Spraying system based on multi-modal vision and artificial intelligence self-correction and control method thereof

Abstract

The invention relates to the technical field of automatic spraying, and particularly discloses a spraying system based on multi-modal vision and artificial intelligence self-correction, comprising a multi-modal vision acquisition module for acquiring color image information, depth information and infrared feature information; the control module is used for generating a spraying sensing model according to the information of the multi-modal visual acquisition module; the control module carries out spraying area identification, track planning and spraying parameter decision making based on the spraying sensing model; the spraying execution module is used for spraying; and the feedback self-correction module iterates the spraying parameter decision in the control module based on a reinforcement learning algorithm according to the difference between the actual coating state information and the expected state. The control method based on multi-modal vision and artificial intelligence self-correction is applied to a spraying system based on multi-modal vision and artificial intelligence self-correction. The scheme is used for solving the problems that the spraying quality of complex workpieces is not high due to the single sensing dimension of an existing spraying system, and the production flexibility is poor due to the rigid control mode.

Description

Technical Field

[0001] This invention relates to the field of automated spraying technology, specifically to a spraying system and its control method based on multimodal vision and artificial intelligence self-correction. Background Technology

[0002] Automated painting robots are widely used in automotive painting production lines. Current mainstream technologies primarily rely on offline programming and preset trajectories. Operators need to pre-determine the robot's motion path and corresponding fixed painting parameters for specific workpiece models through teaching or based on 3D models. During actual operation, the robot strictly replicates this preset program. This approach demonstrates a certain degree of stability and efficiency in standardized, mass production. However, as the manufacturing industry shifts towards multi-variety, small-batch, or customized production, current automated painting robots face limitations and the following problems: First, the system lacks flexibility, and the cost of changeover and debugging is high. Current coating systems heavily rely on preset programs. If the workpiece model, geometry, dimensions, or surface material changes, engineers must reprogram offline, simulate trajectories, and tune parameters. This process is not only time-consuming and labor-intensive but also heavily dependent on the experience of personnel, resulting in long debugging cycles and impacting production line changeover efficiency and capacity utilization.

[0003] Secondly, the limited sensory capabilities make it difficult to handle the challenges of precision coating complex workpieces. Most existing systems rely on a single vision sensor for workpiece positioning or simple contour recognition, lacking the ability to accurately perceive the workpiece's three-dimensional geometric details (such as deep grooves, steep surfaces, and small holes) and surface physical properties (such as reflectivity, roughness, and thermal capacity differences). This results in the system's inability to accurately acquire key geometric information such as surface normals and curvature changes when dealing with complex structures, making it difficult to generate an optimal coating trajectory. The direct consequence is an improper match between the spray gun's posture, distance, and movement speed during the coating process, easily leading to quality problems such as uneven coating thickness, missed areas, runs, or over-spraying in grooves, edges, or curved transition areas, affecting the product's appearance and corrosion resistance.

[0004] Third, the control loop is open-loop, lacking real-time dynamic adjustment and process closed-loop correction capabilities. Most existing systems operate on a "pre-programmed, rigidly executed" open-loop control model. During the spraying process, the unavoidable dynamic changes in the coating deposition state make it impossible to prevent localized overspray or underspray. Without real-time sensing of the coating deposition state (e.g., wet film thickness, uniformity), environmental disturbances (e.g., temperature, humidity changes), or minor deformations and positional shifts of the workpiece itself, timely adjustments to process parameters are impossible. For example, when a thinner paint film is detected in a certain area, traditional systems cannot compensate in real-time during the current spraying cycle; a second spraying is necessary after the initial spraying, which not only reduces efficiency but may also lead to a decrease in coating interfacial adhesion. This "remedial" rather than "preventative" approach results in delayed and reactive quality control.

