Intelligent decision method and system for laser paint removal process parameters

Abstract

The application discloses an intelligent decision method and system for laser paint removal process parameters, and the method can quickly and accurately obtain the best process parameter combination for steel paint removal by fusing finite element simulation, machine vision detection and intelligent optimization algorithm, completely changes a traditional experience trial-and-error mode, and realizes intelligent decision of laser paint removal process parameters on a steel surface.

Description

Technical Field

[0001] This invention relates to the field of laser paint removal technology, specifically to an intelligent decision-making method and system for laser paint removal process parameters. Background Technology

[0002] Against the backdrop of the nation's in-depth implementation of the green transformation and high-quality development strategy for the manufacturing industry, laser cleaning technology, as a breakthrough innovation in the field of clean production, has become a current research hotspot due to its significant advantages such as non-contact processing, no chemical pollution, and precise controllability. In the field of material surface treatment, it focuses on solving common needs such as non-destructive cleaning of metal substrates to remove rust layers from metal components, precise removal of composite paint layers, and surface treatment for mold regeneration. In the field of precision component maintenance, it overcomes technical bottlenecks in non-destructive cleaning of microelectronic devices and maintenance of precision aerospace parts. In the field of special environment cleaning, it has been successfully applied to high-risk scenarios such as decontamination of radioactive pollutants and decommissioning of nuclear facilities.

[0003] Laser cleaning technology, with its non-contact processing, selective removal, and environmental friendliness, has become a research hotspot for green paint removal on steel surfaces, offering new avenues for intelligent maintenance and green upkeep of industrial equipment. However, despite its significant advantages, the setting of process parameters often heavily relies on human experience when facing complex and variable real-world conditions. This leads to unstable cleaning quality and maximizing efficiency. Therefore, how to introduce intelligent methods to break through traditional trial-and-error approaches and achieve intelligent decision-making for cleaning process parameters has become a core challenge hindering the large-scale industrial application of this technology.

[0004] In practical engineering applications, laser cleaning involves extremely complex physical and chemical mechanisms. Laser cleaning encompasses numerous core process parameters (such as laser power, scanning speed, energy density, pulse width, repetition frequency, and spot diameter), which exhibit highly nonlinear coupling relationships, directly impacting the final cleaning quality, substrate safety, and processing efficiency. Currently, the selection of laser cleaning processes in industrial settings still heavily relies on human experience and repeated trial and error by technicians. With the explosive growth in demand for green cleaning across various industries, skilled workers on the front lines face significant technical challenges. Lacking systematic theoretical guidance, practitioners struggle to grasp the inherent laws governing the synergy of multiple parameters in a short period, easily leading to incomplete cleaning or irreversible thermal damage to the substrate due to parameter overshoot. This has become a major industry pain point limiting the large-scale promotion of this technology. Therefore, there is an urgent need to introduce advanced technologies to enable practitioners to learn and master laser cleaning processes more quickly. Summary of the Invention

[0005] To address these issues, this invention proposes an intelligent decision-making method and system for laser paint removal process parameters.

[0006] According to one aspect of the present invention, an intelligent decision-making method for laser paint removal process parameters is proposed, comprising the following steps: S1. Establish a laser paint removal simulation model. Input different combinations of laser paint removal process parameters corresponding to the material combination of the base paint layer into the laser paint removal simulation model for simulation. Filter to obtain the simulation process parameter dataset. S2, Based on the simulation process parameter dataset, conduct corresponding laser paint removal experiments, collect paint removal effect images and perform preprocessing, construct a laser paint removal defect detection model, and train the laser paint removal defect detection model using the preprocessed paint removal effect images. S3. Based on the simulation process parameter dataset, select multiple sets of laser paint removal process parameter combinations and set corresponding process parameter combination gradient experiments. Obtain the paint removal effect image to be detected corresponding to the process parameter combination gradient experiment. Detect the paint removal effect image to be detected based on the trained laser paint removal defect detection model to filter the experimental process parameter dataset. S4, based on the relative parameter prediction error between the experimental process parameter dataset and the simulation process parameter dataset, the simulation process parameter dataset is corrected, and a better cleaning parameter set is obtained by combining the corrected simulation process parameter dataset and the experimental process parameter dataset. S5, Obtain the physical correlation features of the paint layer and the substrate, calculate the mean and standard deviation of the corresponding physical correlation features based on the optimal cleaning parameter set, perform Z-score standardization on the physical correlation features to obtain standardized input features, the physical correlation features include at least the thermal property parameters of the paint layer, the thermal property parameters of the substrate, and derived features calculated based on the thermal property parameters of the paint layer and the substrate; S6, establish a process parameter prediction model combination, use the standardized input features as the input of the process parameter prediction model combination, and use the particle swarm optimization algorithm to find the model parameter combination of the process parameter prediction model combination; S7. Using the process parameters predicted after the model parameters are determined, the laser paint removal process parameter combination of the material combination of the substrate paint layer to be treated is predicted.

[0007] This application combines simulation models, machine vision technology, and particle swarm optimization prediction to establish a parameter prediction model, which can significantly reduce the number of trials and errors and experimental costs, improve the efficiency and accuracy of process parameter screening, and adapt to different combinations of paint layers and substrate materials to meet diverse laser paint removal needs.

[0008] Specifically, the process of acquiring and preprocessing the paint removal effect image in S2 includes: converting the paint removal effect image from the BGR color space to the HSV color space; performing multi-scale Gaussian filtering on the brightness channel of the paint removal effect image to obtain a blurred image; calculating the average value of the blurred image; calculating the gamma value of each pixel based on the average value; using the gamma value to dynamically stretch the grayscale of the brightness channel to correct for uneven lighting; and then merging the processed brightness channel with the hue channel and saturation channel and converting it back to the BGR color space.

[0009] By converting the BGR to HSV color space and performing multi-scale Gaussian filtering and dynamic grayscale stretching only on the luminance channel, uneven illumination can be effectively corrected, the contrast of defect areas can be enhanced, and color information can be preserved, thereby improving the quality of image preprocessing and enhancing the training stability and detection accuracy of the defect detection model.

[0010] Specifically, the laser paint removal defect detection model described in S2 includes a backbone network, a feature fusion module, and a multi-scale output layer connected in sequence. The backbone network starts with the Focus layer as the input, and then passes through multiple convolutional layers for downsampling and progressively expands the channel dimension. Cross-stage local connection units are set between adjacent convolutional layers. The multi-scale feature maps of the paint removal effect image are extracted by combining the multiple convolutional layers with the cross-stage local connection units. At the end of the backbone network, a spatial pyramid pooling layer is set. The spatial pyramid pooling layer performs multi-scale pooling and fusion on the multi-scale feature maps to enhance the feature representation of irregular defect areas. The feature fusion module includes a feature pyramid structure and a feature aggregation structure. The feature pyramid structure is used to receive multi-scale feature maps of different scales output by the backbone network. After enlarging the deep feature map to the same resolution as the shallow feature map, the deep feature map and the shallow feature map are spliced ​​and compressed along the channel dimension and input into the cross-stage local connection unit for feature optimization to obtain a multi-scale fused feature map. The feature aggregation structure downsamples the shallow feature map in the multi-scale fused feature map through a convolutional layer, then concatenates it with the feature map of the corresponding level, and inputs it again into the cross-stage local connection unit for feature optimization to obtain a bidirectional fused feature map. The multi-scale output layer is connected to the bidirectional fused feature maps of different scales output by the feature fusion module to perform multi-scale detection and output defect detection results. The loss function of the laser paint removal defect detection model is a combination of a hard saturated noise excitation function with ReLU as the basic excitation function and a hard saturated noise excitation function with SeLU as the basic excitation function.

[0011] The backbone network employs Focus, cross-stage local connectivity, and spatial pyramid pooling to efficiently extract multi-scale features. The feature fusion module achieves bidirectional fusion of deep and shallow layers through pyramid and aggregation structures, enhancing the expressive power for defects of different sizes and irregularities, and improving the accuracy and robustness of defect detection.

