A method and system for tracing surface particle defects of an automobile trim panel

Abstract

The application provides a method for tracing a surface particle defect of an automobile decorative plate, and has the characteristics that: dust defect images of the plate are continuously acquired and taken as defects to be diagnosed, spatial distribution features and time sequence distribution features thereof are extracted, and a space-time distribution data set is formed; physical and chemical attribute analysis is performed on the dust defects, the physical and chemical attributes are fused with the space-time distribution feature data set, and a multi-modal feature vector is formed; the multi-modal feature vector is matched with a preset pollution source database, a suspected pollution source is acquired, a simulation model is preset, the space-time distribution data set is input into the simulation model, dynamic simulation of dust generation and deposition of the suspected pollution source is performed, and virtual defect results are compared and verified with the dust defect images; if the verification results are consistent, the suspected pollution source is determined to be a real pollution source, and if the verification results are inconsistent, layered interference verification is performed on the defect part to determine a position of a dust layer.

Description

Technical Field

[0001] This invention relates to the field of sheet metal, and in particular to a method and system for tracing the source of surface particle defects in automotive decorative panels. Background Technology

[0002] As a key component of automotive interior and exterior trim, the surface coating quality of automotive trim panels directly affects the vehicle's appearance and market competitiveness. Surface dust particle defects are one of the main causes of product defects during the coating process. On continuous coating production lines for automotive trim panels, the sources of dust particle pollution are complex and dispersed, mainly including four categories: First, the paint itself; if the paint is not filtered or is not thoroughly filtered during mixing and conveying, particles of varying sizes will remain, directly forming surface defects after coating. Second, the substrate; if residual metal shavings, wood chips, or surface unevenness are not removed during the pretreatment stage, particle defects will appear after coating. Third, the coating equipment; wear and corrosion of rollers and conveyor chains, and dust accumulation on fan impellers, will all generate particles that adhere to the panel surface. Fourth, the production environment; dust suspended in the factory air settles onto the coating surface during the coating and curing stages, forming difficult-to-eliminate particle defects.

[0003] To address the aforementioned particulate defects, the industry's traditional quality control solution involves deploying high-precision sensors after each process in the coating production line for online detection, attempting to intercept defects at the source. However, this solution has significant technical drawbacks: First, it is costly, as deploying high-precision sensors throughout the entire process requires substantial investment in equipment purchase and deployment, significantly increasing the construction cost of the production line; second, installation is limited, as coating production lines operate under complex conditions such as high temperature, high humidity, confined spaces, and severe mechanical vibration, and some process areas cannot meet the sensor installation environment requirements, leading to blind spots in detection; third, maintenance is complex, as daily calibration and fault repair of numerous sensors consume significant manpower and time, and the sensors are susceptible to contamination from paint mist and dust in the coating environment, easily leading to a decrease in detection accuracy.

[0004] Due to the aforementioned drawbacks, most manufacturers can only conduct defect detection in the final quality inspection stage. After particle defects are discovered, the traditional method of tracing the source involves large-scale shutdowns and a comprehensive investigation and cleaning of all possible contamination points, including paint supply, sheet pretreatment, coating equipment, and the production environment. This method lacks specificity, resulting in extremely low efficiency, prolonged production line downtime, and significant losses. Furthermore, it fails to accurately pinpoint the true source of contamination, often leading to superficial solutions that fail to address the root cause, resulting in recurring particle defects and severely impacting product yield and production schedule. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for tracing the source of defective particles on the surface of automotive decorative panels, so as to overcome the above-mentioned defects in the existing technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for tracing the source of surface particle defects in automotive trim panels includes: The data acquisition steps involve continuously acquiring dust defect images of the board material, using these images as defects to be diagnosed, and extracting their spatial distribution features and time series distribution features to form a multimodal feature vector. The intelligent matching step matches the multimodal feature vector with a preset pollution source database to obtain suspected pollution sources. A simulation model is preset, and the spatiotemporal distribution dataset is input into the simulation model to perform dynamic simulation of dust generation and deposition of the suspected pollution sources. The virtual defect results are then compared and verified with the dust defect images. The precise location step involves verifying whether the suspected pollution source is the actual pollution source. If the verification results match, the suspected pollution source is determined to be the actual pollution source. If they do not match, a layered interference verification is performed on the defective area to determine the location of the dust in the coating.

[0007] Preferably, the pollution source database includes pollution source entities and at least includes the spatial distribution pattern, expected time series rules, and typical physicochemical property range of dust caused by the corresponding pollution source. The time series distribution characteristics include the trend of defect density changing with production line time and the periodicity of defects.

[0008] Preferably, the dynamic simulation includes: The model initialization step involves obtaining a known pollution source and its corresponding measured defect sample based on the physical model components of the production line. Using the measured defect sample as the calibration target, the environmental dynamic parameters in the digital twin model are adjusted in reverse to minimize the difference between the virtual defect features output by the model and the features of the measured defect sample, thereby obtaining a high-precision digital twin model. The simulation process involves configuring a matching virtual ion release strategy in the high-precision digital twin model based on the suspected contamination source when a defect to be diagnosed is detected. The strategy includes release location, ion physical properties, release time pattern, and release intensity. The simulation then generates a virtual defect result.

[0009] Preferably, the layered interference verification includes: In the multi-band excitation step, a first detection grating and a second detection beam are simultaneously projected onto the surface of the coating to be tested. The first detection beam is a tunable continuous wave laser, and the second detection beam is a wavelength-tunable pulsed laser. In the frequency domain resonance scanning step, linear chirped frequency modulation is applied to the first probe beam so that its frequency is linearly scanned in time, and the scanning range covers the expected resonance frequency range of multiple depths from the coating surface to the substrate; simultaneously, the pulse repetition frequency of the second modulated beam is adjusted accordingly so that it remains coherently locked with the first probe beam in the time domain and frequency domain. The multi-dimensional signal acquisition and calculation steps involve real-time acquisition of photomechanical response signals, photothermal modulation signals, Brillouin scattering signals, and polarization state evolution signals from the coating, and depth calculation is performed using a depth-localization neural optical network model. The results output step outputs the three-dimensional coordinates of each detected particle and its corresponding coating layer.

