A Machine Vision-Based Method and System for Detecting Surface Defects in Coated Steel Plates

Abstract

This invention relates to the field of industrial non-destructive testing technology, specifically a machine vision-based method and system for detecting surface defects in coated steel plates. The method includes: S1: Excitation response acquisition: Simultaneously applying at least two active non-contact physical field excitations and acquiring multimodal response signals during excitation; S2: Feature inversion quantization: Dynamically registering and motion compensating the multimodal response signals to obtain compensated multimodal response signals, and using the compensated multimodal response signals as input to a multi-physical coupling proxy model and a multi-task evaluation network to obtain a corresponding dynamic failure risk score; S3: Determining the detection result: Constructing a structured report of the quantified result based on the dynamic failure risk score. This invention solves the problem of traditional methods being insensitive to deep defects, preventing finished products with intact surfaces but fatal internal defects from entering subsequent processing stages.

Description

Technical Field

[0001] This invention relates to the field of industrial nondestructive testing technology, specifically a method and system for detecting surface defects in coated steel plates based on machine vision. Background Technology

[0002] Coated steel sheets are functional composite materials made by laminating a layer of organic polymer film (such as PVC, PE, PET, etc.) onto the surface of cold-rolled or galvanized steel sheets. They are widely used in home appliance panels, building decoration, food packaging, and automotive interiors. Because they combine the strength of a metal substrate with the corrosion resistance and decorative properties of an organic film, extremely high surface quality is required. However, during the lamination process, factors such as raw material defects, uneven tension control, roller wear, environmental dust, or temperature and humidity fluctuations can easily cause various defects on the steel sheet surface, including scratches, bubbles, stains, wrinkles, delamination, color differences, and uneven film layers. These defects not only affect the product's appearance but may also weaken its protective performance, leading to customer rejection or even quality claims.

[0003] Currently, the detection of surface defects in coated steel sheets mainly relies on manual visual inspection. This method has obvious limitations: on the one hand, human eyes are prone to fatigue, resulting in low detection efficiency and making it difficult to meet the real-time detection requirements of modern high-speed continuous production lines (line speeds can reach 60–120 m / min); on the other hand, manual judgment is highly subjective, standards are not uniform, the rate of missed detection and false detection is high, and the long-term operating costs are high and the management is difficult.

[0004] Chinese invention application CN121068477A discloses an automatic classification device and application method for thermal shock cracks in coatings. The device includes a terahertz wave detection module, a laser-ultrasonic excitation and detection module, a high-precision three-dimensional motion control platform, a multi-modal signal synchronous acquisition and processing unit, an intelligent crack classification and evaluation unit, an environmental simulation and temperature control unit, a protective shell, and a human-machine interface. Terahertz technology characterizes changes in the near-surface dielectric properties of the coating, laser-ultrasonic technology detects the evolution of deep mechanical properties, and a high-temperature environment simulation unit reproduces thermal shock service conditions. A multi-modal feature fusion algorithm and a physical model combined with a machine learning-based inversion model are used to achieve automated quantitative assessment and severity classification of the depth, density, length, and network connectivity of thermal shock cracks in thermal barrier coatings. This invention solves the problems of existing single detection methods being insensitive to vertical cracks, having poor adaptability to high-temperature environments, and having low efficiency in manual assessment.

[0005] However, the above and similar technical solutions still have the following shortcomings: In the process of identifying and detecting coated steel plates of multi-layer composite materials, most of them only detect the two-dimensional optical features of the outermost layer, and cannot effectively detect the deep "subsurface" and "interface" defects. As a result, even in the final product state with a smooth and undamaged surface, there may be fatal defects, which will lead to product failure or even safety accidents during subsequent processing (such as stamping and bending) or use. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for detecting surface defects of coated steel plates based on machine vision, so as to solve the problems mentioned in the background art.

[0007] The technical solution of this invention is: a method for detecting surface defects of coated steel plates based on machine vision, comprising: S1: Excitation response acquisition: Simultaneously apply at least two active non-contact physical field excitations and acquire multimodal response signals during excitation, including but not limited to infrared thermal radiation time series images, laser Doppler surface displacement fields and ultrasonic guided wave propagation time-frequency diagrams; S2: Feature Inversion Quantization: Dynamic registration and motion compensation are performed on the multimodal response signal to obtain the compensated multimodal response signal. The compensated multimodal response signal is then used as input to a multi-physics coupled surrogate model and a multi-task evaluation network, and the corresponding dynamic failure risk score is obtained as output, including: S2.1: Model inversion: Through dynamic registration and motion compensation, the multimodal signal is compensated to obtain the compensated multimodal signal, and the inversion thermal diffusivity and inversion interface shear stiffness are obtained through the compensated multimodal signal. S2.2: Defect assessment: The inverted thermal diffusivity and inverted interface shear stiffness are used as inputs to a multi-task assessment network to perform surface defect segmentation, subsurface defect 3D reconstruction, and interface state quantification. S2.3: Scoring Determination: Based on the defect depth, estimated defect volume, and debonding area obtained during the processing, corresponding geometric risk parameters are set. Based on the final inverted interface shear stiffness and the substrate yield strength of the base steel plate material, corresponding mechanical risk parameters are set. Based on the target stamping speed and target bending radius of the coated steel plate, corresponding process risk parameters are set. At the same time, the geometric risk parameters, mechanical risk parameters, and process risk parameters are combined to determine the corresponding dynamic failure risk score. S3: Determine the test results: Based on the dynamic score of failure risk, construct a structured report of quantitative results, including the test timestamp, coated steel plate ID, report conclusion, defect list, interface status summary, dynamic score of failure risk, handling suggestions and data links.