[0012] Specifically, the defect detection results include category, location, and confidence level, and the category includes over-cleaned, cleaned, and not cleaned; The process of detecting the paint removal effect image based on the trained laser paint removal defect detection model to screen the experimental process parameter dataset specifically includes: If the defect detection result indicates that the defect has been cleaned and the confidence level is greater than the set threshold, then the corresponding combination of experimental process parameters is retained. If the defect detection result is that the machine was not cleaned or was over-cleaned, the corresponding experimental process parameter combination is adjusted to increase or decrease the corresponding energy input.

[0013] Defects are categorized as over-cleaned, cleaned, and not cleaned. Based on the test results and confidence levels, process parameters are screened or adjusted. Unqualified parameter combinations can be automatically eliminated, energy input can be reasonably adjusted, substrate damage or incomplete paint removal can be avoided, and the reliability and effectiveness of experimental process parameter datasets can be improved.

[0014] Specifically, S1 includes: inputting different combinations of laser paint removal process parameters into the laser paint removal simulation model for numerical simulation, obtaining laser paint removal simulation results, and selecting the laser paint removal process parameter combinations corresponding to the laser paint removal simulation results based on a preset evaluation system to obtain the simulation process parameter dataset. The preset evaluation system includes the following four indicators: average paint layer thickness, paint layer removal rate, highest substrate surface temperature, and maximum paint layer burning depth.

[0015] The pre-set evaluation system comprehensively considers four indicators: paint layer thickness, removal rate, maximum substrate temperature, and burning depth. It can select parameter combinations that balance paint removal efficiency and substrate safety from the simulation results, thereby improving the practicality and screening quality of the simulation data.

[0016] Specifically, the laser paint removal process parameter combination includes any combination of at least two process parameters: laser power, laser scanning speed, spot diameter, pulse width, and repetition frequency.

[0017] The combination of process parameters covers the key factors affecting paint removal effect, enhancing the decision-making method's adaptability to complex working conditions and its predictive comprehensiveness.

[0018] Specifically, the process parameter prediction model combination described in S6 includes a combination of a power prediction model and a speed prediction model; The power prediction model uses the standardized input features as independent variables to establish a linear regression relationship for the predicted laser power. The speed prediction model uses the standardized input features as independent variables to establish a linear regression relationship with the natural logarithm of the predicted scanning speed. The weights and biases of the power prediction model and the weights and biases of the velocity prediction model are used as the model parameter combination and uniformly encoded into the particle position vector of the particle swarm optimization algorithm. Using the sum of the squared relative prediction errors of the power prediction model and the velocity prediction model as the fitness function, the combination of model parameters that minimizes the fitness function is found through particle swarm iteration.

[0019] By combining linear regression models and uniformly optimizing weights and biases using particle swarm optimization, the global optimal model parameters can be quickly obtained by minimizing the sum of squared relative errors between the predicted process parameters and the better parameters, thereby improving the accuracy and generalization ability of joint prediction of multiple process parameters.

[0020] Specifically, the thermal properties of the paint layer include at least thermal conductivity. ,density Specific heat capacity ablation temperature Sublimation heat and thickness One of them; The thermal properties of the substrate include at least thermal conductivity. ,density Specific heat capacity and ablation temperature One of them; The derived characteristics are calculated based on the thermal properties of the paint layer and the substrate. The derived characteristics include at least the energy removed per unit area of ​​the paint layer. , substrate thermal inertia Energy-thermal inertia cross term and paint layer thermal resistance One of them, The initial ambient temperature.

[0021] The physical properties, based on the thermophysical parameters of the paint layer and substrate and the mechanism of laser paint removal, can accurately reflect the material's response to laser action, improving the model's interpretability and predictive adaptability to material changes.

[0022] Specifically, the step S4, which involves correcting the simulation process parameter dataset based on the relative parameter prediction error between the experimental process parameter dataset and the simulation process parameter dataset, includes: The model parameters of the laser paint removal simulation model are adjusted based on the relative parameter prediction error to correct the simulation process parameter dataset; or Calculate the average value of the relative parameter prediction error, and multiply the average value by the simulation process parameter dataset to correct the simulation process parameter dataset.

[0023] By adjusting the simulation model parameters or directly correcting the simulation dataset through relative parameter prediction errors, the deviation between simulation and experiment can be effectively reduced, the reliability of simulation process parameter data can be improved, and the training quality of subsequent prediction models can be enhanced.

[0024] According to one aspect of the present invention, an intelligent decision-making system for laser paint removal process parameters applying the method described in any one of the first aspects is proposed, comprising the following modules: The simulation parameter acquisition module is configured to establish a laser paint removal simulation model, input different combinations of laser paint removal process parameters corresponding to the combination of base paint layer parameters and materials into the laser paint removal simulation model for simulation, and filter to obtain a simulation process parameter dataset. The detection model building module is configured to conduct corresponding laser paint removal experiments based on the simulation process parameter dataset, collect paint removal effect images and perform preprocessing, build a laser paint removal defect detection model, and train the laser paint removal defect detection model using the preprocessed paint removal effect images. The experimental parameter acquisition module is configured to select multiple sets of laser paint removal process parameter combinations based on the simulation process parameter dataset and set corresponding process parameter combination gradient experiments, acquire the paint removal effect image to be detected corresponding to the process parameter combination gradient experiments, and detect the paint removal effect image to be detected based on the trained laser paint removal defect detection model to filter the experimental process parameter dataset. The optimal parameter selection module is configured to correct the simulation process parameter dataset based on the relative parameter prediction error between the experimental process parameter dataset and the simulation process parameter dataset, and to obtain an optimal cleaning parameter set based on the corrected simulation process parameter dataset and the experimental process parameter dataset. The feature standardization module is configured to acquire the physical correlation features of the paint layer and the substrate, calculate the mean and standard deviation of the corresponding physical correlation features based on the optimal cleaning parameter set, perform Z-score standardization on the physical correlation features, and obtain standardized input features. The physical correlation features include at least the thermal property parameters of the paint layer, the thermal property parameters of the substrate, and derived features calculated based on the thermal property parameters of the paint layer and the substrate. The prediction model building module is configured to establish a combination of process parameter prediction models, using the standardized input features as input to the combination of process parameter prediction models, and using a particle swarm optimization algorithm to find the combination of model parameters for the combination of process parameter prediction models. The parameter prediction module is configured to predict the combination of laser paint removal process parameters based on the combination of material parameters of the substrate coating layer to be treated by using the process parameters after the model parameters are determined.

[0025] The advantages of this invention are: This invention effectively solves the problems of traditional laser paint removal process parameters relying on manual experience, high trial-and-error costs, and long cycles.

[0026] First, the best process parameter dataset for simulation is selected by pre-screening the simulation model and evaluating multiple indicators, so as to reduce the investment in actual experiments. Secondly, image preprocessing and defect detection models are introduced to achieve automatic judgment of paint removal quality, ensuring that only qualified parameters are included in the experimental optimal process parameter dataset to train the defect detection model. Simulation errors are corrected through experimental data, so that simulation and experimental data are integrated and complementary, thereby improving data reliability. Finally, physical features were constructed and standardized based on the laser paint removal mechanism to enhance the model's sensitivity to differences in material thermal properties. Particle swarm optimization algorithm was used to globally optimize the combination of linear regression models, thereby achieving accurate prediction of the combination of process parameters affecting laser paint removal.

