Steel member surface rust grading and spraying track planning method based on multispectral vision

Abstract

This invention discloses a method for grading corrosion and planning spraying trajectories on steel components based on multispectral vision, belonging to the field of steel structure maintenance technology. The method includes: acquiring a synchronous multispectral image dataset; calculating and generating corrosion component and activity feature maps; generating corrosion level and potential prediction maps using a deductive model; calculating and generating a dynamic spraying demand field based on geometric information; collaboratively optimizing and generating a dynamic spraying trajectory package; and online monitoring of film quality maps and performing real-time correction and iterative optimization. This invention combines multispectral perception with a physical mechanism model, enabling accurate identification of corrosion states through interference, and achieving closed-loop trajectory planning based on dynamic demand and real-time feedback, thus improving the accuracy of corrosion identification, spraying quality, and operational efficiency in complex environments.

Description

Technical Field

[0001] This invention relates to the field of steel structure maintenance technology, and in particular to a method for classifying surface rust and planning spraying trajectories of steel components based on multispectral vision. Background Technology

[0002] Steel components, as core load-bearing structures in modern buildings, bridges, ships, and large equipment, are of paramount importance for safety and durability. However, under natural conditions, the surface of steel components is highly susceptible to electrochemical corrosion, forming a porous and complex rust layer. Rust not only affects the aesthetics of the structure but also continuously erodes the steel substrate, weakening the effective cross-section of the component and reducing its load-bearing capacity, posing a significant hidden danger to structural failure. Therefore, accurate detection and classification of the rust condition on the surface of steel components, and effective anti-corrosion coating maintenance accordingly, are crucial for ensuring the safety of engineering structures and extending their service life.

[0003] Currently, the detection of rust on steel components mainly relies on manual visual comparison with standard images or the use of single machine vision technology. Manual inspection methods are highly subjective, inefficient, and difficult to quantify, especially in large-area or high-altitude operations, where they suffer from poor consistency and high safety risks. While machine vision methods based on traditional visible light images have achieved a degree of automation, their detection principle primarily relies on macroscopic surface features such as rust color and texture, using image segmentation and feature classification algorithms for grading. This method can perform region division and coarse grading of existing rust spots.

[0004] However, existing technologies have significant limitations in the in-depth analysis of corrosion conditions. First, visible light alone cannot penetrate the corrosion surface, making it difficult to obtain key physicochemical information such as the chemical composition, density, and moisture content of corrosion products. Therefore, it cannot effectively distinguish between active corrosion in its early stages and relatively stable passivated corrosion. Second, existing detection methods are usually static, reflecting only the corrosion state at the moment of detection and lacking the ability to predict corrosion development trends. Furthermore, the results of detection and grading are disconnected from subsequent spraying operations. Spraying operations typically employ a uniform thickness "one-size-fits-all" approach, failing to differentiate and refine paint application based on the corrosion level and protection requirements of different areas, resulting in material waste and uneven protective effects. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for classifying and planning the surface corrosion of steel components based on multispectral vision. It employs multispectral image fusion to calculate corrosion components and activity, constructs an evolution model to predict corrosion trends, and combines a closed-loop control strategy of online monitoring and real-time feedback to correct the spraying trajectory. This method enables in-depth quantitative assessment of the corrosion state of steel components and adaptive anti-corrosion spraying operations.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for grading corrosion and planning spraying trajectories on steel component surfaces based on multispectral vision is provided, comprising: acquiring synchronous multispectral image data of the steel component surface to generate a multispectral image dataset; performing cross-modal feature fusion and calculation on the multispectral image dataset to generate corrosion component and activity feature maps; based on the corrosion component and activity feature maps, performing grading and prediction using a preset model for inferring corrosion evolution trends to generate corrosion level and potential prediction maps; calculating dynamic spraying requirements based on the corrosion level and potential prediction maps and the three-dimensional geometric information of the steel component surface to generate a three-dimensional dynamic spraying requirement field; using the three-dimensional dynamic spraying requirement field as the target, co-optimizing the motion trajectory and spraying parameters of the spraying equipment to generate a dynamic spraying trajectory package; executing the dynamic spraying trajectory package to perform spraying operations and monitoring the paint film state after spraying online to generate an instantaneous film quality map; comparing the instantaneous film quality map with the three-dimensional dynamic spraying requirement field to generate a quality error map, and performing real-time correction and iterative optimization of the dynamic spraying trajectory package based on the quality error map.

[0007] Based on the above technical solution, the method for grading and planning the surface corrosion of steel components based on multispectral vision provided in this application adopts a closed-loop control strategy of multispectral image fusion to solve the corrosion components and activity, constructing an evolution model to predict the corrosion trend, and combining online monitoring and real-time feedback to correct the spraying trajectory. This can achieve in-depth quantitative assessment of the corrosion state of steel components and adaptive anti-corrosion spraying operation.

[0008] In conjunction with the first aspect above, in one possible implementation, acquiring synchronous multispectral image data of the steel component surface and generating a multispectral image dataset includes: synchronously acquiring visible light images, shortwave infrared images, and thermal imaging sequence images of the steel component surface; spatially registering and temporally aligning the acquired visible light images, shortwave infrared images, and thermal imaging sequence images; fusing the registered and aligned image data to construct a multispectral tensor field containing the color value, infrared absorption intensity, and thermal change rate information of each pixel, as the multispectral image dataset.

[0009] In conjunction with the first aspect described above, in one possible implementation, performing cross-modal feature fusion and calculation on the multispectral image dataset to generate a corrosion component and activity feature map includes: extracting the absorption spectrum of each pixel from the multispectral image dataset and matching it with a database storing standard spectra of corrosion products to obtain the component ratio characteristics of the corrosion products; combining the thermal imaging sequence data in the multispectral image dataset to analyze the temperature anomaly diffusion pattern to obtain the corrosion activity intensity characteristics; and fusing the component ratio characteristics and the corrosion activity intensity characteristics to generate a corrosion component and activity feature map characterizing the corrosion type composition and electrochemical activity intensity of each pixel.

[0010] In conjunction with the first aspect above, in one possible implementation, the step of classifying and predicting corrosion evolution trends using a pre-set model to generate a corrosion level and potential prediction map includes: analyzing the spatial distribution morphology characteristics of corrosion products in the corrosion component and activity characteristic map, where the spatial distribution morphology characteristics include the boundary irregularity and porosity of the corrosion area; inputting the spatial distribution morphology characteristics, the component ratio characteristics, and the corrosion activity intensity characteristics into the pre-set model for predicting corrosion evolution trends, wherein the model for predicting corrosion evolution trends is a nonlinear mapping model based on electrochemical kinetics; in the model for predicting corrosion evolution trends, the corrosion potential is determined using the component ratio characteristics, the charge transfer impedance is calculated by combining the corrosion activity intensity characteristics and porosity, and the penetration rate of corrosion in the surface and deep layers of the steel component substrate is predicted using an anisotropic diffusion algorithm; and the model for predicting corrosion evolution trends outputs a corrosion level and potential prediction map containing the current corrosion severity classification and the future expansion vector distribution.

[0011] In conjunction with the first aspect above, in one possible implementation, calculating dynamic spraying requirements and generating a three-dimensional dynamic spraying requirement field based on the corrosion level and potential prediction map and the three-dimensional geometric information of the steel component surface includes: determining the basic paint film thickness requirement based on the corrosion level and potential prediction map; acquiring three-dimensional point cloud data of the steel component surface and calculating the surface curvature and normal vector information of the steel component; and calculating the paint adhesion efficiency and sagging trend using a model for simulating the fluid behavior of coatings on different surface morphologies, combined with the basic paint film thickness requirement and the surface curvature and normal vector information. The model used to simulate the fluid behavior of coatings on different morphological surfaces is a semi-empirical fluid physics model. In the model simulating the fluid behavior of coatings on different morphological surfaces, the coating loss caused by rebound is calculated using surface curvature, the effective adhesion rate is calculated by combining the angle between the normal vector and the spray gun spray direction, and the sagging risk index is evaluated based on the ratio of the component of gravity on the inclined plane to the coating viscosity. Based on the coating adhesion efficiency and sagging trend, the basic paint film thickness requirement is compensated and adjusted, and a three-dimensional dynamic spraying demand field including the target paint film thickness, spray dwell time, and allowable flow rate is output.