[0008] Furthermore, the multimodal response signals acquired during excitation include: S1.1: Differential excitation: Subsurface defect-specific response and interface defect-specific response are acquired by pulsed laser thermal excitation and broadband ultrasonic guided wave excitation; S1.2: Sensor array: Based on the excitation positions of the pulsed laser thermal excitation and broadband ultrasonic guided wave excitation, the transmitter and receiver of the infrared thermal imager, laser Doppler vibrometer, and ultrasonic sensor are set up; S1.3: Multimodal response: Through the central controller, during the excitation process, the infrared thermal imager, laser Doppler vibrometer and ultrasonic sensor are simultaneously triggered to acquire the corresponding infrared thermal radiation time sequence image, laser Doppler surface displacement field and ultrasonic guided wave propagation time frequency diagram.

[0009] Furthermore, based on the heating area of ​​the pulsed laser thermal excitation, the infrared thermal imager is set up, and based on the excitation point position of the broadband ultrasonic guided wave excitation, the laser Doppler vibrometer is set up, with the infrared thermal imager and the laser Doppler vibrometer set up side by side. At the same time, based on the coating position of the coated steel plate, the transmitting end and receiving end of the ultrasonic sensor are set up, with the central axes of the transmitting end and the receiving end located on the same straight line.

[0010] Furthermore, the inverted thermal diffusivity and inverted interface shear stiffness are obtained, including: S2.1.1: Compensation Processing: Based on the position coordinates of each signal data in the infrared thermal radiation time series image and the movement direction of the coated steel plate, the absolute position corresponding to each signal data is determined, and based on the absolute position, the corresponding static thermal image is obtained to determine the temperature-time curve corresponding to the absolute position. At the same time, through coordinate transformation, the vibration-time curve corresponding to the absolute position is determined in the laser Doppler surface displacement field, and the ultrasonic signal corresponding to the absolute position is determined in the ultrasonic guided wave propagation time-frequency diagram. S2.1.2: Simulation Construction: Apply heat flux density and force / displacement load to the surface of the three-dimensional coated steel plate model, perform multiphysics coupling simulation, and extract the corresponding simulation temperature curve, simulation vibration waveform, simulation ultrasonic signal and simulation physical interpretability features through thermal conduction equation and elastic wave dynamics equation; S2.1.3: Parameter Determination: The model architecture of the multi-physics coupling surrogate model is set up through a convolutional neural network or a long short-term memory network. The final multi-physics coupling surrogate model is determined by the simulated temperature curve, simulated vibration waveform, simulated ultrasonic signal and simulated physical interpretability features. At the same time, the temperature-time curve, vibration-time curve and ultrasonic signal are used as inputs to the final multi-physics coupling surrogate model. The corresponding thermal diffusivity distribution map, interface shear stiffness distribution map, final inverted thermal diffusivity and final inverted interface shear stiffness are output.

[0011] Furthermore, the heat flux density and force / displacement load are applied at different points on the same surface of the three-dimensional coated steel plate model, and the application points are also applied at the same surface.

[0012] Furthermore, the simulated temperature curve, simulated vibration waveform, and simulated ultrasonic signal are used as inputs to the multi-physics coupling surrogate model. The corresponding predicted physical interpretability features are output, and the feature difference is obtained between the simulated physical interpretability features and the predicted physical interpretability features. Simultaneously, based on the comparison between the feature difference and a preset feature threshold, the corresponding multi-physics coupling surrogate model is determined. Specifically: When the feature difference is less than the preset feature threshold, the corresponding model parameters are the model parameters of the final multi-physics coupled proxy model; otherwise, the corresponding model parameters are fine-tuned by optimizing the gradient direction of the algorithm until the feature difference is less than the preset feature threshold.

[0013] Furthermore, surface defect segmentation, subsurface defect 3D reconstruction, and interface state quantification are performed, including: S2.2.1: Defect segmentation: The trained convolutional neural network segmentation model is obtained by using the optical image of the coated steel plate surface and the corresponding thermal diffusivity distribution map and interface shear stiffness distribution map. The thermal diffusivity distribution map and interface shear stiffness distribution map are used as the input of the trained convolutional neural network segmentation model, and the corresponding defect segmentation mask map is obtained as the output. S2.2.2: Defect Reconstruction: The basic geometric parameters of each defect are determined by the thermal diffusivity distribution map, and the depth distance of each defect is determined by the static thermal map. At the same time, the basic geometric parameters and depth distance of each defect are combined to obtain the corresponding three-dimensional defect visualization. S2.2.3: State Quantization: By using a preset shear stiffness threshold, the shear stiffness value corresponding to each pixel in the interface shear stiffness distribution map is compared and divided to determine the pixels that are less than the preset shear stiffness threshold and set them as debonding pixels. At the same time, the corresponding debonding area is determined according to the number of debonding pixels and the area of ​​a single pixel, and the average shear stiffness value of the corresponding debonding region is determined according to the shear stiffness value of each debonding pixel.

