An AI vision-based intelligent mistake-proofing system and method for automobile parts

Abstract

The application discloses an automobile part intelligent mistake-proofing system and method based on AI vision and real-time data interaction. The system is deployed at an artificial assembly station and is used for preventing mixed assembly errors of multiple appearance-similar variant parts in the same function category. An appearance image of a non-identified part is directly collected by a vision collection unit, and a built-in deep learning model is used to identify the visual feature category of the part. The deep learning model is used to identify the intrinsic feature of the part. Meanwhile, vehicle VIN and station information are acquired in real time, standard assembly requirements are dynamically queried, and cross comparison is performed. If the information is inconsistent, physical blocking, sound and light alarm and remote alarm are triggered immediately. The application significantly improves the mistake-proofing accuracy and reliability, can be perfectly matched with the production line rhythm, and has high system integration and compatibility.

Description

Technical Field

[0001] This invention relates to an intelligent error prevention system and method for automotive parts, specifically to an intelligent error prevention system and method based on AI visual recognition and real-time production data interaction applied to automotive assembly lines, belonging to the interdisciplinary field of industrial manufacturing and artificial intelligence. Background Technology

[0002] Currently, error-proofing technologies for automotive assembly lines mainly fall into two categories: Traditional label scanning and manual error prevention technologies: These widely utilize barcode scanners, RFID readers, and manual visual inspection (refer to the "Automotive Quality Management System Standard"). Their structure mainly consists of a label reading unit (such as a barcode scanner or RFID antenna), a data comparison unit (such as an industrial computer or PLC), and an execution prompting unit (such as indicator lights or buzzers). The principle is to scan the barcode / RFID tags on parts or carriers to obtain label information and compare it with the production order, or to set up manual inspection points at key workstations, relying on operator experience and checklists for visual verification. If the information does not match, an alarm is issued, thus achieving assembly error prevention.

[0003] Conventional machine vision inspection technology: Some high-end production lines have begun to apply visual inspection technologies based on traditional image processing or early machine learning (refer to published patent CN120563378A "Visual Inspection Correction Method and Equipment for Automotive Parts"). Its core structure includes an image acquisition module (fixed industrial camera), an image processing module (based on edge detection, template matching, and other algorithms), and a result determination module. The working principle is to capture images of the parts to be inspected using a camera, extract features such as contours and dimensions using algorithms, and compare them with preset templates to determine whether the parts exist or whether their dimensions are acceptable. However, the inspection targets are mostly "presence" or "defects," and they are usually not deeply correlated or intelligently compared with real-time vehicle information (VIN) and enterprise-level parts databases (such as ERP and MES).

[0004] 2. Problems and shortcomings of existing technologies The shortcomings of traditional label scanning and manual error-proofing technologies are as follows: ① The reliability of error-proofing is fundamentally flawed: it relies entirely on the accuracy of labels or documents and cannot identify mixed material situations where "the label is correct but the parts are installed incorrectly," resulting in blind spots in error-proofing; manual visual inspection is affected by fatigue, experience, and attention, leading to high rates of missed and incorrect inspections; ② Poor flexibility and traceability: it is difficult to adapt to the needs of multi-variety, small-batch, and rapid production changeover; manual record-keeping and traceability are difficult, and the data is incomplete.

[0005] The shortcomings of conventional machine vision inspection technology include: ① Limited recognition capabilities: Traditional image processing algorithms rely on strict lighting, angle, and positioning, are sensitive to background interference, and struggle to robustly distinguish between parts that are visually similar but have different internal structures or subtle features (such as front collision beams of different car models); ② Isolated systems lacking data fusion: Most are independent inspection stations, and the visual recognition results are not bound to the vehicle identification information (VIN) flowing in real time on the production line, nor are they cross-validated with the product design data (BOM) in real time, making it impossible to achieve precise error prevention through "one vehicle, one inspection, one comparison"; ③ Low intelligence and integration: Lack of adaptive feature recognition capabilities based on deep learning; complex system deployment, parameter debugging relying on professional personnel; typically lacks advanced functions such as remote alarms, multi-level responses, data archiving and analysis, making it difficult to integrate into modern intelligent manufacturing management systems; ④ Real-time performance failing to meet cycle time requirements: Some complex algorithms take a long time to process, which may affect the production cycle time of the production line; ⑤ Poor interactivity and deployment convenience: There is no dedicated interactive interface, requiring professional personnel to configure parameters through code, resulting in a long deployment cycle and hindering on-site operation and maintenance in the workshop.