[0027] The overall method has the advantages of high automation, strong adaptability, high decision accuracy and low cost. It can quickly provide optimal laser paint removal process parameters under different combinations of paint layers and substrate materials, which can ensure paint removal efficiency and avoid over-cleaning damage to the substrate. It has important engineering application value for improving laser cleaning quality and promoting intelligent manufacturing. Attached Figure Description

[0028] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0029] Figure 1 A flowchart illustrating an intelligent decision-making method for laser paint removal process parameters according to the present invention is shown. Figure 2 A schematic diagram of the finite element model simulation structure according to the present invention is shown; Figure 3A schematic diagram of a laser paint removal experimental platform according to the present invention is shown; Figure 4 A flowchart illustrating a paint removal effect image preprocessing method according to the present invention is shown; Figure 5 A schematic diagram of a laser paint removal defect detection model according to the present invention is shown. Figure 6 The changes in accuracy and loss during the iterative process of the laser paint removal defect detection model are shown; Figure 7 The diagram shows the paint removal effect obtained by cleaning at different speed nodes; Figure 8 The diagram shows the convergence curve of the fitness function during the parameter fitting process of the PSO algorithm. Figure 9 A schematic diagram illustrating the contribution of physically relevant features to the process parameter prediction model is shown. Figure 10 A schematic diagram of an intelligent decision-making system for laser paint removal process parameters according to the present invention is shown; Figure 11 A schematic diagram of a computer system architecture suitable for implementing the embodiments of this application is shown. Detailed Implementation

[0030] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0031] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] According to one aspect of the present invention, an intelligent decision-making method for laser paint removal process parameters is proposed, such as... Figure 1 As shown, it includes the following steps: S1. Establish a laser paint removal simulation model. Input different combinations of laser paint removal process parameters corresponding to the material combination of the base paint layer into the laser paint removal simulation model for simulation. Filter to obtain the simulation process parameter dataset. S2, Based on the simulation process parameter dataset, conduct corresponding laser paint removal experiments, collect paint removal effect images and perform preprocessing, construct a laser paint removal defect detection model, and train the laser paint removal defect detection model using the preprocessed paint removal effect images. S3. Based on the simulation process parameter dataset, select multiple sets of laser paint removal process parameter combinations and set corresponding process parameter combination gradient experiments. Obtain the paint removal effect image to be detected corresponding to the process parameter combination gradient experiment. Detect the paint removal effect image to be detected based on the trained laser paint removal defect detection model to filter the experimental process parameter dataset. S4, based on the relative parameter prediction error between the experimental process parameter dataset and the simulation process parameter dataset, the simulation process parameter dataset is corrected, and a better cleaning parameter set is obtained by combining the corrected simulation process parameter dataset and the experimental process parameter dataset. S5, obtain the physical correlation features between the paint layer and the substrate, calculate the mean and standard deviation of the corresponding physical correlation features based on the optimal cleaning parameter set, perform Z-score standardization on the physical correlation features to obtain standardized input features, the physical correlation features include at least the thermal property parameters of the paint layer, the thermal property parameters of the substrate, and derived features calculated based on the thermal property parameters of the paint layer and the substrate; S6, establish a process parameter prediction model combination, use the standardized input features as the input of the process parameter prediction model combination, and use the particle swarm optimization algorithm to find the model parameter combination of the process parameter prediction model combination; S7. Using the process parameters predicted after the model parameters are determined, the laser paint removal process parameter combination of the material combination of the substrate paint layer to be treated is predicted.

[0033] Specifically, the laser paint removal process parameter combination includes any combination of at least two process parameters: laser power, laser scanning speed, spot diameter, pulse width, and repetition frequency. Considering that the laser cleaning process involves numerous parameters and there are complex nonlinear coupling relationships between them, in order to reduce simulation complexity, focus on core influencing factors, ensure the comparability of simulation results, and reduce computational load, a strategy of "fixing secondary parameters and optimizing key variables" is adopted.

[0034] In a specific embodiment, the present invention first fixes three parameters: spot diameter, pulse width, and repetition frequency, and selects laser power and laser scanning speed as core variables. In other embodiments of the present invention, two or more of the above five parameters can be arbitrarily selected as variables according to actual needs, while the remaining parameters remain fixed or use default values.

[0035] During laser cleaning, the amount of heat energy generated by the interaction between the pulsed laser and the substrate surface determines the cleaning quality and the substrate surface quality. Different laser cleaning process parameters result in different amounts of heat energy; therefore, observing the temperature field changes caused by this heat energy can provide a range for selecting the laser cleaning process. This application uses Comsol software to conduct simulation experiments on the temperature field changes during laser paint removal. Figure 2 As shown, during the paint removal simulation, the finite element model was designed as a two-layer heat conduction model. The upper layer consisted of paint layers of varying thicknesses, each 1mm × 1mm, while the lower carbon steel substrate was 1mm × 1mm × 0.5mm. To ensure computational efficiency and numerical simulation accuracy, an ultra-fine finite element mesh was used.

[0036] The purpose of finite element simulation is to reveal experimental laws that are difficult to measure, reduce the number of experiments, and lower research costs. Therefore, it is necessary to select appropriate process parameters for simulation.

[0037] In a specific embodiment, based on laboratory experimental conditions, the selection of process parameter ranges was explored for subsequent paint removal experiments and data acquisition. The relevant process parameters for laser paint removal simulation were set as follows: laser power range of 10W-250W, laser pulse width of 200ns, spot diameter of 40μm, fixed laser repetition frequency of 40kHz, and scanning speed of 500mm / s-2500mm / s. During laser cleaning, the energy distribution of the laser is not uniform, but rather acts on the paint surface in a manner similar to heat flow. Therefore, this experiment uses a Gaussian pulse function to simulate the laser.

[0038] Before conducting simulation analysis, it is necessary to establish a quantitative and qualitative evaluation system for cleaning effects. Based on different cleaning quality levels, three process states can be identified: (1) Incomplete removal of residue, with paint residue remaining on the substrate surface; (2) Thorough cleaning, with no paint residue and the substrate intact; (3) Over-removal, with the paint layer completely removed but the substrate suffering physical damage. To objectively evaluate different cleaning effects, an evaluation index system containing four parameters must be established. The specific index system is as follows: Average residual paint thickness : In the formula, R(x) represents the planned paint layer thickness distribution to be removed, W is the width dimension parameter of the paint layer material, and L is the total length of the longitudinal scanning path during laser processing. The average residual paint layer thickness is a key parameter for quantitatively evaluating the effectiveness of laser cleaning. The average residual paint layer thickness on the surface after cleaning is obtained by integrating the residual paint layer thickness distribution during the cleaning process, and then combining it with the scanning area width and scanning distance.

[0039] (2) Paint removal rate : Paint removal rate is a parameter reflecting the efficiency of laser cleaning; it is the proportion of paint removed after cleaning. Its calculation is based on the change in paint thickness before and after cleaning, specifically using the ratio of the initial paint thickness to the residual paint thickness to measure the degree of cleaning.

[0040] (3) Maximum surface temperature of substrate Tmax: The highest surface temperature of the substrate is a key quantitative indicator for judging thermal damage to the substrate, with the melting point of the substrate as the critical threshold. When Tmax is below the melting point, the substrate only experiences a temperature rise without irreversible melting damage; if the peak temperature exceeds the melting point critical value, a melting phase transition will occur on the surface of the substrate, indicating that the process parameters have exceeded the limits.

[0041] (4) Maximum ablation depth of paint layer Dmax.

[0042] The maximum ablation depth of the paint layer is an important indicator for assessing the damage to the substrate during laser cleaning. It represents the deepest part of the paint layer removed in the cleaning area and reflects the impact of laser ablation on the substrate surface during cleaning.

[0043] To overcome the problems of relying on experience and high trial-and-error costs in selecting process parameters, this application conducted numerous numerical experiments using a finite element simulation model, obtaining data on cleaning thresholds and damage boundaries under different working conditions. To facilitate access and management by the decision-making platform, the scattered simulation results need to be organized into a structured process parameter database. For common substrate and paint layer combinations in engineering, this application first established an existing material simulation parameter library. This library mainly includes three typical substrates: Q235 steel, 42CrMo steel, and 304 stainless steel, as well as three mainstream paint layers: epoxy primer, polyurethane paint, and acrylic paint. Typical material combination process parameter ranges were compiled. Table 1 shows the different power and speed windows corresponding to different paint layer thicknesses and substrate types.

[0044] Table 1. Optimal combinations of simulation process parameters for existing materials. Considering the diversity of paint material formulations and the slight fluctuations in the alloy composition of the substrate in actual engineering, this application constructs a series of virtual material combinations in the COMSOL simulation environment by individually changing a key thermophysical parameter. This library is not based on specific physical materials, but rather generates virtual working condition data. Some data is shown in Table 2. The virtual parameter library contains a large number of non-standard parameter combinations. Although these data do not represent a specific real paint, they represent the variation law of material properties within a certain range, providing the necessary training set and boundary constraints for the Particle Swarm Optimization (PSO) algorithm. This method, while keeping other physical properties unchanged, makes incremental adjustments to the target parameter within a 10% range (e.g., adjusting the thermal conductivity of the epoxy primer from 0.2 W / (m•K) to 0.22 W / (m•K)), thereby creating virtual working conditions covering a wider range of parameters.

