A painting block piece closed-loop inspection and control system and method based on distributed vision

Abstract

The application discloses a kind of based on distributed vision's coating block piece closed loop inspection control system and method, it is related to automobile intelligent manufacturing and automatic quality detection technical field.The method includes the following steps: detecting vehicle information, and according to vehicle information and vehicle real-time position acquisition multi-angle vehicle body image, specifically, in the process of uniform speed of vehicle to be detected through detection area, according to the change situation of real-time position trigger distributed image acquisition;Multi-angle vehicle body image is analyzed using visual inspection model to block piece misloading, missing loading;According to the analysis result, man-machine collaborative closed loop inspection is carried out, wherein the vehicle that is unqualified to analysis result is manually reviewed, and manual review result is added to the training sample of visual inspection model to carry out model iteration optimization.The application uses distributed imaging to automatically detect misloading and missing loading of rough sealing block piece on automobile coating production line, and is deeply integrated with production execution system to realize closed loop quality control.

Description

Technical Field

[0001] This invention relates to the field of intelligent automotive manufacturing and automated quality inspection technology, and in particular to a closed-loop inspection and control system and method for paint blockage parts based on distributed vision. Background Technology

[0002] In industrial painting fields such as automobile manufacturing, the correct installation of plugs is a crucial step in ensuring the vehicle body's sealing performance, corrosion resistance, and overall quality. As a cutting-edge area in machine vision, deep learning-based detection systems have gradually become a hot topic in research and application. The core of these methods lies in utilizing models such as deep convolutional neural networks to perform end-to-end feature learning and pattern recognition on a large number of labeled industrial image samples.

[0003] Compared to traditional template matching algorithms, deep learning models can automatically extract hierarchical features from data, ranging from low-level edges to high-level semantics, through multi-level nonlinear transformations. However, current deep learning models have fixed detection parameters, lacking adaptive and efficient detection solutions for different vehicle models. Furthermore, the detection methods, such as process optimization, parameter learning, and instruction execution, are relatively fragmented, affecting the overall detection accuracy. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a closed-loop inspection and control system and method for coating blockages based on distributed vision. This system utilizes distributed imaging to automatically detect misinstallation and omissions of coarse sealing blockages on automotive coating production lines, and deeply integrates with the production execution system to achieve closed-loop quality control.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: The first aspect of this invention provides a closed-loop inspection and control method for coating blockages based on distributed vision, comprising the following steps: The vehicle model information is detected, and multi-angle images of the vehicle body are collected based on the vehicle model information and the real-time position of the vehicle. Specifically, as the vehicle to be detected passes through the detection area at a constant speed, distributed image acquisition is triggered based on the changes in the real-time position. Visual inspection models were used to analyze the misinstallation and omission of parts in multi-angle vehicle body images; Based on the analysis results, a closed-loop inspection and control system involving human and machine collaboration is implemented. Vehicles that fail the analysis results are manually reviewed, and the results of the manual review are added to the training samples of the visual inspection model for iterative optimization.

[0006] Furthermore, the multi-angle vehicle body images include polarization orthogonal images and polarization parallel images, and polarization fusion and image enhancement preprocessing operations are performed on the multi-angle vehicle body images.

[0007] Furthermore, the RFID reader is used to dynamically identify the vehicle model information of vehicles entering the inspection station. When a vehicle enters, the RFID reader reads the code information in the disc at the bottom of the skid to obtain unique vehicle model information.

[0008] Furthermore, image acquisition is performed using a distributed industrial camera group, which specifically includes a top camera group, a left camera group, and a right camera group. During the distributed image acquisition process, which is triggered by real-time position changes, the industrial camera group is set to acquire images three times based on the position changes.

[0009] Furthermore, the specific steps for analyzing the misinstallation and omission of parts in multi-angle vehicle body images using a visual inspection model are as follows: Obtain known labeled data to form training samples; Construct a visual detection model based on the YOLOv8s architecture; The visual detection model is trained using training samples; The trained visual detection model is used to detect multi-angle vehicle images in real time.

[0010] A second aspect of the present invention provides a closed-loop inspection and control system for coating blockages based on distributed vision, comprising: The perception layer is used to detect vehicle model information and acquire multi-angle images of the vehicle body based on the vehicle model information and the real-time position of the vehicle. Specifically, as the vehicle to be detected passes through the detection area at a constant speed, distributed image acquisition is triggered based on the changes in the real-time position. The control layer is used to analyze the misinstallation and omission of parts in multi-angle vehicle body images using a visual inspection model; The execution and interaction layer is used for closed-loop inspection and control based on the analysis results. Among them, vehicles with unqualified analysis results are manually reviewed, and the manual review results are added to the training samples of the visual inspection model for iterative optimization of the model.

[0011] Furthermore, the perception layer includes an image preprocessing module for performing polarization fusion and image enhancement preprocessing operations on multi-angle vehicle body images, which include polarization orthogonal images and polarization parallel images.

[0012] Furthermore, the RFID reader is used to dynamically identify the vehicle model information of vehicles entering the inspection station. When a vehicle enters, the RFID reader reads the code information in the disc at the bottom of the skid to obtain unique vehicle model information.

[0013] Furthermore, image acquisition is performed using a distributed industrial camera group, which specifically includes a top camera group, a left camera group, and a right camera group. During the distributed image acquisition process, which is triggered by real-time position changes, the industrial camera group is set to acquire images three times based on the position changes.

[0014] Furthermore, in the control layer, the specific steps for analyzing the misinstallation and omission of parts in multi-angle vehicle body images using a visual inspection model are as follows: Obtain known labeled data to form training samples; Construct a visual detection model based on the YOLOv8s architecture; The visual detection model is trained using training samples; The trained visual detection model is used to detect multi-angle vehicle images in real time.

[0015] The above one or more technical solutions have the following beneficial effects: This invention discloses a closed-loop inspection and control system and method for paint congestion parts based on distributed vision. It designs a distributed camera layout and a moving-and-shoot triggering mechanism based on absolute position, allowing image acquisition to be synchronized with vehicle movement. This achieves seamless matching with the production cycle, eliminating time losses caused by vehicle stops or robot scanning, and reducing single-vehicle inspection time from tens of seconds in traditional methods to less than one production cycle. By introducing polarized optics technology to suppress reflection at the source, and combining it with image preprocessing technology, the powerful feature extraction and anti-interference capabilities of the deep learning AI model significantly outperform fixed-template matching algorithms in terms of accuracy and robustness in identifying changes in lighting, complex backgrounds, and novel defects, thus achieving stable and reliable high-precision inspection. Furthermore, through dynamic linkage between RFID vehicle model identification and PLC parameter library, the system achieves one-click software switching of inspection parameters, requiring no hardware adjustments or manual calibration, seamlessly adapting to multi-vehicle mixed-line production, minimizing changeover time and costs. This invention establishes a closed-loop data mechanism of "detection-decision-feedback-optimization" through PDA manual review and low-confidence sample feedback. This enables the AI ​​model to continuously learn and evolve, effectively combating performance degradation and achieving long-term benefits of increasing accuracy with use. The solution not only possesses high robustness, flexibility, precision, and efficiency in detecting coating blockages, but also achieves closed-loop self-optimization, overcoming the comprehensive bottlenecks of the dispersed, isolated, rigid, and passive detection modes in existing technologies.

