Automobile front fingerboard selective post-processing method based on visual inspection

Abstract

The invention discloses an automobile front fingerboard selective post-processing method based on visual inspection, and belongs to the technical field of intelligent manufacturing of automobile parts, and the method comprises the following steps: 1) collecting surface and internal defect data of a front fingerboard by using a 2D linear array camera, a 3D laser scanner and a thermal infrared imager, and extracting defect feature vectors containing 18 features; 2) inputting the defect feature vector into an improved Transform neural network, and outputting an optimal post-processing process scheme including a process type, a tool type, process parameters and a processing path; according to the invention, through multi-modal visual fusion detection and 18-dimensional feature extraction, the detection accuracy of tiny and complex defects is high; the intelligent decision-making model realizes accurate matching of defects and processes, the material removal amount error control is accurate, and the coating thickness uniformity is improved; a processing path is optimized through a genetic algorithm, and the postprocessing time of a single front fingerboard is shortened in combination with quick remodeling of the robot; and the reinforcement learning automatic optimization process reduces the manual debugging time.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent manufacturing of automobile parts, and in particular relates to a selective post-processing method of an automobile front finger beam based on visual detection. Background Art

[0002] Currently, the post-processing of automotive front finger beams mostly adopts a unified process, such as sandblasting, grinding or spraying all front finger beams, resulting in material waste in areas without defects.

[0003] Traditional 2D visual inspection can only determine the presence or absence of defects, but cannot determine key parameters such as defect depth and curvature. Furthermore, inspection results are independent of post-processing, failing to effectively translate inspection data into a basis for adjusting process parameters. This forces processing equipment to operate according to fixed parameters, making it difficult to adapt to the processing needs of different defects.

[0004] In existing technologies, matching defects with post-processing processes relies on manual experience and lacks a quantitative mapping relationship. When encountering new defects or product model changes, process adjustment cycles are long, making it difficult to meet the rapid production iteration requirements of the automotive industry. Summary of the Invention

[0005] The purpose of the present invention is to solve the above-mentioned problems and to propose a selective post-processing method for automobile front finger beam based on visual inspection.

[0006] In order to achieve the above-mentioned object, the present invention adopts the following technical solution: a selective post-processing method of an automobile front finger beam based on visual inspection, which comprises the following steps:

[0007] 1) Using a 2D linear array camera, a 3D laser scanner, and an infrared thermal imager to collect surface and internal defect data of the front finger beam, a defect feature vector containing 18 features was extracted;

[0008] 2) Input the defect feature vector into the improved Transformer neural network and output the optimal post-processing process plan, including process type, tool type, process parameters and processing path;

[0009] 3) Post-processing is performed using a six-axis robot and end-of-line tool changer, with dynamic adjustment of process parameters achieved using force control sensors and a visual servo system.

[0010] 4) Perform secondary inspection on the processed front finger beam, and use reinforcement learning algorithm to optimize the process matching model based on the inspection results.

[0011] As a further description of the above technical solution:

[0012] The 18-dimensional defect feature vector includes four types of features: defect type, location coordinates, geometric parameters and material damage degree.

[0013] As a further description of the above technical solution:

[0014] The improved Transformer neural network is trained using 100,000 sets of defect sample data.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0016] 1. In the present invention, multimodal visual fusion detection and 18-dimensional feature extraction are used to achieve high accuracy in detecting small and complex defects; the intelligent decision-making model achieves precise matching of defects and processes, accurately controls the error of material removal, and improves the uniformity of coating thickness.

[0017] 2. In the present invention, the processing path is optimized by genetic algorithm, combined with the rapid changeover of robots, and the post-processing time of a single front finger beam is shortened; reinforcement learning automatically optimizes the process, reduces manual debugging time, and improves product changeover efficiency.

[0018] 3. In the present invention, the closed-loop quality control system is used to improve the product dimensional accuracy qualification rate and reduce the rework rate; precise processing reduces material waste and saves costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The flowchart of a selective post-processing method for automobile front finger beam based on visual inspection. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0021] Example:

[0022] Hardware system construction:

[0023] Testing equipment: Basler ac A12000-20gm line scan camera with a telecentric lens (distortion rate <0.03%); Keyence LJ-G050H 3D laser scanner with a scanning speed of 3000 times / second and a pixel pitch of 0.05mm; and FLIRA655sc infrared thermal imager with a temperature resolution of 0.03°C.

[0024] Execution equipment: A KUKAKR16-2 six-axis robot with a repeatability accuracy of ±0.03mm is used. It is equipped with an ATIQC-11 end-of-tool quick-changer, which supports fast switching between grinding wheels of different specifications, SATAjet5000B spray guns, and FroniusTPS2700 welding guns.