[0012] In conjunction with the first aspect above, in one possible implementation, taking the three-dimensional dynamic spraying demand field as the target, the motion trajectory and spraying parameters of the spraying equipment are collaboratively optimized to generate a dynamic spraying trajectory package, which includes: converting the three-dimensional dynamic spraying demand field into an objective function for achieving uniform distribution of paint deposition; using the objective function as a constraint, collaboratively optimizing the spatial path, moving speed, paint output, atomization pressure, and spray gun attitude angle parameters of the spraying equipment; and encapsulating the optimized spatial path, time series, six-degree-of-freedom attitude, and spraying equipment parameters to generate a dynamic spraying trajectory package.

[0013] In conjunction with the first aspect above, in one possible implementation, executing the dynamic spraying trajectory package for spraying operations and monitoring the paint film state online to generate an instantaneous film quality map includes: controlling the spraying equipment to execute the dynamic spraying trajectory package for spraying; simultaneously, using a broadband light source to illuminate the surface of the undried paint film and receiving the interference fringe image formed by the reflected light; analyzing the interference fringe image, inverting and calculating the wet film thickness distribution and leveling state of the paint film, and generating an instantaneous film quality map.

[0014] In conjunction with the first aspect above, in one possible implementation, comparing the instantaneous film-forming quality map with the three-dimensional dynamic spraying demand field to generate a quality error map, and performing real-time correction and iterative optimization of the dynamic spraying trajectory package based on the quality error map includes: comparing the wet film thickness distribution in the instantaneous film-forming quality map with the target paint film thickness in the three-dimensional dynamic spraying demand field point by point to generate a thickness error distribution map; inputting the thickness error distribution map as a state input to a preset intelligent decision-making model, wherein the intelligent decision-making model is a reinforcement learning decision-making model; in the intelligent decision-making model, by extracting the error mean and variance features in the thickness error distribution map, mapping and generating trajectory parameter correction amounts for subsequent unexecuted paths, wherein the trajectory parameter correction amounts include a moving speed correction ratio, a paint output correction ratio, and a path overlap rate adjustment value; and updating the execution parameters of subsequent path segments in the dynamic spraying trajectory package in real time according to the adjustment instructions generated by the trajectory parameter correction amounts.

[0015] In conjunction with the first aspect described above, in one possible implementation, the intelligent decision-making model further performs offline training based on historical operational data to optimize the model used to predict the corrosion evolution trend and the model used to simulate the fluid behavior of coatings on different morphological surfaces. This includes: acquiring a historical causal chain sample set composed of historical quality error maps, corresponding corrosion component and activity feature maps, and corrosion level and potential prediction maps; using a gradient descent algorithm, with the goal of optimizing the loss function between prediction deviation and actual thickness error, adjusting the neuron weights within the model predicting the corrosion evolution trend to optimize the mapping accuracy of the model predicting the corrosion evolution trend to the corrosion expansion potential; and iteratively correcting the curvature influence coefficient and sagging influence coefficient in the model simulating the fluid behavior of coatings on different morphological surfaces by analyzing the statistical correlation of thickness deviations under specific geometric features in the historical quality error maps, thereby achieving adaptive compensation optimization of the physical behavior of coating deposition.

[0016] Secondly, a system for grading corrosion and planning spraying trajectories on steel components based on multispectral vision is provided, comprising: a multispectral image acquisition module for acquiring synchronous multispectral image data of the steel component surface and generating a multispectral image dataset; a cross-modal feature analysis module for performing cross-modal feature fusion and calculation on the multispectral image dataset to generate corrosion component and activity feature maps; a corrosion grading and prediction module for grading and predicting corrosion based on the corrosion component and activity feature maps using a preset model for extrapolating corrosion evolution trends, generating a corrosion level and potential prediction map; and a spraying demand calculation module for calculating the corrosion level and potential prediction map based on the corrosion level and potential prediction map and the surface of the steel component. The system utilizes three-dimensional geometric information to calculate dynamic spraying requirements and generate a three-dimensional dynamic spraying requirement field. A trajectory planning and optimization module is used to collaboratively optimize the motion trajectory and spraying parameters of the spraying equipment based on the three-dimensional dynamic spraying requirement field, generating a dynamic spraying trajectory package. An execution and online monitoring module is used to execute the dynamic spraying trajectory package for spraying operations and monitor the paint film state after spraying online, generating an instantaneous film quality map. A real-time correction and iterative optimization module is used to compare the instantaneous film quality map with the three-dimensional dynamic spraying requirement field, generate a quality error map, and perform real-time correction and iterative optimization of the dynamic spraying trajectory package based on the quality error map.

[0017] Compared with the prior art, the present invention has the following advantages: This invention integrates multi-spectral visual information, including visible light, short-wave infrared, and thermal imaging, to perform cross-modal feature calculations on corrosion products on the surface of steel components. It surpasses the limitations of traditional single-vision methods that can only identify color and texture, and deeply analyzes the chemical composition and electrochemical activity intensity of corrosion. This enables a deep and accurate quantitative assessment of the corrosion state from macroscopic appearance to microscopic physicochemical essence, providing a solid data foundation for subsequent scientific protection.

[0018] This invention combines static grading of corrosion with dynamic evolution trend prediction based on electrochemical kinetics. It not only assesses the current severity of damage but also proactively identifies high-risk areas for future corrosion expansion. This predictive analysis shifts corrosion prevention strategies from passive repair to proactive prevention, generating differentiated and targeted protection needs based on the potential direction and speed of corrosion development, thereby achieving more effective protection for steel components throughout their entire lifecycle.

[0019] This invention constructs a complete closed-loop intelligent control system encompassing perception, planning, execution, and feedback. By monitoring the instantaneous film quality during the spraying process online and comparing the actual results with the requirements of dynamic planning in real time, it can dynamically correct subsequent spraying trajectories and parameters. This adaptive online correction capability effectively compensates for uncertainties in real-world working conditions, ensuring the uniformity and accuracy of the final anti-corrosion coating. Furthermore, by iteratively optimizing the core prediction model using historical data, the system possesses self-learning and continuous evolution capabilities.

[0020] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A structural architecture diagram of a multispectral vision-based steel component surface rust classification and spraying trajectory planning system provided in an embodiment of this application; Figure 2 A flowchart illustrating the method for grading corrosion on steel components and planning spraying trajectories based on multispectral vision, provided in this application embodiment; Figure 3 These are the surface temperature dynamic response curves under different corrosion states provided in the embodiments of this application.

[0023] Figure 4 This is a thermal map showing the instantaneous film quality distribution after spraying, provided in an embodiment of this application. Detailed Implementation

[0024] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0025] The method for grading corrosion on steel components and planning spraying trajectories based on multispectral vision provided in this application can be applied to, for example... Figure 1 In the multispectral vision-based steel component surface rust classification and spraying trajectory planning system 100 shown, such as... Figure 1 As shown, the system includes: a multispectral image acquisition module for acquiring synchronous multispectral image data of the steel component surface and generating a multispectral image dataset; a cross-modal feature analysis module for performing cross-modal feature fusion and calculation on the multispectral image dataset to generate a corrosion component and activity feature map; a corrosion grading and prediction module for grading and predicting corrosion based on the corrosion component and activity feature map using a preset model for extrapolating corrosion evolution trends, generating a corrosion level and potential prediction map; and a spraying demand calculation module for calculating dynamic spraying demand based on the corrosion level and potential prediction map and the three-dimensional geometric information of the steel component surface. The system generates a three-dimensional dynamic spraying demand field; a trajectory planning and optimization module is used to coordinately optimize the motion trajectory and spraying parameters of the spraying equipment based on the three-dimensional dynamic spraying demand field, generating a dynamic spraying trajectory package; an execution and online monitoring module is used to execute the dynamic spraying trajectory package to perform spraying operations and monitor the paint film status after spraying online, generating an instantaneous film quality map; a real-time correction and iterative optimization module is used to compare the instantaneous film quality map with the three-dimensional dynamic spraying demand field, generate a quality error map, and perform real-time correction and iterative optimization of the dynamic spraying trajectory package based on the quality error map.

[0026] like Figure 2As shown in the embodiments of this application, a method for grading corrosion on steel component surfaces and planning spraying trajectories based on multispectral vision is provided, including: acquiring synchronous multispectral image data of the steel component surface to generate a multispectral image dataset; performing cross-modal feature fusion and calculation on the multispectral image dataset to generate corrosion component and activity feature maps; based on the corrosion component and activity feature maps, performing grading and prediction through a preset model for inferring corrosion evolution trends to generate corrosion level and potential prediction maps; calculating dynamic spraying requirements based on the corrosion level and potential prediction maps and the three-dimensional geometric information of the steel component surface to generate a three-dimensional dynamic spraying requirement field; using the three-dimensional dynamic spraying requirement field as the target, co-optimizing the motion trajectory of the spraying equipment and spraying parameters to generate a dynamic spraying trajectory package; executing the dynamic spraying trajectory package to perform spraying operations and monitoring the paint film state after spraying online to generate an instantaneous film quality map; comparing the instantaneous film quality map with the three-dimensional dynamic spraying requirement field to generate a quality error map, and performing real-time correction and iterative optimization of the dynamic spraying trajectory package based on the quality error map.