[0016] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a diagram of the closed-loop inspection and control system for coating blockages based on distributed vision in Embodiment 2 of the present invention. Figure 2 This is a schematic diagram of the camera layout and imaging range in Embodiment 2 of the present invention; Figure 3 This is an actual effect diagram of the camera taking pictures in Embodiment 2 of the present invention; Figure 4 This is a hardware timing diagram of the closed-loop inspection and control system for coating blockages based on distributed vision in Embodiment 2 of the present invention; Figure 5 This is the detection logic diagram of the closed-loop inspection and control system for coating blockages based on distributed vision in Embodiment 2 of the present invention. Detailed Implementation

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] Example 1: Embodiment 1 of the present invention provides a closed-loop inspection and control method for coating blockages based on distributed vision, comprising the following steps: S1: Detect vehicle model information and acquire multi-angle images of the vehicle body based on the vehicle model information and the real-time position of the vehicle. Specifically, as the vehicle to be detected passes through the detection area at a constant speed, distributed image acquisition is triggered based on the changes in the real-time position.

[0022] In one specific implementation, the coating plug in this embodiment refers to a process component installed on parts such as the front fender, door sill, and center floor plate after the electrophoretic coating of the car body is completed, used to temporarily or permanently seal workpiece holes. Its core functions include sealing, waterproofing, shock absorption, and corrosion protection. The plug patches tested include: hot-melt plugs, CVT butyl patches, and circular shock-absorbing pads. In the coating process, these types of plugs, gaskets, and patches used for sealing can all be referred to as coating plugs.

[0023] This step is implemented based on the perception layer, which includes an information identification unit, a vehicle body position sensing unit, an image acquisition unit, and an image preprocessing module. The information identification unit is a vehicle model RFID reader / writer, used to identify vehicle model or code information, sending it to the vehicle model-photography process parameter database to query the photography scheme, and also sending it to the PLC programmable logic controller (PLC). This information, along with information collected by the photoelectric / proximity sensor group, serves as a trigger signal for vehicle entry. The vehicle body position sensing unit includes a photoelectric / proximity sensor group and a roller encoder. The roller encoder collects absolute position signals and sends them to the PLC. The image acquisition unit includes a distributed industrial camera group (12 units in this embodiment), controlled by the PLC generating photography signals based on different vehicle models and times. This includes 8 cameras on the top, 2 on the left, and 2 on the right. The image acquisition unit also includes an industrial vision light source, used to provide light. It works in conjunction with the cameras and is controlled by lighting control signals generated by the PLC. The image data acquired by the image acquisition unit is input into the image preprocessing module, which performs polarization fusion and image enhancement preprocessing operations on the multi-angle vehicle body images to obtain processed image data.

[0024] Specifically, the perception layer is used to detect vehicle model information and acquire multi-angle images of the vehicle body based on the vehicle model information and the vehicle's real-time position. Specifically, as the vehicle to be detected passes through the detection area at a constant speed, distributed image acquisition is triggered based on changes in its real-time position.

[0025] In one specific implementation, the information identification unit includes an RFID reader / writer for dynamically identifying vehicle model information entering the inspection station. When a vehicle enters, the RFID reader / writer reads the code information from the disc at the bottom of the skid to obtain unique vehicle model information, thereby selecting the corresponding preset photographic inspection scheme for that model. This ensures that each vehicle can be photographed in the optimal way, with imaging covering all components of the vehicle body.

[0026] In one specific implementation, the vehicle body position sensing unit consists of a sensor (photoelectric sensor) and a high-precision encoder installed at the inspection station of the painting and rough sealing production line. The sensor is used to detect the arrival and departure of the vehicle, while the encoder rotates with the spindle of the roller bed to provide the absolute position signal of the vehicle, enabling precise triggering.

[0027] In one specific implementation, the image acquisition unit consists of multiple industrial cameras spatially distributed via a gantry structure, used to acquire images of the moving vehicle body from different angles. The image acquisition unit includes a distributed industrial camera group, composed of 12 industrial cameras arranged in an optimized spatial distribution, such as... Figure 2 As shown, the system is fixed on a gantry spanning the production line, specifically comprising a top camera group (8 units) responsible for capturing images of the upper and lower positions of the front fenders, the front, middle, and rear floor panels, and the door sill area; a left camera group (2 units); and a right camera group (2 units) responsible for capturing images of the trunk, left and right rear panels, and side panels. The captured images are as follows: Figure 3 As shown in the diagram, the design concept of this layout is based on the installation location of the blockage parts throughout the vehicle. Since these blockage parts cover various locations on the vehicle, a gantry spanning the production line is used as the camera mounting device, with the vehicle as the inspection center, and 12 industrial cameras distributed around it. During the inspection process, the system will, according to a preset plan, call different cameras at different shooting positions as image acquisition units to complete the blockage part image acquisition step by step, thereby achieving comprehensive, blind-spot-free coverage of the moving vehicle body blockage parts, adapting to different vehicle models. To improve image quality, this unit also integrates a collaboratively controllable industrial vision light source. During the image acquisition process, the light source will be turned on according to a preset plan to provide support, thereby acquiring clearer and moderately bright images, ensuring high-quality imaging from the source.

[0028] In this embodiment, the industrial vision light source that can be controlled collaboratively adopts centralized control by PLC and utilizes the distributed execution architecture of the light source controller to achieve precise collaboration between the light source and the camera.

[0029] Specifically, to achieve high-quality imaging, this embodiment constructs a collaborative control architecture combining a PLC master controller and a light source controller slave controller. Each industrial camera is equipped with an independent light source module. All light source modules are uniformly connected to the control system through the light source controller. The light source controller is connected to the PLC via a high-speed I / O interface, receiving trigger commands from the PLC and precisely controlling the lighting timing, lighting duration, and brightness level of the light source according to preset parameters.

[0030] The multi-angle vehicle body images include polarization orthogonal images and polarization parallel images. The image preprocessing module performs polarization fusion and image enhancement preprocessing operations on the multi-angle vehicle body images to obtain the processed image data.

[0031] Specifically, an adjustable polarizing filter is installed in front of the lens of each industrial camera, and a polarizing film is installed in front of the corresponding light source. The polarization directions of the two are set to be orthogonal to physically suppress specular reflection. During image acquisition, the camera captures two images for each shooting point: the first is an orthogonally polarized image (strongly suppressing reflection), and the second is a parallel polarized image (normal brightness). After acquisition, the two images are fused pixel-level to generate a fused image that effectively eliminates reflective spots while fully preserving the details of the obstruction. Adaptive contrast adjustment is applied to the fused image to make the grayscale difference between the obstruction area and the vehicle background more obvious. Edge sharpening is performed to enhance the contour features of the obstruction. Based on a preset target brightness range, overall brightness equalization is performed on the image to ensure consistent brightness across images captured from different points. After the above image preprocessing, higher quality image data can be obtained.