[0014] Furthermore, obtain the corresponding three-dimensional defect visualization, including: S2.2.2.1: Positioning Measurement: By using a preset pixel threshold, the pixel values ​​of each pixel in the thermal diffusivity distribution map are compared and divided. Based on the division results, the corresponding binary mask image is obtained. At the same time, based on the binary mask image, the corresponding suspected defect area is determined. The planar contour corresponding to the suspected defect area is determined by an edge detection algorithm or a contour tracking algorithm. S2.2.2.2: Depth Determination: Based on the suspected defect area, extract the average temperature change curve of the suspected defect area and the average temperature change curve of the intact area from the static thermal map, obtain the relative temperature value at each time, and determine the maximum relative temperature value based on the relative temperature value. At the same time, determine the defect depth corresponding to the suspected defect area through the maximum relative temperature value. S2.2.2.3: Volume Determination: Based on the planar contour, set the corresponding three-dimensional shape of the defect and set the length and width dimensions of the three-dimensional shape of the defect. At the same time, based on the defect depth, set the height dimension of the three-dimensional shape of the defect, obtain the corresponding three-dimensional defect view, and determine the corresponding estimated volume of the defect.

[0015] Furthermore, pixels with values ​​below the preset pixel threshold are designated as suspected defective region pixels, and the thermal diffusion coefficient of the suspected defective region pixels is set to 1. Pixels with values ​​not below the preset pixel threshold are designated as intact region pixels, and the thermal diffusion coefficient of the intact region pixels is set to 0. The marker colors corresponding to the suspected defective region and the intact region are different. Simultaneously, based on the marker colors and thermal diffusion coefficient values ​​corresponding to the suspected defective region and the intact region, the corresponding binary mask images are obtained.

[0016] A machine vision-based surface defect detection system for coated steel plates utilizes any one of the machine vision-based surface defect detection methods described above.

[0017] This invention provides an improved method and system for detecting surface defects in coated steel plates based on machine vision, which has the following improvements and advantages compared with the prior art: Firstly, this invention, by simultaneously applying pulsed laser thermal excitation and broadband ultrasonic guided wave excitation, can excite the thermal diffusion effect of subsurface defects and the guided wave propagation effect of interface defects, respectively. At the same time, by combining an infrared thermal imager, a laser Doppler vibrometer, and an ultrasonic sensor to simultaneously acquire multimodal response signals, it can simultaneously detect surface defects, subsurface defects, and interface debonding defects. This solves the problem of traditional methods being insensitive to deep defects and prevents finished products with intact surfaces but fatal internal defects from flowing into subsequent processing stages. Secondly, this invention uses a multi-physics coupling proxy model to invert the compensated multimodal signals into interpretable physical parameters, and combines it with a multi-task evaluation network to perform surface defect segmentation, sub-surface defect 3D reconstruction and interface state quantification, thereby reducing the risk of subjective misjudgment and providing reliable data support for process optimization. Thirdly, the failure risk dynamic scoring mechanism of this invention combines geometric risk parameters, mechanical risk parameters and process risk parameters, so as not only to predict the failure risk of defects in subsequent processing, but also to dynamically adapt to the process requirements of different production lines, avoiding a one-size-fits-all judgment standard. Attached Figure Description

[0018] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the surface defect detection method for coated steel plates in this invention; Figure 2 This is a schematic diagram of signal acquisition in the excitation response acquisition method of the present invention; Figure 3 This is a flowchart illustrating the multimodal signal compensation processing procedure in this invention. Figure 4 This is a schematic diagram illustrating the process of determining the multi-physics coupled proxy model in this invention; Figure 5 This is a schematic diagram of the processing flow of the multi-task evaluation network in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] refer to Figure 1 This embodiment provides a machine vision-based method for detecting surface defects in coated steel plates, which specifically includes the following steps: Step S1: Excitation Response Acquisition. This involves simultaneously applying at least two active non-contact physical field excitations with different depth detection capabilities on a continuously moving coated steel sheet production line. These include, but are not limited to, pulsed laser thermal excitation for stimulating deep thermal diffusion effects and broadband ultrasonic guided wave excitation for stimulating interface guided wave propagation effects. In other words, the corresponding subsurface defect-specific responses and interface defect-specific responses are acquired through pulsed laser thermal excitation and broadband ultrasonic guided wave excitation.

[0021] Furthermore, during the acquisition of specific responses, corresponding multimodal response signals are simultaneously acquired, including but not limited to infrared thermal radiation time-series images for characterizing defect depth information, laser Doppler surface displacement fields for characterizing interface bonding states, and ultrasonic guided wave propagation time-frequency diagrams. Specifically, an infrared thermal imager is used to capture temperature difference changes on the surface of the coated steel plate, obtaining the corresponding infrared thermal radiation time-series images. Simultaneously, a laser Doppler vibrometer is used to measure the vibration displacement field on the surface of the coated steel plate caused by ultrasonic wave propagation, obtaining the corresponding laser Doppler surface displacement field. An ultrasonic sensor is used to receive ultrasonic signals that penetrate or propagate within the coated steel plate, obtaining the corresponding ultrasonic guided wave propagation time-frequency diagram.

[0022] Step S2: Feature Inversion Quantization. This involves preprocessing the infrared thermal radiation time-series image, laser Doppler surface displacement field, and ultrasonic guided wave propagation time-frequency diagram obtained in Step S1 through dynamic registration and motion compensation to obtain the corresponding compensated signal data. Simultaneously, the obtained compensated signal data is used as input to the set multi-physics coupled surrogate model, and the output is the corresponding physical interpretability features, including the inverted thermal diffusivity coefficient and the inverted interface shear stiffness.