[0005] Fourth, existing intelligent approaches suffer from limitations such as shallow technological levels and late control intervention. Although some research has attempted to introduce artificial intelligence technology into the spraying field, such as the "Spraying System and Method Based on Artificial Intelligence Algorithm" disclosed in existing patent publication number 25982A, the main function of visual recognition before spraying is limited to the rough "positioning and marking" of the spraying units. The role of its artificial intelligence model is to perform quality assessment and feature analysis on the collected spraying effect images after spraying, and to optimize some parameters of subsequent operations through supervised learning based on historical data samples. This mode is essentially an "experience optimization" or "post-compensation" mechanism based on result feedback. However, post-compensation methods rely on a large number of post-spraying samples and cannot make rapid self-adjustments, still making it difficult to meet the high-quality spraying requirements of complex workpieces. Summary of the Invention

[0006] The present invention aims to provide a spraying system and its control method based on multimodal vision and artificial intelligence self-correction, so as to solve the problems of low spraying quality of complex workpieces due to the single perception dimension and poor production flexibility due to the rigid control mode of the existing spraying system.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A spraying system based on multimodal vision and artificial intelligence self-correction includes: A multimodal vision acquisition module is used to simultaneously acquire color image information, depth information, and infrared feature information of the surface of the object being sprayed; The control module, connected to the multimodal vision acquisition module, is used to receive and fuse the color image information, depth information, and infrared feature information to generate a spraying perception model. The spraying perception model at least represents the three-dimensional geometric information of the spraying object and the multimodal visual features spatially associated with it. The control module performs spraying area identification, trajectory planning, and spraying parameter decision based on the spraying perception model. A spraying execution module, connected to the control module, is used to perform spraying operations; The feedback self-correction module is connected to the multimodal vision acquisition module and the control module. It is used to trigger the multimodal vision acquisition module to perform result detection after spraying, obtain actual coating state information, and iteratively optimize the spraying parameter decision strategy in the control module based on the difference between the actual coating state information and the expected state, using a reinforcement learning algorithm.

[0008] Preferably, as an improvement, it also includes a data storage and learning module for storing historical operation data and providing model training support for the control module and the feedback self-correction module.

[0009] Preferably, as an improvement, the multimodal vision acquisition module includes an RGB sensor and an infrared sensor disposed at the end of the spraying execution module, and a depth sensor for acquiring depth information; The infrared feature information collected by the infrared sensor is used to provide information reflecting the reflectivity change and heat distribution characteristics of the sprayed surface, so as to help determine the trend of paint layer thickness change and spray uniformity.

[0010] Preferably, as an improvement, the control module includes: The AI ​​recognition unit processes the spraying perception model based on a deep learning network to segment the spraying area; The parameter optimization unit maintains and executes a dynamic mapping relationship, which takes the three-dimensional geometric information and multimodal visual features as input states and spraying parameters as output actions. The motion control unit plans the spraying trajectory and posture based on the three-dimensional geometric information of the spraying area and the spraying perception model.

[0011] Preferably, as an improvement, the AI ​​recognition unit uses a deep learning network that takes the features extracted from the fused color image information, depth information, and infrared feature information as input, and outputs pixel-level segmentation results of the sprayed area through an encoder-decoder structure; wherein, the deep learning network suppresses background areas that have a preset distance difference from the sprayed object through the depth information, and enhances the boundary features of the sprayed area through multimodal feature fusion, thereby achieving effective suppression of irrelevant backgrounds.

[0012] Preferably, as an improvement, the parameter optimization unit can initially set the spraying parameters based on at least one of the preset physical model parameters, historical spraying data, or a lookup table method; and dynamically correct them during the spraying process based on the real-time acquired spraying perception model information.

[0013] Preferably, as an improvement, the dynamic mapping relationship maintained in the parameter optimization unit is driven by the reinforcement learning algorithm in the feedback self-correction module and is continuously updated online based on historical operation data.

[0014] Preferably, as an improvement, the feedback self-calibration module includes: The state comparison unit is used to obtain the actual coating state information by analyzing the changes in the spraying perception model before and after spraying. The error correction unit integrates a reinforcement learning algorithm engine, which constructs a reward function based on the difference between the actual coating state information and the desired spraying, and updates the spraying parameter decision strategy in the control module accordingly.