[0027] It should be noted that by integrating multimodal visual information such as visible light, infrared, and thermal imaging, a deep physicochemical state analysis of the corrosion on the surface of steel components is performed. This not only identifies the current level of corrosion but also predicts its future evolution potential through a dynamic model. Combining this two-dimensional corrosion health state map with the three-dimensional geometric model of the steel component generates a non-uniform three-dimensional dynamic spraying demand field, which precisely defines the required protection level for different locations. The motion trajectory and spraying process parameters of the spraying robot are collaboratively optimized to form a set of dynamically changing execution instructions. During the spraying operation, online monitoring technology is used to perceive the instantaneous film-forming state of the paint film in real time and compare it with the preset dynamic spraying demand to generate an error signal. This error signal drives the intelligent decision-making model to proactively correct and dynamically adjust subsequent unexecuted trajectories and parameters in real time, ultimately constructing a complete closed-loop intelligent control process from accurate perception and predictive planning to adaptive execution.

[0028] In one possible implementation of the embodiments of this application, combined with Figure 2 The process of acquiring synchronous multispectral image data of the steel component surface and generating a multispectral image dataset includes: synchronously acquiring visible light images, short-wave infrared images, and thermal imaging sequence images of the steel component surface; performing spatial registration and temporal alignment on the acquired visible light images, short-wave infrared images, and thermal imaging sequence images; fusing the registered and aligned image data to construct a multispectral tensor field containing the color value, infrared absorption intensity, and thermal change rate information of each pixel, as the multispectral image dataset.

[0029] In some implementations, during the multispectral image data acquisition stage, a hardware synchronous trigger source drives the visible light sensor, short-wave infrared sensor, and long-wave infrared thermal imager to perform frame-synchronous acquisition. The engineering goal of acquiring synchronous multispectral image data of the steel component surface is to establish a joint sensing field encompassing geometric characterization and physical composition. This leverages the differences in sensitivity of different spectral bands to iron oxide composition, moisture content, and thermal diffusivity to overcome the accuracy limitations of single-modal images in identifying corrosion thickness. The visible light image is responsible for extracting pixel-level R, G, and B color features and spatial texture features. The short-wave infrared image typically operates in the 900nm to 1700nm band to extract the infrared absorption features of corrosion products. The thermal imaging sequence records the dynamic temperature feedback of the steel component after active thermal excitation. The synchronization error of the sensor array is suppressed to within 2ms. High-precision external trigger frequency control of each photoelectric conversion component ensures the consistency of the observation field during dynamic movement. After acquiring the original image sequence, spatial registration and temporal alignment are performed to eliminate the impact of physical optical axis offset, focal length misalignment, and shutter phase difference on data consistency. First, a projection transformation matrix is ​​generated using a multispectral calibration board to normalize the short-wave infrared coordinate system and the thermal imaging coordinate system into the visible light pixel space, ensuring that the mapping error of the same physical coordinate point in different bands is less than 0.5 pixels. Then, data frames are extracted based on time-stamped signals, extracting the thermal distribution slices corresponding to the visible light trigger moment from the continuously acquired thermal imaging video stream, eliminating multimodal feature misalignment caused by device movement through time synchronization. For the registered data, normalization processing addresses the issue of large differences in dynamic range between different sensors, providing a dimensionally consistent base map for subsequent channel fusion. During the generation of the multispectral image dataset, the processed spatiotemporally aligned data is fused into a multispectral tensor field through dimension stacking. Discrete sensor physical quantities are converted into computable computational units, reflecting the electrochemical characteristics of the corrosion layer through channel correlation. Each element in this tensor field is composed of the spectral absorption vector, color vector, and temperature response component at that pixel coordinate. For the extraction of thermal change rate information, continuous inversion of the thermal imaging sequence is performed based on the temporal difference method to capture the characteristics of heat diffusion obstruction caused by the microporous structure within the corrosion layer. Multispectral tensor field It is constructed by the following formula: ; in, Represents a multispectral tensor field. This represents the color-corrected intensity tensor of the three visible light channels. The infrared absorbance tensor represents the infrared absorbance under a specific spectral line. This represents the thermal feature evolution tensor after spatiotemporal alignment. The above operation is not a numerical weighted addition, but rather refers to the concatenation across multidimensional array dimensions and the point-by-point association of physical quantities. (For...) The quantization of the components is obtained by extracting the instantaneous derivatives of the thermal imaging sequence: ; in, This represents the rate of thermal change of the current pixel. This represents the grayscale response value of the current frame at that pixel. Indicates backwards The grayscale response value at the same coordinates corresponding to the frame. Indicates the first Frame and the The time difference between frames. and Preprocessing is performed using the non-uniformity correction parameters of the infrared camera to make it conform to a physically linear thermal radiation response. The time difference is calculated from the system time deviation of the hardware counter. This time-difference analysis method can accurately obtain the dynamic thermal inertia of the steel component surface excited by temperature rise, serving as a key indicator for determining the depth and adhesion of the rust layer. The results will form a multispectral image dataset with a three-dimensional array structure, directly mapping the physical degradation state of the steel component surface, providing data support for subsequent rust level classification. Figure 3 This study demonstrates the temperature evolution process of steel component surfaces under active thermal excitation. By comparing the thermal hysteresis differences between healthy areas and areas with different degrees of corrosion, it provides physical criteria for extracting temperature change acceleration characteristics and determining the depth of the corrosion layer.

[0030] For example, in the acquisition of typical rusted areas on the surface of a steel box girder of a cross-sea bridge, a hardware synchronous trigger source drives a visible light sensor, a short-wave infrared sensor, and a long-wave infrared thermal imager to perform frame-synchronized acquisition, ensuring that the synchronization error is suppressed to within 1.5ms to maintain the consistency of the observation field of view. After acquiring the original image sequence, a preset projection transformation matrix is ​​used to normalize the short-wave infrared and thermal imaging coordinate systems to the visible light pixel space, achieving a spatial registration error of less than 0.4 pixels. Time alignment is then performed by extracting thermal distribution slices corresponding to the visible light trigger moment using time-stamped signals. Subsequently, a multispectral tensor field is constructed by dimension stacking. For a typical pixel visible light component After color correction, the normalized color vector is extracted, including the short-wave infrared component. The infrared absorbance was measured to be 0.85 at a specific spectral line of 1400 nm. This was based on the thermal characteristic evolution component. Quantization is performed using temporal difference calculations on the thermal imaging sequence after active thermal excitation. If the point is in the current frame... grayscale response value The value is 155, backtracking forward. The grayscale response value of the corresponding coordinates of the frame The time difference between the two frames is 120. The value is 0.2s. Substituting this into the formula... The rate of thermal change at that point was calculated. Finally, the physical quantities of each channel are correlated and stitched together point by point to form a multispectral image dataset with a three-dimensional array structure. This dataset is then analyzed based on the dynamic thermal inertia of each pixel. Value and infrared absorption rate The coupling of values ​​accurately maps the depth of the corrosion layer and the electrochemical degradation state at that location.

[0031] In one possible implementation, combining Figure 2 The process of performing cross-modal feature fusion and calculation on the multispectral image dataset to generate a corrosion component and activity feature map includes: extracting the absorption spectrum of each pixel from the multispectral image dataset and matching it with a database storing standard spectra of corrosion products to obtain the component ratio characteristics of the corrosion products; combining the thermal imaging sequence data in the multispectral image dataset to analyze the temperature anomaly diffusion mode and obtain the corrosion activity intensity characteristics; and fusing the component ratio characteristics and the corrosion activity intensity characteristics to generate a corrosion component and activity feature map characterizing the corrosion type composition and electrochemical activity intensity of each pixel.