[0006] In summary, existing error-proofing technologies for automotive assembly line parts have significant shortcomings in terms of the reliability of error-proofing principles, the accuracy of identification, the degree of system integration, and the real-time and intelligent response. They cannot meet the high standards of the automotive manufacturing industry for "zero-defect" production and full-process traceability. There is an urgent need for a deep intelligent error-proofing technology solution that integrates high-precision AI visual recognition, real-time production data interaction, intelligent early warning, and convenient interaction. Summary of the Invention

[0007] The purpose of this invention is to address the problems existing in the prior art by proposing an intelligent error prevention system for automotive parts based on AI vision and real-time data interaction. The specific technical solution is as follows: It includes: The vision acquisition unit is configured in the component assembly fixtures of the automotive discrete manufacturing field. It involves mixed-line assembly fixtures for similar parts with multiple state variations that require manual identification and judgment, including at least the automotive front collision beam assembly station, automotive headlight bracket assembly station, automotive engine compartment side beam assembly station, and automotive front and rear floor side beam assembly station. This tool is used to capture real-time images of unmarked automotive parts to be inspected at assembly stations, and supports multiple source access methods, including industrial cameras, network streaming cameras, local image files, and video files. This system can train a YOLO object detection and classification model by adding a dataset of target parts, configure the ROI region and feature code extraction rules for the corresponding parts, and extend it to error-proof detection of other automotive parts such as headlight brackets and front and rear floor side beams; The core processing unit is an industrial control computer equipped with the system software. It has a built-in trained AI vision model, which is used to preprocess the appearance images acquired by the vision acquisition unit, perform feature recognition, and output visual category results and confidence scores. The data interaction unit includes a PLC interface and a database interface. The PLC interface is used to communicate with the PLC of the production line in real time to obtain the unique identifier VIN number of the vehicle to be assembled. The database interface is used to connect to the enterprise's production MES database and query the standard component information corresponding to the vehicle based on the VIN number. It also uses a preset string processing algorithm to extract a combination of fields with clear meaning from the part number of the standard component and use it as the standard feature code of the part. The human-computer interaction unit provides a graphical user interface (GUI) and a WeChat robot interface for configuring system parameters, monitoring operating status, performing visual recognition of AI models, and visually displaying the comparison results of recognition results with standard feature codes in the MES database. The WeChat robot interface is used to send real-time alarm notifications to designated personnel when parts are misassembled or system malfunctions occur. The intelligent comparison unit is used to automatically compare the visual category results output by the core processing unit with the feature codes in the standard component information obtained from the data interaction unit to determine whether the detection result is normal. A multi-level early warning unit is used to execute preset early warning measures when the intelligent comparison unit determines that the detection result is abnormal; The data recording unit is used to automatically save complete process data for each test and generate traceable quality reports.

[0008] Furthermore, the vision acquisition unit is an industrial camera, which is set at a preset position at the assembly station of the component to be inspected to ensure that the field of view covers the area of ​​the component to be inspected, so as to capture the front collision beam image of the assembly station in real time; the industrial camera is connected to the core processing unit via USB or network, and supports multiple input sources, including local USB industrial cameras, network RTSP streaming media cameras, and local image or video files.

[0009] Furthermore, the vision unit includes at least one industrial camera, which is positioned diagonally above the assembly station of the part to be inspected, at a 45° installation angle to the horizontal direction, and at a vertical height of 2 meters from the ground, to ensure that the appearance image of the part to be inspected can be captured completely and clearly. The visual acquisition unit supports the collaborative operation of multiple cameras, and the installation height and angle of the cameras can be adjusted for different workstations to achieve synchronous acquisition of multi-angle images of parts.

[0010] Furthermore, the early warning measures of the multi-level early warning unit include: a. Hard interception: An emergency stop command is written to the PLC controlling the elevator via the Siemens S7 communication protocol. After receiving the command, the PLC forces the elevator to stop at its current position, physically blocking the car body from flowing out. b. On-site audible and visual alarm: Control the on-site three-color lights to enter red flashing mode via serial port commands to warn on-site operators; c. Remote instant notification: Call the WeChat Work robot interface to send structured alarm messages, which should include at least the anomaly type, occurrence time, VIN number, visual result, standard part number and on-site image information; d. Local Recording: Highlight error details in the GUI interface and automatically save the abnormal data, i.e. the content contained in step c, to the specified folder to generate a traceable text report.