[0010] Preferably, the layered interference verification includes: The material feature library component steps involve pre-establishing a multi-physical magnetic field response feature database for each coating of the automotive trim panel, including the response thresholds of each layer to various stimuli, feature product markers, and fingerprint features. The defect area preprocessing step involves performing multimodal initial detection on the automotive trim panel area containing dust defects, recording the initial three-dimensional morphology, composition characteristics, and acoustic response of the defects. The intelligent layered cleaning process generates the optimal cleaning strategy based on the response feature database and the initial state through a decision algorithm. According to the generated cleaning strategy, the corresponding stimulus combination is applied, and the surface of the car trim panel is monitored in real time during each cleaning step. The defect exposure and depth positioning steps involve pausing cleaning when the preset depth is reached and using an image recognition strategy to determine dust levels. If the dust is exposed, the results of the currently removed coating are recorded to determine the depth and location of the defect. If the dust is not exposed, the next layer of cleaning is continued.

[0011] Preferably, the layered interference verification also includes a defect tracing step, in which the defect undergoes micro-area laser-induced breakdown spectroscopy analysis, the obtained elemental composition spectrum is matched with a potential pollution source database, the matching degree is calculated and possible pollution sources are output, and based on the depth, location, morphological characteristics and composition information of the defect, combined with the production process parameters of the automotive trim panel, the production process that the defect may be introduced into is inferred.

[0012] Preferably, when a composite defect is detected, the process proceeds to the contamination chain reasoning step, including: The feature separation sub-step extracts the multimodal feature set of the defect, decouples the feature set, and separates multiple independent feature subsets; The pollution chain generation sub-step is based on the production line process flow. The components include a directed causal model containing pollution source nodes, transmission path nodes, process link nodes, and defect morphology nodes. Each feature subset is mapped to a specific defect morphology node in the causal graph. Based on the connection relationship between nodes and time constraints, multiple possible pollution propagation chain hypotheses are automatically inferred and generated. Each hypothesis consists of a series of ordered nodes. The multi-agent simulation sub-step generates a simulation scenario for each pollution chain in the digital twin model. In each scenario, the corresponding virtual pollution source agent and pollutant particle agent group are initialized according to the node order and type of the hypothesis chain. Simulation of all scenarios is run synchronously to simulate the dynamic release, migration, interaction and final deposition process of multi-source pollutants in the virtual production line environment, and generate virtual composite defect morphology corresponding to each hypothesis. In the comparative evaluation sub-step, the virtual composite defect morphology output by each simulation scenario is matched and calculated with the multimodal features to obtain the contamination chain hypothesis with the highest comprehensive similarity score as the optimal diagnostic result.

[0013] Preferably, the system also includes equipment failure maintenance steps. When the equipment is detected as a source of pollution, the system collects multi-dimensional time-series features of the dust defects it generates in real time, and simultaneously obtains the historical maintenance data of the equipment. This data is then input into the equipment health assessment model, which generates the equipment health index. Based on the equipment health index, the system predicts the effective service life of the equipment and the time window for the next failure, and generates equipment maintenance data.

[0014] The pollution sources are identified as equipment pollution sources, environmental pollution sources, paint pollution sources, and board material pollution sources.

[0015] A traceability system for surface particle defects in automotive trim panels includes: The data acquisition module continuously acquires images of dust defects on the board material, uses these images as defects to be diagnosed, and extracts their spatial distribution features and time series distribution features to form a spatiotemporal distribution dataset. The multimodal analysis module performs physicochemical property analysis on dust defects and fuses these physicochemical properties with a spatiotemporal distribution feature dataset to form a multimodal feature vector. The intelligent matching module matches the multimodal feature vector with a preset pollution source database to obtain suspected pollution sources. It also has a preset simulation model. The spatiotemporal distribution dataset is input into the simulation model to perform dynamic simulation of dust generation and deposition of the suspected pollution sources. The virtual defect results are compared and verified with the dust defect images. The precise positioning module determines the suspected pollution source as the actual pollution source if the verification results match; otherwise, it performs layered interference verification on the defective area to determine the location of the dust in the coating.

[0016] The beneficial effects of this invention are as follows: The multimodal analysis step integrates the physicochemical properties and spatiotemporal distribution characteristics of dust defects to form a multimodal feature vector, realizing the feature correlation between defect appearance and defect essence. By analyzing the physicochemical properties such as particle composition and particle size, the range of pollution sources can be initially narrowed down, avoiding the blindness of traditional source tracing with indiscriminate investigation and improving the targeting of source tracing. The intelligent matching step relies on a preset pollution source database to complete the matching of suspected pollution sources, and uses a simulation model to dynamically simulate the dust generation and precipitation process of suspected pollution sources. The virtual defect results are compared and verified with actual dust defect images, which can quickly eliminate unrelated pollution sources, greatly reducing the workload of manual investigation and shortening the source tracing cycle. In the case of mismatch between simulation verification and actual source tracing, the layered interference verification of the precise positioning step can clarify the coating location where dust is located, further locking the process node where pollution occurs, solving the technical pain point of traditional source tracing that cannot distinguish between pollution sources inside and outside the coating, and realizing the precise positioning of pollution sources. It eliminates the need to deploy a large number of high-precision sensors throughout the entire process, significantly reducing equipment purchase and maintenance costs. At the same time, its precise traceability capability can avoid large-scale shutdowns for investigation, minimizing downtime and significantly reducing downtime losses. In addition, by quickly locating the actual source of pollution, it can eliminate the root causes of defects, improve the coating yield of automotive trim panels, and ensure the stable and efficient operation of the production line. Attached Figure Description Figure 1 This is an overall flowchart of the present invention; Figure 2 This is the layered interference path diagram of the present invention; Figure 3 This is a pollution chain reasoning flowchart of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings: A method for tracing the source of surface particle defects in automotive trim panels includes: The data acquisition steps involve continuously acquiring images of dust defects on the board material, using these images as defects to be diagnosed, and extracting their spatial distribution features and temporal series distribution features to form a spatiotemporal distribution dataset. This is achieved by continuously photographing the automotive trim panel and analyzing the images to determine if dust is present. When dust is present, it is used as a defect to be diagnosed, and its spatiotemporal distribution features and spatial series distribution features are extracted. The spatial distribution features include the defect's location, morphology, and density.