[0045] Table 2 Optimal Combinations of Simulation Process Parameters for Virtual Materials Through simulation design using the controlled variable method, the evolution of the temperature field and ablation depth of the aforementioned virtual material under different laser process parameters was simulated. This process verified the influence mechanism of thermophysical property sensitivity on the cleaning threshold, generating multiple expanded datasets containing mapping relationships between "material properties - optimal parameters - cleaning effect". These data enable the subsequently established process parameter decision model to learn more complex nonlinear characteristics, rather than a limited number of material combinations, thus significantly enhancing the model's reasoning and fitting capabilities when faced with unknown material combinations.

[0046] In a specific embodiment, the laser paint removal experimental platform constructed in this application is as follows: Figure 3 As shown, 1 represents the image acquisition platform, 2 the laser emitter, 3 the industrial camera, 4 the auxiliary light source, and 5 the laser control PC. The industrial camera, laser emitter, auxiliary light source, USB acquisition card, signal amplification module, and PC are used to synchronously extract and process various types of data. When the laser emitter processes the sample surface, the auxiliary light source provides uniform illumination, and the industrial camera acquires real-time images of the processing. The image data is transmitted to the PC for storage and analysis. Based on the optimal combination of process parameters obtained from finite element simulation, the laser parameter range set in this experimental platform completely covers the simulation parameter range. Using the optimal simulation parameters as a benchmark, system experiments are conducted by expanding parameter combinations within their neighborhood.

[0047] Based on the analysis of the macroscopic and microscopic morphological characteristics of samples under various paint removal conditions, the following judgment criteria can be summarized: When the metal surface exhibits a metallic luster, the microscopic morphology shows relatively uniform and smooth characteristics, and the carbon and oxygen content remains at a low level (approximately 4%), while the roughness Sa value is approximately 0.6 μm, it indicates that the paint on the sample has been completely removed, achieving the expected effect. When the oxygen content suddenly rises to 11%, and the surface roughness also increases accordingly, it indicates that the sample has been over-cleaned, and the surface has been damaged.

[0048] This laser paint removal experiment acquired over 1200 images of the cleaning results at a resolution of 1980×1440 pixels. After removing images with substrate defects or obvious similarities, the remaining images were classified according to the aforementioned criteria and then uniformly cropped to 768×765 pixels. A total of 1000 images were ultimately obtained for training the subsequent paint removal effect detection algorithm, covering three categories: uncleaned, cleaned, and over-cleaned.

[0049] During laser paint removal, uneven illumination distribution in the acquired image can result from laser-irradiated areas or reflections from the substrate metal after scanning. Cleaned samples exhibit poor visual quality, while uncleaned samples are overexposed, leading to significant information loss and hindering the learning process of subsequent detection algorithms. Therefore, surface illumination intensity correction of the acquired paint removal images to improve image quality is of considerable practical significance.

[0050] like Figure 4 Compared to adjusting the illumination components of the RGB three channels of an image F(x,y), the HSV color space, composed of three independent components—hue (H), saturation (S), and lightness (V)—is closer to the correction requirements. Multi-scale Gaussian filtering is first applied to the luminance component to preserve subtle textures and edge information while smoothing larger-scale paint layer structures and noise, improving the quality of the acquired illumination component I(x,y). The average value of the average blurred image is calculated, and the gamma γ value of each pixel is calculated based on this average value for dynamic grayscale stretching. The processed luminance channel replaces the original image's luminance channel to obtain the corrected luminance component V'. Hue and saturation are not subject to unnecessary processing. After merging, a new paint-removed image F'(x,y) in the BGR color space is generated. This correction results in adaptive illumination correction of the overall luminance characteristics of the image, a more uniform luminance distribution, better reflection of the image's visual effect, and avoidance of color distortion.

[0051] The laser paint removal defect detection model designed in this application, based on a convolutional neural network, is as follows: Figure 5 As shown, by combining convolutional neural networks, efficient detection and real-time evaluation of the laser paint removal process are achieved. The research framework consists of three parts: a backbone network, a feature fusion module (Neck), and a multi-scale output, with the following structure: Specifically, in S5... The backbone network starts with a Focus layer as input and progressively expands the channel dimension (64→128→256→512→1024) through 3×3 convolutional kernels and stride 2 operations. Combined with the C3 module (i.e., cross-stage local connection unit) for cross-layer connections, it achieves multi-scale feature extraction. The initial stage (64→128 channels) focuses on the details of paint layer edges and heat-affected zones. The mid-to-deep layers (256→512 channels) enhance the local feature reuse capability through residual connections and grouped convolutions in the C3 module. Finally, multi-scale grid features (4×4, 8×8, 16×16) are fused through Spatial Pyramid Pooling (SPP) layers to enhance the model's accuracy in detecting irregular defects. Through progressive abstraction, from high-resolution details to global semantic information, the backbone network provides multi-granular feature representations for subsequent modules.

[0052] The feature fusion module achieves multi-scale feature fusion through cross-layer concatenation and upsampling operations. The feature fusion module includes a feature pyramid structure and a feature aggregation structure. The feature pyramid structure is used to receive multi-scale feature maps of different scales output by the backbone network. Specifically, the 1024-channel feature map at the end of the backbone network is compressed to 512 channels by a 1×1 convolution, then upsampled using bilinear interpolation to achieve the same resolution as the shallow 512-channel feature map. The two are then concatenated along the channel dimension (Concat, resulting in 1024 channels), and the channel redundancy is compressed to 512 channels using a 1×1 convolution. This is then input into the C3 module to optimize weight allocation. Subsequently, the 512-channel feature map is compressed to 256 channels again by a 1×1 convolution, and upsampled using bilinear interpolation to restore it to a higher resolution. This is then concatenated with the shallow 256-channel feature map (resulting in 512 channels), compressed to 256 channels using a 1×1 convolution, and input into the C3 module. By dynamically integrating shallow details and deep semantics, background noise interference is effectively suppressed, the spatial correlation of defective regions is strengthened, and a multi-scale fused feature map is obtained.

[0053] The feature aggregation structure receives shallow features from the multi-scale fused feature map, downsamples them through a 3×3 convolutional layer with a stride of 2, and then concatenates them with the feature map of the corresponding level (from the middle layer of the feature pyramid structure or the backbone network) along the channel dimension. It is then input into the cross-stage local connection unit (C3 module) for feature optimization, passing the high-resolution location information of the shallow layer to the deep layer step by step to obtain a bidirectional fused feature map, thereby improving the localization accuracy of the bounding box of the large target.

[0054] The multi-scale output layer is connected to the 256-channel, 512-channel, and 1024-channel bidirectional fused feature maps output by the feature fusion module, respectively. It further aggregates defect information using 3×3 convolutions and the C3 module to construct a feature pyramid for multi-scale detection. The output layer maps feature channels to detection parameters using 1×1 convolutions. The defect detection result includes category, location, and confidence level, where the category includes over-cleaned, cleaned, and uncleaned. The output format is as follows: Where na is the number of anchor boxes and nc is the number of defect categories. The output layer optimizes boundary localization accuracy through a dynamic threshold segmentation algorithm, supporting sub-pixel-level edge residue recognition.

[0055] The entire framework reduces the number of parameters through lightweight design, enabling rapid real-time inference while ensuring detection accuracy, providing efficient and reliable technical support for online quality monitoring of laser cleaning processes.

[0056] Choosing the appropriate activation function in laser cleaning image segmentation tasks has a significant impact on network performance. Traditional ReLU and NRReLU activation functions are effective at handling nonlinear features, but their hard truncation leads to the loss of some gradient information, easily causing problems such as gradient vanishing or underfitting. In contrast, using noisy activation functions NHRReLU and NHSeLU can introduce a degree of randomness, which helps to address these issues.

[0057] Wherein, NHReLU is the hard-saturated noise excitation function with ReLU as the basic excitation function h(x), i.e. NHSeLU is a hard-saturated noise excitation function with SeLU as the fundamental excitation function h(x), i.e. In the formula, It is a hyperparameter; It is an activation function. It is noise sampled from a normal distribution.