[0011] Furthermore, the AI ​​vision model is an architecture-based target detection and classification model. For automotive parts inspection tasks, by introducing anchor frame optimization of specific size, multi-scale feature fusion enhancement, and transfer learning training strategies, the model's generalization ability and recognition stability in complex industrial environments are significantly improved, enabling real-time accurate positioning and high-precision classification of target parts.

[0012] Furthermore, in the intelligent comparison unit, if the feature code mapped by the visual category result is consistent with the feature code extracted from the standard component information and the confidence level is higher than a preset threshold, the detection result is determined to be normal; if the feature code is inconsistent, the component to be detected is not detected, or the confidence level is lower than the preset threshold, the detection result is determined to be abnormal.

[0013] Furthermore, the data recording unit is used to automatically save complete process data for each test. The complete process data includes the original image, the processed image, the AI ​​recognition result, the comparison record, the VIN, and the timestamp, in order to generate a traceable electronic quality archive.

[0014] PLC communication module: It adopts the snap7 communication library and communicates with the production line PLC in real time via the Siemens S7 protocol based on TCP / IP to obtain the VIN number data stored in the DB block; Database query module: Connects to the enterprise MES production database through Oracle Instant Client and oracledb library, and executes standard query process: Query table UDA_EXTEND_WORK_ORDER_V to get WIP_ENTITY_NAME_W based on VIN number, and then query table ERP_ORDER_BOM_V to get PART_NUM standard part number; Feature extraction module: Extracts feature codes from standard part numbers using preset regular expression rules, including: Main rule: r'-ED(\d{2})-', used to match part numbers in the format XXXX-ED12-XXXX; Alternative rule: r'ED(\d{2})', used to match part numbers in the format ED12 or XXXXED12XXXX; If extraction fails, the system will trigger an enterprise WeChat robot interface to push a system abnormality alarm, notifying relevant personnel that "feature code extraction failed, please check the VIN number or database connection", and prompting the operator to intervene in the human-computer interaction interface.