[0021] The characteristics of time series distribution include the trend of defect density changing with production line time and the periodicity of defects.

[0022] The multimodal analysis step involves analyzing the physicochemical properties of dust defects and fusing these properties with a spatiotemporal distribution feature dataset to form a multimodal feature vector. Specialized equipment is used to detect the multidimensional physicochemical properties of dust and generate the multimodal feature vector.

[0023] The intelligent matching step matches multimodal feature vectors with a pre-defined pollution source database to identify potential pollution sources. A pre-defined simulation model is used, and the spatiotemporal distribution dataset is input into the model to dynamically simulate dust generation and deposition from the potential pollution sources. The virtual defect results are then compared and verified with dust defect images. A weighted cosine similarity algorithm combined with hierarchical clustering is used to complete feature matching. First, the similarity of each dimension of the multimodal feature vector is calculated with the corresponding indicators of the defect feature set in the pollution source database. Then, a comprehensive matching degree is calculated based on pre-defined feature weight coefficients. A comprehensive matching degree threshold is set, and pollution source entities that meet the threshold requirements are selected. A list of potential pollution sources is generated by sorting the matching degrees from high to low. For example, if a defect's multimodal feature vector shows an iron element content of 72%, a sheet-like morphology, concentration at the left roller station, and a high frequency of occurrence over 24 hours, then its comprehensive matching degree with the conveyor roller wear pollution source in the database reaches 92%, making it the first potential pollution source. The simulated virtual defect results are compared with the actual collected dust defect images in multiple dimensions. The comparison indicators include defect spatial location clustering, particle size distribution ratio, density level, time period regularity and microscopic morphology similarity. The overall consistency ≥85% is set as the verification pass standard to preliminarily determine the rationality of the suspected pollution source.

[0024] The pollution source database includes pollution source entities and at least the spatial distribution patterns, expected time series rules, and typical physicochemical properties of dust caused by the corresponding pollution sources. A structured pollution source database is constructed, with its core consisting of two main modules: a pollution source entity database and a defect feature matching database. The pollution source entity database is divided into four subcategories based on pollution type: paint (paint impurities, resin agglomerates), board substrate (metal shavings, wood chips, substrate residual dust), coating equipment (roller wear debris, chain rust particles, fan dust), and production environment (workshop suspended dust, air conditioning filter dust). Each subcategory corresponds to a unique pollution source code. Each pollution source entity is bound to a unique defect feature set, which contains at least three types of core data: first, spatial distribution patterns, such as paint impurity defects being randomly and uniformly distributed, and roller wear debris defects being concentrated on the contact side of the board conveyor; second, expected time series rules, such as fan dust accumulation defects showing a 24-hour periodic peak, and paint replacement batch impurity defects showing a sudden increase within 1 hour after paint replacement; and third, typical physicochemical property ranges, such as the Fe element content of chain rust particles being greater than 65%, the fibrous morphology of wood chip particles being greater than 90%, and the specific functional group matching degree of paint resin impurities being greater than 85%. At the same time, weight coefficients are assigned to each feature to achieve the basic data support for accurate matching.

[0025] Dynamic simulation includes: The model initialization steps involve creating an initial digital twin model based on the physical model components of the production line. Known pollution sources and corresponding measured defect samples are then acquired. Using these measured defect samples as calibration targets, the environmental dynamic parameters in the digital twin model are adjusted in reverse to minimize the difference between the virtual defect features output by the model and the features of the measured defect samples, resulting in a high-precision digital twin model. Based on the actual layout of the coating production line, a 1:1 initial digital twin model is constructed. This model encompasses three core modules: first, a workstation space module, accurately mapping the three-dimensional coordinates and relative positions of workstations such as pretreatment, spraying, curing, and conveying; second, an equipment process module, recording process parameters such as conveyor roller speed, spraying pressure, paint mist flow rate, and curing oven temperature; and third, an environmental dynamic module, including environmental parameters such as workshop airflow speed and direction, temperature and humidity distribution, dust suspension concentration, and particulate matter settling rate, achieving a full-element digital replication of the physical scene of the production line. By selecting known pollution sources from the production line's history, such as known paint impurity pollution sources and corresponding measured defect samples, and using the spatiotemporal distribution characteristics, defect location, density, time period, and physicochemical properties (particle size and composition) of the measured samples as calibration targets, a gradient descent inverse optimization algorithm is used to iteratively adjust key parameters of the model's environmental dynamics module, such as airflow velocity, particulate matter settling coefficient, and paint mist diffusion coefficient. After each adjustment, virtual defect features are output and compared with the measured samples, and the mean square error between the two is calculated until the error is less than a threshold, ultimately forming a high-precision digital twin model that can accurately reproduce the pollution process.