[0023] Furthermore, the acquired physical interpretability features are used as input to the multi-task evaluation network, and the output is the corresponding dynamic score of failure risk. Specifically, the multi-task evaluation network in this embodiment can perform multiple core tasks simultaneously. The core tasks in this embodiment include, but are not limited to, surface defect segmentation, subsurface defect 3D reconstruction, and interface state quantification. That is, through surface defect segmentation, surface defects of the coated steel plate, such as scratches and pits, are identified and delineated. Through subsurface defect 3D reconstruction, the infrared thermal radiation time series image, laser Doppler surface displacement field, and ultrasonic guided wave propagation time-frequency map are fused to construct a three-dimensional model of the defects in the coated steel plate and determine the corresponding volume and spatial location. Through interface state quantification, the inverted thermal diffusivity coefficient and the inverted interface shear stiffness are combined to classify the interface bonding state and determine the area of ​​the debonding region.

[0024] Step S3: Determine the test results. This involves combining the dynamic failure risk score obtained in Step S2 to construct a corresponding quantitative result structured report. It is worth noting that the quantitative result structured report in this embodiment includes a test timestamp, coated steel plate ID, report conclusions, a defect list (including the defect type, location, size, volume / area for each defect), an interface state summary (total debonding area, average bonding strength, debonding state), a failure risk dynamic score, processing suggestions, and data links.

[0025] This embodiment also provides a machine vision-based surface defect detection system for coated steel plates, which uses the aforementioned machine vision-based surface defect detection method for coated steel plates.

[0026] In this embodiment, pulsed laser thermal excitation and broadband ultrasonic guided wave excitation are used to acquire subsurface defect-specific responses and interface defect-specific responses, and multimodal response signals during the specific response process are acquired simultaneously. (Reference) Figure 2 This embodiment provides a method for acquiring excitation response, which specifically includes the following steps: Step S1.1: Differential Excitation. This involves acquiring the specific response of subsurface defects through pulsed laser thermal excitation and acquiring the specific response of interface defects through broadband ultrasonic guided wave excitation. Specifically, during the testing of the coated steel plate, at least two different excitation sources are simultaneously emitted, including pulsed laser thermal excitation and broadband ultrasonic guided wave excitation.

[0027] Furthermore, a high-frequency pulsed laser (such as an Nd:YAG laser or a fiber laser) emits a laser pulse with a preset energy and width (which can be specifically set according to actual needs, so it is not specifically described in this embodiment) onto the coated steel plate to be tested. Through the set diffractive optical elements or microlens array and beam expander system (composed of a set of convex and concave lenses), a short-term, uniformly shaped heating area is formed on the surface of the coated steel plate.

[0028] Furthermore, using a broadband ultrasonic excitation source (such as a laser ultrasonic excitation system (such as a Q-switched pulse laser) and an air-coupled ultrasonic transducer (such as a piezoelectric composite material transducer)), ultrasonic waves of a preset intensity (which can be specifically set according to actual needs, so it is not specifically described in this embodiment) are emitted to the coated steel plate to be tested in a non-contact excitation manner. The waves are amplified by a power amplifier, the energy is maintained by an impedance matching network, and then sent to a waveform generator. In other words, the waveform generator generates a corresponding narrowband pulse (such as a Hanning window modulated sine wave) or sharp pulse excitation electrical signal.

[0029] Step S1.2: Sensor Array. Based on the heating area formed by the high-frequency pulsed laser in step S1.1, an infrared thermal imager is set up directly behind the heating area (i.e., along the movement direction of the coated steel plate), and fixed at a distance of 0.5m-1m from the surface of the coated steel plate by an adjustable bracket, so that the field of view of the infrared thermal imager can cover the entire heating area.

[0030] Furthermore, based on the excitation point position of the excitation electrical signal in step S1.1, a laser Doppler vibration meter is set up near the excitation point position of the excitation electrical signal (for example, 1cm-10cm), and the laser Doppler vibration meter and the infrared thermal imager are set up side by side so that the detection area of ​​the coated steel plate by the laser Doppler vibration meter is the same as the detection area of ​​the infrared thermal imager.

[0031] Furthermore, depending on the coating position of the coated steel plate, the transmitting end of the ultrasonic sensor, i.e., the ultrasonic excitation source (such as an air-coupled ultrasonic transducer), is set on the coated side of the coated steel plate, at a distance of 5mm-20mm from the surface of the coated steel plate, and vertically aligned with the corresponding detection area. At the same time, the receiving end of the ultrasonic sensor (such as another air-coupled ultrasonic transducer) is set on the back side of the coated steel plate (i.e., the opposite side of the coated side), and the central axes of the transmitting end and the receiving end of the ultrasonic sensor are on the same straight line.

[0032] Step S1.3: Multimodal Response. This involves electrically connecting the high-frequency pulsed laser, broadband ultrasonic excitation source, and air-coupled ultrasonic transducer from Step S1.1 to the infrared thermal imager, laser Doppler vibrometer, and ultrasonic sensor from Step S1.2 via the central controller. Specifically, the central controller simultaneously sends trigger signals to the high-frequency pulsed laser, broadband ultrasonic excitation source, air-coupled ultrasonic transducer, infrared thermal imager, laser Doppler vibrometer, and ultrasonic sensor, causing them to start operating simultaneously. This allows for the synchronous acquisition of corresponding pulsed laser thermal excitation, broadband ultrasonic guided wave excitation, infrared thermal radiation time-series images, laser Doppler surface displacement field, and ultrasonic guided wave propagation time-frequency diagrams.