[0015] A smart error prevention method for automotive parts based on AI vision and real-time data interaction includes the following steps: Step 1: System initialization and configuration, start the core processing software on the industrial control computer, and load the pre-trained YOLO object detection and classification model; Configure system parameters via the graphical user interface (GUI): Camera parameters: Select camera type, set screen rotation angle, define ROI (Region of Interest); PLC communication parameters: Set the PLC's IP address, rack number, slot number, DB block number, and VIN data offset address; Database connection parameters: Set the database address, port, service name, username, and password; Warning parameters: Set the three-color light control port, the Enterprise WeChat robot Webhook URL, and the list of alarm recipients; The system automatically tests the connection status of each hardware component, including the camera, PLC, database, and serial port LED. If the test fails, a specific fault message is displayed. Automatic retry is supported. If the retry fails, manual repair guidance is provided. After initialization, "System ready" is displayed in the GUI interface. Step 2: VIN Triggering and Image Acquisition. The system continuously monitors the VIN number data area at offset address 486 in the specified DB block (DB23) of the PLC through the PLC interface. When it detects that the currently read VIN number is inconsistent with the last valid VIN number recorded by the system, and the newly read VIN number is not an empty string, a complete error prevention detection process is automatically triggered. The image acquisition unit immediately controls the camera to capture a high-definition image of the current workstation (resolution not less than 1280×720 pixels, clear image, and identifiable feature areas of the parts to be inspected). If the acquisition fails or the image quality does not meet the set threshold, the system will automatically retry up to 3 times. If it fails three times in a row, the system will record the abnormality and send an "image acquisition failure" alarm notification through the enterprise WeChat robot interface, while prompting the operator to intervene on the human-machine interface. Step 3: AI recognition and database query, rotate the acquired image, supporting four fixed angle rotations of 0°, 90°, 180° and 270° to ensure that the parts maintain a positive posture in the image, and then crop the image according to the defined ROI area; The preprocessed image is input into a detection model optimized based on the YOLOv11 architecture. The model outputs detection boxes with confidence scores. An exact bounding box is defined as a rectangular region that accurately matches the target location with at least a 50% Intersection over Union (IoU) and has a detection confidence score higher than a set threshold (default 0.5). The component regions within this bounding box are extracted, and the YOLO-optimized classification model outputs visual classification results and corresponding confidence scores. Simultaneously, using the current VIN number, the core processing unit connects to the database interface of the data interaction unit to perform a VIN number query, obtaining the standard component part number corresponding to the vehicle. From the returned part number, the system extracts its feature code according to preset regular expression rules. The regular expressions used in this system mainly adapt to the "ED + two digits" format, such as the pattern r'-ED(\d{2})-' or r'ED(\d{2})', which can accommodate most standard part numbers. If the part number format changes, the system supports updating the regular expression rules through configuration to adapt to different encoding structures; if feature code extraction fails, the system will record the exception, push an alarm, skip the subsequent comparison process, and mark this detection as "ED code extraction failed". At the same time, it supports manual intervention by the operator in the graphical interface. Step 4: Intelligent comparison, The intelligent comparison unit acquires the visual category results and confidence scores from the AI ​​recognition output, as well as the feature codes extracted from the database query. The system compares the feature codes mapped from the visual results with the feature codes extracted from the part number to see if they are completely consistent. If they are completely consistent and the confidence score is higher than the set threshold, the detection is considered successful. If the comparison is inconsistent, the system will immediately trigger a serious alarm through the WeChat robot interface to notify relevant personnel that "the parts do not match or the assembly is incorrect," and send an emergency stop command to the assembly line control system through the PLC interface to prevent the vehicle from leaving the current workstation. Step 5: Early Warning and Data Recording Under normal circumstances, the core processing unit sends a "detection passed" signal to the human-machine interaction unit, the GUI interface displays "detection passed", the green indicator light stays on, and the detection record is saved to the specified folder in local storage; When an abnormal situation occurs: the core processing unit immediately sends an abnormal signal to the multi-level early warning unit, executes multi-level early warning, and automatically saves the abnormal data packet to a designated folder in local storage, generates a traceable text report, and the save path is simultaneously displayed in the result area of ​​the GUI interface; the abnormal data packet includes at least the original image acquired, the image after the marked detection box, the VIN number, the visual classification result and confidence level, the queried part number and the extracted ED code, the comparison result, the timestamp, and the description of the abnormal type; Step 6: Manual Handling and Reset. Authorized personnel, after username and password verification, will proceed to the workstation to check and resolve the anomaly based on the lights, WeChat alarm prompts, and the elevator's shutdown status. After resolution, the authorized personnel will click the "Reset" button in the GUI interface. The system will clear the alarm status, the tri-color lights will stop flashing, and the system will return to standby mode. The system will continue to monitor the PLC's VIN number for changes, preparing for the next test. During initialization and operation, the system will continuously or periodically test the connection status of various hardware components (including cameras, PLCs, databases, and tri-color lights). If a test fails, the system will clearly display the faulty module in the GUI status bar and result text box. For automatically recoverable faults, the system will automatically attempt reconnection according to a preset strategy. If automatic retry fails, the system will provide detailed diagnostic information in the GUI interface (such as specific error information for database connection errors). The system allows manual intervention: operators can manually trigger reconnection tests for specific modules in the "Camera Settings" or "System Control" areas. Only when all necessary hardware connections are restored to normal will the system status be updated to "Ready," allowing the testing task to continue.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Significantly improved accuracy and reliability of error prevention Compared to traditional barcode / RFID scanning (which suffers from the fundamental flaw of "correct label, wrong part") and manual visual inspection (with a miss / mistake rate as high as 5-10%), this invention utilizes deep learning-based AI visual recognition technology to directly identify the features of key components such as the front collision beam. Combined with a 99% positioning accuracy and 98% classification accuracy, it fundamentally eliminates misassembly problems caused by incorrect labeling or human negligence. Through real-time cross-verification of the VIN number with standard database information, a dual verification mechanism of "visual recognition + data comparison" is constructed, reducing the error detection rate to below 0.1%, effectively intercepting assembly errors and preventing defective vehicles from entering the next process.

[0017] 2. Real-time performance perfectly matches production line cycle time. The entire system process (image acquisition, AI recognition, database query, and result comparison) can be completed in under 1.5 seconds, with a single-frame inference speed of 30ms and a database query response time of less than 200ms, meeting the high-speed online inspection requirements of automotive assembly lines. Compared to traditional post-production sampling or offline inspection methods, it achieves "real-time detection and real-time early warning," ensuring real-time quality control without affecting production flow.