[0026] The simulation process involves configuring a matching virtual ion release strategy in a high-precision digital twin model when a defect to be diagnosed is detected, based on the suspected contamination source. This strategy includes release location, ion physical properties, release time pattern, and release intensity. A simulation is then performed to generate virtual defect results. For the selected suspected contamination sources, a dedicated virtual particle release strategy is configured in the high-precision digital twin model. This strategy must perfectly match the characteristics of the contamination source and the defect features, specifically including four aspects: First, the release location, corresponding to the actual workstation of the contamination source, such as setting the contact area between the left conveyor roller and the plate for roller wear contamination sources, and setting paint impurities for the area below the spray gun nozzle; second, the particle physical properties, matching the physicochemical parameters in the multimodal feature vector, such as particle size 20-50μm, flaky morphology, iron content 72%, and hardness HV350; third, the release time pattern, conforming to the defect time sequence pattern, such as 24-hour periodic release or concentrated release 1 hour after shift change; and fourth, the release intensity, corresponding to the defect density level. A digital twin model is launched for dynamic simulation, with the simulation cycle consistent with the actual production cycle. The entire process of virtual particle generation, transport, diffusion, and sedimentation is tracked in real time. The attachment position, density, morphological distribution, and time change of particles on the decorative panel surface are recorded. Finally, virtual defect images and feature datasets consistent with the actual detection format are generated, providing a basis for subsequent comparative verification.

[0027] The precise location step involves several steps. If the verification results match, the suspected contamination source is determined to be the actual contamination source. If they do not match, a layered interference verification is performed on the defective area to determine the location of the dust on the coating. If the overall consistency between the virtual defect result and the actual defect image is greater than a threshold, the suspected contamination source is directly determined to be the actual contamination source. Key information such as the specific type of contamination source, the location of the contamination, and the pattern of contamination triggering time is simultaneously output, providing a precise basis for targeted rectification on the production side. If the overall consistency is less than a threshold, the layered interference verification process is initiated to pinpoint the process node where the contamination occurred by precisely identifying the location of dust particles on the coating.

[0028] Layered interference verification includes: In the multi-band excitation step, a first probe grating and a second probe beam are simultaneously projected onto the surface of the coating under test. The first probe beam is a tunable continuous-wave laser, and the second probe beam is a wavelength-tunable pulsed laser. The wavelength of the first probe beam, a tunable continuous-wave laser, can be specifically adjusted according to the coating material of the decorative panel, such as polyurethane topcoat or epoxy primer. The wavelength adjustment range of the second probe beam is consistent with that of the first probe beam. The pulse width is set to 10 ns, and the single-pulse energy is controlled at 1 μJ to ensure that the photothermal and photomechanical responses of the coating are excited without damaging the coating structure. The frequency domain resonance scanning step involves applying linear chirped frequency modulation to the first probe beam, causing its frequency to scan linearly in time, covering the expected resonance frequency range at multiple depths from the coating surface to the substrate. Simultaneously, the pulse repetition frequency of the second modulation beam is adjusted accordingly to maintain coherent locking with the first probe beam in both the time and frequency domains. The linear chirped frequency modulation of the first probe beam is achieved by applying linear chirped frequency modulation to the first probe beam through a built-in acousto-optic modulator. The modulation frequency scanning range is set to 10MHz-10GHz, and the scanning rate is 1GHz / ms. The scanning range accurately covers the expected resonance frequency range at multiple depths from the coating surface to the substrate, achieving targeted excitation of coatings at different depths. The pulse repetition frequency of the second probe beam is synchronously adjusted based on the chirped modulation pattern of the first probe beam. A pulse signal generator synchronously adjusts the pulse repetition frequency of the second probe beam in 10MHz steps to ensure that the pulse repetition frequency of the second probe beam is synchronized with the instantaneous frequency of the first probe beam in the time domain and coherently locked in the frequency domain. This guarantees that both beams can produce stable interference effects at different depths of the coating. During frequency domain scanning, a front-end photodetector monitors the initial resonance feedback signal of the coating in real time. When a sudden change in resonance signal intensity is detected, the frequency scan is automatically paused and the corresponding frequency value is recorded. This delineates key depth ranges for subsequent multidimensional signal acquisition, improving detection efficiency.

[0029] The multi-dimensional signal acquisition and calculation process involves real-time acquisition of photomechanical response signals, photothermal modulation signals, Brillouin scattering signals, and polarization state evolution signals from the coating. Depth calculation is performed using a depth-localization neural optical network model. A multi-channel signal acquisition module synchronously acquires four types of characteristic signals generated after the coating is stimulated, with a sampling frequency set to 100 MS / s to ensure the integrity and timeliness of the signal data: Photomechanical response signal: Micro-vibration displacement signals at different depths of the coating are acquired using a laser Doppler vibrometer, reflecting the bonding state between particles and the coating substrate and the physical morphology of the particles; Photothermal modulation signal: Instantaneous temperature change signals of the coating are acquired using an infrared thermal imager, utilizing the difference in thermal conductivity between dust particles and the coating resin to distinguish the boundary between particles and the coating substrate; Brillouin scattering signal: Frequency shift signals of scattered light are acquired using a Brillouin optical time-domain analyzer, determining the medium type at the detection point as dust particles or coating resin based on the magnitude of the frequency shift; Polarization state evolution signal: Polarization state change data of reflected light are acquired using a polarization state analyzer, accurately locating the three-dimensional boundary of the particles based on the difference in refractive index between the particles and the coating. The deep localization neural optical network model is a pre-trained dedicated model. The training dataset contains multi-dimensional signal samples of different coating levels and different types of dust particles. The model input is the fusion data of four types of feature signals, and the output is the depth coordinates of the detection point. After the signal acquisition is completed, the real-time acquired signal data is input into the model. The model performs depth feature extraction and coordinate calculation through multi-layer convolution and attention mechanisms, while adaptively filtering out signal noise, and controlling the localization error within ±0.2μm.