[0033] In this embodiment, a multi-physical coupling proxy model is used to obtain corresponding physical interpretability features. These features are then used as input to a multi-task evaluation network, which outputs a dynamic failure risk score. (Reference) Figures 3-5 This embodiment provides a feature inversion quantization method, which specifically includes the following steps: Step S2.1: Model Inversion. This involves dynamically registering and compensating for motion to obtain the multimodal signals (i.e., infrared thermal radiation time-series images, laser Doppler surface displacement field, and ultrasonic guided wave propagation time-frequency diagrams) acquired in Step S1.3, resulting in compensated multimodal signals. Simultaneously, the corresponding inverted thermal diffusivity and inverted interface shear stiffness are obtained from the compensated multimodal signals. Details are as follows: Step S2.1.1: Compensation Processing. This involves setting a photoelectric sensor at the starting point of the inspection area of ​​the coated steel sheet production line to establish the time and spatial zero points of the inspection process. Simultaneously, a rotary encoder is installed on the production line rollers to determine the corresponding movement distance based on the number of pulses emitted by the encoder. In other words, by setting the time zero point, spatial zero point, and movement distance, the position coordinates corresponding to each signal data point in the multimodal signal are determined.

[0034] Furthermore, based on the position coordinates corresponding to each signal data point in the infrared thermal radiation time-series image and the movement direction of the coated steel plate, the absolute position corresponding to each signal data point is determined (for example, if the position coordinates of the 100th frame image are 10 meters, and the movement direction of the coated steel plate is to the right, then the absolute position of the 100th frame image is 10 meters, and the 101st frame image is located to the right of the 100th frame image). Simultaneously, based on the absolute position corresponding to each signal data point, the absolute positions of all signal data points are sequentially connected to obtain the corresponding static thermal map.

[0035] Furthermore, based on the absolute position corresponding to each signal data point, the temperature-time curve corresponding to each absolute position is determined by acquiring the static thermal map. Simultaneously, through coordinate transformation (i.e., a conventional coordinate transformation algorithm, which is not specifically described in this embodiment), the position coordinates of the absolute position of each signal data point in the laser Doppler surface displacement field and the ultrasonic guided wave propagation time-frequency diagram are obtained. Then, using the laser Doppler surface displacement field and the ultrasonic guided wave propagation time-frequency diagram, the vibration-time curve and ultrasonic signal corresponding to each position coordinate are determined.

[0036] Step S2.1.2: Simulation Construction. This involves using finite element analysis software (such as COMSOL Multiphysics or ABAQUS) to create a three-dimensional model of the coated steel plate. This model includes material properties (such as the density, specific heat capacity, thermal conductivity, elastic modulus, and Poisson's ratio of the base steel), corresponding properties of the coating layer (such as thermophysical properties (including but not limited to thermal conductivity, specific heat capacity, density, and thermal diffusivity), mechanical and acoustic properties (including but not limited to elastic modulus, Poisson's ratio, density, sound velocity, and attenuation coefficient), surface and interface properties (including but not limited to optical reflectivity and thermal infrared emissivity), and interface characteristic parameters (including but not limited to interface shear stiffness / interface toughness)). Interface properties refer to interface shear stiffness. Geometric structure (such as the dimensions and thickness of the steel plate, and the thickness of the coating layer) and defect parameters (such as preset subsurface defects and preset interface defects, which can be specifically preset according to actual needs, and are not specifically described in this embodiment) are also included.

[0037] Furthermore, multiphysics coupling simulations were performed based on the established three-dimensional coated steel plate model. It is noteworthy that during the multiphysics coupling simulation, heat flux density and force / displacement loads were applied to the surface of the three-dimensional coated steel plate model as corresponding thermal and acoustic excitations. It is also important to note that the application surfaces and application points corresponding to the heat flux density and force / displacement loads belong to different application points on the same surface.

[0038] Specifically, in the process of multiphysics coupled simulation, the thermal conductivity equation and the elastic wave dynamics equation are solved (it is worth noting that the solution of the thermal conductivity equation and the elastic wave dynamics equation in this embodiment is a conventional solution process, so it is not specifically described in this embodiment), and the coupling effect between the thermal conductivity equation and the elastic wave dynamics equation (such as thermal expansion generating stress, and stress changes affecting heat conduction) is used to extract the corresponding simulated temperature curves, simulated vibration waveforms, and simulated ultrasonic signals. It is worth noting that during the coupled solution process, the corresponding simulated physical interpretability characteristics can be output, namely the simulated inverted thermal diffusivity coefficient and the simulated inverted interface shear stiffness.

[0039] Step S2.1.3: Parameter Determination. This involves using a convolutional neural network or a long short-term memory network as the model architecture for the multi-physics coupling surrogate model. Simultaneously, the simulated temperature curve, simulated vibration waveform, and simulated ultrasonic signal extracted in step S2.1.2 are used as inputs to the multi-physics coupling surrogate model. Through processing by the multi-physics coupling surrogate model, the corresponding predictive physical interpretability features are obtained, namely, the predicted inversion thermal diffusivity coefficient and the predicted inversion interface shear stiffness.