[0018] 3. High system integration and strong compatibility The system deeply integrates various elements of the industrial site: it supports USB and RTSP network camera access to adapt to different workstation hardware conditions; it communicates stably with mainstream PLCs through standardized protocols to acquire production information in real time; and it directly connects to enterprise production databases such as Oracle to achieve real-time linkage between design data and physical production. It provides multi-dimensional outputs such as local audible and visual alarms, graphical interface display, remote push notifications via WeChat, and automatic data archiving, meeting all needs from on-site operation to remote management.

[0019] 4. Flexible deployment and excellent maintainability Based on a cross-platform (Windows / Mac) architecture and modular design, the system can be quickly deployed on different models of industrial PCs. Through a graphical user interface (GUI), all key parameters (camera settings, PLC address, database connection, ROI area, warning rules, etc.) can be easily configured and adjusted without requiring professional personnel to modify the code. When the production line layout changes or new part types are added, the system can be quickly adapted by updating the model and configuration, significantly reducing the cost and technical threshold of system maintenance and upgrades.

[0020] 5. Outstanding economic and quality benefits Intelligent error prevention can be achieved through software system integration and standard industrial interfaces without requiring large-scale physical modifications to existing production lines. This avoids the enormous costs associated with recalls and reduces production line downtime. The system automatically generates complete and traceable electronic inspection records, providing data support for quality analysis, process improvement, and problem tracing, significantly enhancing enterprise quality management and product reputation. Attached Figure Description

[0021] The invention will now be further described with reference to the accompanying drawings.

[0022] Figure 1 This is a schematic diagram of the system architecture of the present invention.

[0023] Figure 2 This is a flowchart of the present invention. Detailed Implementation Example

[0024] This embodiment provides an AI vision-based intelligent error prevention system for automotive parts. This system is applicable to automotive part assembly fixtures that do not rely on physical markings and can directly extract features from the appearance of the part itself, and these assembly stations meet the following characteristics: Component installation is completed manually or semi-automatically. Components of the same functional category may have multiple visually distinguishable variations (such as with / without accessories, left / right parts, different colors, different hole layouts, etc.). Variations can easily be confused due to their visual similarity, and the consequences of mispackaging can be severe. Example

[0025] The AI ​​vision-based intelligent error prevention system for automotive parts in this embodiment is applied to the assembly station of the front collision beam of an automobile.

[0026] In practical implementation, the following hardware, software, and parameter configurations are adopted: Core Processor Unit: Utilizing the Advantech UNO-2484G industrial computer, equipped with an Intel Core i7 processor, 16GB of RAM, and an NVIDIA GTX 4060 GPU, running Windows 10 / 11 operating system, and containing the system's core processing software. This core processing software is an AI vision model for beam localization and classification optimized based on YOLO11, used for preprocessing, feature recognition, and outputting visual category results and confidence scores from the appearance images acquired by the vision acquisition unit.

[0027] This AI vision model is an object detection and classification model based on architectural improvements. For automotive parts inspection tasks, it first optimizes anchor frames by pre-setting specific aspect ratios based on the actual dimensions of the parts. Then, it enhances the model through multi-scale feature fusion. By introducing an attention mechanism and constructing a multi-scale detection module, it balances local details with global semantic information, thereby improving the ability to identify minute defects and complex backgrounds. Furthermore, the model employs a transfer learning training strategy, fine-tuning only a few parameters to retain the representational advantages of the pre-trained model, significantly improving its generalization ability and recognition stability in complex industrial environments. In automotive parts inspection tasks, the model achieves a mean accuracy (mAP50) of 99.6%, 5.5 percentage points higher than the benchmark model, and a detection speed of 33.5 FPS, enabling real-time accurate localization and high-precision classification of target parts.

[0028] The confidence thresholds are set as follows: the detection confidence threshold is set to 0.5 by default, and the classification confidence threshold is set to 0.7 by default. Both thresholds can be dynamically adjusted in the "System Settings" of the system's graphical user interface, with an adjustment range of 0.1 to 0.9. The system considers a recognition result valid only when the confidence of the detection box is ≥0.5 and the classification confidence of its internal region is ≥0.7.

[0029] Visual acquisition unit: An industrial camera is used, specifically an 8K ultra-high-definition 5-50mm (12MP) UVC driverless industrial camera, which connects to the industrial computer via a USB 3.0 interface. The camera is fixed to the assembly fixture at an angle 2 meters above the ground using a bracket, with the lens tilted at a 45° angle to the horizontal, providing a complete field of view covering the assembly area of ​​the front collision beam.