[0032] In some implementations, extracting pixel absorption spectra from multispectral image datasets allows for precise identification of corrosion products on steel component surfaces using the optical fingerprint of the material. The difference in sensitivity of the infrared band to hydrated iron oxide and non-hydrated oxides enables quantitative stripping of chemical components. Pixel-by-pixel scanning of short-wave infrared channel data in the multispectral tensor field removes the influence of ambient background stray light by constructing a spectral envelope, extracting spectral curves containing the position, depth, and width of reflectance characteristic peaks. The extracted real-time spectral vectors are then compared with the standard spectral features of standard corrosion products in a pre-existing database, such as iron hydroxide, hematite, and magnetite. The spectral matching process calculates the spectral angle between measured and standard values ​​to obtain the corresponding component proportion characteristics. This proportion reflects the density and chemical stability of the corrosion layer; for example, a higher proportion of iron hydroxide typically indicates a more porous corrosion layer. In practical engineering environments, spectral reflectance typically fluctuates between 0.15 and 0.85, and a narrowband filter with a sampling frequency of 50 Hz ensures the purity of the spectral data. Combining thermal imaging sequence data analysis with temperature anomaly diffusion patterns helps detect the physical structural integrity beneath the rust layer. The intensity of electrochemical activity is assessed through the dynamic evolution of heat flow under active thermal excitation. Analyzing the surface temperature rise curves of steel components after heating reveals that active rust regions exhibit significant thermal hysteresis due to their porous structure and moisture accumulation, resulting in reduced local thermal diffusivity. Temperature change acceleration features are obtained by extracting the second derivative of the thermal sequence images along the time axis, serving as the rust activity intensity feature. High activity intensity typically corresponds to areas of severe electrochemical corrosion, where the energy exchange rate deviates by 15% to 30% from the stable region. This dynamic analysis method effectively identifies under-painting corrosion or hidden pitting corrosion that is invisible to the naked eye, providing a deep-dimensional activity criterion for graded prediction. The feature map generated by fusing component ratio features and rust activity intensity features constructs a spatially continuous surface risk matrix, coupling chemical properties with physical activity to provide target weight guidance for the spraying trajectory. Spatial dimension merging of cross-modal features is performed, multiplying and weighting the material composition information and dynamic activity information at each spatial pixel. The fusion process employs a nonlinear normalization algorithm to eliminate the dimensional differences between spectral intensity and temperature change rate, generating a corrosion component and activity feature map characterizing the corrosion type and electrochemical activity intensity of each pixel. The specific feature fusion calculation is expressed by the following formula: ; in, This represents the fusion evaluation value of each pixel in the generated feature map. Represents the component proportion eigenvector. This represents the characteristic value of corrosion activity intensity after normalization. This represents the weighting coefficient indicating the influence of chemical components on corrosion risk. This represents the weighting coefficient of the impact of electrochemical activity on risk. It is obtained by taking the reciprocal of the spectral matching residual; the larger the value, the higher the matching degree. It is obtained by mapping the instantaneous heat flux variance of the pixel through a logarithmic function. and The values ​​of both are in the range of 0 to 1 and the arithmetic sum of both is always 1. The specific value depends on the pre-set corrosion resistance standard level of the service environment of the steel.

[0033] For example, corrosion detection at a bridge node involves cross-modal computation on a multispectral image dataset. First, short-wave infrared channel data is scanned pixel-by-pixel to remove stray light interference, and a spectral curve containing reflection characteristic peaks is extracted. The component proportion characteristics are then calculated by subtracting the reciprocal of the spectral angle residual between the real-time spectral vector at a detection point and the standard spectrum of iron hydroxyl oxide in the database. The value of 0.85 indicates that the rust products at this location are relatively loose. Subsequently, the temperature rise acceleration was analyzed using thermal imaging sequence data. The second derivative of the time axis was extracted from the surface temperature rise curve after heating. If the point exhibits significant thermal hysteresis due to water retention in the pores, the normalized activity intensity characteristic obtained by logarithmic mapping of its instantaneous heat flux variance was determined. The value is 0.78, reflecting strong electrochemical activity; finally, feature space dimensions are merged, and chemical component weights are set. With electrochemical activity weight Substitute into the formula The fusion evaluation value was calculated. This generates a map of rust components and activity characteristics stored in the form of a multi-channel tensor. The high-score points are highlighted in the map to accurately map the distribution of high-risk active rust diseases.

[0034] In one possible implementation, combining Figure 2 The step of classifying and predicting corrosion evolution trends using a pre-set model to generate a corrosion level and potential prediction map includes: analyzing the spatial distribution morphology of corrosion products in the corrosion component and activity characteristic map, wherein the spatial distribution morphology includes the boundary irregularity and porosity of the corrosion area; inputting the spatial distribution morphology, component ratio, and corrosion activity intensity into the pre-set model for predicting corrosion evolution trends, wherein the model for predicting corrosion evolution trends is a nonlinear mapping model based on electrochemical kinetics; in the model for predicting corrosion evolution trends, the corrosion potential is determined using the component ratio, the charge transfer impedance is calculated by combining the corrosion activity intensity and porosity, and the penetration rate of corrosion in the surface and deep layers of the steel component substrate is predicted using an anisotropic diffusion algorithm; and the model for predicting corrosion evolution trends outputs a corrosion level and potential prediction map containing the current corrosion severity classification and future expansion vector distribution.

[0035] In some implementations, the engineering objective of analyzing spatial distribution morphology is to identify the evolution degree and potential expansion paths of pitting corrosion by quantifying the geometric complexity of the rust region. Multi-scale edge detection is performed on rust components and activity feature maps to calculate the geometric dimension of the rust region's edge, yielding the boundary irregularity. Boundary irregularity reflects the unsteady expansion rate of the rust region in the horizontal direction; a higher dimension indicates a more intense local galvanic cell reaction. Simultaneously, texture analysis operators are used to identify the porosity distribution on the surface of the rust product layer, and the porosity density is obtained by statistically analyzing the porosity percentage per unit area. Porosity density characterizes the ease with which corrosive media penetrate the rust layer and reach the metal substrate. Irregularity parameters typically range from 1.1 to 1.8, while porosity density exhibits significant clustering characteristics under specific corrosion cycles. Inputting these features into a model for predicting rust evolution trends establishes a physical-logical mapping from static characterization to dynamic evolution, enabling a nonlinear assessment of corrosion depth. The model for predicting corrosion evolution is a nonlinear mapping model based on electrochemical kinetics, with an integrated corrosion current conservation equation. Extracted component proportions are mapped to corrosion potentials under specific corrosion product combinations; this potential represents the physical pressure difference driving the electrochemical reaction. Combining corrosion activity characteristics and porosity distribution, the charge transfer impedance is calculated using the polarization curve inversion method. The charge transfer impedance represents the resistance to electron transfer; a smaller value indicates a higher corrosion rate. The model was fitted offline using extensive electrochemical impedance spectroscopy experimental data to establish a correlation function between characteristic parameters and physical impedance. An anisotropic diffusion algorithm is used to predict the differences in corrosion penetration rates along the depth and horizontal propagation directions of the substrate. In the model predicting corrosion evolution, a non-uniform spatial step size is used to simulate the advancement of the corrosion front. This algorithm uses a gradient tensor to adjust the weights of the diffusion operator, preferentially propagating corrosion potential along the direction of lowest resistance and largest potential gradient. The anisotropic diffusion algorithm simulates the rapid penetration of corrosion at metal grain boundaries and the slow diffusion at dense products by setting a diffusion direction threshold. By iteratively calculating the physical parameters of each grid node, the model generates a corrosion propagation vector at each pixel, which contains the potential increase in corrosion thickness and the direction of horizontal spread per unit time. The corrosion penetration rate in the depth direction is calculated as follows: ; in, This indicates the penetration rate of the current pixel. Represents the electrochemical kinetic conversion constant. This represents the corrosion potential determined by the compositional characteristics. This represents the pore distribution density value extracted at that pixel. It represents the charge transfer impedance. The Faraday constant, obtained from experiments, and the material density are combined to form a quantity whose dimensions match the physical law of depth change per unit time. and All parameters are real-time parameters calculated based on multispectral data, and It is subject to nonlinear correction based on the activity intensity characteristics. In this formula, As a driving force, with pore density The product of reflects the superposition effect of the corrosive medium and the driving force, while and The division operator and Ohm's law in the logic of electrochemical systems.

[0036] For example, the prediction of ulcerative corrosion at the bottom of a steel structure beam was achieved by performing multi-scale edge detection and texture analysis on the feature map. The result showed that the boundary irregularity of the local corrosion area was 1.65, and the porosity distribution density per unit area was [not specified]. The value was 18%. These characteristics were then input into a nonlinear mapping model based on electrochemical kinetics, and the corrosion potential was obtained using the component ratio characteristic mapping. It is -0.45V, and the charge transfer impedance is corrected for the activity intensity. Measured as Set the electrochemical kinetic conversion constant. for Substitute into the permeation rate formula Calculated depth permeation rate The model further incorporates anisotropic diffusion algorithms, using gradient tensors to simulate the penetration trend of the corrosion front at metal grain boundaries, generating a potential prediction map that includes the current severity level and future expansion vectors.