[0030] The output step provides the 3D coordinates and coating layer level of each detected particle. A depth-localization neural optical network model outputs the 3D spatial coordinates of each particle, where the X and Y axes are calibrated by the grating projection position on the sample surface, and the Z-axis coordinate is determined by the depth detection result, forming complete 3D localization data including surface planar coordinates and depth coordinates. The particle's Z-axis depth coordinate is matched with preset coating depth ranges for decorative panels, such as topcoat (0-20μm), intermediate coat (20-50μm), primer (50-80μm), and substrate (>80μm), to determine the specific coating layer level of each particle. Simultaneously, a visual detection report is generated, including a 3D heatmap of particle distribution, a statistical chart of coating layer assignment, and a boundary morphology diagram of particles and coatings. Based on the coating layer level of the particle, the process node where contamination occurred is pinpointed; for example, if the particle is in the primer layer, it corresponds to the pretreatment or primer spraying process; if it is in the topcoat layer, it corresponds to the topcoat spraying or curing process, providing precise process guidance for the final location of the contamination source.

[0031] Layered interference verification includes: The material feature library construction process involves pre-establishing a multi-physics magnetic field response feature database for each coating of the automotive trim panel. This database includes the response thresholds of each layer to various stimuli, characteristic product markers, and fingerprint features. The database is divided into four sub-databases based on coating type: primer, intermediate coat, topcoat, and clear coat. Each sub-database is bound to the corresponding coating's basic parameters, including thickness, density, adhesion, and multi-physics response features, providing a benchmark for subsequent layered cleaning and positioning. Three typical stimuli—laser, ultrasound, and mild chemical reagents—are selected to cover both physical and chemical action modes. The tolerance thresholds of each coating to different stimuli are tested using the controlled variable method. For example, epoxy primer can tolerate a 633nm laser with a power ≤8mW and a 40kHz pulse frequency. Ultrasonic stimulation lasts ≤15s, and polyurethane topcoat can withstand immersion in weakly alkaline paint remover for ≤30s. Exceeding these thresholds will result in irreversible damage to the coating, such as dissolution and peeling. These thresholds are recorded in the database. Characteristic products of each coating after stimulation are also recorded. For example, polyester intermediate coatings will produce specific resin pyrolysis fragments after mild irradiation with a 1064nm laser, and acrylic varnish will precipitate trace amounts of acrylic monomers upon contact with organic solvents. The composition and morphology of these products are stored as characteristic markers in the database. Fingerprint feature extraction: Specific fingerprint features of each coating under different stimuli are collected. The ultrasonic echo characteristic frequency of polyurethane topcoat is 32kHz, forming a fingerprint database for coating identification. Dynamic database updates: Multiphysics response data of new batches of coatings are collected periodically, and the database is updated using an incremental learning algorithm to ensure a data matching degree ≥99% with coatings used in actual production.

[0032] The defect area preprocessing steps include: performing multimodal initial detection on automotive trim panel areas containing dust defects, recording the initial three-dimensional morphology, compositional characteristics, and acoustic response of the defects; conducting three rounds of multimodal initial detection on the extracted defect area samples to achieve comprehensive acquisition of basic defect information; using a confocal laser scanning microscope to scan the defect area, acquiring the three-dimensional morphology of the defects, including protrusion height, substrate area, edge slope, and boundary characteristics with surrounding coatings, generating a three-dimensional morphology model; compositional characteristic detection: using a portable X-ray energy dispersive spectrometer to perform micro-area compositional scanning, recording the elemental composition and content ratio of the defect area, and initially distinguishing between organic and inorganic particle types; acoustic response detection: using an ultrasonic flaw detector to emit ultrasonic signals, acquiring the ultrasonic echo attenuation coefficient and phase change of the defect area to determine the degree of adhesion between the defect and the coating. Initial data integration and archiving: binding the three types of detection data with unique sample codes to generate an initial defect state archive, serving as the decision input for subsequent intelligent layered cleaning.

[0033] The intelligent layered cleaning process, based on a response feature database and initial state, generates an optimal cleaning strategy through a decision algorithm. Following this strategy, corresponding stimulus combinations are applied. Real-time monitoring of the automotive trim panel surface is conducted at each cleaning step. Based on the coating response thresholds and initial defect state files in the material feature library, a fuzzy decision algorithm generates a layered cleaning strategy. The algorithm inputs coating type, defect 3D morphology, and acoustic response parameters, and outputs a combination of stimulus type, intensity, and duration. Stimulus combinations are applied strictly according to the generated strategy, treating the coating layer by layer from the surface to the inner layers and from weak to strong stimuli. During cleaning, a laser Raman spectrometer and a high-magnification industrial camera are simultaneously used for monitoring. The former determines the current coating type through fingerprint feature peaks, while the latter acquires the coating surface morphology in real time. If a change in coating fingerprint features is detected, such as a shift from topcoat feature peaks to intermediate coat feature peaks, the current stimulus is immediately stopped, and the process proceeds to the next layer, ensuring precise and controllable cleaning and preventing cross-layer damage.

[0034] The defect exposure and depth localization steps involve pausing cleaning when a preset depth is reached. An image recognition strategy is used to determine dust levels. If dust is exposed, the current removed coating is recorded, and the depth of the defect is determined. If no dust is exposed, cleaning continues to the next layer. The preset depth is based on the standard thickness of each coating in the material feature library. Each completed layer of cleaning represents a preset depth node. Upon reaching the preset depth node, cleaning is paused. An AI image recognition model identifies the cleaned surface, using particle morphology and contrast characteristics to determine if dust defects are exposed. If a defect is identified, the current coating layer is immediately recorded. For example, if a defect is exposed after cleaning the intermediate coating, it is determined to be located at the interface between the intermediate and primer coatings, and the precise depth of the defect is determined using a 3D coordinate analyzer. If no defect is identified, the next layer solution from the cleaning strategy library is called to continue cleaning until the defect is fully exposed.