[0040] Furthermore, based on the obtained predicted physical interpretability features and the simulated physical interpretability features obtained in step S2.1.2, the corresponding feature differences are determined. Simultaneously, the determined feature differences are compared with a preset feature threshold (which can be set according to actual needs, and is not specifically described in this embodiment). Based on the comparison result, the corresponding multi-physics coupling proxy model is determined. Specifically: When the determined feature difference is less than the preset feature threshold, the corresponding model parameters (such as network connection weights) become the model parameters of the final multi-physics coupled proxy model. Conversely, when the determined feature difference is not less than the preset feature threshold, the corresponding model parameters (such as network connection weights) are fine-tuned by optimizing the gradient direction of an algorithm (such as the Adam algorithm) until the determined feature difference is less than the preset feature threshold.

[0041] In other words, using the determined final multiphysics coupling surrogate model, the temperature-time curve, vibration-time curve, and ultrasonic signal obtained in step S2.1.1 are used as inputs to the final multiphysics coupling surrogate model, and the outputs are the corresponding final physical interpretability features, namely the final inverted thermal diffusivity coefficient and the final inverted interface shear stiffness. It is worth noting that when obtaining the corresponding final inverted thermal diffusivity coefficient and the final inverted interface shear stiffness, corresponding thermal diffusivity coefficient distribution maps and interface shear stiffness distribution maps are also obtained. That is, the corresponding final inverted thermal diffusivity coefficient is extracted from the thermal diffusivity coefficient distribution map, and the corresponding final inverted interface shear stiffness is extracted from the interface shear stiffness distribution map.

[0042] Step S2.2: Defect Assessment. The final physical interpretability features obtained in step S2.1.3 (i.e., the final inverted thermal diffusivity and the final inverted interface shear stiffness) are used as input to the multi-task assessment network for surface defect segmentation, subsurface defect 3D reconstruction, and interface state quantification. Specifically: Step S2.2.1: Defect Segmentation. This involves training a convolutional neural network (CNN) segmentation model using optical images of the coated steel plate surface and corresponding thermal diffusivity and interface shear stiffness distribution maps. It's important to note that during training, all defect contours in the corresponding thermal diffusivity and interface shear stiffness distribution maps are labeled, and different pixel regions correspond to different defects. For example, the pixel region for scratch defects is set to red, the pixel region for dent defects is set to green, and the background pixel region is set to black.

[0043] Furthermore, the thermal diffusivity distribution map and interface shear stiffness distribution map obtained in step S2.1.3 are used as inputs to the trained convolutional neural network segmentation model, and the corresponding label map, i.e. the defect segmentation mask map, is output.

[0044] Step S2.2.2: Defect Reconstruction. This involves determining the basic geometric parameters of each defect using the thermal diffusivity distribution map obtained in step S2.1.3. The depth distance of each defect is determined using the static thermal image obtained in step S2.1.1. In other words, the determined basic geometric parameters and depth distance of each defect are combined to obtain a corresponding 3D defect visualization. Details are as follows: Step S2.2.2.1: Positioning Measurement. This involves comparing and classifying the pixel values ​​of each pixel in the thermal diffusivity distribution map obtained in step S2.1.3 using a preset pixel threshold (which can be set according to actual needs, and is not specifically described in this embodiment). Pixels with values ​​below the preset pixel threshold are designated as suspected defect area pixels and marked with the same color (e.g., white), with their corresponding thermal diffusivity value set to 1. Simultaneously, pixels with values ​​not below the preset pixel threshold are marked with the same color (e.g., black), but with a different color from the suspected defect area pixels, and their corresponding thermal diffusivity value is set to 0. In other words, a corresponding binary mask image is obtained based on the marked color and thermal diffusivity value of each pixel.

[0045] Furthermore, based on the acquired binary mask image, all interconnected suspected defect region pixels (i.e., white pixel groups) are identified, and the corresponding suspected defect regions are obtained. Simultaneously, using edge detection algorithms (such as the Canny algorithm) or contour tracking algorithms, the planar contours corresponding to the suspected defect regions are determined, and based on these determined planar contours, the corresponding planar projected area, centroid position, and minimum bounding rectangle are obtained.

[0046] Step S2.2.2.2: Depth Determination. Based on the suspected defect areas identified in Step S2.2.2.1, the average temperature change curves of the suspected defect areas and the intact areas are extracted from the static thermal image obtained in Step S2.1.1. Simultaneously, based on the temperature values ​​at each moment in the average temperature change curves of the suspected defect areas and the intact areas, the relative temperature value at each moment is determined, specifically: in: This is a relative temperature value. Let be the average temperature of the suspected defect region at time t. Let be the average temperature of the intact region at time t.

[0047] Furthermore, based on the relative temperature value determined at each moment, the corresponding maximum relative temperature value is determined, and based on the maximum relative temperature value, the defect depth corresponding to the suspected defect area is determined, specifically: in: For defect depth, The thermal diffusivity of the base steel plate material. This represents the relative value of the maximum temperature.

[0048] Step S2.2.2.3: Volume Determination. Based on the planar contour obtained in Step 2.2.2.1 (i.e., the corresponding planar projected area, centroid position, and minimum bounding rectangle), the corresponding 3D shape of the defect is set. Simultaneously, the length and width dimensions of the 3D shape of the defect are set according to the length and width dimensions of the planar contour. Furthermore, the height dimension of the 3D shape of the defect is set according to the defect depth obtained in Step 2.2.2., thereby obtaining the corresponding 3D defect visualization. In other words, the estimated volume of the defect is determined based on the obtained 3D defect visualization.