[0030] Data interaction unit: Includes PLC communication: The production line PLC is a Siemens S7-1500 series. The system uses the snap7 communication library and communicates with the PLC in real time via the Siemens S7 protocol (based on TCP / IP). Fixed communication parameters are: IP address 172.168.10.4, rack number 0, slot number 1. The monitored VIN number data is located in DB block number 23, offset address 486.

[0031] The multi-level early warning unit includes: hard interception: writing emergency stop commands to the PLC controlling the elevator via PROFINET.

[0032] Audible and visual alarm: It connects to the aforementioned industrial control computer via RS232 / RS485 serial port, receives serial port commands, and realizes the flashing control of the three-color warning light (red / yellow / green). If an abnormality occurs, the red light will flash.

[0033] Remote notification: The industrial control computer connects to the enterprise intranet via Ethernet to access the MES database and call the WeChat robot interface; configure the WeChat robot Webhook URL so that alarm messages can be directly pushed to the quality engineer group for timely handling.

[0034] Human-computer interaction: Provides a graphical user interface (GUI) and a WeChat Work robot interface for configuring system parameters, monitoring operating status, performing visual recognition of AI models, and visually displaying the comparison results of recognition results with standard feature codes in the MES database; the WeChat Work robot interface is used to send real-time alarm notifications to designated personnel when parts are misassembled or system malfunctions occur. Intelligent comparison is used to automatically compare the visual category result output by the core processing unit with the feature code in the standard component information obtained from the data interaction unit to determine whether the detection result is normal. If the feature code mapped by the visual category result is consistent with the feature code extracted from the standard component information and the confidence level is higher than a preset threshold, the detection result is determined to be normal. If the feature code is inconsistent, no component to be detected is detected, or the confidence level is lower than the preset threshold, the detection result is determined to be abnormal.

[0035] Data recording is used to automatically save complete process data for each test. The complete process data includes the original image, the processed image, the AI ​​recognition result, the comparison record, the VIN and the timestamp, ultimately realizing the function of generating a traceable electronic quality archive.