[0035] Layered interferometry verification also includes a defect tracing step. Micro-area laser-induced breakdown spectroscopy (LIBS) analysis is performed on the defects. The obtained elemental composition spectra are matched with a potential pollution source database to calculate the matching degree and output possible pollution sources. Based on the depth, morphological characteristics, and composition information of the defects, combined with the production process parameters of automotive trim panels, the production stages that the defects may have introduced are inferred. Micro-area LIBS detection is performed on exposed dust particles to collect the full elemental composition spectra of the particles, accurately obtaining the content percentage and characteristic spectral lines of each element. The LIBS elemental spectra are compared with the characteristic spectra of potential pollution sources such as paint impurities, equipment wear debris, substrate residue, and environmental dust. A cosine similarity algorithm is used to calculate the matching degree, outputting the top three possible pollution sources. For example, if Fe accounts for 75% and Zn accounts for 12%, the matching degree with the conveyor chain corrosion pollution source reaches 91%, making it the primary suspect source. By combining the depth, location, morphological characteristics, and composition information of defects, and linking the production process parameters of automotive trim panels, the spraying time of each coating, equipment station, and environmental cleaning cycle, a comprehensive inference is made: if the defect is located in the primer layer and the composition is base metal shavings, it is inferred that the pollution occurred in the pretreatment process of the panel before primer spraying; if the defect is located on the topcoat surface and the composition is environmental dust, it is inferred that the pollution occurred in the workshop environmental control process during the topcoat curing stage, and finally a complete pollution source tracing report is output.

[0036] When a composite defect is detected, the contamination chain inference step is initiated. When a composite defect is detected on the decorative panel surface (i.e., the defect contains two or more independent contaminants), the contamination chain inference step is automatically triggered. Through feature decoupling, causal modeling, multi-agent simulation, and comparative evaluation, the precise tracing of the multi-source causes and propagation paths of composite defects is achieved, including: The feature separation sub-step extracts the multimodal feature set of the defect, decouples the feature set, and separates multiple independent feature subsets. It then integrates all data from previous data acquisition, multimodal analysis, and hierarchical interference verification to construct a multimodal feature set of the composite defect. This set contains four core features: 1) spatial distribution features, including the location clustering and relative spacing of each contaminant component; 2) time series features, including the occurrence cycle and abrupt change sequence of defects of different components; 3) physicochemical property features, including the elemental proportion, microstructure, and thermal decomposition temperature of each component; and 4) coating depth features, including the coating layer and depth coordinates of different components. An algorithm combining independent component analysis and density clustering is used to decouple the feature set. The specific process is as follows: first, the mixed high-dimensional feature data is decomposed into independent feature components using the ICA algorithm; then, the feature components are clustered using the DBSCAN algorithm. Based on core indicators such as the elemental spectrum and morphological characteristics of the components, the feature set is split into multiple independent feature subsets, each corresponding to a single contaminant component in the composite defect. For example, the feature set of a certain composite defect can be decoupled into a feature subset of metal scrap with 75% Fe element content, flaky morphology, and primer layer depth, and a feature subset of paint impurities with 88% resin functional group matching degree, spherical morphology, and topcoat layer depth. The integrity of each feature subset after decoupling is checked, and noisy features and redundant data are removed to ensure that each subset can independently characterize the core attributes of a contaminant, providing accurate input for the subsequent generation of the contamination chain.

[0037] The pollution chain generation sub-steps, based on the production line's process flow, construct a directed causal model comprising pollution source nodes, transmission path nodes, process step nodes, and defect morphology nodes. Each feature subset is mapped to a specific defect morphology node in the causal graph. Based on the connections between nodes and time constraints, multiple possible pollution propagation chain hypotheses are automatically generated, each hypothesis consisting of a series of ordered nodes. Based on the complete automotive trim panel coating process—pretreatment-primer spraying-primer curing-intermediate coat spraying-intermediate coat curing-topcoat spraying-topcoat curing-finished product inspection—a directed causal model is constructed, containing four types of nodes: pollution source nodes: covering all potential pollution sources such as the paint supply system, substrate, conveyor rollers, and workshop environment; transmission path nodes: including contaminant migration paths such as conveyor belts, workshop airflow channels, paint mist delivery pipelines, and equipment contact surfaces; process step nodes: corresponding to each core process in coating production, with each node bound to the process's time window, process parameters, and other information; and defect morphology nodes: containing typical defect characteristics of different pollution components (such as the flaky morphology of metal scraps and the spherical morphology of paint impurities). At the same time, directed connections are established for nodes based on production logic, and time constraints are added to each connection. Feature subset mapping: Each decoupled feature subset is precisely mapped to the corresponding defect morphology node in the causal model according to its defect morphology characteristics. For example, the metal scrap feature subset is mapped to the metal scrap defect morphology node, and the paint impurity feature subset is mapped to the paint impurity defect morphology node. Pollution propagation chain hypothesis generation: Based on the directed connections between nodes and time constraints, multiple possible pollution propagation chain hypotheses are automatically generated using a Bayesian network inference algorithm. Each hypothesis consists of a series of ordered nodes, fully covering the entire chain from pollution source to transmission path, process step, and defect morphology. For example, for the aforementioned composite defect, two core hypothesis chains can be generated: The multi-agent simulation sub-step involves generating a simulation scenario for each pollution chain hypothesis in the digital twin model. Within each scenario, based on the node order and type of the hypothesis chain, the corresponding virtual pollution source agent and pollutant particle agent group are initialized. Simultaneously, simulations of all scenarios are run to simulate the dynamic release, migration, interaction, and final deposition process of multi-source pollutants in the virtual production line environment, generating virtual composite defect morphology corresponding to each hypothesis. In the high-precision digital twin model, a dedicated simulation scenario is generated for each pollution chain hypothesis, and two types of core agents are defined: one corresponding to the pollution source node in the hypothesis chain, possessing pollutant release rules consistent with real pollution sources; and the other corresponding to the pollutants from each pollution source, with each particle agent bound to the physicochemical properties of a corresponding feature subset, such as the flaky morphology and Fe element ratio of metal scrap particles, and the spherical morphology and resin composition of paint impurity particles, while also assigning dynamic behavior rules for particle migration, collision, and adhesion. The simulation scenarios corresponding to all pollution chain assumptions are launched and run synchronously according to the actual rhythm of the coating production line to fully simulate the dynamic behavior of multi-source pollutants: In the release stage, each pollution source agent releases pollutant particle groups according to the time sequence and intensity of the assumption chain; in the migration stage, the particle agents migrate along the transmission path nodes of the assumption chain, affected by environmental factors such as workshop airflow and equipment movement; in the interaction stage, different types of particle agents interact with each other through collision, adsorption, and encapsulation, forming the prototype of composite defects; in the deposition stage, the particle agents deposit at the specified coating depth on the board within the time window of the corresponding process node, finally generating a virtual composite defect morphology corresponding to each assumption chain, including the positional distribution, morphological characteristics, and hierarchical information of each component.