[0058] The improvements not only enhance the model's adaptability to noisy samples but also improve its ability to identify segmentation boundaries, thereby improving network performance and the accuracy of cleaning effect detection. Therefore, replacing traditional activation functions with NHRReLU and NHSeLU improves the algorithm's reliability and accuracy. By using the noise activation functions NHReLU and NHSeLU, output prediction can be achieved at two scales, rather than only at the final upsampling layer, effectively addressing the issue of target region size variations in laser paint removal images. Therefore, the laser paint removal defect detection model in this application not only enhances the model's adaptability to samples of different sizes but also improves the stability of feature extraction.

[0059] After experimental testing and optimization, the following settings were determined: Adam as the optimizer, batch size of 64, and learning rate of 1×10⁻⁶. -4 The model was iterated 100 times, with a learning rate in the range of [20, 80] and a gamma parameter of 0.1.

[0060] Figure 6 (a) and (b) in the figure show the changes in accuracy and loss during the model iteration process, respectively. The model reached its maximum accuracy of 99.56% in the 87th iteration, and the loss curve also decreased slowly, which proves that the laser paint removal defect detection model designed in this application has ideal fitting ability. The test set also has an accuracy of 99.35%, which proves that the model can fit the unknown data distribution and has ideal generalization ability, which fully demonstrates the effectiveness of the model in laser paint removal detection.

[0061] To verify the accuracy and reliability of the process parameters obtained from finite element simulation and to study the deviation between the simulation model and the actual physical field, this application designed a systematic physical experiment verification scheme. By comparing the theoretical optimal value of the simulation with the actual cleaning effect, the most suitable process window for the steel substrate-paint layer system was selected, and a high-fidelity experimental parameter database was established.

[0062] In a specific embodiment, polyurethane paint (30μm) and Q235 steel were used as a typical combination. The experiment employed a controlled variable method, using the theoretically optimal parameters obtained from simulation experiments as the center point, and setting a gradient variation range around it. When the laser power was 105W, the ablation depth of the paint layer was closest to the paint layer thickness (30μm) at a scanning speed of 500mm / s, and the substrate temperature did not reach the damage threshold. Therefore, this experiment selected 105W as a fixed power value and designed a scanning speed gradient experiment, selecting three key speed nodes: 100mm / s, 500mm / s, and 1000mm / s. This design can cover all working conditions from low speed and high overlap to high speed and low energy. The cleaning effect obtained at the three speed nodes is as follows: Figure 7 As shown.

[0063] When the scanning speed remains constant, the ablation depth is optimal at a power of 105W. Therefore, this experiment selected a fixed scanning speed of 500mm / s and designed a laser power gradient experiment at this speed, choosing four power nodes: 95W, 100W, 105W, and 110W. The purpose of this design is to accurately find the inflection point of the cleaning effect: the low-power segment is used to verify whether the paint layer can be completely removed; the theoretically optimal segment is used to verify the actual cleaning quality; and the high-power segment is used to verify whether the substrate will be damaged due to over-ablation or expansion of the heat-affected zone as predicted by simulation.

[0064] An industrial camera captures real-time images of the surface of the cleaning area, and the resulting image data stream is fed into a visual inspection model. Using the previously obtained model weights, forward inference is performed on each captured image frame. The model outputs three evaluation results: cleaning state classification, defect region location, and confidence level. The effectiveness of the set of parameters is determined based on the model's output.

[0065] If the model determines that the material has been cleaned and the confidence level is greater than the set threshold (e.g., 0.85), then the parameter combination is considered to have reached the optimal cleaning state expected by the simulation. If the model determines that the material has not been cleaned, then the energy input of this parameter combination is insufficient and the power or speed needs to be adjusted. If the model determines that the material has been over-cleaned (substrate damage), then the parameter combination has exceeded the material's damage threshold.

[0066] Through the above experimental scheme, this application can obtain the actual cleaning process parameters for Q235 steel and polyurethane coating. Real-time feedback from the image monitoring model is used to establish an experimental dataset containing process parameters, cleaning images, and model evaluation results. This dataset is cross-compared with previous simulated process parameter data. The degree of agreement between the model evaluation results and simulation predictions is used to correct the thermophysical parameters in the simulation model, thereby achieving a closed-loop improvement from "theoretical simulation" to "experimental verification" and selecting highly reliable training samples.

[0067] In the process of establishing a laser paint removal process parameter library, the finite element simulation model can efficiently simulate paint removal data, significantly reducing trial and error costs. However, due to the inevitable differences between the theoretical model and the actual physical process, error correction must be performed on the optimal process parameters predicted by simulation and the optimal parameters obtained from actual experiments.

[0068] This application uses Q235 steel, 42CrMo steel, and 304 stainless steel as substrate materials to compare process parameters for laser paint removal of epoxy primer, polyurethane paint, and acrylic paint (thickness range of 20μm~40μm). The optimal laser power and optimal scanning speed measured experimentally are used as benchmark values, and the accuracy of the simulation data is evaluated by calculating the relative parameter prediction error. The calculation formula is as follows: Relative parameter prediction error = ×100%.

[0069] Calculations and statistics were performed on the 18 sets of comparative data extracted. The results showed that the laser power obtained through simulation was relatively good, with a relative error of 6.31%. The power prediction error was within 10% for most operating conditions. The largest deviation occurred when removing 30μm polyurethane paint from the surface of 42CrMo steel. The optimal simulated power prediction was 205W, while the optimal experimental power was 155W, resulting in a relative error of 32.26%. The average relative error for the best scanning speed was 5.09%. After 18 sets of comparisons, the simulated and experimental speeds were the same in 15 operating conditions, with an error of 0%. A few operating conditions with significant deviations mainly occurred when removing a 30μm thick paint layer. When removing 30μm polyurethane paint from the surface of 304 stainless steel, the recommended simulation speed was 1000mm / s, while the optimal experimental speed was 1500mm / s, with an error of 33.33%.

[0070] In one embodiment, the average power relative parameter prediction error of 6.31% and the average scanning speed relative parameter prediction error of 5.09% obtained from the error analysis of this application are used to correct the simulated process parameter database. Due to limitations in experimental sample conditions, virtual materials cannot be tested in physical experiments; therefore, an equivalent and optimal combination of process parameters is obtained through error conversion.

[0071] In one embodiment, the model parameters of the laser paint removal simulation model are adjusted based on the relative parameter prediction error to correct the simulation process parameter dataset.

[0072] The experimentally measured process parameter combinations and the corrected virtual simulation process parameter combinations are summarized to obtain a database of optimal process parameters.

[0073] This application obtained recommended process parameters for different substrates and paint layer systems through simulation and experiments. However, in engineering applications, it is impossible to conduct complete experiments for every material combination. To enable the parameter library to infer unknown combinations, this application establishes a gray-box parameter fitting method based on the Particle Swarm Optimization (PSO) algorithm. This method does not directly use material names for empirical lookup, but instead transforms the thermophysical parameters of the substrate and paint layer into calculable features, and then uses PSO to optimize the predicted model parameters, thereby realizing the mapping from material properties to laser power and scanning speed.

[0074] Specifically, the process parameter prediction model combination described in S6 includes a combination of a power prediction model and a speed prediction model; The power prediction model uses the standardized input features as independent variables to establish a linear regression relationship for the predicted laser power. The speed prediction model uses the standardized input features as independent variables to establish a linear regression relationship with the natural logarithm of the predicted scanning speed. The weights and biases of the power prediction model and the weights and biases of the velocity prediction model are used as the model parameter combination and uniformly encoded into the particle position vector of the particle swarm optimization algorithm. Using the sum of the squared relative prediction errors of the power prediction model and the velocity prediction model as the fitness function, the combination of model parameters that minimizes the fitness function is found through particle swarm iteration.