[0036] This embodiment provides an intelligent error prevention method for automotive parts based on AI vision, which includes the following steps: Step 1: System initialization and configuration, start the core processing software on the industrial control computer, and load the pre-trained YOLO object detection and classification model; Configure system parameters via the graphical user interface (GUI): Camera parameters: Select USB camera type, set screen rotation angle to 180 degrees, and define ROI (region of interest) that can encompass the entire crash beam; PLC communication parameters: Set the PLC's IP address to 172.168.10.4, rack number to 0, slot number to 1, DB block number to 23, and VIN data offset address to 486; Database connection parameters: Set the database address, port, service name, username, and password; Warning parameters: Set the three-color light control port, the Enterprise WeChat robot Webhook URL, and the list of alarm recipients; The system automatically tests the connection status of each hardware component, including the camera, PLC, database, and serial port LED. If the test fails, a specific fault message is displayed. Automatic retry is supported. If the retry fails, manual repair guidance is provided. After initialization, "System ready" is displayed in the GUI interface. Step 2: VIN Triggering and Image Acquisition. The system continuously monitors the VIN number data area at offset address 486 in the specified DB block (DB23) of the PLC through the PLC interface. When it detects that the currently read VIN number is inconsistent with the last valid VIN number recorded by the system, and the newly read VIN number is not an empty string, a complete error prevention detection process is automatically triggered. The image acquisition unit immediately controls the camera to capture a high-definition image of the current workstation (resolution not less than 1280×720 pixels, clear image, and identifiable feature areas of the parts to be inspected). If the acquisition fails or the image quality does not meet the set threshold, the system will automatically retry up to 3 times. If it fails three times in a row, the system will record the abnormality and send an "image acquisition failure" alarm notification through the enterprise WeChat robot interface, while prompting the operator to intervene on the human-machine interface. Step 3: AI recognition and database query, rotate the acquired image, supporting four fixed angle rotations of 0°, 90°, 180° and 270° to ensure that the parts maintain a positive posture in the image, and then crop the image according to the defined ROI area; The preprocessed image is input into a detection model optimized based on the YOLOv11 architecture. The model outputs detection boxes with confidence scores. An exact bounding box is defined as a rectangular region that accurately matches the target location with at least a 50% Intersection over Union (IoU) and has a detection confidence score higher than a set threshold (default 0.5). The component regions within this bounding box are extracted, and the YOLO-optimized classification model outputs visual classification results and corresponding confidence scores. Simultaneously, using the current VIN number, the core processing unit connects to the database interface of the data interaction unit to perform a VIN number query, obtaining the standard component part number corresponding to the vehicle. From the returned part number, the system extracts its feature code according to preset regular expression rules. The regular expressions used in this system mainly adapt to the "ED + two digits" format, such as the pattern r'-ED(\d{2})-' or r'ED(\d{2})', which can accommodate most standard part numbers. If the part number format changes, the system supports updating the regular expression rules through configuration to adapt to different encoding structures; if feature code extraction fails, the system will record the exception, push an alarm, skip the subsequent comparison process, and mark this detection as "ED code extraction failed". At the same time, it supports manual intervention by the operator in the graphical interface. Step 4: Intelligent comparison, The intelligent comparison unit acquires the visual category results and confidence scores from the AI ​​recognition output, as well as the feature codes extracted from the database query. The system compares the feature codes mapped from the visual results with the feature codes extracted from the part number to see if they are completely consistent. If they are completely consistent and the confidence score is higher than the set threshold, the detection is considered successful. If the comparison is inconsistent, the system will immediately trigger a serious alarm through the WeChat robot interface to notify relevant personnel that "the parts do not match or the assembly is incorrect," and send an emergency stop command to the assembly line control system through the PLC interface to prevent the vehicle from leaving the current workstation. Step 5: Early Warning and Data Recording Under normal circumstances, the core processing unit sends a "detection passed" signal to the human-machine interaction unit, the GUI interface displays "detection passed", the green indicator light stays on, and the detection record is saved to the specified folder in local storage; When an abnormal situation occurs: the core processing unit immediately sends an abnormal signal to the multi-level early warning unit, executes multi-level early warning, and automatically saves the abnormal data packet (the original image collected, the image after the marked detection box, the VIN number, the visual classification result and confidence level, the queried part number and the extracted ED code, the comparison result, the timestamp, and the abnormal type description) to the specified folder of local storage, generates a traceable text report, and the save path is synchronously displayed in the result area of ​​the GUI interface; Step 6: Manual Handling and Reset. Authorized personnel, after username and password verification, will proceed to the workstation to check and handle any abnormalities based on the lights, WeChat alarm prompts, and the elevator's shutdown status. After handling, the authorized personnel will click the "Reset" button in the GUI interface. The system will clear the alarm status, the tri-color lights will stop flashing, and the system will return to standby mode. The system will continue to monitor changes in the PLC's VIN number, preparing for the next test. During initialization and operation, the system will continuously or periodically test the connection status of various hardware components (including cameras, PLCs, databases, and tri-color lights). If a test fails, the system will clearly display the faulty module (e.g., "Camera connection failed," "PLC communication abnormality") in the GUI status bar and result text box. For automatically recoverable faults (e.g., network fluctuations, PLC momentary unresponsiveness), the system will automatically attempt reconnection according to a preset strategy (maximum 3 retries, 2-second intervals). If automatic retry fails, the system will provide detailed diagnostic information in the GUI interface (e.g., specific error information for database connection errors). The system allows manual intervention: operators can manually trigger reconnection tests for specific modules in the "Camera Settings" or "System Control" areas. The system status will only be updated to "Ready" once all necessary hardware connections have been restored, allowing the testing task to continue. Example

[0037] The AI ​​vision-based intelligent error prevention system and method for automotive parts proposed in this application can also be used in the installation process of headlight brackets.

[0038] This embodiment is applied to the installation station of automotive headlight brackets to prevent operators from picking up the wrong left or right bracket type. Compared with the previous embodiment, the camera installation method and height remain unchanged in this embodiment, and the installation angle is 20° for better cooperation. The core processing unit and data interaction unit remain unchanged. Only the dataset is changed and the object detection and classification model is retrained to adapt to the new workstation, fully demonstrating the modularity and versatility of the technical solution. Dataset modification: The original fixed camera at the workstation is used, keeping the installation position, angle, and lighting conditions unchanged, and images containing headlight brackets are captured. The dataset must include different working conditions: normal installation, lighting changes (such as shadows, reflections), partial occlusion, etc. Model retraining: The detection architecture of the previous example is used, and the output is modified to include the headlight bracket type code.

[0039] In addition to the examples described above, the present invention may have other embodiments. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection of this claim.

