fluctuations are the core cause of decreased interfacial bonding strength. The speed anomaly index in the raw material feeding stage is 1.8, higher than other parameter dimensions, but still far lower than the peak value in the casting and laminating stage, indicating that the anomalies are mainly concentrated in the speed control module of the casting and laminating stage. From the perspective of quality traceability logic, this heat map directly verifies the conclusion of tracing the source of the layered quality parameter set: the periodic fluctuation of the speed parameter in the casting and laminating stage corresponding to the interfacial bonding layer is the core process root cause of the "interlayer peeling" defect, providing a precise direction for subsequent equipment parameter optimization and process stability control.

[0108] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for quality traceability of EVA-coated nonwoven fabric products, characterized in that, The method includes: By deploying a sensor network on the production site, the timing process parameters of the raw material feeding, melt blending, casting and laminating and winding stages that constitute the EVA laminated nonwoven fabric production process chain are collected synchronously, and a unique original parameter stream is generated for each independent production roll. The original parameter flow is decomposed into a multi-scale component to separate the baseline component representing long-term process drift, the rhythm component representing periodic fluctuations, and the disturbance component representing instantaneous anomalies. The baseline component, rhythm component, and disturbance component are mapped according to the physical structure hierarchy of the laminated nonwoven fabric, and respectively associated with the raw material layer, interface bonding layer, and surface layer, generating a layered quality parameter set, specifically including: The baseline component of the melt blending process is mapped to the raw material layer quality parameters that affect the overall uniformity. The rhythmic components of the casting and lamination process are mapped to interfacial bonding layer quality parameters that affect the bonding strength between the EVA film and the nonwoven substrate. The disturbance components of each stage are mapped to surface quality parameters that lead to surface defects or local weak points in strength. The mapped raw material layer quality parameters, interface bonding layer quality parameters, and surface layer quality parameters are organized according to the same production volume unique identifier code to form a structured layered quality parameter set. Based on the hierarchical quality parameter set of historical production volumes and final product inspection records, a defect pattern sample library is established. Unsupervised clustering methods are used to identify common potential defect evolution paths from this sample library, specifically including: From historical production records, select production rolls with final finished product inspection records showing non-conforming products, extract the stratified quality parameter set corresponding to the production rolls, and form an initial defect sample set; For each sample in the initial defect sample set, the hierarchical quality parameters are sliced ​​at equal time intervals from the start of production to the time when the defect is detected, forming a series of parameter state snapshots arranged in chronological order. All parameter state snapshots are input into an unsupervised clustering model, which assigns the snapshots to different clusters based on their proximity in the multidimensional feature space. The transition sequences of a series of snapshots belonging to the same historical defect sample are analyzed between different clusters. Frequently occurring transition sequences are extracted and defined as a potential defect evolution path from the early process state to the final defect state. A path feature template is established for each path. For the current production roll, the dynamic matching degree between its hierarchical quality parameter set and the early stages of each potential defect evolution path is calculated in real time. When the dynamic matching degree exceeds a preset sensitivity threshold, a quality traceability process for the current production roll is triggered. The quality traceability process includes parameter backtracking and root cause localization of the process links involved in the matched potential defect evolution path.

2. The method for quality traceability of EVA-coated nonwoven fabric products according to claim 1, characterized in that, Through a sensor network deployed on the production site, the timing process parameters of the raw material feeding, melt blending, casting and laminating, and winding stages of the EVA laminated nonwoven fabric production process chain are collected synchronously, generating a raw parameter stream uniquely bound to each independent production roll, specifically including: During the raw material feeding stage, batch codes, feeding times, and initial temperature parameters of different batches of EVA particles and non-woven fabric substrates were collected. In the melt blending process, the continuous variation curves of temperature, melt pressure and melt viscosity in each zone of the screw of the blending equipment are collected; In the casting and lamination process, data on die temperature distribution, cooling roller temperature, traction speed, and online thickness spectrum are collected. During the winding process, the results of visual inspection of winding tension, roll diameter changes, and surface defects are collected. Each independent production roll is assigned a globally unique identifier, and the time-series process parameters of all stages are spliced ​​and packaged using timestamps as indexes to form the original parameter stream of the production roll.

3. The method for quality traceability of EVA-coated nonwoven fabric products according to claim 2, characterized in that, The original parameter flow is decomposed into multiple scales to separate the baseline component representing long-term process drift, the rhythm component representing periodic fluctuations, and the disturbance component representing instantaneous anomalies, specifically including: An adaptive sliding window mean filtering method is used to process the original parameter stream, filter out high-frequency noise, and the resulting smooth trend line is used as the baseline component characterizing long-term process drift. The baseline component is subtracted from the original parameter stream to obtain the residual sequence. The residual sequence is subjected to spectral analysis to identify the frequency components related to the equipment operation cycle and feeding cycle. The periodic components are extracted and reconstructed as rhythmic components characterizing periodic fluctuations. The rhythmic components are further removed from the residual sequence, and the remaining signal is the disturbance component that represents the instantaneous anomaly. A dynamic threshold is set for the disturbance component to mark the abnormal pulse.