[0032] like Figure 4 As shown, the specific data acquisition process includes: At time t0, the system is in standby preparation state. At times t1 and t2, the skid carrying the electrophoresis vehicle body begins to enter the inspection station. Sensors 1 and 2 sequentially detect the vehicle body's arrival, and the signals sequentially change from low to high, indicating that the vehicle has fully entered the inspection area. Simultaneously, the encoder starts working, and its output absolute position value linearly increases from the initial value, indicating that the vehicle is continuously moving forward with the production line, providing the system with a precise reference for the vehicle's real-time position. At time t3, when the vehicle moves to the RFID identification area, the RFID reader is triggered to read the vehicle model information, and the signal changes from low to high. At time t4, the RFID reading is complete, and the vehicle continues to move forward. The RFID transmits the acquired information to the PLC via a signal. Upon receiving this information, the PLC immediately dynamically calls and loads a photography scheme that perfectly matches the vehicle model from its internal vehicle model-photography process parameter library. This scheme predefines the trigger position, the number of cameras used, and their corresponding camera parameters for each inspection step, realizing a leap from "fixed program" to "vehicle model adaptation," greatly improving system flexibility. Suitable for production line scenarios involving the mixed production of multiple vehicle models. At time t5, the vehicle continues to move forward, and the RFID signal changes from high to low, awaiting the next stage of the process.

[0033] S2: Use a visual inspection model to analyze the misinstallation and omission of parts in multi-angle vehicle body images.

[0034] This step is implemented based on the control layer, which includes a vehicle model-photography process parameter library, a PLC (Programmable Logic Controller), an industrial switch, an AI vision inspection and analysis system, an intelligent inspection unit, and a model iteration and optimization module. The vehicle model-photography process parameter library sends corresponding preset parameters to the PLC based on real-time vehicle model information and real-time vehicle location. The PLC then sends trigger signals or vehicle model information to the AI ​​vision inspection and analysis system via the industrial switch. The AI ​​vision inspection and analysis system analyzes the processed image data obtained from the image preprocessing module based on model inference support provided by the intelligent inspection unit, obtaining defect location, type, and execution degree information, which is sent back to the PLC. The system then sends the inspection result (OK or NG) to the industrial switch and sends the generated low-confidence or defect samples to the model iteration and optimization module.

[0035] Specifically, the control layer is used to analyze the misinstallation and omission of parts in multi-angle vehicle body images using a visual inspection model. It is also used to retrieve pre-stored detection parameters for the vehicle model based on the vehicle model information, and to accurately trigger the corresponding camera to take pictures during the uniform movement of the vehicle body based on the real-time position signal of the encoder.

[0036] In one specific implementation, the PLC (Programmable Logic Controller) is connected to the information recognition unit, the roller encoder, and the image acquisition unit. The PLC internally stores a vehicle model-photography process parameter library, containing preset camera parameter configurations for each vehicle model. Upon receiving signals from the sensing layer, the PLC dynamically schedules the entire inspection process according to preset logic.

[0037] Specifically, the PLC internally stores a preset database of vehicle model photography parameters. This database predefines a complete photography process scheme for each vehicle model, including: the total number of detection steps, the absolute trigger position of each step (coordinates based on the encoder zero point), a list of camera IDs to be called at each step, and the optical parameters of each camera (such as exposure time, gain, and light source brightness). After receiving the vehicle model information identified by RFID, the PLC immediately calls the information from the corresponding vehicle model photography parameter database. While the vehicle is moving at a constant speed, it compares the absolute position fed back by the encoder with the preset trigger position in the database in real time. When the positions match, it accurately triggers the corresponding camera group to take pictures.

[0038] In one specific implementation, the AI ​​visual inspection and analysis system includes an industrial computer core with a deep learning visual inspection model trained on the YOLOv8s architecture, which performs intelligent analysis on the acquired images to detect mis-installation or omission of parts.

[0039] Specifically, the steps for analyzing the misinstallation and omission of parts in multi-angle vehicle body images using a visual inspection model are as follows: First, obtain known labeled data to form training samples. These training samples can be expanded from manually reviewed results or previously identified low-confidence samples. Second, construct a visual inspection model based on the YOLOv8s architecture. Third, train the visual inspection model using the training samples. Finally, use the trained visual inspection model to detect multi-angle vehicle body images in real time.

[0040] Specifically, it includes the following steps: Step 1: Construct training samples and perform data partitioning and preprocessing.

[0041] Training samples were constructed using historical image databases, manual verification results, and low-confidence samples, including positive samples (correctly installed) and negative samples (missing, incorrectly installed, or warped edges). These samples were then divided into training, validation, and test sets in a 7:2:1 ratio. Online enhancement preprocessing operations such as random rotation, scaling, brightness adjustment, and flipping were performed on the training set.

[0042] Step 2: Build and initialize the visual detection model.

[0043] The YOLOv8s architecture was selected, pre-trained weights were loaded, and the model's feature extraction capability was preserved. The last layer of the model (classification layer) was modified so that the number of output categories matched the number of component types (e.g., normal, missing, incorrect).

[0044] Step 3: Train and test the visual detection model using training samples.

[0045] Train the model using the training set, evaluate it on the validation set after each round, and save the best model. Evaluate the final model's performance using the test set, and deploy it once it meets the target.

[0046] Step 4: After going live, continue incremental learning.

[0047] Regularly collect new review samples, load the current model weights, freeze the underlying layer, fine-tune the upper layer, and replace the deployment after verification.

[0048] The visual inspection model in this embodiment can accurately identify and locate the installation status of obstructed components (normal, incorrectly installed, or missing). This embodiment uses YOLOv8s as the base model because it achieves an excellent balance between accuracy and speed. Its superior feature extraction capabilities and lightweight structure are particularly suitable for the high requirements of real-time detection and accuracy in this scenario. Its innovation lies in integrating a model iteration and optimization module. During each system run, low-confidence data is fed back, allowing the model to continuously learn and iterate, ensuring that the model has a large-scale training dataset for both positive and negative sample recognition.

[0049] Specifically, the visual inspection model consists of a model scheduling layer, a model inference layer, and a result fusion layer: The model scheduling layer manages the timing of model calls for different image capture steps and different camera images. It is responsible for receiving images and allocating processing threads.

[0050] The model inference layer is used in the core YOLOv8s model and is responsible for object detection and classification of a single ROI sub-image.

[0051] The result fusion layer is used to summarize the detection results of each step and each point according to spatial location.

[0052] This embodiment of the visual detection model adopts a strategy of step-by-step independent processing and result spatial fusion. The first step is to receive image batches step by step. The system processes the images acquired in each imaging step as an independent batch.

[0053] For example: Front of the vehicle: Acquire Cam1-Cam3 images → Form a batch.

[0054] Middle of the vehicle: Collect Cam4-Cam8 → Form the second batch.

[0055] Rear of the vehicle: Cam9-Cam12 → forming the third batch.

[0056] The second step involves parallel processing within a batch. For each batch of multiple images, the system initiates multi-threaded parallel inference using the YOLOv8s model: each image is cropped into multiple ROI sub-images based on a pre-stored ROI template. The model performs independent object detection and classification for each ROI sub-image. Each ROI outputs: presence of a blockage, blockage type, and confidence level. The third step temporarily stores the batch results. After each step's analysis is complete, the system temporarily stores the detection results in memory and binds them to the current vehicle's VIN code: 1. Record the detection status of all ROIs at point 1. 2. Record the detection status of all ROIs at point 2. 3. Record the detection status of all ROIs at point 3.

[0057] The fourth step is to fuse all results based on the spatial dimension. Once all three steps are completed (the vehicle leaves the inspection area), the result fusion layer is activated: according to the predefined spatial partitions in the ROI template (front fender, sill, trunk, etc.), the inspection results of the three steps are spatially stitched together to generate a complete vehicle blockage inspection report, including the inspection status of each spatial location.