[0038] In the comparative evaluation sub-step, the virtual composite defect morphology output from each simulation scenario is matched with multimodal features to obtain the pollution chain hypothesis with the highest comprehensive similarity score as the optimal diagnostic result. The virtual composite defect morphology output from each simulation scenario is then matched with the multimodal features of the actual composite defects in multiple dimensions. The pollution propagation chain hypothesis with the highest comprehensive similarity score is selected as the final diagnostic result. Simultaneously, all pollution sources, transmission paths, process steps, and their timing within the chain are output, and a visualized pollution chain map is generated, providing accurate end-to-end evidence for collaborative remediation of multi-source pollution in production.

[0039] It also includes equipment failure maintenance steps. When equipment is detected as a pollution source, multi-dimensional time-series characteristics of the dust defects it generates are collected in real time, and historical maintenance data of the equipment is simultaneously acquired and input into the equipment health assessment model. The model obtains the equipment health index, and based on the equipment health index, the effective service life of the equipment and the time window of the next failure are predicted, generating equipment maintenance data. When the coating equipment is determined to be the core pollution source through the aforementioned source tracing steps, such as wear of conveyor rollers, chain corrosion, and dust shedding from the fan, the equipment failure maintenance steps are automatically triggered. Through multi-dimensional data collection, health status assessment, and lifespan and failure prediction, a precise maintenance plan is generated, achieving proactive management of equipment failures. Relying on the defect detection and data acquisition modules already deployed on the production line, multi-dimensional time-series characteristics of the dust defects generated by the equipment are collected in real time, including defect occurrence frequency characteristics, defect density evolution characteristics, and defect physicochemical property time-series characteristics. The complete historical data of the equipment is simultaneously retrieved from the production line equipment management system to construct an equipment operation and maintenance dataset. The equipment health assessment model is a hybrid model based on Long Short-Term Memory (LSTM) networks combined with gradient boosting decision trees. It synchronously inputs real-time collected defect time-series feature datasets and retrieved equipment operation and maintenance (O&M) datasets into the pre-trained model. The model outputs a health index for the equipment through feature association analysis and multi-dimensional weight calculation. Based on the equipment health index and core parameters from the O&M data—service life, cumulative runtime, historical failure frequency, and the equipment's design life baseline—the model calculates the equipment's effective remaining service life using a remaining life prediction algorithm. Based on the evolution trend of defect time-series features and the temporal patterns of historical failures, the model defines the time window for the next failure. Based on the equipment health level, remaining life, and failure time window, it generates differentiated maintenance strategies through a rule engine and expert knowledge base.

[0040] Pollution sources include equipment pollution sources, environmental pollution sources, paint pollution sources, and board material pollution sources.

[0041] A traceability system for surface particle defects in automotive trim panels includes: The data acquisition module continuously acquires images of dust defects on the board material, uses these images as defects to be diagnosed, and extracts their spatial distribution features and time series distribution features to form a spatiotemporal distribution dataset. The multimodal analysis module performs physicochemical property analysis on dust defects and fuses these physicochemical properties with a spatiotemporal distribution feature dataset to form a multimodal feature vector. The intelligent matching module matches multimodal feature vectors with a preset pollution source database to obtain suspected pollution sources. It also has a preset simulation model. The spatiotemporal distribution dataset is input into the simulation model to perform dynamic simulation of dust generation and deposition of suspected pollution sources, and the virtual defect results are compared and verified with dust defect images. The precise positioning module determines the suspected pollution source as the actual pollution source if the verification results match; otherwise, it performs layered interference verification on the defective area to determine the location of the dust in the coating.

[0042] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for tracing the source of surface particle defects in automotive decorative panels, characterized in that, include: The data acquisition steps involve continuously acquiring dust defect images of the board material, using these images as defects to be diagnosed, and extracting their spatial distribution features and time series distribution features to form a multimodal feature vector. The intelligent matching step matches the multimodal feature vector with a preset pollution source database to obtain suspected pollution sources. A simulation model is preset, and the spatiotemporal distribution dataset is input into the simulation model to perform dynamic simulation of dust generation and deposition of the suspected pollution sources. The virtual defect results are then compared and verified with the dust defect images. The precise location step involves verifying whether the suspected pollution source is the actual pollution source. If the verification results match, the suspected pollution source is determined to be the actual pollution source. If they do not match, a layered interference verification is performed on the defective area to determine the location of the dust in the coating.

2. The method for tracing the source of surface particle defects in automotive decorative panels according to claim 1, characterized in that, The pollution source database includes pollution source entities and at least the spatial distribution patterns, expected time series rules, and typical physicochemical properties range of dust caused by the corresponding pollution sources. The time series distribution characteristics include the trend of defect density changing with production line time and the periodicity of defects.

3. The method for tracing the source of surface particle defects in automotive decorative panels according to claim 1, characterized in that, The dynamic simulation includes: The model initialization step involves obtaining a known pollution source and its corresponding measured defect sample based on the physical model components of the production line. Using the measured defect sample as the calibration target, the environmental dynamic parameters in the digital twin model are adjusted in reverse to minimize the difference between the virtual defect features output by the model and the features of the measured defect sample, thereby obtaining a high-precision digital twin model. The simulation process involves configuring a matching virtual ion release strategy in the high-precision digital twin model based on the suspected contamination source when a defect to be diagnosed is detected. The strategy includes release location, ion physical properties, release time pattern, and release intensity. The simulation then generates a virtual defect result.