[0037] In one possible implementation, combining Figure 2Based on the corrosion level and potential prediction map and the three-dimensional geometric information of the steel component surface, the dynamic spraying requirements are calculated, and a three-dimensional dynamic spraying requirement field is generated. This includes: determining the basic paint film thickness requirement based on the corrosion level and potential prediction map; acquiring three-dimensional point cloud data of the steel component surface and calculating the surface curvature and normal vector information; calculating the paint adhesion efficiency and sagging trend using a model for simulating the fluid behavior of paint on different morphological surfaces, combined with the basic paint film thickness requirement and the surface curvature and normal vector information, wherein the model for simulating the fluid behavior of paint on different morphological surfaces is a semi-empirical fluid physics model; in the model simulating the fluid behavior of paint on different morphological surfaces, the paint loss caused by rebound is calculated using the surface curvature, the effective adhesion rate is calculated using the angle between the normal vector and the spray gun spray direction, and the sagging risk index is evaluated based on the ratio of the component of gravity on the inclined plane to the paint viscosity; and the basic paint film thickness requirement is compensated and adjusted based on the paint adhesion efficiency and sagging trend, and a three-dimensional dynamic spraying requirement field containing the target paint film thickness, spray residence time, and allowable flow rate is output.

[0038] In some implementations, determining the required base paint film thickness based on the rust level and potential prediction map quantifies corrosion risk into initial protective layer design parameters, ensuring differentiated initial coating protection for areas with different rust states. The rust level and extended potential value of each pixel in the prediction map are converted into an initial target dry film thickness value through a preset nonlinear mapping function. This function is formulated based on the performance of the anti-corrosion coating and engineering standards; typically, the required base paint film thickness for high-risk areas is 30% to 50% higher than that for conventional areas. Simultaneously, high-density three-dimensional point cloud data of the steel component surface is acquired through laser scanning or structured light technology. This point cloud data is then meshed, and a neighborhood surface fitting algorithm is used to calculate the surface curvature and normal vector information of each triangular facet, providing geometric input for subsequent hydrodynamic corrections. Surface curvature describes the degree of surface bending, while the normal vector defines the local orientation of the surface. The engineering purpose of using a model to simulate the fluid behavior of coatings on surfaces with different morphologies is to predict and compensate for uneven coating deposition caused by complex geometries, avoiding quality defects caused by over- or under-spraying. This model is a semi-empirical fluid physics model, the core of which is to establish the coupling relationship between the trajectory of paint particles and surface geometry. In the model, surface curvature is used to quantify the rebound effect of paint particles impacting a convex surface; the greater the curvature, the higher the paint loss due to the rebound. Simultaneously, the model calculates the angle between the theoretical spray direction of the spray gun and the surface normal vector. The cosine of this angle is directly related to the proportion of paint effectively adhering to the surface per unit time, i.e., the effective adhesion rate. For assessing sagging tendency, the model generates a dimensionless sagging risk index by calculating the ratio of the tangential component of gravity on the inclined surface to the preset viscosity parameter of the paint. If this index exceeds a critical threshold, such as 0.8, it means that there is a high sagging risk in that area. Compensating for the base paint film thickness requirement generates a physically achievable final spraying target that considers all process influencing factors. Based on the calculation results of paint adhesion efficiency and sagging tendency, the base paint film thickness requirement is iteratively corrected. Areas with low adhesion efficiency require increased spray volume to achieve the target thickness, while areas with high sagging risk require limiting the wet film thickness of a single spray. The final target paint film thickness The correction calculation is performed using the following formula: ; in, This indicates the target paint film thickness after compensation adjustment. This indicates the required base paint film thickness based on the corrosion level. This indicates the angle between the direction of the spray jet and the surface normal vector. This represents the average curvature of a local surface. This represents the rebound coefficient, which is related to the physical properties of the coating. The calculation is based on the planned spray gun posture and the obtained surface normal vector information. It is obtained by real-time calculation from 3D point cloud data. This is an empirical constant, calibrated through spraying experiments under specific coating and process parameters. It ensures that on inclined or curved surfaces, more coating is applied to compensate for losses caused by geometric projection and bounce.

[0039] For example, the calculation of the spraying requirements for the arc-shaped transition zone on the web of a steel structure bridge first determines that the area is in a high-risk state based on the corrosion level and potential prediction map, and then sets the basic paint film thickness requirements. for The surface curvature at that location was then calculated using 3D point cloud data. for And determine the angle between the spray jet direction and the surface normal vector based on the planned posture. The angle is 15°; a semi-empirical fluid physics model is used to introduce the experimentally calibrated coating rebound coefficient. Substitute into the compensation adjustment formula The final target paint film thickness is calculated through iterative correction. Meanwhile, the model assesses the sagging risk index based on the ratio of the tangential component of gravity to the viscosity of the coating. If the index does not exceed the critical threshold of 0.8, the output is a three-dimensional dynamic spraying demand field that includes the target thickness, theoretical spraying residence time, and maximum allowable flow rate.

[0040] In one possible implementation, combining Figure 2 Taking the three-dimensional dynamic spraying demand field as the target, the motion trajectory and spraying parameters of the spraying equipment are collaboratively optimized to generate a dynamic spraying trajectory package. This includes: converting the three-dimensional dynamic spraying demand field into an objective function to achieve uniform distribution of paint deposition; using the objective function as a constraint, collaboratively optimizing the spatial path, moving speed, paint output, atomization pressure, and spray gun attitude angle parameters of the spraying equipment; and encapsulating the optimized spatial path, time series, six-degree-of-freedom attitude, and spraying equipment parameters to generate a dynamic spraying trajectory package.

[0041] In some implementations, transforming the 3D dynamic spraying demand field into an objective function constitutes a quantified optimization problem, enabling the motion behavior of the spraying equipment to directly serve the final paint film quality target. The target paint film thickness value at each spatial point in the 3D dynamic spraying demand field is mapped to the paint volume that the spraying equipment should deposit at that point per unit time through a paint deposition model. This model considers factors such as the spray gun's spray pattern, paint flow rate, and spraying distance, typically using a Gaussian or elliptic distribution model. The final objective function is constructed as a minimization function, whose value is the root mean square error between the desired paint film thickness distribution and the actual simulated deposition thickness distribution. The smaller the error, the closer the planned trajectory and parameters are to the ideal state. Collaborative optimization of the multidimensional parameters of the spraying equipment addresses the strong coupling problem between robot path planning and spraying process parameter settings, seeking to meet paint film quality requirements while balancing operational efficiency and equipment motion smoothness. The optimization algorithm uses the objective function as a constraint and employs heuristic search methods such as multi-objective genetic algorithms or particle swarm optimization to globally optimize the six degrees of freedom of the spraying equipment: the XYZ coordinates on the spatial path, the ABC angles of the spray gun posture, and process parameters such as movement speed, paint output, and atomization pressure. Optimization of the spatial path aims to ensure the spray gun maintains the optimal spraying distance and angle with the workpiece surface. Optimization of the movement speed is directly related to the paint deposition per unit area; the speed is reduced in areas requiring greater thickness and increased conversely. The coordinated adjustment of paint output and atomization pressure affects the spray pattern and the atomization effect of the paint, thus influencing the coating uniformity. During the optimization process, kinematic constraints, such as joint velocity and acceleration limits, are set to ensure that the generated trajectory is physically executable. Encapsulating the optimization results into a dynamic spraying trajectory package generates a complete set of operation instructions that can be directly parsed and executed by the robot controller. After the optimization algorithm converges, the output is a series of discrete path points. Each path point is associated with a precise timestamp, a set of six-DOF attitude parameters (e.g., X, Y, Z, A, B, C in Cartesian coordinates), and a set of instantaneous spraying equipment parameters (e.g., flow rate). ,pressure These discrete data points are interpolated and smoothed to generate continuous spatial paths and time series. Then, all information is encoded and encapsulated according to the communication protocol format of the robot control system to form a dynamic spraying trajectory package. This trajectory package differs from traditional fixed-parameter paths; its internal process parameters dynamically change with spatial position. Objective function The structure is as follows: ; in, Let the objective function be the one to be minimized. Represents discrete points on the surface of a steel component The target paint film thickness is provided by the three-dimensional dynamic spraying demand field. This indicates the results calculated using a coating deposition model at discrete points. The actual thickness of the paint film deposited at that location. It is the total number of discrete points on the surface. The calculation depends on a series of variables to be optimized, such as the path, speed, attitude, and spraying parameters of the spraying equipment. It is a complex nonlinear function, and its specific form is: ; in, This represents the coating deposition rate function, whose input is time. The position of the spray gun tip at that time Movement speed Paint output atomization pressure and spray gun posture The integration interval is the entire spraying operation time. For coating efficiency, For point Radial distance from the center of the spray pattern To match the atomization pressure and the diffusion coefficient related to the spraying distance, This is for compensation of the angle between the jet direction and the surface normal. The sampling period of the control system. By minimizing The optimization algorithm can drive all variables to be optimized to adjust in the direction that best satisfies the target thickness distribution.