[0058] The fifth step is to report and display the results. The integrated, complete report is reported to the MES system for quality traceability. It is also pushed to a large visualization screen to display the OK / NG status of each area. If any point is NG, an audible and visual alarm is triggered.

[0059] In one specific implementation, the model iteration optimization module is used to perform incremental learning and optimization of the AI ​​model based on low-confidence samples and manual review results.

[0060] Specifically, the first step is to screen samples based on confidence levels.

[0061] In this embodiment, the system sets a confidence threshold of 0.85. When the confidence level of the visual detection model for the detection result is lower than the threshold, the sample is defined as a low-confidence sample and enters the queue for review. Samples with a confidence level higher than the threshold are considered high-confidence samples and are directly added to the database as positive samples.

[0062] Next, manual PDA on-site verification and classification were conducted. Quality inspectors used PDAs to conduct on-site verification of low-confidence samples, and the verification results were divided into four categories: True defect: AI correctly identifies the defect, but a defect actually exists → add to the negative sample set. False positive: AI misjudged, but there is actually no defect → Add to the positive sample set (to correct the misjudgment). Missed detection: AI failed to detect the defect, indicating an actual flaw → Add to negative sample set → Add to negative sample set (for intensive training) New defect types: novel defects not seen by AI → labeled separately and added to a dedicated training set.

[0063] like Figure 4 As shown, the specific process of distributed trigger imaging based on absolute position includes: at times t6, t8, and t10: the vehicle moves forward at a constant speed. At time t6, the vehicle reaches the first imaging station, and the PLC sends a trigger signal to industrial camera group 1. At time t7, the vehicle continues to move, and at time t8, industrial camera group 2 is triggered. At time t9, the vehicle continues to move, and at time t10, industrial camera group 3 is triggered. At time t11, image acquisition is completed, and the camera group signal changes from high level to low level. This embodiment completes the image acquisition of the entire vehicle body component in three steps, therefore the industrial camera group completes three trigger signals. During this process, the PLC compares the absolute position value fed back by the encoder with the trigger point in the preset scheme in real time. Therefore, the PLC also synchronously generates three trigger signals to control the camera to take pictures. Different vehicle models can be configured accordingly according to their needs.

[0064] This embodiment pre-sets different numbers of photo-taking steps and corresponding locations depending on the vehicle model.

[0065] The following explanation uses the three-step photography method to collect images of all blocked parts on the vehicle as an example.

[0066] The first step is to trigger the camera to take pictures of the front of the vehicle (front fender, engine compartment).

[0067] The second step is to place a trigger camera in the middle of the vehicle body (floor, door sill, left and right side panels) to take pictures.

[0068] The third step is to trigger the camera to take pictures at the rear of the vehicle (trunk, rear panel).

[0069] The specific trigger location for taking photos and the combined photo-taking method of camera access are pre-stored in the vehicle model-photography process parameter library. When multiple vehicle models are produced on mixed production lines, the body length and the location of the plugs vary depending on the vehicle model. The parameters can be flexibly configured according to different vehicle models to ensure comprehensive collection of plug data and avoid missed shots.

[0070] This process clearly demonstrates that multiple cameras do not operate simultaneously, but are precisely triggered in a time-sharing and zone-based manner based on the vehicle's real-time location. This distributed imaging mechanism replaces the traditional method of scanning with cameras carried by mechanical rotating tables or robots, significantly improving detection efficiency and reducing mechanical complexity. It ensures comprehensive image acquisition of vehicle body components while greatly reducing complex control processes. During this process, industrial vision light sources work in conjunction with the cameras to ensure uniform and stable illumination. The acquired images are first processed by an image preprocessing module, which applies reflection suppression algorithms and image enhancement algorithms based on polarization optics technology. Figure 2 (Component diagram) This is a specific optical layout (such as the combined use of polarized light sources and polarizing filters) used to suppress strong reflections from the vehicle body, ensuring the image quality input to the AI ​​system from the source and laying the foundation for high-precision analysis.

[0071] like Figure 4 As shown, the specific AI visual inspection and analysis process includes: At time t6: Image acquisition at the first camera location is completed, and the AI ​​server simultaneously begins batch analysis. This mode significantly improves processing speed, enabling the identification and classification of the installation status (normal / incorrect / missing / defective) of obstructed parts within milliseconds, thus meeting the cycle time requirements of high-speed production lines and achieving real-time alarms and vehicle interception. At time t12: The vehicle begins to leave the inspection station, the sensor 1 signal changes from high to low, the AI ​​analysis is completed, and the results, including defect type, location, and confidence level, are sent to the PLC. At time t13: The vehicle completely leaves the inspection station, the sensor 2 signal changes from high to low, and the PLC simultaneously sends the results to the visualization screen for real-time display and uploads them to the MES system for archiving. This achieves real-time visualization and digital traceability of the inspection results.

[0072] The above design in this embodiment is to match the high-speed production cycle. The vehicle travels continuously at a constant speed. If we wait for all data points to be collected before performing unified analysis, the processing time would accumulate, exceeding the cycle time window. This embodiment adopts a simultaneous data collection and analysis mode. While the first data point is being analyzed, the vehicle is moving to the second data point and completing data collection, achieving time overlap between data collection and analysis. The analysis processes for the three data points are independent and not progressive, each responsible for covering different spatial areas of the vehicle body.

[0073] The specific steps are as follows: First, the images collected from different locations were analyzed separately.

[0074] Image acquisition at point 1 completed → AI analysis of this batch of images is immediately triggered (based on the ROI library corresponding to this point).

[0075] Image acquisition at point 2 completed → AI analysis of this batch of images is triggered immediately.

[0076] Image acquisition at point 3 is complete → AI analysis of this batch of images is immediately triggered. The analysis process for each point is completely independent and executed in parallel in its own thread.

[0077] Next, the analysis results will be temporarily saved.

[0078] After each location analysis is completed, the detection results (missing / incorrectly installed / normal) are temporarily stored in memory and bound to the VIN code. They will be uploaded collectively after the entire location analysis process is finished.

[0079] Finally, the temporarily stored results are merged and reported.

[0080] Once the analysis of all three locations is complete (the vehicle leaves the inspection area), the system aggregates and merges the inspection results from the three locations to generate a comprehensive vehicle blockage inspection report. This report includes: the inspection status, defect type, and confidence level for each location and each ROI area. The merged, complete report is then submitted to the MES system and displayed on a large visualization screen.

[0081] S3: Conduct closed-loop inspection and control based on the analysis results. For vehicles that fail the analysis, manual review is performed, and the manual review results are added to the training samples of the visual inspection model for iterative optimization of the model.

[0082] This step is implemented based on the execution and interaction layer, which includes a central visualization screen on the production line, a production line audible and visual early warning unit, a data interaction unit, and a production line MES system. The PLC (Programmable Logic Controller) sends inspection reports and quality data to the central visualization screen on the production line for visual display, and sends alarm or status control signals to the production line audible and visual early warning unit. The data interaction unit performs PDA quality inspection through a manual re-inspection terminal, then sends the manual re-inspection instructions and results back to the production line MES system, and sends the corresponding manually reviewed samples back to the AI ​​vision inspection and analysis system and model iteration optimization module. After receiving the manual re-inspection instructions and results, the production line MES system, in conjunction with the inspection results and alarm signals generated by the PLC and the synchronization signals of the production cycle, executes the instructions and records them in the process parameter library.