[0015] Preferably, as an improvement, the coating state information includes the coating thickness distribution. The state comparison unit generates an error matrix characterizing the spraying deviation by comparing the actual detected coating thickness distribution with the expected coating thickness model. The reinforcement learning algorithm engine in the error correction unit uses the error matrix as the basis for calculating its reward function, evaluates the spraying effect of the current combination of spraying parameters, updates the spraying parameter decision strategy accordingly, and outputs the adjusted spraying parameters. The adjusted spraying parameters are fed back to the spraying execution module for subsequent spraying operations.

[0016] Preferably, as an improvement, the state space of the reinforcement learning algorithm engine includes at least the surface normal vector, surface curvature, and infrared features of the current spraying provided by the spraying perception model.

[0017] This invention also provides a control method based on multimodal vision and artificial intelligence self-correction, applied to the aforementioned spraying system based on multimodal vision and artificial intelligence self-correction, comprising the following steps: S1. Multimodal data acquisition and fusion: Simultaneously acquire color image information, depth information and infrared feature information of the sprayed object to generate a spraying perception model; S2. Area identification and trajectory planning: Based on the spraying perception model, identify the spraying area and plan the spraying trajectory and spraying posture; S3. Parameter Determination and Spraying Execution: Based on the multimodal visual features provided by the spraying perception model, the initial spraying parameters are determined, and the spraying execution module is controlled to perform operations along the planned trajectory. During the process, the parameters are dynamically adjusted according to the real-time perception features. S4. Online detection of spraying effect: After the spraying is completed, step S1 is executed again to obtain the spraying perception model after spraying and to calculate the actual coating state information. S5. AI Optimization Self-Correction: Compare the actual coating state information obtained in step S4 with the expected state, use reinforcement learning algorithm, update the spraying parameter decision strategy based on the reward signal generated by the comparison result, and apply the updated spraying parameter decision strategy to subsequent spraying operations.

[0018] In summary, the present invention has the following advantages: 1. This invention constructs a "spraying perception model" that encompasses spatial geometry, surface properties, and process conditions by simultaneously fusing three different and complementary data streams: color image information (providing texture, color, and boundary information), depth information (providing accurate three-dimensional geometric information), and infrared features (reflecting surface reflectivity, heat distribution, and paint layer state trends). Through the complementary fusion of multimodal information, this model can effectively reduce the influence of single visual information on factors such as illumination and material reflection, thereby improving the reliability of spraying area identification and spraying result evaluation.

[0019] 2. This invention significantly advances the intervention point of artificial intelligence from the "post-spraying evaluation" in current conventional technologies to "pre-spraying decision-making" and "in-spraying control." The AI ​​recognition unit in the control module directly uses the information displayed by the fused spraying perception model as input, achieves pixel-level fine segmentation of the spraying area based on a deep learning network, effectively utilizes depth information to suppress irrelevant backgrounds, and enhances boundary features through multimodal fusion to improve the accuracy of subsequent spraying.

[0020] 3. The parameter optimization unit of this invention uses a dynamic mapping relationship where the spraying perception model is the "input state" and the spraying parameters are the "output action." This allows the system to "think" and generate a customized spraying trajectory and initial parameter strategy based on a deep understanding of the current workpiece before the spraying operation begins. This changes the traditional mode of "one program continuously acting on one product" or "requiring multiple manual trial and error parameter adjustments." This invention enables the system to quickly obtain the optimal spraying trajectory and spraying posture based on real-time perception, even for complex curved surfaces and multiple vehicle models, greatly shortening changeover time and supporting flexible manufacturing modes for multi-variety, small-batch, and even mixed-line production.

[0021] 4. The feedback self-correction module of this invention does not simply compare the post-coating image with a standard image and linearly adjust a few parameters. Instead, it constructs a complete "evaluation-learning-optimization" reinforcement learning closed loop. The principle is as follows: the actual coating state after spraying (obtained by re-running the multimodal visual acquisition module) is compared with the desired state, quantifying and generating an error matrix containing multi-dimensional evaluation indicators. This error matrix is ​​then used to drive the reinforcement learning algorithm engine for policy evaluation and updates. The reinforcement learning engine's "input state space" contains rich real-time perception information (such as surface normals, curvature, infrared features, etc.), while its "output action space" is the combination of spraying parameters to be optimized (such as spray gun angle, distance, flow rate, speed, etc.). Through continuous iteration, the system autonomously learns how to take the optimal spraying action under different input states, enabling the invention to form a dynamic mapping from spraying state to spraying parameter strategy. This achieves online, adaptive, and self-correcting updates of spray gun parameters, ensuring that the system can achieve uniform coverage of complex areas such as grooves, edges, and curved surfaces, significantly reducing thickness deviation, improving the consistency of coating appearance and performance, and reducing overspray, atomization waste, and rework rate, which helps to reduce the cost of paint and energy. Furthermore, under automatic update and correction, it can autonomously make and optimize spraying strategies based on real-time conditions, reducing the reliance on highly skilled workers in production.