[0075] In a specific embodiment, this application employs a gray box modeling approach: first, physical characteristics are constructed based on heat conduction and thermal ablation theories; then, PSO is used to search for the parameter combination that minimizes the prediction error in the model parameter space. Let the i-th sample be determined by the substrate material, paint layer material, and paint layer thickness, and its actual recommended process parameters be P. i and V i The goal of parameter fitting is to learn a mapping relationship on an existing sample set, enabling the model to predict the corresponding P and V based on new material thermophysical parameters. This problem can be represented as: x i →(P i V i ), where x iThe input feature vector consists of the material's thermal properties and physical derivatives. Due to the small size of the training data, directly using high-capacity black-box models such as deep neural networks can easily lead to overfitting and make it difficult to explain the relationship between the prediction results and the laser paint removal mechanism.

[0076] In other embodiments of the present invention, the process parameter prediction model can be constructed as a high-capacity black-box model such as a deep neural network, with the corresponding network parameters as particle position vectors, according to actual needs.

[0077] This modeling approach has two advantages over traditional empirical lookup methods. First, the model input consists of continuous thermal property parameters. When new paint formulations or substrate materials appear in the database, as long as the corresponding parameters such as thermal conductivity, density, specific heat capacity, and thickness are available, extrapolation predictions can be made using the model. Second, the intermediate features of the model have clear physical meanings, and the prediction results can be interpreted from the perspectives of paint layer energy removal, substrate heat dissipation capacity, and paint layer thermal resistance, facilitating subsequent experimental corrections and engineering applications. Therefore, the algorithm established in this application is not simply aimed at obtaining lower errors on existing samples, but rather provides a scalable intelligent decision-making mechanism. As the number of experimental and simulation samples in the database continues to increase, the model parameters can be refitted, and the recommendation capability of the parameter library can be updated synchronously.

[0078] The laser paint removal process is essentially the result of the combined effects of laser energy input, paint layer endothermic ablation, and substrate heat dissipation response. In order to make the fitting model reflect this physical process, this application does not directly use the material name as the model input, but extracts thermophysical parameters from both the paint layer and the substrate.

[0079] Specifically, the physical related features mentioned in S5 include at least the thermal properties of the paint layer, the thermal properties of the substrate, and derived features calculated based on the thermal properties of the paint layer and the substrate; The thermal properties of the paint layer include at least thermal conductivity. ,density Specific heat capacity ablation temperature Sublimation heat and thickness One of them; The thermal properties of the substrate include at least thermal conductivity. ,density Specific heat capacity and ablation temperature One of them; The derived characteristics are calculated based on the thermal properties of the paint layer and the substrate. The derived characteristics include at least the energy removed per unit area of ​​the paint layer. , substrate thermal inertia Energy-thermal inertia cross term and paint layer thermal resistance One of them, The initial ambient temperature.

[0080] Through the above feature construction, the model input is no longer an isolated material parameter, but a gray box feature system that includes the difficulty of paint removal, the thermal response of the substrate, and the resistance to heat transfer.

[0081] The energy E for paint removal per unit area is the core variable in the feature system of this application. It combines paint density, thickness, specific heat capacity, ablation temperature, and heat of sublimation into a single energy scale, reflecting the minimum heat input level required to complete paint removal. Generally speaking, The larger the value, the more energy the paint layer per unit area needs to absorb, and the laser power or energy input recommended by the model will be increased accordingly.

[0082] Base thermal inertia This describes the diffusion and buffering capacity of heat after it enters the substrate. For substrates with high thermal conductivity and volumetric heat capacity, the heat input by the laser diffuses more easily from the paint layer interface into the substrate interior, weakening the local temperature rise process of the paint layer. Therefore, considering only the thermal properties of the paint layer itself is insufficient to determine the process parameters; the thermal response of the substrate also needs to be incorporated into the model.

[0083] Energy-thermal inertia cross term This study further characterizes the coupling effect between the difficulty of paint removal and the heat dissipation capacity of the substrate. The same paint layer may require different power and scanning speeds on different substrates, and different paint layers on the same substrate will exhibit different cleaning windows. By constructing cross terms, the model can describe some nonlinear coupling relationships while maintaining a relatively simple expression. Paint layer thermal resistance... This reflects the heat transfer resistance effect jointly determined by the paint layer thickness and thermal conductivity. When the thermal resistance is high, the accumulation and diffusion of heat within the paint layer will change, thus affecting the ablation threshold and effective cleaning speed. These physically related characteristics together constitute the input basis of the PSO gray box fitting model of this application.

[0084] After feature construction is complete, the input features are Z-score standardized using the mean and standard deviation of the training set to obtain the standardized features Z. The purpose of standardization is to eliminate the differences in dimensions and orders of magnitude between different physical quantities, so that subsequent PSO searches within a relatively balanced parameter space.

[0085] In one embodiment, the power parameter prediction model uses a linear expression: ,in For power model weights, This serves as a bias for the power model. Since scan speed is a positive variable and engineering evaluation focuses more on relative error, this application models the speed using a log-linear model: ,in For the velocity model weights, This is used to bias the velocity model. This form ensures that the predicted velocity is positive, while also allowing the model to better describe the variation of velocity on a scale.

[0086] The key to PSO fitting lies in transforming model parameters into particle positions and defining a fitness function that reflects the process recommendation error. In this application, the parameters of the power parameter prediction model and the velocity parameter prediction model are uniformly denoted as: .

[0087] If the standardized input feature dimension is d, then the power parameter prediction model contains d weights and 1 bias, and the velocity parameter prediction model also contains d weights and 1 bias. Therefore, the dimension of the PSO particle is 2(d+1). The particle's position represents a complete set of model parameters, and the particle's quality is determined by the prediction error of this set of parameters on the training samples. Since the numerical ranges of laser power and scanning velocity are different, if ordinary squared error is used directly, the scanning velocity term will dominate the objective function due to its larger value. To ensure that the two output variables have the same proportional meaning during the optimization process, this application uses the relative error sum of squares as the fitness function and adds an L2 regularization term to limit the parameters from being too large. .

[0088] Where P and V are the actual power and the actual scan speed, respectively. and These are the predicted values ​​from the corresponding process parameter prediction models, where λ is the regularization coefficient and N is the sample size. This objective function considers power error, speed error, and model complexity simultaneously, preventing the model from fitting only one output variable well while neglecting another.

[0089] Under this objective function, the task of PSO is to find the parameter vector that minimizes the loss in the model parameter space. Unlike directly using laser power and scanning speed as particle dimensions, the particles in this application represent predicted model parameters. Therefore, after fitting, the model can automatically calculate the recommended power and scanning speed based on the new material thermophysical property input.

[0090] From the perspective of particle encoding, the first half of a particle corresponds to the power prediction model parameters, and the second half corresponds to the velocity prediction model parameters. Each particle position update changes the weights of both the power and velocity models simultaneously; therefore, the fitness function must calculate the prediction errors for both P and V. This avoids the model improving power prediction at the expense of velocity prediction, and also avoids deviating from power recommendations by solely pursuing velocity accuracy.

[0091] This application uses relative error, rather than absolute error, as the main component of the objective function because power and scan speed differ significantly in numerical magnitude. Power is typically tens to hundreds of watts, while scan speed is typically hundreds to thousands of millimeters per second. If the sum of squared absolute errors were used, the speed term would numerically dominate, weakening the power fit. The relative error form unifies both to a percentage scale, which better aligns with the actual error evaluation methods recommended in process specifications. The regularization term is used to limit the magnitude of model parameters, preventing the particle swarm optimization from searching for extreme parameters that overly rely on a few features on small sample data.

[0092] For fitting tasks with a limited number of samples and high feature dimensions, such as those described in this application, regularization terms help improve model stability and enable it to maintain good generalization ability on test samples.

[0093] like Figure 8 As shown, the fitness function of PSO decreases slowly in the early stages of iteration, indicating that the particle swarm is still in a large-scale global search phase. In the middle and later stages, some particles gradually approach the optimal parameter region, and the fitness function decreases significantly, eventually stabilizing. Although the convergence trajectories of multiple searches differ, they all show an overall downward trend, indicating that the constructed fitness function can effectively guide the particle swarm to move towards the model parameter region with smaller errors.