[0083] Specifically, the execution and interaction layer is used for closed-loop inspection and control through human-machine collaboration based on the analysis results. This includes manually reviewing vehicles that fail the analysis, and adding the manual review results to the training samples of the visual inspection model for iterative optimization. It also reports the inspection results to the MES system and executes release, alarm, or interception operations based on the results. The execution and interaction layer interacts with the upper-level manufacturing execution system via an industrial switch. All layers are connected through an industrial network, and data and control flows are as follows: Figure 1 As shown, this constitutes a complete information loop.

[0084] In one specific implementation, the Manufacturing Execution System (MES) is communicatively connected to the AI ​​vision inspection and analysis system to receive and record inspection results, thereby enabling quality traceability and closed-loop management.

[0085] In one specific implementation, the production line audible and visual early warning unit includes an audible and visual alarm. When the AI ​​vision inspection and analysis system determines that the status is unqualified, it triggers the audible and visual alarm through the PLC programmable logic controller to issue a directional alarm.

[0086] like Figure 4As shown, the specific execution and interaction process includes: After the AI ​​model determines the result, the process branches out: if the result is "OK", the vehicle passes normally and the process ends. If the result is "NG", the closed-loop optimization process begins. Real-time control: The PLC immediately triggers an audible and visual alarm and can control the line to stop, thereby intercepting the vehicle. Manual review: Through the quality inspection PDA, workers are notified to conduct on-site review and confirmation of the NG parts. The results of the manual review (whether it confirms a defect or corrects the AI's misjudgment) will serve as extremely valuable high-quality samples, along with the low-confidence samples determined by the AI ​​itself, and will be fed back to the model iteration and optimization module of the AI ​​system. The system uses this on-site data to incrementally train the deep learning model periodically or triggered by events. This enables the system to continuously learn new defect features and continuously optimize the judgment threshold, thus possessing the adaptive and self-learning ability to become more accurate with use, effectively combating model degradation and forming a strong technical barrier. At time t15, the current inspection process is completely completed, the encoder position is reset or the offset is recorded, and the system prepares for the next inspection cycle.

[0087] Example 2: Embodiment 2 of the present invention provides a closed-loop inspection and control system for coating blockages based on distributed vision, such as... Figures 1 to 5 As shown, it includes a perception layer, a control layer, and an execution and interaction layer.

[0088] The perception layer includes an information identification unit, a vehicle body position sensing unit, an image acquisition unit, and an image preprocessing module. The information identification unit is a vehicle model RFID reader / writer, used to identify vehicle model or code information, sending it to the vehicle model-photography process parameter database to query the photography scheme, and also sending it to the PLC programmable logic controller (PLC). This information, along with information collected by the photoelectric / proximity sensor group, serves as a trigger signal for vehicle entry. The vehicle body position sensing unit includes a photoelectric / proximity sensor group and a roller encoder. The roller encoder collects absolute position signals and sends them to the PLC. The image acquisition unit includes a distributed industrial camera group (12 units in this embodiment), controlled by the PLC generating photography signals based on different vehicle models and times. This includes 8 cameras on the top, 2 on the left, and 2 on the right. The image acquisition unit also includes an industrial vision light source, used to provide light. It works in conjunction with the cameras and is controlled by lighting control signals generated by the PLC. The image data acquired by the image acquisition unit is input to the image preprocessing module, which performs polarization fusion and image enhancement preprocessing on multi-angle vehicle body images to obtain processed image data.

[0089] The control layer includes a vehicle model-photography process parameter library, a PLC (Programmable Logic Controller), an industrial switch, an AI vision inspection and analysis system, an intelligent inspection unit, and a model iteration and optimization module. The vehicle model-photography process parameter library sends corresponding preset parameters to the PLC based on real-time vehicle model information and location. The PLC then sends trigger signals or vehicle model information to the AI ​​vision inspection and analysis system via the industrial switch. The AI ​​vision inspection and analysis system analyzes the processed image data obtained from the image preprocessing module based on model inference support provided by the intelligent inspection unit, obtaining defect location, type, and execution degree information, which is sent back to the PLC. The system then sends the inspection result (OK or NG) to the industrial switch and sends the generated low-confidence or defect samples to the model iteration and optimization module.

[0090] The execution and interaction layer includes a central visualization screen for the production line, a production line audible and visual early warning unit, a data interaction unit, and a production line MES system. The PLC (Programmable Logic Controller) sends inspection reports and quality data to the central visualization screen for display and sends alarm or status control signals to the production line audible and visual early warning unit. The data interaction unit performs PDA quality inspection via a manual re-inspection terminal, then sends the manual re-inspection instructions and results back to the production line MES system, and sends the corresponding manually reviewed samples back to the AI ​​vision inspection and analysis system and model iteration optimization module. After receiving the manual re-inspection instructions and results, the production line MES system, in conjunction with the inspection results and alarm signals generated by the PLC and the synchronization signals of the production cycle, executes the instructions and records them in the process parameter library.

[0091] The details are as follows: The perception layer is used to detect vehicle model information and acquire multi-angle images of the vehicle body based on the vehicle model information and the vehicle's real-time position. Specifically, as the vehicle to be detected passes through the detection area at a constant speed, distributed image acquisition is triggered based on changes in its real-time position.

[0092] In one specific implementation, the information identification unit includes an RFID reader / writer for dynamically identifying vehicle model information entering the inspection station. When a vehicle enters, the RFID reader / writer reads the code information from the disc at the bottom of the skid to obtain unique vehicle model information, thereby selecting the corresponding preset photographic inspection scheme for that model. This ensures that each vehicle can be photographed in the optimal way, with imaging covering all components of the vehicle body.

[0093] In one specific implementation, the vehicle body position sensing unit consists of a sensor (photoelectric sensor) and a high-precision encoder installed at the inspection station of the painting and rough sealing production line. The sensor is used to detect the arrival and departure of the vehicle, while the encoder rotates with the spindle of the roller bed to provide the absolute position signal of the vehicle, enabling precise triggering.

[0094] In one specific implementation, the image acquisition unit comprises multiple industrial cameras spatially distributed via a gantry structure, used to acquire images of the moving vehicle body from different angles. The image acquisition unit includes a distributed industrial camera group, which is used for image acquisition. The distributed industrial camera group consists of 12 industrial cameras arranged in an optimized spatial distribution, such as... Figure 2 As shown, the distributed industrial camera group, fixed on the gantry spanning the production line, specifically includes a top camera group (8 units) responsible for capturing images of the upper and lower positions of the front fenders, the front, middle and rear floor panels, and the door sill area of ​​the vehicle body; a left camera group (2 units); and a right camera group (2 units) responsible for capturing images of the vehicle body's trunk, left and right rear panels, and side panels. The captured images are as follows: Figure 3 As shown in the diagram, the design concept of this layout is based on the installation location of the blockage parts throughout the vehicle. Since these blockage parts cover various locations on the vehicle, a gantry spanning the production line is used as the camera mounting device, with the vehicle as the inspection center, and 12 industrial cameras distributed around it. During the inspection process, the system will, according to a preset plan, call different cameras at different shooting positions as image acquisition units to complete the blockage part image acquisition step by step, thereby achieving comprehensive, blind-spot-free coverage of the moving vehicle body blockage parts, adapting to different vehicle models. To improve image quality, this unit also integrates a collaboratively controllable industrial vision light source. During the image acquisition process, the light source will be turned on according to a preset plan to provide support, thereby acquiring clearer and moderately bright images, ensuring high-quality imaging from the source.