[0022] 5. This system organically integrates multimodal perception and reinforcement learning, enabling the reinforcement learning self-correction mechanism to make the system a "living" system capable of learning from multiple painting operation experiences. It can achieve continuous optimization of painting quality without the need for complex dynamic modeling, making it particularly suitable for painting applications involving multiple vehicle models, multiple curved surfaces, and frequent changes in working conditions. It boasts high engineering practicality and a high level of intelligence. Furthermore, the longer the system runs and the more data it accumulates, the more optimized its painting strategy becomes, and the stronger its ability to cope with abnormal working conditions becomes. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall structure of the spraying system based on multimodal vision and artificial intelligence self-correction according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram illustrating the relationship between the multimodal vision acquisition module, the spraying execution module, and the workpiece to be sprayed, as shown in this embodiment of the invention.

[0025] Figure 3 This is a schematic diagram illustrating the principle of the feedback self-calibration module in an embodiment of the present invention.

[0026] Figure 4 This is a flowchart illustrating the control method based on multimodal vision and artificial intelligence self-correction according to an embodiment of the present invention.

[0027] Figure description: 1. Robotic arm; 2. Spray gun; 3. RGB camera; 4. Infrared thermal imager; 5. Depth sensor; 10. Object to be sprayed. Detailed Implementation

[0028] The following detailed description illustrates the specific implementation methods: The basic implementation examples are as follows: Figures 1 to 4 As shown.

[0029] Combination Figure 1 A spraying system based on multimodal vision and artificial intelligence self-correction includes a multimodal vision acquisition module, a control module, a spraying execution module, a feedback self-correction module, and a data storage and learning module.

[0030] I. Multimodal Vision Acquisition Module: The multi-source data acquired by the multimodal vision acquisition module is transmitted to the control module for spraying area identification, spraying parameter calculation, and spraying process feedback control.

[0031] Specifically, the multimodal vision acquisition module consists of an RGB sensor (such as an RGB camera), a depth sensor (such as a depth camera), and an infrared sensor (such as an infrared thermal imager), used to simultaneously acquire color image information, depth information, and infrared feature information of the surface of the object being sprayed.

[0032] An RGB camera and an infrared thermal imager are mounted on the end of the robotic arm of the spraying execution module, while a depth camera is fixed above the spraying area.

[0033] Among them, RGB image data is used to extract the surface texture and color features of the sprayed object to assist in the initial positioning and coverage detection of the sprayed area.

[0034] The depth information collected by the depth camera is used to construct a three-dimensional point cloud model (i.e., three-dimensional geometric information, three-dimensional data model) of the object to be sprayed, and to calculate the spatial position, curvature and surface normal vector of the sprayed surface, providing a geometric basis for the dynamic adjustment of the spray gun distance and spraying angle.

[0035] Infrared feature data collected by infrared thermal imagers are used to reflect the reflectivity changes and heat distribution characteristics of the sprayed surface, in order to help determine the trend of paint layer thickness change and spray uniformity online.

[0036] The visual shooting direction of the RGB camera and infrared thermal imager is parallel to the spraying direction of the spray gun.

[0037] II. Control Module: The control module connects to the multimodal vision acquisition module to receive and fuse color image information, depth information, and infrared feature information to generate a unified spraying perception model. This spraying perception model at least represents the three-dimensional geometric information of the sprayed object (such as surface normals, curvature, etc.) and its spatially associated multimodal visual features (such as RGB texture, infrared reflection features, etc.). Based on this spraying perception model, the control module performs spraying area identification and segmentation, trajectory planning, and spraying parameter decision-making. The control module integrates a GPU acceleration unit for running subsequent AI inference tasks.