[0094] As shown in Table 3, this application uses existing experimental and simulated process parameter combinations as training data to test the obtained process parameter prediction model under different substrate materials. MAPE is used to measure the average deviation of the predicted values ​​from the actual values. The Q235, 42CrMo, and 304SS substrate samples all maintained low prediction errors. Overall, the average relative errors of power and speed are around 10% or less, which meets the accuracy requirements for preliminary intelligent recommendation of the parameter library. The process parameter prediction model established using the PSO algorithm can quickly and accurately obtain the optimal combination of process parameters, completely changing the traditional trial-and-error approach and realizing intelligent decision-making for laser paint removal process parameters on steel surfaces.

[0095] Table 3 Prediction errors of test samples with different substrate materials Figure 9 The results show that the energy required to remove paint layer per unit area, paint layer thickness, energy-thermal inertia cross term, and paint layer thermal resistance all contribute significantly to both the power and velocity models. This result is consistent with the laser paint removal mechanism: the energy required for paint layer removal determines the cleaning threshold, substrate thermal inertia affects interface temperature accumulation, and paint layer thermal resistance affects the heat distribution within the paint layer. Therefore, the physical features constructed in this application not only improve the fitting accuracy but also enhance the interpretability of the model results.

[0096] like Figure 10 As shown, according to one aspect of the present invention, an intelligent decision-making system for laser paint removal process parameters applying the method described in any one of the first aspects is proposed, comprising the following modules: The simulation parameter acquisition module 101 is configured to establish a laser paint removal simulation model, input different laser paint removal process parameter combinations corresponding to the substrate paint layer parameter material combination into the laser paint removal simulation model for simulation, and filter to obtain the simulation process parameter dataset. The detection model construction module 102 is configured to conduct corresponding laser paint removal experiments based on the simulation process parameter dataset, collect paint removal effect images and perform preprocessing, construct a laser paint removal defect detection model, and train the laser paint removal defect detection model using the preprocessed paint removal effect images. The experimental parameter acquisition module 103 is configured to select multiple sets of laser paint removal process parameter combinations based on the simulation process parameter dataset and set corresponding process parameter combination gradient experiments, acquire the paint removal effect image to be detected corresponding to the process parameter combination gradient experiments, and detect the paint removal effect image to be detected based on the trained laser paint removal defect detection model to filter the experimental process parameter dataset. The optimal parameter screening module 104 is configured to correct the simulation process parameter dataset based on the relative parameter prediction error between the experimental process parameter dataset and the simulation process parameter dataset, and to obtain an optimal cleaning parameter set based on the corrected simulation process parameter dataset and the experimental process parameter dataset. The feature standardization module 105 is configured to acquire the physical correlation features between the paint layer and the substrate, calculate the mean and standard deviation of the corresponding physical correlation features based on the optimal cleaning parameter set, perform Z-score standardization on the physical correlation features, and obtain standardized input features. The physical correlation features include at least the thermal property parameters of the paint layer, the thermal property parameters of the substrate, and derived features calculated based on the thermal property parameters of the paint layer and the substrate. The prediction model building module 106 is configured to establish a combination of process parameter prediction models, using the standardized input features as input to the combination of process parameter prediction models, and using a particle swarm optimization algorithm to find the combination of model parameters for the combination of process parameter prediction models. The parameter prediction module 107 is configured to predict the combination of laser paint removal process parameters by using the process parameters after the model parameters are determined.

[0097] The following is for reference. Figure 11 It shows a schematic diagram of the structure of a computer system 1100 suitable for implementing electronic devices according to embodiments of the present application. Figure 11 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0098] like Figure 11 As shown, the computer system 1100 includes a central processing unit (CPU) 1101, which performs various appropriate actions and processes based on programs stored in read-only memory (ROM) 1102 or programs loaded from storage portion 1108 into random access memory (RAM) 1103. The RAM 1103 also stores various programs and data required for the operation of the system 1100. The CPU 1101, ROM 1102, and RAM 1103 are interconnected via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0099] The following components are connected to I / O interface 1105: an input section 1106 including a keyboard, mouse, etc.; an output section 1107 including a liquid crystal display (LCD) and speakers, etc.; a storage section 1108 including a hard disk, etc.; and a communication section 1109 including a network interface card such as a LAN card and a modem, etc. The communication section 1109 performs communication processing via a network such as the Internet. A drive 1110 is also connected to I / O interface 1105 as needed. Removable media 1111, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1110 as needed so that computer programs read from them can be installed into storage section 1108 as needed.

[0100] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts are implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program is downloaded and installed from a network via communication section 1109, and / or installed from removable medium 1111. When the computer program is executed by central processing unit (CPU) 1101, it performs the functions defined in the methods of this application.

[0101] It should be noted that the computer-readable storage medium of this application is a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium is, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium is any tangible medium that contains or stores a program used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium includes a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals take various forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium or any computer-readable storage medium other than a computer-readable storage medium may transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wireline, optical fiber, RF, etc., or any suitable combination thereof.

[0102] Computer program code for performing the operations of this application is written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code executes entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer is connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or connected to an external computer (e.g., via the Internet using an Internet service provider).

[0103] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram represents a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually execute substantially in parallel, and they may sometimes execute in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, is implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0104] The modules described in the embodiments of this application are implemented in software or hardware.

[0105] In another aspect, this application also provides a computer-readable storage medium, which is included in the electronic device described in the above embodiments; or it exists independently and is not assembled into the electronic device. The computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: S1, establish a laser paint removal simulation model, input different laser paint removal process parameter combinations corresponding to the substrate paint layer parameter material combination into the laser paint removal simulation model for simulation, and filter to obtain a simulation process parameter dataset; S2, conduct corresponding laser paint removal experiments based on the simulation process parameter dataset, collect paint removal effect images and perform preprocessing, construct a laser paint removal defect detection model, and train the laser paint removal defect detection model using the preprocessed paint removal effect images; S3, select multiple sets of laser paint removal process parameter combinations based on the simulation process parameter dataset and set corresponding process parameter combination gradient experiments, obtain the paint removal effect images to be detected corresponding to the process parameter combination gradient experiments, and detect the paint removal effect images to be detected based on the trained laser paint removal defect detection model to filter the experimental process parameter dataset; S4, based on the experimental process parameter dataset and The relative parameter prediction error of the simulation process parameter dataset is corrected. Based on the corrected simulation process parameter dataset and the experimental process parameter dataset, a better cleaning parameter set is obtained; S5, the physical correlation features of the paint layer and the substrate are obtained. Based on the better cleaning parameter set, the mean and standard deviation of the corresponding physical correlation features are calculated, and the physical correlation features are Z-score standardized to obtain standardized input features. The physical correlation features include at least the thermal properties of the paint layer, the thermal properties of the substrate, and derived features calculated based on the thermal properties of the paint layer and the substrate; S6, a process parameter prediction model combination is established. The standardized input features are used as the input of the process parameter prediction model combination, and the model parameter combination of the process parameter prediction model combination is found by the particle swarm optimization algorithm; S7, using the process parameter prediction model combination after determining the model parameters, the laser paint removal process parameter combination of the substrate paint layer parameter material combination to be treated is predicted.

[0106] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A smart decision-making method for laser paint removal process parameters, characterized in that, Includes the following steps: S1. Establish a laser paint removal simulation model. Input different combinations of laser paint removal process parameters corresponding to the material combination of the base paint layer into the laser paint removal simulation model for simulation. Filter to obtain the simulation process parameter dataset. S2, Based on the simulation process parameter dataset, conduct corresponding laser paint removal experiments, collect paint removal effect images and perform preprocessing, construct a laser paint removal defect detection model, and train the laser paint removal defect detection model using the preprocessed paint removal effect images. S3. Based on the simulation process parameter dataset, select multiple sets of laser paint removal process parameter combinations and set corresponding process parameter combination gradient experiments. Obtain the paint removal effect image to be detected corresponding to the process parameter combination gradient experiment. Detect the paint removal effect image to be detected based on the trained laser paint removal defect detection model to filter the experimental process parameter dataset. S4, based on the relative parameter prediction error between the experimental process parameter dataset and the simulation process parameter dataset, the simulation process parameter dataset is corrected, and a better cleaning parameter set is obtained by combining the corrected simulation process parameter dataset and the experimental process parameter dataset. S5, obtain the physical correlation features between the paint layer and the substrate, calculate the mean and standard deviation of the corresponding physical correlation features based on the optimal cleaning parameter set, perform Z-score standardization on the physical correlation features to obtain standardized input features, the physical correlation features include at least the thermal property parameters of the paint layer, the thermal property parameters of the substrate, and derived features calculated based on the thermal property parameters of the paint layer and the substrate; S6, establish a process parameter prediction model combination, use the standardized input features as the input of the process parameter prediction model combination, and use the particle swarm optimization algorithm to find the model parameter combination of the process parameter prediction model combination; S7. Using the process parameters predicted after the model parameters are determined, the laser paint removal process parameter combination of the material combination of the substrate paint layer to be treated is predicted.