[0049] Step S2.2.3: State Quantization. This involves comparing and dividing the shear stiffness value of each pixel in the interface shear stiffness distribution map obtained in step S2.1.3 using a preset shear stiffness threshold (which can be set according to actual needs, such as 50 MPa) to determine all debonding areas. Specifically, when the shear stiffness value is less than the preset shear stiffness threshold, the corresponding pixel is the debonded pixel.

[0050] Furthermore, the number of debonded pixels and the area of ​​each pixel are combined to determine the corresponding debonded area. Simultaneously, the average shear stiffness value of the debonded region is determined based on the shear stiffness value corresponding to each debonded pixel.

[0051] Step S2.3: Scoring Determination. Based on the defect depth obtained in Step S2.2.2.2, the estimated defect volume obtained in Step S2.2.2.3, and the debonding area obtained in Step S2.2.3, the corresponding geometric risk parameters are determined. Simultaneously, based on the final inverted interface shear stiffness and the substrate yield strength of the base steel plate material obtained in Step S2.1.3, the corresponding mechanical risk parameters are determined. Finally, based on the target stamping speed and target bending radius of the coated steel plate in subsequent stamping processes, the corresponding process risk parameters are determined.

[0052] Furthermore, by combining the determined geometric risk parameters, mechanical risk parameters, and process risk parameters, a corresponding dynamic score for failure risk is determined, specifically as follows: in: Dynamic scoring for failure risk, Geometric risk weights Let the geometric risk function be... Let the mechanical risk function be... For process risk function, Estimate the volume of the defect. For defect depth, For the debonding area, For mechanical risk weights, To ultimately invert the interface shear stiffness, The yield strength of the substrate. As a process risk weight, For the target stamping speed, The target bending radius.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting surface defects in coated steel plates based on machine vision, characterized in that, Including: S1: Excitation response acquisition: Simultaneously apply at least two active non-contact physical field excitations and acquire multimodal response signals during excitation, including but not limited to infrared thermal radiation time series images, laser Doppler surface displacement fields and ultrasonic guided wave propagation time-frequency diagrams; S2: Feature Inversion Quantization: Dynamic registration and motion compensation are performed on the multimodal response signal to obtain the compensated multimodal response signal. The compensated multimodal response signal is then used as input to a multi-physics coupled surrogate model and a multi-task evaluation network, and the corresponding dynamic failure risk score is obtained as output, including: S2.1: Model inversion: Through dynamic registration and motion compensation, the multimodal signal is compensated to obtain the compensated multimodal signal, and the inversion thermal diffusivity and inversion interface shear stiffness are obtained through the compensated multimodal signal. S2.2: Defect assessment: The inverted thermal diffusivity and inverted interface shear stiffness are used as inputs to a multi-task assessment network to perform surface defect segmentation, subsurface defect 3D reconstruction, and interface state quantification. S2.3: Scoring Determination: Based on the defect depth, estimated defect volume, and debonding area obtained during the processing, corresponding geometric risk parameters are set. Based on the final inverted interface shear stiffness and the substrate yield strength of the base steel plate material, corresponding mechanical risk parameters are set. Based on the target stamping speed and target bending radius of the coated steel plate, corresponding process risk parameters are set. At the same time, the geometric risk parameters, mechanical risk parameters, and process risk parameters are combined to determine the corresponding dynamic failure risk score. S3: Determine the test results: Based on the dynamic score of failure risk, construct a structured report of quantitative results, including the test timestamp, coated steel plate ID, report conclusion, defect list, interface status summary, dynamic score of failure risk, handling suggestions and data links.

2. The method for detecting surface defects of coated steel plates based on machine vision according to claim 1, characterized in that, The multimodal response signals acquired during excitation include: S1.1: Differential excitation: Subsurface defect-specific response and interface defect-specific response are acquired by pulsed laser thermal excitation and broadband ultrasonic guided wave excitation; S1.2: Sensor array: Based on the excitation positions of the pulsed laser thermal excitation and broadband ultrasonic guided wave excitation, the transmitter and receiver of the infrared thermal imager, laser Doppler vibrometer, and ultrasonic sensor are set up; S1.3: Multimodal response: Through the central controller, during the excitation process, the infrared thermal imager, laser Doppler vibrometer and ultrasonic sensor are simultaneously triggered to acquire the corresponding infrared thermal radiation time sequence image, laser Doppler surface displacement field and ultrasonic guided wave propagation time frequency diagram.

3. The method for detecting surface defects of coated steel plates based on machine vision according to claim 2, characterized in that, The infrared thermal imager is set according to the heating area of ​​the pulsed laser thermal excitation, and the laser Doppler vibrometer is set according to the excitation point position of the broadband ultrasonic guided wave excitation. The infrared thermal imager and the laser Doppler vibrometer are set side by side. At the same time, the transmitting end and receiving end of the ultrasonic sensor are set according to the coating position of the coated steel plate, and the central axes of the transmitting end and the receiving end are located on the same straight line.