[0038] Specifically, the control module includes an AI recognition unit, a parameter optimization unit, and a motion control unit.

[0039] The AI ​​recognition unit processes the spraying perception model based on improved U-Net or Transformer deep learning networks to achieve pixel-level segmentation of the sprayed area. This deep learning network takes a fused color image, depth, and infrared features as input, and outputs the segmentation result through an encoder-decoder structure. The network utilizes depth information to suppress background areas with preset distance differences from the sprayed object and enhances the boundary features of the sprayed area through multimodal feature fusion, thereby effectively suppressing irrelevant backgrounds.

[0040] The parameter optimization unit maintains and executes a dynamic mapping relationship model. This dynamic mapping relationship model takes three-dimensional geometric information and multimodal visual features (including surface normals, curvature, and infrared reflection features) as input states, and spraying parameters (including spray distance, spraying angle, spraying flow rate, and spray gun movement speed) as output actions.

[0041] The parameter optimization unit can initially set the spraying parameters based on at least one of the following: preset physical model parameters, historical spraying data, or a lookup table method; and dynamically correct them during the spraying process based on the real-time acquired spraying perception model. This dynamic mapping relationship model is driven by a reinforcement learning algorithm in the feedback self-correction module and can be continuously updated online based on historical operation data.

[0042] The motion control unit generates the optimal spraying trajectory covering the spraying area based on the spraying area contour and corresponding 3D point cloud data output by the AI ​​recognition unit, and combines the surface normal vector to correct the spray gun posture.

[0043] III. Spraying Execution Module: The spraying execution module is connected to the control module and is used to complete the automated spraying operation on the object to be sprayed according to the spraying control commands output by the control module.

[0044] Specifically, the spraying execution module includes a multi-degree-of-freedom robotic arm and an intelligent spray gun mounted at its end. The robotic arm controls the spatial position and movement speed of the spray gun; the intelligent spray gun integrates a spraying parameter adjustment unit, which receives control signals and adjusts the spraying angle, spray distance, and spraying flow rate in real time.

[0045] IV. Feedback Self-Correction Module: The feedback self-calibration module is connected to the multimodal vision acquisition module and the control module to monitor the spraying results in real time and correct the spraying parameters based on the reinforcement learning algorithm.

[0046] Specifically, the feedback self-correction module includes a state comparison unit and an error correction unit. In this embodiment, the state comparison unit compares images, and is therefore also called a real-time image comparison unit.

[0047] The state comparison unit (real-time image comparison unit) is used to obtain actual coating state information, such as coating thickness distribution, by analyzing the changes in the coating perception model before and after spraying. This state comparison unit generates an error matrix characterizing spraying deviations by comparing the actual detected coating thickness distribution with the expected coating thickness model.

[0048] The error correction unit integrates a reinforcement learning algorithm engine. This engine uses the error matrix generated by the state comparison unit as the basis for calculating its reward function, evaluates the spraying effect of the current combination of spraying parameters, and updates the spraying parameter decision strategy in the control module accordingly, outputting the adjusted spraying parameters (including spray gun distance, spraying angle, spraying flow rate, and spray gun movement speed). The adjusted spraying parameters are fed back to the spraying execution module for subsequent spraying operations, thereby achieving self-learning and self-correction of the spraying parameters.

[0049] V. Data Storage and Learning Module: The data storage and learning module is used to store historical operation data (including spraying parameters, environmental data, and spraying effects), and to provide the data support required for the training and evolution of the AI ​​recognition unit in the control module and the reinforcement learning algorithm engine in the feedback self-correction module.

[0050] This embodiment also provides a control method based on multimodal vision and artificial intelligence self-correction, applied to the aforementioned spraying system, including the following steps: S1. Multimodal data acquisition and fusion: Simultaneously acquire color image information, depth information and infrared feature information of the sprayed object, and fuse three-dimensional geometric information with multimodal visual features to generate a spraying perception model.