2. The intelligent decision-making method for laser paint removal process parameters according to claim 1, characterized in that, The acquisition and preprocessing of the paint removal effect image described in S2 specifically includes: converting the paint removal effect image from the BGR color space to the HSV color space; performing multi-scale Gaussian filtering on the brightness channel of the paint removal effect image to obtain a blurred image; calculating the average value of the blurred image; calculating the gamma value of each pixel based on the average value; using the gamma value to dynamically stretch the grayscale of the brightness channel to correct for uneven lighting; and then merging the processed brightness channel with the hue channel and saturation channel and converting it back to the BGR color space.

3. The intelligent decision-making method for laser paint removal process parameters according to claim 1, characterized in that, The laser paint removal defect detection model described in S2 specifically includes a backbone network, a feature fusion module, and a multi-scale output layer connected in sequence. The backbone network starts with the Focus layer as the input, and then passes through multiple convolutional layers for downsampling and progressively expands the channel dimension. Cross-stage local connection units are set between adjacent convolutional layers. The multi-scale feature maps of the paint removal effect image are extracted by combining the multiple convolutional layers with the cross-stage local connection units. At the end of the backbone network, a spatial pyramid pooling layer is set. The spatial pyramid pooling layer performs multi-scale pooling and fusion on the multi-scale feature maps to enhance the feature representation of irregular defect areas. The feature fusion module includes a feature pyramid structure and a feature aggregation structure; The feature pyramid structure is used to receive multi-scale feature maps of different scales output by the backbone network. After enlarging the deep feature map to the same resolution as the shallow feature map, the deep feature map and the shallow feature map are spliced ​​and compressed along the channel dimension and input into the cross-stage local connection unit for feature optimization to obtain a multi-scale fused feature map. The feature aggregation structure downsamples the shallow feature map in the multi-scale fused feature map through a convolutional layer, then concatenates it with the feature map of the corresponding level, and inputs it again into the cross-stage local connection unit for feature optimization to obtain a bidirectional fused feature map. The multi-scale output layer is connected to the bidirectional fused feature maps of different scales output by the feature fusion module to perform multi-scale detection and output defect detection results. The loss function of the laser paint removal defect detection model is a combination of a hard saturated noise excitation function with ReLU as the basic excitation function and a hard saturated noise excitation function with SeLU as the basic excitation function.

4. The intelligent decision-making method for laser paint removal process parameters according to claim 3, characterized in that, The defect detection results include category, location, and confidence level, and the category includes over-cleaned, cleaned, and not cleaned. The process of detecting the paint removal effect image based on the trained laser paint removal defect detection model to screen the experimental process parameter dataset specifically includes: If the defect detection result indicates that the defect has been cleaned and the confidence level is greater than the set threshold, then the corresponding combination of experimental process parameters is retained. If the defect detection result is that the machine was not cleaned or was over-cleaned, the corresponding experimental process parameter combination is adjusted to increase or decrease the corresponding energy input.

5. The intelligent decision-making method for laser paint removal process parameters according to claim 1, characterized in that, S1 specifically includes: inputting different combinations of laser paint removal process parameters into the laser paint removal simulation model for numerical simulation, obtaining laser paint removal simulation results, and selecting the laser paint removal process parameter combinations corresponding to the laser paint removal simulation results based on a preset evaluation system to obtain the simulation process parameter dataset. The preset evaluation system includes the following four indicators: average paint layer thickness, paint layer removal rate, highest substrate surface temperature, and maximum paint layer burning depth.

6. The intelligent decision-making method for laser paint removal process parameters according to claim 1, characterized in that, The laser paint removal process parameter combination includes any combination of at least two process parameters: laser power, laser scanning speed, spot diameter, pulse width, and repetition frequency.

7. The intelligent decision-making method for laser paint removal process parameters according to claim 6, characterized in that, The process parameter prediction model combination described in S6 includes a combination of a power prediction model and a speed prediction model; The power prediction model uses the standardized input features as independent variables to establish a linear regression relationship for the predicted laser power. The speed prediction model uses the standardized input features as independent variables to establish a linear regression relationship with the natural logarithm of the predicted scanning speed. The weights and biases of the power prediction model and the weights and biases of the velocity prediction model are used as the model parameter combination and uniformly encoded into the particle position vector of the particle swarm optimization algorithm. Using the sum of the squared relative prediction errors of the power prediction model and the velocity prediction model as the fitness function, the combination of model parameters that minimizes the fitness function is found through particle swarm iteration.

8. The intelligent decision-making method for laser paint removal process parameters according to claim 1, characterized in that, The thermophysical parameters of the paint layer include at least thermal conductivity. ,density Specific heat capacity ablation temperature Sublimation heat and thickness One of them; The thermal properties of the substrate include at least thermal conductivity. ,density Specific heat capacity and ablation temperature One of them; The derived characteristics are calculated based on the thermal properties of the paint layer and the substrate. The derived characteristics include at least the energy removed per unit area of ​​the paint layer. , substrate thermal inertia Energy-thermal inertia cross term and paint layer thermal resistance One of them, The initial ambient temperature.

9. The intelligent decision-making method for laser paint removal process parameters according to claim 1, characterized in that, The correction of the simulated process parameter dataset based on the relative parameter prediction error between the experimental process parameter dataset and the simulated process parameter dataset, as described in S4, specifically includes: The model parameters of the laser paint removal simulation model are adjusted based on the relative parameter prediction error to correct the simulation process parameter dataset; or Calculate the average value of the relative parameter prediction error, and multiply the average value by the simulation process parameter dataset to correct the simulation process parameter dataset.

10. An intelligent decision-making system for laser paint removal process parameters, characterized in that, include The following modules: The simulation parameter acquisition module is configured to establish a laser paint removal simulation model, input different combinations of laser paint removal process parameters corresponding to the combination of base paint layer parameters and materials into the laser paint removal simulation model for simulation, and filter to obtain a simulation process parameter dataset. The detection model building module is configured to conduct corresponding laser paint removal experiments based on the simulation process parameter dataset, collect paint removal effect images and perform preprocessing, build a laser paint removal defect detection model, and train the laser paint removal defect detection model using the preprocessed paint removal effect images. The experimental parameter acquisition module is configured to select multiple sets of laser paint removal process parameter combinations based on the simulation process parameter dataset and set corresponding process parameter combination gradient experiments, acquire the paint removal effect image to be detected corresponding to the process parameter combination gradient experiments, and detect the paint removal effect image to be detected based on the trained laser paint removal defect detection model to filter the experimental process parameter dataset. The optimal parameter selection module is configured to correct the simulation process parameter dataset based on the relative parameter prediction error between the experimental process parameter dataset and the simulation process parameter dataset, and to obtain an optimal cleaning parameter set based on the corrected simulation process parameter dataset and the experimental process parameter dataset. The feature standardization module is configured to acquire the physical correlation features of the paint layer and the substrate, calculate the mean and standard deviation of the corresponding physical correlation features based on the optimal cleaning parameter set, perform Z-score standardization on the physical correlation features, and obtain standardized input features. The physical correlation features include at least the thermal property parameters of the paint layer, the thermal property parameters of the substrate, and derived features calculated based on the thermal property parameters of the paint layer and the substrate. The prediction model building module is configured to establish a combination of process parameter prediction models, using the standardized input features as input to the combination of process parameter prediction models, and using a particle swarm optimization algorithm to find the combination of model parameters for the combination of process parameter prediction models. The parameter prediction module is configured to predict the combination of laser paint removal process parameters based on the combination of material parameters of the substrate coating layer to be treated by using the process parameters after the model parameters are determined.

Images