4. The method for detecting surface defects of coated steel plates based on machine vision according to claim 1, characterized in that, Obtaining the inverted thermal diffusivity and the inverted interface shear stiffness includes: S2.1.1: Compensation Processing: Based on the position coordinates of each signal data in the infrared thermal radiation time series image and the movement direction of the coated steel plate, the absolute position corresponding to each signal data is determined, and based on the absolute position, the corresponding static thermal image is obtained to determine the temperature-time curve corresponding to the absolute position. At the same time, through coordinate transformation, the vibration-time curve corresponding to the absolute position is determined in the laser Doppler surface displacement field, and the ultrasonic signal corresponding to the absolute position is determined in the ultrasonic guided wave propagation time-frequency diagram. S2.1.2: Simulation Construction: Apply heat flux density and force / displacement load to the surface of the three-dimensional coated steel plate model, perform multiphysics coupling simulation, and extract the corresponding simulation temperature curve, simulation vibration waveform, simulation ultrasonic signal and simulation physical interpretability features through thermal conduction equation and elastic wave dynamics equation; S2.1.3: Parameter Determination: The model architecture of the multi-physics coupling surrogate model is set up through a convolutional neural network or a long short-term memory network. The final multi-physics coupling surrogate model is determined by the simulated temperature curve, simulated vibration waveform, simulated ultrasonic signal and simulated physical interpretability features. At the same time, the temperature-time curve, vibration-time curve and ultrasonic signal are used as inputs to the final multi-physics coupling surrogate model. The corresponding thermal diffusivity distribution map, interface shear stiffness distribution map, final inverted thermal diffusivity and final inverted interface shear stiffness are output.

5. The method for detecting surface defects of coated steel plates based on machine vision according to claim 4, characterized in that, The heat flux density and force / displacement load are applied at different points on the same surface of the three-dimensional coated steel plate model.

6. The method for detecting surface defects of coated steel plates based on machine vision according to claim 4, characterized in that, The simulated temperature curve, simulated vibration waveform, and simulated ultrasonic signal are used as inputs to the multi-physics coupling surrogate model. The corresponding predicted physical interpretability features are output, and the feature difference is obtained between the simulated physical interpretability features and the predicted physical interpretability features. Simultaneously, based on a comparison between the feature difference and a preset feature threshold, the corresponding multi-physics coupling surrogate model is determined. Specifically: When the feature difference is less than the preset feature threshold, the corresponding model parameters are the model parameters of the final multi-physics coupled proxy model; otherwise, the corresponding model parameters are fine-tuned by optimizing the gradient direction of the algorithm until the feature difference is less than the preset feature threshold.

7. The method for detecting surface defects of coated steel plates based on machine vision according to claim 1, characterized in that, Surface defect segmentation, subsurface defect 3D reconstruction, and interface state quantification are performed, including: S2.2.1: Defect segmentation: The trained convolutional neural network segmentation model is obtained by using the optical image of the coated steel plate surface and the corresponding thermal diffusivity distribution map and interface shear stiffness distribution map. The thermal diffusivity distribution map and interface shear stiffness distribution map are used as the input of the trained convolutional neural network segmentation model, and the corresponding defect segmentation mask map is obtained as the output. S2.2.2: Defect Reconstruction: The basic geometric parameters of each defect are determined by the thermal diffusivity distribution map, and the depth distance of each defect is determined by the static thermal map. At the same time, the basic geometric parameters and depth distance of each defect are combined to obtain the corresponding three-dimensional defect visualization. S2.2.3: State Quantization: By using a preset shear stiffness threshold, the shear stiffness value corresponding to each pixel in the interface shear stiffness distribution map is compared and divided to determine the pixels that are less than the preset shear stiffness threshold and set them as debonding pixels. At the same time, the corresponding debonding area is determined according to the number of debonding pixels and the area of ​​a single pixel, and the average shear stiffness value of the corresponding debonding region is determined according to the shear stiffness value of each debonding pixel.

8. The method for detecting surface defects of coated steel plates based on machine vision according to claim 7, characterized in that, Obtain the corresponding 3D defect visualization, including: S2.2.2.1: Positioning Measurement: By using a preset pixel threshold, the pixel values ​​of each pixel in the thermal diffusivity distribution map are compared and divided. Based on the division results, the corresponding binary mask image is obtained. At the same time, based on the binary mask image, the corresponding suspected defect area is determined. The planar contour corresponding to the suspected defect area is determined by an edge detection algorithm or a contour tracking algorithm. S2.2.2.2: Depth Determination: Based on the suspected defect area, extract the average temperature change curve of the suspected defect area and the average temperature change curve of the intact area from the static thermal map, obtain the relative temperature value at each time, and determine the maximum relative temperature value based on the relative temperature value. At the same time, determine the defect depth corresponding to the suspected defect area through the maximum relative temperature value. S2.2.2.3: Volume Determination: Based on the planar contour, set the corresponding three-dimensional shape of the defect and set the length and width dimensions of the three-dimensional shape of the defect. At the same time, based on the defect depth, set the height dimension of the three-dimensional shape of the defect, obtain the corresponding three-dimensional defect view, and determine the corresponding estimated volume of the defect.

9. The method for detecting surface defects of coated steel plates based on machine vision according to claim 8, characterized in that, Pixels with pixel values ​​below the preset pixel threshold are designated as suspected defective region pixels, and their thermal diffusion coefficient is set to 1. Pixels with pixel values ​​not below the preset pixel threshold are designated as intact region pixels, and their thermal diffusion coefficient is set to 0. The marker colors corresponding to the suspected defective region and the intact region are different. Based on the marker colors and thermal diffusion coefficient values ​​corresponding to the suspected defective region and the intact region, the corresponding binary mask images are obtained.

10. A machine vision-based surface defect detection system for coated steel plates, characterized in that, The method for detecting surface defects of coated steel plates based on machine vision, as described in any one of claims 1-9, was used.

Images