[0051] S2. Region Recognition and Trajectory Planning: Based on the spraying perception model, a deep learning network is used to identify the spraying area, and the optimal motion trajectory and spraying posture of the spray gun are planned by combining the three-dimensional geometric information of the spraying object and multimodal visual features.

[0052] S3. Parameter Determination and Spraying Execution: Based on the 3D geometric information and multimodal visual features provided by the spraying perception model, the initial spraying parameters (including spray gun distance, spraying angle, spraying flow rate, and spray gun movement speed) are determined through a dynamic mapping relationship model, and the spraying execution module is controlled to operate along the planned trajectory. During the spraying process, dynamic corrections are made based on feedback from the feedback self-correction module.

[0053] S4. Online detection of spraying effect: After spraying is completed, step S1 is executed again to obtain the spraying perception model. The actual coating thickness distribution is calculated using a thickness estimation algorithm (based on reflection model or structured light ranging) or geometric measurement method, which serves as the basis for evaluating the spraying effect.

[0054] S5. AI-based self-correction optimization: The actual coating thickness distribution obtained in step S4 is compared with the desired coating thickness model to generate an error matrix. Using a reinforcement learning algorithm, this error matrix is ​​used as a reward signal to update the input state to a coating parameter decision-making strategy that takes coating parameters as the output action. The updated coating parameter decision-making strategy is then applied to subsequent coating operations on similar workpieces, forming a continuously optimized process loop.

[0055] The spraying system in this embodiment utilizes a multimodal vision-based spraying perception model throughout the entire process, solving the problem of insufficient spraying accuracy for complex workpieces caused by a single perception dimension. Before spraying, it integrates RGB, depth, and infrared feature information to complete fine-grained region segmentation, boundary extraction, and 3D contour and normal vector acquisition, which are then directly used for high-precision trajectory generation and parameter calculation. During spraying, it dynamically senses state changes, providing real-time input for AI reinforcement learning self-correction. After spraying, it evaluates the spraying thickness and constructs an error matrix. The multimodal vision acquisition module not only achieves "perception" but also directly participates in "decision and control," providing accurate information assurance for uniform spraying of complex curved surfaces.

[0056] Furthermore, this system and method construct a reinforcement learning-based self-correcting closed loop by quantifying the deviation between the actual coating and the desired coating target. This solves the problems of poor production flexibility caused by rigid control modes and insufficient real-time adaptive capability caused by shallow intelligence levels. The mechanism inputs error matrix information into the reinforcement learning algorithm engine of the feedback self-correcting module, forming a dynamic mapping from the spraying state to the parameter strategy, realizing online, adaptive, and strategy-level updates of spray gun parameters. This not only enables the system to automatically adapt to different workpieces and working conditions without relying on fixed procedures and manual intervention, greatly improving production flexibility, but also achieves real-time dynamic adjustment and continuous self-evolution of the spraying process through AI-driven closed-loop optimization.

[0057] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A spraying system based on multimodal vision and artificial intelligence self-correction, characterized in that, include: A multimodal vision acquisition module is used to simultaneously acquire color image information, depth information, and infrared feature information of the surface of the object being sprayed; The control module, connected to the multimodal vision acquisition module, is used to receive and fuse the color image information, depth information, and infrared feature information to generate a spraying perception model. The spraying perception model at least represents the three-dimensional geometric information of the spraying object and the multimodal visual features spatially associated with it. The control module performs spraying area identification, trajectory planning, and spraying parameter decision based on the spraying perception model. A spraying execution module, connected to the control module, is used to perform spraying operations; The feedback self-correction module is connected to the multimodal vision acquisition module and the control module. It is used to trigger the multimodal vision acquisition module to perform result detection after spraying, obtain actual coating state information, and iteratively optimize the spraying parameter decision strategy in the control module based on the difference between the actual coating state information and the expected state, using a reinforcement learning algorithm.

2. The spraying system based on multimodal vision and artificial intelligence self-correction according to claim 1, characterized in that, It also includes a data storage and learning module, which stores historical job data and provides model training support for the control module and the feedback self-correction module.