[0042] For example, the spraying plan for the welding area of ​​a steel component with a complex curved surface involves determining the target paint film thickness at each spatial point in a three-dimensional dynamic spraying demand field. Transform into an optimization objective, if a certain discrete point The target thickness is A collaborative optimization algorithm is used to globally optimize the six-degree-of-freedom attitude and process parameters of the spraying equipment. In the objective function... In the calculation, the actual deposition thickness at this point is calculated using a coating deposition model. Set the spraying efficiency At a certain sampling time Paint output Due to atomization pressure Determined diffusion coefficient It is 15mm, the radial distance from the center of the spray pattern. 10mm, spray gun posture Defined angle of incidence The sampling period of the control system is 20°. The value is 0.02s. Substitute this value into the coating deposition rate function to calculate the contribution at this moment. This is obtained by offline summation of the entire operation time. If the simulated total thickness for The root mean square error term generated at this point is The optimization algorithm minimizes the total error. The driving parameters converge, and the generated dynamic spraying trajectory package encapsulates a complete instruction sequence containing the timestamp of each path point, the six degrees of freedom posture, and the synchronous triggering spraying parameters, ensuring that the robot controller can achieve precise adaptive spraying with "one policy for each location".

[0043] In one possible implementation, combining Figure 2 The process of executing the dynamic spraying trajectory package for spraying and monitoring the paint film state after spraying online to generate an instantaneous film quality map includes: controlling the spraying equipment to execute the dynamic spraying trajectory package for spraying; simultaneously, using a broadband light source to illuminate the surface of the undried paint film and receiving the interference fringe image formed by the reflected light; analyzing the interference fringe image, inverting and calculating the wet film thickness distribution and leveling state of the paint film, and generating an instantaneous film quality map.

[0044] In some implementations, controlling the spraying equipment to execute a dynamic spraying trajectory package transforms optimized virtual instructions into actual spraying actions in the physical world, achieving automated protective treatment of steel component surfaces. The dynamic spraying trajectory package is loaded into the robot controller, which parses the instructions line by line, converting path points and postures in Cartesian space into rotation angles of each joint through inverse kinematics calculations. Simultaneously, the controller sends real-time spraying parameter instructions to the spraying execution unit via a high-speed bus, precisely controlling the paint output of the servo pump and the atomization pressure of the pneumatic pressure regulating valve. The entire execution process is performed under closed-loop servo control, ensuring that the tracking error between the robot's actual motion trajectory and posture and the planned values ​​is less than 0.5 mm, and the flow control accuracy is maintained within ±2% of the set value. Simultaneous online monitoring of the paint film state using optical interferometry allows for real-time acquisition of the physical characteristics of the wet film after spraying, providing immediate feedback data for process quality control and dynamic correction. An optical measuring head is integrated into the end effector of the spraying robot, containing a broadband light source with a wavelength range of 400 nm to 1000 nm and a high-speed CCD camera. A light source illuminates the freshly sprayed, uncured paint film surface at a specific angle. Because the light is reflected at the air-coating interface and the coating-substrate interface, these two reflected beams interfere, forming an image of alternating bright and dark interference fringes. A CCD camera captures these interference fringes at a rate of at least 100 frames per second. The advantages of this method are its non-contact, high-speed nature and ability to instantaneously image dynamically flowing wet films. Analyzing the interference fringe image to calculate the wet film thickness involves converting the optical signal into quantified physical dimensional information, thus obtaining a precise paint film thickness distribution. Fourier transform spectral analysis is used to process the captured interference fringe image. A one-dimensional Fourier transform is performed on the interference signal at each pixel in the image to obtain the spectrum in the frequency domain. The phase difference between different frequency components in the interference spectrum is directly related to the optical path difference, which depends on the wet film thickness and the refractive index of the coating. By extracting the dominant frequency peak position in the spectrum and combining it with the known refractive index parameters of the coating, the wet film thickness value corresponding to that pixel is accurately inverted. Simultaneously, by analyzing the clarity and morphological changes of interference fringes, the leveling status of the paint film is assessed. For example, blurred or distorted fringes may indicate uneven surface tension or the onset of sagging. Wet film thickness. The calculation formula is: ; in, This represents the calculated wet film thickness. This represents the wavelength corresponding to the dominant frequency peak in the interference spectrum. It is the order of the interference fringes at that wavelength. It is the wet film coating at wavelength The refractive index below, It is the angle of incidence of the light source. It is obtained by finding the peaks after performing a Fourier transform on the interference signal. Determining the value usually requires combining multi-wavelength analysis or phase unwrapping algorithms to eliminate ambiguity. It is a material constant and needs to be calibrated in advance through experiments. These are the fixed geometric parameters of the measurement system. This formula, based on the fundamental physical principles of thin-film interferometry, ensures the accuracy of thickness inversion. Figure 4 The instantaneous film quality distribution thermal map, generated based on optical interference fringe inversion, visually presents the consistency of wet film thickness in the working area with pixel-level precision. This map is used for subsequent comparison with the demand field and to trigger real-time correction commands.

[0045] For example, in the on-site spraying operation and film formation monitoring at the corner of a bridge steel component, a robot controller analyzes trajectory package instructions in real time and drives the servo mechanism to ensure that the motion trajectory tracking error is less than 0.5 mm. Simultaneously, an optical measuring head monitors the freshly sprayed wet film online. A broadband light source with wavelengths from 400 nm to 1000 nm illuminates the paint film surface, and a CCD camera captures interference fringe images at a rate of 100 frames per second. The interference signal is extracted using Fourier transform spectral analysis. For a typical pixel within the monitoring area, peak-finding processing is used to obtain the wavelength corresponding to the dominant frequency peak in the interference spectrum. The wavelength is 650 nm, and the interference fringe order is determined using a phase unwrapping algorithm. Given that the refractive index of the epoxy coating is 1, The incident angle of the light source in the measurement system is 1.52. The angle is 30°. Substitute this into the wet film thickness calculation formula. The instantaneous wet film thickness at that point was calculated. A pseudo-color instantaneous film formation quality map is generated based on the thickness values ​​of each pixel and precisely registered with the surface coordinates. This map intuitively represents the physical thickness distribution and leveling state of the current region and serves as the benchmark input for quality error comparison.

[0046] In one possible implementation, combining Figure 2The process involves comparing the instantaneous film-forming quality map with the three-dimensional dynamic spraying demand field to generate a quality error map, and then performing real-time correction and iterative optimization of the dynamic spraying trajectory package based on the quality error map. This includes: comparing the wet film thickness distribution in the instantaneous film-forming quality map with the target paint film thickness in the three-dimensional dynamic spraying demand field point by point to generate a thickness error distribution map; inputting the thickness error distribution map as a state input to a preset intelligent decision-making model, wherein the intelligent decision-making model is a reinforcement learning decision-making model; in the intelligent decision-making model, by extracting the mean and variance features of the error in the thickness error distribution map, mapping and generating trajectory parameter correction amounts for subsequent unexecuted paths, wherein the trajectory parameter correction amounts include a moving speed correction ratio, a paint output correction ratio, and a path overlap rate adjustment value; and updating the execution parameters of subsequent path segments in the dynamic spraying trajectory package in real time according to the adjustment instructions generated by the trajectory parameter correction amounts.