4. The method for tracing the source of surface particle defects in automotive decorative panels according to claim 1, characterized in that, The hierarchical interference verification includes: In the multi-band excitation step, a first detection grating and a second detection beam are simultaneously projected onto the surface of the coating to be tested. The first detection beam is a tunable continuous wave laser, and the second detection beam is a wavelength-tunable pulsed laser. In the frequency domain resonance scanning step, linear chirped frequency modulation is applied to the first probe beam so that its frequency is linearly scanned in time, and the scanning range covers the expected resonance frequency range of multiple depths from the coating surface to the substrate; simultaneously, the pulse repetition frequency of the second modulated beam is adjusted accordingly so that it remains coherently locked with the first probe beam in the time domain and frequency domain. The multi-dimensional signal acquisition and calculation steps involve real-time acquisition of photomechanical response signals, photothermal modulation signals, Brillouin scattering signals, and polarization state evolution signals from the coating, and depth calculation is performed using a depth-localization neural optical network model. The results output step outputs the three-dimensional coordinates of each detected particle and its corresponding coating layer.

5. The method for tracing the source of surface particle defects in automotive decorative panels according to claim 1, characterized in that, The hierarchical interference verification includes: The material feature library component steps involve pre-establishing a multi-physical magnetic field response feature database for each coating of the automotive trim panel, including the response thresholds of each layer to various stimuli, feature product markers, and fingerprint features. The defect area preprocessing step involves performing multimodal initial detection on the automotive trim panel area containing dust defects, recording the initial three-dimensional morphology, composition characteristics, and acoustic response of the defects. The intelligent layered cleaning process generates the optimal cleaning strategy based on the response feature database and the initial state through a decision algorithm. According to the generated cleaning strategy, the corresponding stimulus combination is applied, and the surface of the car trim panel is monitored in real time during each cleaning step. The defect exposure and depth positioning steps involve pausing cleaning when the preset depth is reached and using an image recognition strategy to determine dust levels. If the dust is exposed, the results of the currently removed coating are recorded to determine the depth and location of the defect. If the dust is not exposed, the next layer of cleaning is continued.

6. The method for tracing the source of surface particle defects in automotive decorative panels according to claim 5, characterized in that, The layered interference verification also includes a defect tracing step, in which the defect undergoes micro-area laser-induced breakdown spectroscopy analysis, the obtained elemental composition spectrum is matched with a potential pollution source database, the matching degree is calculated and possible pollution sources are output, and based on the depth, location, morphological characteristics and composition information of the defect, combined with the production process parameters of the automotive trim panel, the production process that the defect may be introduced into is inferred.

7. The method for tracing the source of surface particle defects in automotive decorative panels according to claim 1, characterized in that, When a composite defect is detected, the contamination chain reasoning step is initiated, including: The feature separation sub-step extracts the multimodal feature set of the defect, decouples the feature set, and separates multiple independent feature subsets; The pollution chain generation sub-step is based on the production line process flow. The components include a directed causal model containing pollution source nodes, transmission path nodes, process link nodes, and defect morphology nodes. Each feature subset is mapped to a specific defect morphology node in the causal graph. Based on the connection relationship between nodes and time constraints, multiple possible pollution propagation chain hypotheses are automatically inferred and generated. Each hypothesis consists of a series of ordered nodes. The multi-agent simulation sub-step generates a simulation scenario for each pollution chain in the digital twin model. In each scenario, the corresponding virtual pollution source agent and pollutant particle agent group are initialized according to the node order and type of the hypothesis chain. Simulation of all scenarios is run synchronously to simulate the dynamic release, migration, interaction and final deposition process of multi-source pollutants in the virtual production line environment, and generate virtual composite defect morphology corresponding to each hypothesis. In the comparative evaluation sub-step, the virtual composite defect morphology output by each simulation scenario is matched and calculated with the multimodal features to obtain the contamination chain hypothesis with the highest comprehensive similarity score as the optimal diagnostic result.

8. The method for tracing the source of surface particle defects in automotive decorative panels according to claim 1, characterized in that, It also includes equipment failure maintenance steps. When the equipment is detected as a pollution source, the multi-dimensional time series characteristics of the dust defects it generates are collected in real time, and the historical maintenance data of the equipment is obtained simultaneously and input into the equipment health assessment model. The equipment health index is obtained through the model. Based on the equipment health index, the effective service life of the equipment and the time window of the next failure are predicted, and equipment maintenance data is generated.

9. The method for tracing the source of surface particle defects in automotive decorative panels according to claim 1, characterized in that, The pollution sources include equipment pollution sources, environmental pollution sources, paint pollution sources, and board material pollution sources.

10. A traceability system for surface particle defects in automotive trim panels, characterized in that, include: The data acquisition module continuously acquires images of dust defects on the board material, uses these images as defects to be diagnosed, and extracts their spatial distribution features and time series distribution features to form a spatiotemporal distribution dataset. The multimodal analysis module performs physicochemical property analysis on dust defects and fuses these physicochemical properties with a spatiotemporal distribution feature dataset to form a multimodal feature vector. The intelligent matching module matches the multimodal feature vector with a preset pollution source database to obtain suspected pollution sources. It also has a preset simulation model. The spatiotemporal distribution dataset is input into the simulation model to perform dynamic simulation of dust generation and deposition of the suspected pollution sources. The virtual defect results are compared and verified with the dust defect images. The precise positioning module determines the suspected pollution source as the actual pollution source if the verification results match; otherwise, it performs layered interference verification on the defective area to determine the location of the dust in the coating.

Images