[0025] Software system implementation:

[0026] Defect detection software: Develop image and point cloud processing programs based on OpenCV and PCL libraries to implement image denoising, threshold segmentation, point cloud registration, and other functions; use the TensorFlow framework to build a deep learning model to complete defect identification and feature extraction;

[0027] Process decision-making software: Use PyTorch to build an improved Transformer neural network, and use Python to write a genetic algorithm program to implement path planning; develop a human-computer interaction interface to facilitate process parameter setting and system monitoring.

[0028] SO1: Multimodal visual inspection and defect feature extraction:

[0029] Data acquisition: At the inspection station on the production line, a 2D linear array camera (resolution 12000×1200, frame rate 200fps) is deployed to capture surface texture images of the front finger beam and identify scratches and cracks ≥0.1mm. A 3D laser scanner (accuracy ±0.02mm) acquires the three-dimensional coordinates of defects and measures geometric parameters such as depression depth and protrusion height. An infrared thermal imager detects sub-surface defects and captures temperature anomalies caused by internal damage.

[0030] Feature extraction: Construct an 18-dimensional defect feature vector that includes defect type (eight types, including scratches, dents, and cracks), location coordinates (X, Y, Z), geometric parameters (six items, including length, depth, and curvature radius), and material damage (four items, including coating damage level and metal deformation degree), to comprehensively describe defect information;

[0031] SO2: Defect-Process Intelligent Matching:

[0032] Defect classification: Defects are divided into three levels: Level I (serious defects, such as dents and through cracks with a depth of ≥1.5mm), Level II (moderate defects, scratches 0.5mm≤ and <1.5mm in depth), and Level III (minor defects, surface flaws with a depth of <0.5mm), to clearly define the priority of handling;

[0033] Intelligent decision-making model: Using an improved Transformer neural network, it inputs defect feature vectors and outputs the optimal post-processing solution, including process type (grinding, repair welding, spraying, polishing, etc.), tool type (grinding wheels of different specifications, spray guns, welding guns), process parameters (grinding pressure 0.1-0.8MPa, spray flow rate 30-250ml / min, welding current 80-200A), and processing path planning strategy. The model was trained on 100,000 sets of defect sample data, achieving a process matching accuracy of 98%.

[0034] SO3: Flexible post-processing execution:

[0035] Path planning: For irregular defects, we use genetic algorithm-based path optimization technology, combining the defect contour with the robot kinematic model to generate the shortest, collision-free processing path, which improves efficiency by 40% compared to traditional path planning.

[0036] Tool adaptive control: The contact force during the processing is fed back in real time by a force control sensor (accuracy ±0.05N), dynamically adjusting the grinding pressure and welding force. The position of the spray gun and welding gun is precisely controlled using a visual servo system (positioning accuracy ±0.01mm) to ensure process execution accuracy.

[0037] SO4: Closed-loop quality control:

[0038] Secondary inspection: After the treatment is completed, the front finger beam is inspected again using the multimodal vision system to compare the changes in defect characteristics before and after treatment to determine whether the treatment effect meets the standards;

[0039] Process optimization: The test results and processing data are stored in the database. Using a reinforcement learning algorithm, the process matching model is automatically optimized every 50 products processed to continuously improve process adaptability.

[0040] The present invention builds an integrated system of "intelligent detection - precise decision-making - flexible execution - closed-loop feedback", which includes:

[0041] Multimodal visual inspection module: Integrates 2D linear array cameras, 3D laser scanners, and infrared thermal imagers to achieve multi-dimensional detection of front finger beam defects;

[0042] Defect-process intelligent decision module: Based on deep learning algorithms, it establishes a quantitative mapping relationship between defect characteristics and post-processing processes;

[0043] Flexible execution module: equipped with a six-axis robot and end-of-line tool quick-change system to achieve precise execution of various processes;

[0044] Quality re-inspection and feedback module: conduct secondary inspection on the processed front finger beam to optimize process parameters.

[0045] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A selective post-processing method for automobile front finger beam based on visual inspection, characterized by: The steps include: 1) Using a 2D linear array camera, a 3D laser scanner, and an infrared thermal imager to collect surface and internal defect data of the front finger beam, a defect feature vector containing 18 features was extracted; 2) Input the defect feature vector into the improved Transformer neural network and output the optimal post-processing process plan, including process type, tool type, process parameters and processing path; 3) Post-processing is performed using a six-axis robot and end-of-line tool changer, with dynamic adjustment of process parameters achieved using force control sensors and a visual servo system. 4) Perform secondary inspection on the processed front finger beam, and use reinforcement learning algorithm to optimize the process matching model based on the inspection results.

2. The method for selective post-processing of a front finger beam of an automobile based on visual inspection according to claim 1, characterized in that: The 18-dimensional defect feature vector includes four types of features: defect type, location coordinates, geometric parameters and material damage degree.

3. The method for selective post-processing of automobile front finger beam based on visual inspection according to claim 1 is characterized in that: The improved Transformer neural network is trained using 100,000 sets of defect sample data.

Images