[0047] In some implementations, comparing the instantaneous film-forming quality map with the three-dimensional dynamic spraying demand field achieves online closed-loop control of spraying quality. By quantifying the deviation between the actual film-forming effect and the theoretical target, precise error signals are provided for real-time correction. First, the real-time acquired instantaneous film-forming quality map is precisely registered spatially with the pre-stored three-dimensional dynamic spraying demand field. Then, the two registered data fields are subtracted point-by-point to generate a thickness error distribution map that intuitively reflects the spraying quality deviation. Positive values ​​in the map indicate overspray, and negative values ​​indicate underspray; the distribution pattern reveals the systematic or random nature of the error. Inputting the thickness error distribution map into a preset intelligent decision-making model utilizes the nonlinear mapping capability of artificial intelligence algorithms to establish an intelligent correlation between observation errors and control commands, enabling adaptive adjustment of the spraying process. This intelligent decision-making model is an offline-trained reinforcement learning decision-making model. Instead of directly inputting the entire error map, statistical features are extracted from the error distribution corresponding to the recently sprayed local area, and the mean and variance of the error are calculated. The mean error reflects the overall thickness deviation trend in the region, while the variance error characterizes the uniformity of the paint film. These two features constitute a low-dimensional state vector, serving as input to the reinforcement learning model. Within the intelligent decision-making model, the input state vector is mapped to an action vector, i.e., the trajectory parameter correction, through its policy network. This correction is not applied to the completed path but is proactively applied to subsequent path segments. The trajectory parameter correction specifically includes a movement speed correction ratio, a paint output correction ratio, and a path overlap adjustment value. For example, when persistent underspray is detected, i.e., the mean error is negative, the model may output a negative speed correction ratio, instructing the robot to reduce subsequent movement speed to increase paint deposition. When the variance error is large, the model may increase the path overlap adjustment value to improve uniformity. The correction ratio is typically limited to between -15% and +15% to ensure the stability of the control system. The trajectory parameter update process can be represented by the following formula: ; in, It represents the set of action vectors output by the model, i.e., trajectory parameter corrections. This represents the trained policy network in the intelligent decision-making model. This represents the state vector extracted from the thickness error distribution map, whose main components are the mean and variance of the error. Let... For the original planned speed, The updated parameters are based on the original planned paint output. and Calculated in the following way: ; ; in, and It is an action vector The system includes components for movement speed correction and paint output correction. These adjustment commands are sent to the robot controller in real time, covering the preset parameters for the corresponding path segments in the dynamic painting trajectory package.

[0048] For example, in the closed-loop quality control of a complex curved surface spraying operation on a steel component, the instantaneous film quality map obtained from online monitoring is spatially registered with the three-dimensional dynamic spraying demand field and subtracted point by point to generate a thickness error distribution map. If the average wet film thickness measured in a local area that has just been sprayed is... The target thickness requirement is The average thickness error of this region is then calculated to be... The variance is The state vector containing the mean and variance. The input is fed into a pre-defined reinforcement learning decision model, and processed through a policy network. Mapping is performed to obtain action vectors. Assuming the trajectory parameter correction amount output by the model is the same as the speed correction ratio. and paint output correction ratio For the path segments that will be executed next, if their originally planned speed... The speed is 0.4 m / s, and the original planned paint output is... If the flow rate is 280 ml / min, then the corrected execution parameters are calculated based on the updated formula: New speed New paint output These adjustment instructions are sent to the robot controller in real time and override the original trajectory package parameters.

[0049] In one possible implementation, combining Figure 2 The intelligent decision-making model also performs offline training based on historical operation data to optimize the model used to predict the corrosion evolution trend and the model used to simulate the fluid behavior of coatings on different morphological surfaces. This includes: acquiring a historical causal chain sample set consisting of historical quality error maps, corresponding corrosion component and activity feature maps, and corrosion level and potential prediction maps; using the gradient descent algorithm, with the goal of optimizing the loss function between prediction deviation and actual thickness error, adjusting the neuron weights inside the model for predicting the corrosion evolution trend to optimize the mapping accuracy of the model for predicting the corrosion evolution trend to the corrosion expansion potential; and iteratively correcting the curvature influence coefficient and sagging influence coefficient in the model for simulating the fluid behavior of coatings on different morphological surfaces by analyzing the statistical correlation of thickness deviation under specific geometric features in the historical quality error map, thereby achieving adaptive compensation optimization of the physical behavior of coating deposition.

[0050] In some implementations, acquiring and utilizing historical causal chain sample sets for offline training enables the self-evolution of the system's core prediction model. Through in-depth mining of historical operation data, the intrinsic parameters of the models used to predict rust evolution trends and to simulate the fluid behavior of coatings on different surface morphologies are corrected and optimized, thereby improving the overall accuracy of the system from perception to execution. Complete data records for each operation are continuously accumulated in the background database. The initial rust component and activity feature map, the rust level and potential prediction map generated based on this map, and the final quality error map generated after spraying are considered as complete causal units, forming the historical causal chain sample set. Optimizing the model predicting rust evolution trends using the gradient descent algorithm corrects the model's systematic bias in assessing rust development potential. A loss function is constructed with the goal of minimizing the correlation between the model's prediction results and the final actual spraying error. If the model predicts a region with high expansion potential, leading the system to plan a thicker paint film, but the final quality error map shows that the region is still under-sprayed, this indicates that the model's assessment of the expansion potential for this type of rust is still too low. The gradient descent algorithm gradually reduces prediction bias by calculating the partial derivative of the loss function with respect to the weights of neurons within the model and fine-tuning the weights in the opposite direction of the gradient. The process of updating the model weights can be represented as: ; in, This represents the updated neuron weight matrix. This is the current weight matrix. It is the learning rate, a hyperparameter that controls the update step size, and its value is usually between 0.001 and 0.01. It is a loss function Regarding weight The gradient. Loss function. The thickness compensation requirement is calculated by comparing the difference between the predicted incremental coating demand in the rust level and potential prediction map and the actual thickness compensation requirement in the quality error map. Iterative correction of the model simulating the fluid behavior of coatings on different surface morphologies, based on historical data analysis, enables this semi-empirical physical model to adapt to the complex deposition behavior under different coating rheological properties and specific geometric features. Regions with specific geometric features, such as acute angles, grooves, or large-curvature convex surfaces, are selected from the historical quality error map, and the thickness deviations in these regions are statistically analyzed. By calculating the average thickness error under specific curvature or normal vector angles, the accuracy of the preset curvature influence coefficient and sagging influence coefficient in the model can be evaluated. If overspray is found to be prevalent in a region within a certain curvature range, the curvature influence coefficient corresponding to that curvature range will be increased accordingly, thereby reducing the theoretical coating demand in that region in future trajectory planning. The iterative correction of the coefficients is as follows: ; in, These are corrected coefficients, such as the curvature influence coefficient. This is the current coefficient. It is the iteration rate factor. This is the average normalized thickness error statistically obtained under this geometric feature. Through continuous negative feedback adjustment, the fitting accuracy of the fluid behavior model to the real physical process is continuously improved.

[0051] For example, in the phased model optimization of a cross-sea bridge steel structure corrosion protection project, after completing multiple batches of operations, an initial feature map, potential prediction map, and final quality error map are extracted from the background database to form a historical causal chain sample set. For the model used to predict the corrosion evolution trend, the prediction bias is minimized using the gradient descent algorithm, and the weights of the neurons within the current model are... Set the learning rate to 0.45. The value is 0.01, and the loss function is calculated. gradient of weights The value is 2.5, and the weight update formula is used. Calculate the updated weights To correct the assessment bias of the potential for expansion of specific rust types, for models used to simulate the fluid behavior of coatings on different surface morphologies, large curvature convex regions were selected for statistical analysis, such as curvature... The average normalized thickness error under this feature The current curvature influence coefficient is 0.15. Set the iteration rate factor to 0.12. The value is 0.2, which is then substituted into the correction formula. Obtain the corrected coefficients This negative feedback adjustment improves the accuracy of the physical model in fitting real deposition behavior.

[0052] It should be noted that all equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for grading corrosion and planning spraying trajectory on steel component surfaces based on multispectral vision, characterized in that, The method includes: Acquire synchronous multispectral image data of the steel component surface and generate a multispectral image dataset; Cross-modal feature fusion and calculation are performed on the multispectral image dataset to generate corrosion components and activity feature maps; Based on the corrosion components and activity characteristic map, a pre-set model for inferring the corrosion evolution trend is used to classify and predict the corrosion level and potential prediction map. Based on the corrosion level and potential prediction map and the three-dimensional geometric information of the steel component surface, the dynamic spraying requirements are calculated, and a three-dimensional dynamic spraying requirement field is generated. Taking the three-dimensional dynamic spraying demand field as the target, the motion trajectory of the spraying equipment and the spraying parameters are optimized in a coordinated manner to generate a dynamic spraying trajectory package; The dynamic spraying trajectory package is executed to perform the spraying operation, and the paint film status after spraying is monitored online to generate an instantaneous film quality map; The instantaneous film formation quality map is compared with the three-dimensional dynamic spraying demand field to generate a quality error map, and the dynamic spraying trajectory package is corrected and iteratively optimized in real time based on the quality error map.