[0095] In this embodiment, the industrial vision light source that can be controlled collaboratively adopts centralized control by PLC and utilizes the distributed execution architecture of the light source controller to achieve precise collaboration between the light source and the camera.

[0096] Specifically, to achieve high-quality imaging, this embodiment constructs a collaborative control architecture combining a PLC master controller and a light source controller slave controller. Each industrial camera is equipped with an independent light source module. All light source modules are uniformly connected to the control system through the light source controller. The light source controller is connected to the PLC via a high-speed I / O interface, receiving trigger commands from the PLC and precisely controlling the lighting timing, lighting duration, and brightness level of the light source according to preset parameters.

[0097] The multi-angle vehicle body images include polarization orthogonal images and polarization parallel images. The image preprocessing module performs polarization fusion and image enhancement preprocessing operations on the multi-angle vehicle body images to obtain the processed image data.

[0098] Specifically, an adjustable polarizing filter is installed in front of the lens of each industrial camera, and a polarizing film is installed in front of the corresponding light source. The polarization directions of the two are set to be orthogonal to physically suppress specular reflection. During image acquisition, the camera captures two images for each shooting point: the first is an orthogonally polarized image (strongly suppressing reflection), and the second is a parallel polarized image (normal brightness). After acquisition, the two images are fused pixel-level to generate a fused image that effectively eliminates reflective spots while fully preserving the details of the obstruction. Adaptive contrast adjustment is applied to the fused image to make the grayscale difference between the obstruction area and the vehicle background more obvious. Edge sharpening is performed to enhance the contour features of the obstruction. Based on a preset target brightness range, overall brightness equalization is performed on the image to ensure consistent brightness across images captured from different points. After the above image preprocessing, higher quality image data can be obtained.

[0099] like Figure 4 As shown, the specific data acquisition process includes: At time t0, the system is in standby preparation state. At times t1 and t2, the skid carrying the electrophoresis vehicle body begins to enter the inspection station. Sensors 1 and 2 sequentially detect the vehicle body's arrival, and the signals sequentially change from low to high, indicating that the vehicle has fully entered the inspection area. Simultaneously, the encoder starts working, and its output absolute position value linearly increases from the initial value, indicating that the vehicle is continuously moving forward with the production line, providing the system with a precise reference for the vehicle's real-time position. At time t3, when the vehicle moves to the RFID identification area, the RFID reader is triggered to read the vehicle model information, and the signal changes from low to high. At time t4, the RFID reading is complete, and the vehicle continues to move forward. The RFID transmits the acquired information to the PLC via a signal. Upon receiving this information, the PLC immediately dynamically calls and loads a photography scheme that perfectly matches the vehicle model from its internal vehicle model-photography process parameter library. This scheme predefines the trigger position, the number of cameras used, and their corresponding camera parameters for each inspection step, realizing a leap from "fixed program" to "vehicle model adaptation," greatly improving system flexibility. Suitable for production line scenarios involving the mixed production of multiple vehicle models. At time t5, the vehicle continues to move forward, and the RFID signal changes from high to low, awaiting the next stage of the process.

[0100] The control layer is used to analyze the misinstallation and omission of parts in multi-angle vehicle body images using a visual inspection model. It is also used to retrieve pre-stored detection parameters for the vehicle model based on the vehicle model information, and to accurately trigger the corresponding camera to take pictures during the uniform movement of the vehicle body based on the real-time position signal of the encoder.

[0101] In one specific implementation, the PLC (Programmable Logic Controller) is connected to the information recognition unit, the roller encoder, and the image acquisition unit. The PLC internally stores a vehicle model-photography process parameter library, containing preset camera parameter configurations for each vehicle model. Upon receiving signals from the sensing layer, the PLC dynamically schedules the entire inspection process according to preset logic.

[0102] Specifically, the PLC internally stores a preset database of vehicle model photography parameters. This database predefines a complete photography process scheme for each vehicle model, including: the total number of detection steps, the absolute trigger position of each step (coordinates based on the encoder zero point), a list of camera IDs to be called at each step, and the optical parameters of each camera (such as exposure time, gain, and light source brightness). After receiving the vehicle model information identified by RFID, the PLC immediately calls the information from the corresponding vehicle model photography parameter database. While the vehicle is moving at a constant speed, it compares the absolute position fed back by the encoder with the preset trigger position in the database in real time. When the positions match, it accurately triggers the corresponding camera group to take pictures.

[0103] In one specific implementation, the AI ​​visual inspection and analysis system includes an industrial computer core with a deep learning visual inspection model trained on the YOLOv8s architecture. This model performs intelligent analysis of mis-installed or missing parts in the acquired images. Specifically, the steps for analyzing mis-installed or missing parts in multi-angle vehicle body images using the visual inspection model are as follows: First, acquire known labeled data to form training samples. These training samples can be expanded from manually reviewed results or previously identified low-confidence samples. Second, construct a visual inspection model based on the YOLOv8s architecture. Third, train the visual inspection model using the training samples. Finally, use the trained visual inspection model to inspect real-time multi-angle vehicle body images.

[0104] Specifically, it includes the following steps: Step 1: Construct training samples and perform data partitioning and preprocessing.

[0105] Training samples were constructed using historical image databases, manual verification results, and low-confidence samples, including positive samples (correctly installed) and negative samples (missing, incorrectly installed, or warped edges). These samples were then divided into training, validation, and test sets in a 7:2:1 ratio. Online enhancement preprocessing operations such as random rotation, scaling, brightness adjustment, and flipping were performed on the training set.

[0106] Step 2: Build and initialize the visual detection model.

[0107] The YOLOv8s architecture was selected, pre-trained weights were loaded, and the model's feature extraction capability was preserved. The last layer of the model (classification layer) was modified so that the number of output categories matched the number of component types (e.g., normal, missing, incorrect).

[0108] Step 3: Train and test the visual detection model using training samples.

[0109] Train the model using the training set, evaluate it on the validation set after each round, and save the best model. Evaluate the final model's performance using the test set, and deploy it once it meets the target.

[0110] Step 4: After going live, continue incremental learning.

[0111] Regularly collect new review samples, load the current model weights, freeze the underlying layer, fine-tune the upper layer, and replace the deployment after verification.

[0112] The visual inspection model in this embodiment can accurately identify and locate the installation status of obstructed components (normal, incorrectly installed, or missing). This embodiment uses YOLOv8s as the base model because it achieves an excellent balance between accuracy and speed. Its superior feature extraction capabilities and lightweight structure are particularly suitable for the high requirements of real-time detection and accuracy in this scenario. Its innovation lies in integrating a model iteration and optimization module. During each system run, low-confidence data is fed back, allowing the model to continuously learn and iterate, ensuring that the model has a large-scale training dataset for both positive and negative sample recognition.