4. The method for quality traceability of EVA-coated nonwoven fabric products according to claim 3, characterized in that, For the current production roll, the dynamic matching degree between its hierarchical quality parameter set and the early stages of each potential defect evolution path is calculated in real time, specifically including: The hierarchical quality parameter set of the current production volume from the start to the current moment is acquired in real time, and it is divided into a parameter state snapshot sequence with the same interval as the potential defect evolution path definition in time order; The current parameter status snapshot sequence is compared stage by stage with the path feature template of each potential defect evolution path; The comparison process calculates the multidimensional feature distance between the current sequence and the corresponding path template at the center of the stage cluster at each stage, and performs a weighted summation according to the time weight to obtain the overall dynamic matching score between the current production state and the potential defect evolution path. Maintain a real-time dynamic matching score for each potential defect evolution path.

5. The method for quality traceability of EVA-coated nonwoven fabric products according to claim 4, characterized in that, When the dynamic matching degree exceeds a preset sensitivity threshold, a quality traceability process is triggered for the current production roll. This quality traceability process includes parameter backtracking and root cause localization of the process steps involved in the potential defect evolution path of the matched roll, specifically including: Preset sensitive thresholds for the evolution paths of potential defects of different severity levels; The system monitors the dynamic matching score of all potential defect evolution paths in real time. When any score exceeds its corresponding sensitivity threshold, the system automatically triggers the quality traceability process. The tracing process first identifies the potential defect evolution path that exceeds the matching limit, analyzes the feature template of the potential defect evolution path, and determines the defect type and key impact stage it represents. Based on the key impact stage, trace back to the raw material layer, interface bonding layer, or surface layer parameters corresponding to the stratified quality parameter set. Further linking back to the original parameter stream, the system pinpoints the specific production process, equipment, or even sensor that generated the abnormal parameter data, and generates a preliminary root cause analysis report that includes the time range, location, abnormal parameter values, and associated defect types.

6. The method for quality traceability of EVA-coated nonwoven fabric products according to claim 5, characterized in that, The process employs an adaptive sliding window mean filtering method to process the original parameter stream, filtering out high-frequency noise. The resulting smooth trend line serves as the baseline component characterizing long-term process drift. Specifically, this includes: First, set an initial sliding window length for each process parameter timing data. The sliding window length is predefined based on the physical change inertia of the parameter and the sampling frequency. During data processing, the standard deviation of the data within the sliding window is calculated in real time to measure the degree of fluctuation of the data within the window; The standard deviation within the current window is compared with the standard deviation sequence of historical windows, and the length of the next sliding window is dynamically adjusted. If the fluctuations intensify, the window length is appropriately reduced to improve the response speed to trend changes. If the fluctuations are stable, the window length is maintained or appropriately increased to enhance the smoothing effect. Within a defined sliding window, calculate the arithmetic mean of all time-series data points covered by the window, and assign the arithmetic mean to the baseline component value at the center of the window. The sliding window is moved point by point along the time axis, and the mean value within the window is repeatedly calculated and assigned to the center point. Finally, a smooth trend line that runs through the entire production roll time axis and filters out high-frequency fluctuations is obtained. The smooth trend line is the baseline component that characterizes long-term process drift.

7. The method for quality traceability of EVA-coated nonwoven fabric products according to claim 6, characterized in that, The process involves inputting all parameter state snapshots into an unsupervised clustering model. The unsupervised clustering model assigns the snapshots to different clusters based on their proximity in the multidimensional feature space. Specifically, this includes: Construct a multi-dimensional feature space whose dimensions consist of all key feature indicators of the raw material layer quality parameters, interface bonding layer quality parameters, and surface layer quality parameters contained in the hierarchical quality parameter set; For all parameter state snapshots extracted from the historical defect sample set, their corresponding feature index values ​​are standardized to eliminate the influence of different parameter units and numerical ranges, and mapped to a coordinate point in the multidimensional feature space. Unsupervised clustering models employ density-based clustering algorithms to scan the entire feature space and automatically find regions where data points are densely distributed. The clustering algorithm sets a neighborhood radius and a minimum point threshold, and groups the core points in the feature space that contain no less than the minimum point threshold within the neighborhood radius and the points whose density is reachable into the same cluster; Through this process, all input parameter state snapshots are automatically divided into several different clusters based on their spatial clustering in the multidimensional feature space, with each cluster representing a type of process state that is similar in quality parameter characteristics.

8. A quality traceability system for EVA-coated nonwoven fabric products, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the EVA-coated nonwoven fabric product quality traceability method as described in any one of claims 1 to 7.

Images