2. The method for grading corrosion and planning spraying trajectory of steel components based on multispectral vision according to claim 1, characterized in that, The process of acquiring synchronous multispectral image data of the steel component surface and generating a multispectral image dataset includes: Simultaneously acquire visible light images, shortwave infrared images, and thermal imaging sequences of the steel component surface; Spatial registration and temporal alignment are performed on the acquired visible light images, shortwave infrared images, and thermal imaging sequence images; The registered and aligned image data are fused to construct a multispectral tensor field containing the color value, infrared absorption intensity, and thermal change rate information of each pixel, which serves as a multispectral image dataset.

3. The method for grading corrosion and planning spraying trajectory of steel components based on multispectral vision according to claim 1, characterized in that, Cross-modal feature fusion and solving are performed on the multispectral image dataset to generate corrosion component and activity feature maps, including: The absorption spectrum of each pixel is extracted from the multispectral image dataset and matched with a database containing standard spectra of corrosion products to obtain the component ratio characteristics of the corrosion products. By combining the thermal imaging sequence data in the multispectral image dataset, the temperature anomaly diffusion pattern is analyzed to obtain the corrosion activity intensity characteristics; By integrating the component ratio characteristics and the corrosion activity intensity characteristics, a corrosion component and activity feature map is generated that characterizes the corrosion type composition and electrochemical activity intensity of each pixel.

4. The method for grading corrosion and planning spraying trajectory of steel components based on multispectral vision according to claim 3, characterized in that, The step of classifying and predicting corrosion using a pre-set model for extrapolating corrosion evolution trends, and generating a corrosion level and potential prediction map, includes: Analyze the spatial distribution morphology of the rust components and active feature diagrams, including the boundary irregularity and porosity of the rusted area; The spatial distribution morphology characteristics, the component ratio characteristics, and the corrosion activity intensity characteristics are input together into a preset model for predicting the corrosion evolution trend. The model for predicting the corrosion evolution trend is a nonlinear mapping model based on electrochemical kinetics. In the model for predicting the corrosion evolution trend, the corrosion potential is determined by the component ratio characteristics, the charge transfer impedance is calculated by combining the corrosion activity intensity characteristics and pore distribution density, and the penetration rate of corrosion in the surface and deep layers of the steel component substrate is predicted by an anisotropic diffusion algorithm. The model that predicts the evolution trend of corrosion outputs a corrosion level and potential prediction map that includes the current severity level of corrosion and the future expansion vector distribution.

5. The method for grading corrosion and planning spraying trajectory of steel components based on multispectral vision according to claim 1, characterized in that, Based on the corrosion level and potential prediction map and the three-dimensional geometric information of the steel component surface, the dynamic spraying requirements are calculated, and a three-dimensional dynamic spraying requirement field is generated, including: The required base paint film thickness is determined based on the corrosion level and potential prediction diagram. Obtain the three-dimensional point cloud data of the steel component surface, and calculate the surface curvature and normal vector information of the steel component; By using a model to simulate the fluid behavior of coatings on different morphological surfaces, and combining the basic paint film thickness requirements with the surface curvature and normal vector information, the coating adhesion efficiency and sagging trend are calculated. The model used to simulate the fluid behavior of coatings on different morphological surfaces is a semi-empirical fluid physics model. In the model simulating the fluid behavior of coatings on different morphological surfaces, the coating loss caused by rebound is calculated using surface curvature, the effective adhesion rate is calculated by combining the angle between the normal vector and the spray gun spray direction, and the sagging risk index is evaluated based on the ratio of the component of gravity on the inclined plane to the coating viscosity. Based on the coating adhesion efficiency and sagging trend, the basic paint film thickness requirement is compensated and adjusted, and a three-dimensional dynamic spraying requirement field including the target paint film thickness, spraying residence time and allowable flow rate is output.

6. The method for rust classification and spraying trajectory planning of steel components based on multispectral vision according to claim 1, characterized in that, Taking the aforementioned three-dimensional dynamic spraying demand field as the target, the motion trajectory of the spraying equipment and the spraying parameters are collaboratively optimized to generate a dynamic spraying trajectory package, including: The three-dimensional dynamic spraying demand field is converted into an objective function to achieve uniform distribution of paint deposition. Using the objective function as a constraint, the spatial path, moving speed, paint output, atomization pressure, and spray gun attitude angle parameters of the spraying equipment are optimized collaboratively. The optimized spatial path, time series, six-degree-of-freedom attitude, and spraying equipment parameters are encapsulated to generate a dynamic spraying trajectory package.

7. The method for grading corrosion and planning spraying trajectory of steel components based on multispectral vision according to claim 1, characterized in that, The process of executing the dynamic spraying trajectory package to perform the spraying operation and monitoring the paint film status online after spraying, generating an instantaneous film quality map includes: The spraying equipment is controlled to execute the dynamic spraying trajectory package for spraying. During the spraying process, a broadband light source is used to illuminate the surface of the undried paint film, and the interference fringe image formed by the reflected light is received. By analyzing the interference fringe image, the wet film thickness distribution and leveling state of the paint film are calculated, and an instantaneous film quality map is generated.

8. The method for rust classification and spraying trajectory planning of steel components based on multispectral vision according to claim 1, characterized in that, The instantaneous film formation quality map is compared with the three-dimensional dynamic spraying demand field to generate a quality error map. Based on the quality error map, the dynamic spraying trajectory package is corrected and iteratively optimized in real time, including: The wet film thickness distribution in the instantaneous film formation quality map is compared point by point with the target paint film thickness in the three-dimensional dynamic spraying demand field to generate a thickness error distribution map. The thickness error distribution map is used as a state input and input into a preset intelligent decision-making model, wherein the intelligent decision-making model is a reinforcement learning decision-making model; In the intelligent decision-making model, by extracting the mean and variance features of the thickness error distribution map, a trajectory parameter correction amount is generated for subsequent unexecuted paths. The trajectory parameter correction amount includes the moving speed correction ratio, the paint output correction ratio, and the path overlap rate adjustment value. The execution parameters of subsequent path segments in the dynamic spraying trajectory package are updated in real time based on the adjustment instructions generated according to the trajectory parameter correction amount.

9. The method for grading corrosion and planning spraying trajectory of steel components based on multispectral vision according to claim 8, characterized in that, The intelligent decision-making model also performs offline training based on historical operation data to optimize the model used to predict the corrosion evolution trend and the model used to simulate the fluid behavior of coatings on different surface morphologies, including: Obtain a historical causal chain sample set consisting of historical quality error maps, corresponding corrosion component and activity characteristic maps, and corrosion grade and potential prediction maps; Using the gradient descent algorithm, with the goal of optimizing the loss function between the prediction bias and the actual thickness error, the neuron weights inside the model that predicts the corrosion evolution trend are adjusted to optimize the mapping accuracy of the model that predicts the corrosion evolution trend on the corrosion expansion potential. By analyzing the statistical correlation of thickness deviation under specific geometric features in historical quality error graphs, the curvature influence coefficient and sagging influence coefficient of the simulated coating in the model of fluid behavior on different morphological surfaces are iteratively corrected to achieve adaptive compensation optimization of the physical behavior of coating deposition.

10. A system for classifying rust on steel components and planning spraying trajectories based on multispectral vision, characterized in that, The system is used in the method for rust classification and spraying trajectory planning of steel components based on multispectral vision as described in any one of claims 1-9, and the system comprises: The multispectral image acquisition module is used to acquire synchronous multispectral image data of the steel component surface and generate a multispectral image dataset. The cross-modal feature analysis module is used to perform cross-modal feature fusion and calculation on the multispectral image dataset to generate corrosion components and activity feature maps; The corrosion grading and prediction module is used to grade and predict corrosion based on the corrosion components and activity characteristic map, using a preset model for extrapolating corrosion evolution trends, and to generate corrosion grade and potential prediction map. The spraying demand calculation module is used to calculate dynamic spraying demand based on the corrosion level and potential prediction map and the three-dimensional geometric information of the steel component surface, and generate a three-dimensional dynamic spraying demand field. The trajectory planning and optimization module is used to coordinate the motion trajectory and spraying parameters of the spraying equipment with the target of the three-dimensional dynamic spraying demand field, and generate a dynamic spraying trajectory package. The execution and online monitoring module is used to execute the dynamic spraying trajectory package to perform the spraying operation, monitor the paint film status after spraying online, and generate an instantaneous film quality map. The real-time correction and iterative optimization module is used to compare the instantaneous film formation quality map with the three-dimensional dynamic spraying demand field, generate a quality error map, and perform real-time correction and iterative optimization of the dynamic spraying trajectory package based on the quality error map.

Images