[0113] Specifically, the visual inspection model consists of a model scheduling layer, a model inference layer, and a result fusion layer: The model scheduling layer manages the timing of model calls for different image capture steps and different camera images. It is responsible for receiving images and allocating processing threads.

[0114] The model inference layer is used in the core YOLOv8s model and is responsible for object detection and classification of a single ROI sub-image.

[0115] The result fusion layer is used to summarize the detection results of each step and each point according to spatial location.

[0116] This embodiment of the visual detection model adopts a strategy of step-by-step independent processing and result spatial fusion. The first step is to receive image batches step by step. The system processes the images acquired in each imaging step as an independent batch.

[0117] For example: Front of the vehicle: Acquire Cam1-Cam3 images → Form a batch.

[0118] Middle of the vehicle: Collect Cam4-Cam8 → Form the second batch.

[0119] Rear of the vehicle: Cam9-Cam12 → forming the third batch.

[0120] The second step involves parallel processing within a batch. For each batch of multiple images, the system initiates multi-threaded parallel inference using the YOLOv8s model: each image is cropped into multiple ROI sub-images based on a pre-stored ROI template. The model performs independent object detection and classification for each ROI sub-image. Each ROI outputs: presence of a blockage, blockage type, and confidence level. The third step temporarily stores the batch results. After each step's analysis is complete, the system temporarily stores the detection results in memory and binds them to the current vehicle's VIN code: 1. Record the detection status of all ROIs at point 1. 2. Record the detection status of all ROIs at point 2. 3. Record the detection status of all ROIs at point 3.

[0121] The fourth step is to fuse all results based on the spatial dimension. Once all three steps are completed (the vehicle leaves the inspection area), the result fusion layer is activated: according to the predefined spatial partitions in the ROI template (front fender, sill, trunk, etc.), the inspection results of the three steps are spatially stitched together to generate a complete vehicle blockage inspection report, including the inspection status of each spatial location.

[0122] The fifth step is to report and display the results. The integrated, complete report is reported to the MES system for quality traceability. It is also pushed to a large visualization screen to display the OK / NG status of each area. If any point is NG, an audible and visual alarm is triggered.

[0123] In one specific implementation, the model iteration optimization module is used to perform incremental learning and optimization of the AI ​​model based on low-confidence samples and manual review results.

[0124] Specifically, the first step is to screen samples based on confidence levels.

[0125] In this embodiment, the system sets a confidence threshold of 0.85. When the confidence level of the visual detection model for the detection result is lower than the threshold, the sample is defined as a low-confidence sample and enters the queue for review. Samples with a confidence level higher than the threshold are considered high-confidence samples and are directly added to the database as positive samples.

[0126] Next, manual PDA on-site verification and classification were conducted. Quality inspectors used PDAs to conduct on-site verification of low-confidence samples, and the verification results were divided into four categories: True defect: AI correctly identifies the defect, but a defect actually exists → add to the negative sample set. False positive: AI misjudged, but there is actually no defect → Add to the positive sample set (to correct the misjudgment). Missed detection: AI failed to detect the defect, indicating an actual flaw → Add to negative sample set → Add to negative sample set (for intensive training) New defect types: novel defects not seen by AI → labeled separately and added to a dedicated training set.

[0127] In this embodiment, during the distributed image acquisition process triggered by real-time location changes, the industrial camera group is configured to acquire images three times based on the location changes. Figure 4 As shown, the specific process of distributed trigger imaging based on absolute position includes: at times t6, t8, and t10: the vehicle moves forward at a constant speed. At time t6, the vehicle reaches the first imaging station, and the PLC sends a trigger signal to industrial camera group 1. At time t7, the vehicle continues to move, and at time t8, industrial camera group 2 is triggered. At time t9, the vehicle continues to move, and at time t10, industrial camera group 3 is triggered. At time t11, image acquisition is completed, and the camera group signal changes from high level to low level. This embodiment completes the image acquisition of the entire vehicle body component in three steps, therefore the industrial camera group completes three trigger signals. During this process, the PLC compares the absolute position value fed back by the encoder with the trigger point in the preset scheme in real time. Therefore, the PLC also synchronously generates three trigger signals to control the camera to take pictures. Different vehicle models can be configured accordingly according to their needs.

[0128] This embodiment pre-sets different numbers of photo-taking steps and corresponding locations depending on the vehicle model.

[0129] The following explanation uses the three-step photography method to collect images of all blocked parts on the vehicle as an example.

[0130] The first step is to trigger the camera to take pictures of the front of the vehicle (front fender, engine compartment).

[0131] The second step is to place a trigger camera in the middle of the vehicle body (floor, door sill, left and right side panels) to take pictures.

[0132] The third step is to trigger the camera to take pictures at the rear of the vehicle (trunk, rear panel).

[0133] The specific trigger location for taking photos and the combined photo-taking method of camera access are pre-stored in the vehicle model-photography process parameter library. When multiple vehicle models are produced on mixed production lines, the body length and the location of the plugs vary depending on the vehicle model. The parameters can be flexibly configured according to different vehicle models to ensure comprehensive collection of plug data and avoid missed shots.

[0134] This process clearly demonstrates that multiple cameras do not operate simultaneously, but are precisely triggered in a time-sharing and zone-based manner based on the vehicle's real-time location. This distributed imaging mechanism replaces the traditional method of scanning with cameras carried by mechanical rotating tables or robots, significantly improving detection efficiency and reducing mechanical complexity. It ensures comprehensive image acquisition of vehicle body components while greatly reducing complex control processes. During this process, industrial vision light sources work in conjunction with the cameras to ensure uniform and stable illumination. The acquired images are first processed by an image preprocessing module, which applies reflection suppression algorithms and image enhancement algorithms based on polarization optics technology. Figure 2(Component diagram) This is a specific optical layout (such as the combined use of polarized light sources and polarizing filters) used to suppress strong reflections from the vehicle body, ensuring the image quality input to the AI ​​system from the source and laying the foundation for high-precision analysis.

[0135] like Figure 4 As shown, the specific AI visual inspection and analysis process includes: At time t6: Image acquisition at the first camera location is completed, and the AI ​​server simultaneously begins batch analysis. This mode significantly improves processing speed, enabling the identification and classification of the installation status (normal / incorrect / missing / defective) of obstructed parts within milliseconds, thus meeting the cycle time requirements of high-speed production lines and achieving real-time alarms and vehicle interception. At time t12: The vehicle begins to leave the inspection station, the sensor 1 signal changes from high to low, the AI ​​analysis is completed, and the results, including defect type, location, and confidence level, are sent to the PLC. At time t13: The vehicle completely leaves the inspection station, the sensor 2 signal changes from high to low, and the PLC simultaneously sends the results to the visualization screen for real-time display and uploads them to the MES system for archiving. This achieves real-time visualization and digital traceability of the inspection results.

[0136] The above design in this embodiment is to match the high-speed production cycle. The vehicle travels continuously at a constant speed. If we wait for all data points to be collected before performing unified analysis, the processing time would accumulate, exceeding the cycle time window. This embodiment adopts a simultaneous data collection and analysis mode. While the first data point is being analyzed, the vehicle is moving to the second data point and completing data collection, achieving time overlap between data collection and analysis. The analysis processes for the three data points are independent and not progressive, each responsible for covering different spatial areas of the vehicle body.

[0137] The specific steps are as follows: First, the images collected from different locations were analyzed separately.