3. The spraying system based on multimodal vision and artificial intelligence self-correction according to claim 1, characterized in that, The multimodal vision acquisition module includes an RGB sensor and an infrared sensor located at the end of the spraying execution module, and a depth sensor for acquiring depth information; The infrared feature information collected by the infrared sensor is used to provide information reflecting the reflectivity changes and thermal distribution characteristics of the sprayed surface.

4. The spraying system based on multimodal vision and artificial intelligence self-correction according to claim 3, characterized in that, The control module includes: The AI ​​recognition unit processes the spraying perception model based on a deep learning network to segment the spraying area; The parameter optimization unit maintains and executes a dynamic mapping relationship, which takes the three-dimensional geometric information and multimodal visual features as input states and spraying parameters as output actions. The motion control unit plans the spraying trajectory and posture based on the three-dimensional geometric information of the spraying area and the spraying perception model.

5. A spraying system based on multimodal vision and artificial intelligence self-correction according to claim 4, characterized in that, The AI ​​recognition unit uses a deep learning network that takes the features extracted from the fused color image information, depth information, and infrared feature information as input, and outputs pixel-level segmentation results of the sprayed area through an encoder-decoder structure. The deep learning network suppresses background areas that have a preset distance difference from the sprayed object through the depth information, and enhances the boundary features of the sprayed area through multimodal feature fusion.

6. A spraying system based on multimodal vision and artificial intelligence self-correction according to claim 4, characterized in that, The parameter optimization unit can initially set the spraying parameters based on at least one of the preset physical model parameters, historical spraying data, or lookup table method; and dynamically correct them during the spraying process based on the real-time acquired spraying perception model.

7. A spraying system based on multimodal vision and artificial intelligence self-correction according to claim 6, characterized in that, The dynamic mapping relationship maintained in the parameter optimization unit is driven by the reinforcement learning algorithm in the feedback self-correction module and is continuously updated online based on historical job data.

8. A spraying system based on multimodal vision and artificial intelligence self-correction according to claim 1, characterized in that, The feedback self-calibration module includes: The state comparison unit is used to obtain the actual coating state information by analyzing the changes in the spraying perception model before and after spraying. The error correction unit integrates a reinforcement learning algorithm engine, which constructs a reward function based on the difference between the actual coating state information and the desired spraying, and updates the spraying parameter decision strategy in the control module accordingly.

9. A spraying system based on multimodal vision and artificial intelligence self-correction according to claim 8, characterized in that, The coating state information includes the coating thickness distribution. The state comparison unit generates an error matrix characterizing the spraying deviation by comparing the actual detected coating thickness distribution with the expected coating thickness model. The reinforcement learning algorithm engine in the error correction unit uses the error matrix as the basis for calculating its reward function, evaluates the spraying effect of the current combination of spraying parameters, updates the spraying parameter decision strategy accordingly, and outputs the adjusted spraying parameters. The adjusted spraying parameters are fed back to the spraying execution module for subsequent spraying operations.

10. A control method based on multimodal vision and artificial intelligence self-correction, characterized in that, The spraying system based on multimodal vision and artificial intelligence self-correction, applied to any one of claims 1-9, comprises the following steps: S1. Multimodal data acquisition and fusion: Simultaneously acquire color image information, depth information and infrared feature information of the sprayed object to generate a spraying perception model; S2. Area identification and trajectory planning: Based on the spraying perception model, identify the spraying area and plan the spraying trajectory and spraying posture; S3. Parameter Determination and Spraying Execution: Based on the multimodal visual features provided by the spraying perception model, the initial spraying parameters are determined, and the spraying execution module is controlled to perform operations along the planned trajectory. During the process, the parameters are dynamically adjusted according to the real-time perception features. S4. Online detection of spraying effect: After the spraying is completed, step S1 is executed again to obtain the spraying perception model after spraying and to calculate the actual coating state information. S5. AI Optimization Self-Correction: Compare the actual coating state information obtained in step S4 with the expected state, use reinforcement learning algorithm, update the spraying parameter decision strategy based on the reward signal generated by the comparison result, and apply the updated spraying parameter decision strategy to subsequent spraying operations.

Images