[0138] Image acquisition at point 1 completed → AI analysis of this batch of images is immediately triggered (based on the ROI library corresponding to this point).

[0139] Image acquisition at point 2 completed → AI analysis of this batch of images is triggered immediately.

[0140] Image acquisition at point 3 is complete → AI analysis of this batch of images is immediately triggered. The analysis process for each point is completely independent and executed in parallel in its own thread.

[0141] Next, the analysis results will be temporarily saved.

[0142] After each location analysis is completed, the detection results (missing / incorrectly installed / normal) are temporarily stored in memory and bound to the VIN code. They will be uploaded collectively after the entire location analysis process is finished.

[0143] Finally, the temporarily stored results are merged and reported.

[0144] Once the analysis of all three locations is complete (the vehicle leaves the inspection area), the system aggregates and merges the inspection results from the three locations to generate a comprehensive vehicle blockage inspection report. This report includes: the inspection status, defect type, and confidence level for each location and each ROI area. The merged, complete report is then submitted to the MES system and displayed on a large visualization screen.

[0145] The execution and interaction layer is used for closed-loop inspection and control through human-machine collaboration based on the analysis results. This includes manually reviewing vehicles that fail the analysis, and adding the manual review results to the training samples of the visual inspection model for iterative optimization. It also reports the inspection results to the MES system and executes release, alarm, or interception operations based on the results. The execution and interaction layer interacts with the upper-level manufacturing execution system via an industrial switch. All layers are connected through an industrial network, and data and control flows are as follows: Figure 1 As shown, this constitutes a complete information loop.

[0146] In one specific implementation, the Manufacturing Execution System (MES) is communicatively connected to the AI ​​vision inspection and analysis system to receive and record inspection results, thereby enabling quality traceability and closed-loop management.

[0147] In one specific implementation, the production line audible and visual early warning unit includes an audible and visual alarm. When the AI ​​vision inspection and analysis system determines that the status is unqualified, it triggers the audible and visual alarm through the PLC programmable logic controller to issue a directional alarm.

[0148] like Figure 4 As shown, the specific execution and interaction process includes: After the AI ​​model determines the result, the process branches out: If the result is "OK", the vehicle passes normally and the process ends. If the result is "NG", the closed-loop optimization process begins. Real-time control: The PLC immediately triggers an audible and visual alarm and can control the line to stop, thereby intercepting the vehicle. Manual review: Through the quality inspection PDA, workers are notified to conduct on-site review and confirmation of the NG parts. The results of the manual review (whether it confirms a defect or corrects the AI's misjudgment) will serve as extremely valuable high-quality samples, along with the low-confidence samples determined by the AI ​​itself, and will be fed back to the model iteration and optimization module of the AI ​​system. The system uses this on-site data to incrementally train the deep learning model periodically or triggered by events. This allows the system to continuously learn new defect features and continuously optimize the judgment threshold, thus possessing the adaptive and self-learning ability to become more accurate with use, effectively combating model degradation and forming a strong technical barrier. At time t15, the current inspection process is completely completed, the encoder position is reset or the offset is recorded, and the system prepares for the next inspection cycle.

[0149] The steps and methods involved in the above embodiment two correspond to those in embodiment one. For specific implementation details, please refer to the relevant description section of embodiment one.

[0150] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0151] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc.

[0152] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A closed-loop inspection and control method for coating blockages based on distributed vision, characterized in that, Includes the following steps: The vehicle model information is detected, and multi-angle images of the vehicle body are collected based on the vehicle model information and the real-time position of the vehicle. Specifically, as the vehicle to be detected passes through the detection area at a constant speed, distributed image acquisition is triggered based on the changes in the real-time position. Visual inspection models were used to analyze the misinstallation and omission of parts in multi-angle vehicle body images; Based on the analysis results, a closed-loop inspection and control system involving human and machine collaboration is implemented. Vehicles that fail the analysis results are manually reviewed, and the results of the manual review are added to the training samples of the visual inspection model for iterative optimization.

2. The closed-loop inspection and control method for coating blockages based on distributed vision as described in claim 1, characterized in that, Multi-angle vehicle body images include polarization orthogonal images and polarization parallel images. Polarization fusion and image enhancement preprocessing operations are performed on the multi-angle vehicle body images.

3. The closed-loop inspection and control method for coating blockages based on distributed vision as described in claim 1, characterized in that, RFID readers are used to dynamically identify vehicle model information entering the inspection station. When a vehicle enters, the RFID reader reads the code information in the disc at the bottom of the skid to obtain unique vehicle model information.

4. The closed-loop inspection and control method for coating blockages based on distributed vision as described in claim 1, characterized in that, Image acquisition is performed using a distributed industrial camera group, which specifically includes a top camera group, a left camera group, and a right camera group. During the distributed image acquisition process, which is triggered by real-time position changes, the industrial camera group is set to acquire images three times in total based on the position changes.

5. The closed-loop inspection and control method for coating blockages based on distributed vision as described in claim 1, characterized in that, The specific steps for analyzing the misinstallation and omission of parts in multi-angle vehicle body images using a visual inspection model are as follows: Obtain known labeled data to form training samples; Construct a visual detection model based on the YOLOv8s architecture; The visual detection model is trained using training samples; The trained visual detection model is used to detect multi-angle vehicle images in real time.

6. A closed-loop inspection and control system for coating blockages based on distributed vision, characterized in that, include: The perception layer is used to detect vehicle model information and acquire multi-angle images of the vehicle body based on the vehicle model information and the real-time position of the vehicle. Specifically, as the vehicle to be detected passes through the detection area at a constant speed, distributed image acquisition is triggered based on the changes in the real-time position. The control layer is used to analyze the misinstallation and omission of parts in multi-angle vehicle body images using a visual inspection model; The execution and interaction layer is used for closed-loop inspection and control based on the analysis results. Among them, vehicles with unqualified analysis results are manually reviewed, and the manual review results are added to the training samples of the visual inspection model for iterative optimization of the model.

7. The closed-loop inspection and control system for coating blockages based on distributed vision as described in claim 6, characterized in that, The perception layer includes an image preprocessing module for performing polarization fusion and image enhancement preprocessing operations on multi-angle vehicle images, which include polarization orthogonal images and polarization parallel images.

8. The closed-loop inspection and control system for coating blockages based on distributed vision as described in claim 6, characterized in that, RFID readers are used to dynamically identify vehicle model information entering the inspection station. When a vehicle enters, the RFID reader reads the code information in the disc at the bottom of the skid to obtain unique vehicle model information.

9. The closed-loop inspection and control system for coating blockages based on distributed vision as described in claim 6, characterized in that, Image acquisition is performed using a distributed industrial camera group, which specifically includes a top camera group, a left camera group, and a right camera group. During the distributed image acquisition process, which is triggered by real-time position changes, the industrial camera group is set to acquire images three times in total based on the position changes.

10. The closed-loop inspection and control system for coating blockages based on distributed vision as described in claim 6, characterized in that, In the control layer, the specific steps for analyzing the misinstallation and omission of parts in multi-angle vehicle body images using a visual inspection model are as follows: Obtain known labeled data to form training samples; Construct a visual detection model based on the YOLOv8s architecture; The visual detection model is trained using training samples; The trained visual detection model is used to detect multi-angle vehicle images in real time.

Images