Claims

1. An intelligent error prevention system for automotive parts based on AI vision and real-time data interaction, characterized in that, include: The vision acquisition unit is configured in the component assembly fixtures of the automotive discrete manufacturing field. It involves mixed-line assembly fixtures for similar parts with multiple state variations that require manual identification and judgment, including at least the automotive front collision beam assembly station, automotive headlight bracket assembly station, automotive engine compartment side beam assembly station, and automotive front and rear floor side beam assembly station. This tool is used to capture real-time images of unmarked automotive parts to be inspected at assembly stations, and supports multiple source access methods, including industrial cameras, network streaming cameras, local image files, and video files. This system can train a YOLO object detection and classification model by adding a dataset of target parts, configure the ROI region and feature code extraction rules for the corresponding parts, and extend it to error-proof detection of other automotive parts such as headlight brackets and front and rear floor side beams; The core processing unit is an industrial control computer equipped with the system software. It has a built-in trained AI vision model, which is used to preprocess the appearance images acquired by the vision acquisition unit, perform feature recognition, and output visual category results and confidence scores. The data interaction unit includes a PLC interface and a database interface. The PLC interface is used to communicate with the PLC of the production line in real time to obtain the unique identifier VIN number of the vehicle to be assembled. The database interface is used to connect to the enterprise's production MES database and query the standard component information corresponding to the vehicle based on the VIN number. It also uses a preset string processing algorithm to extract a combination of fields with clear meaning from the part number of the standard component and use it as the standard feature code of the part. The human-computer interaction unit provides a graphical user interface (GUI) and a WeChat robot interface for configuring system parameters, monitoring operating status, performing visual recognition of AI models, and visually displaying the comparison results of recognition results with standard feature codes in the MES database. The WeChat robot interface is used to send real-time alarm notifications to designated personnel when parts are misassembled or system malfunctions occur. The intelligent comparison unit is used to automatically compare the visual category results output by the core processing unit with the feature codes in the standard component information obtained from the data interaction unit to determine whether the detection result is normal. A multi-level early warning unit is used to execute preset early warning measures when the intelligent comparison unit determines that the detection result is abnormal; The data recording unit is used to automatically save complete process data for each test and generate traceable quality reports.

2. The AI ​​vision-based intelligent error prevention system for automotive parts according to claim 1, characterized in that: The vision acquisition unit is an industrial camera, which is set at a preset position at the assembly station of the part to be inspected to ensure that the field of view covers the area of ​​the part to be inspected, so as to capture the front collision beam image of the assembly station in real time. The industrial camera is connected to the core processing unit via USB or network and supports multiple input sources, including local USB industrial cameras, network RTSP streaming media cameras, and local image or video files.

3. The AI ​​vision-based intelligent error prevention system for automotive parts according to claim 2, characterized in that: The vision unit includes at least one industrial camera, which is installed diagonally above the assembly station of the part to be inspected, at a 45° angle to the horizontal and at a vertical height of 2 meters above the ground, to ensure that the appearance image of the part to be inspected can be captured completely and clearly. The visual acquisition unit supports the collaborative operation of multiple cameras, and the installation height and angle of the cameras can be adjusted for different workstations to achieve synchronous acquisition of multi-angle images of parts.

4. The AI ​​vision-based intelligent error prevention system for automotive parts according to claim 1, characterized in that: The early warning measures of the multi-level early warning unit include: a. Hard interception: An emergency stop command is written to the PLC controlling the elevator via the Siemens S7 communication protocol. After receiving the command, the PLC forces the elevator to stop at its current position, physically blocking the car body from flowing out. b. On-site audible and visual alarm: Control the on-site three-color lights to enter red flashing mode via serial port commands to warn on-site operators; c. Remote instant notification: Call the WeChat Work robot interface to send structured alarm messages, which should include at least the anomaly type, occurrence time, VIN number, visual result, standard part number and on-site image information; d. Local Recording: Highlight error details in the GUI interface and automatically save the abnormal data, i.e. the content contained in step c, to the specified folder to generate a traceable text report.

5. The AI ​​vision-based intelligent error prevention system for automotive parts according to claim 1, characterized in that: The AI ​​vision model is an architecture-based target detection and classification model. For automotive parts inspection tasks, it significantly improves the model's generalization ability and recognition stability in complex industrial environments by introducing anchor frame optimization of specific size, multi-scale feature fusion enhancement, and transfer learning training strategies, thereby achieving real-time accurate positioning and high-precision classification of target parts.

6. The AI ​​vision-based intelligent error prevention system for automotive parts according to claim 1, characterized in that: In the intelligent comparison unit, if the feature code mapped by the visual category result is consistent with the feature code extracted from the standard component information and the confidence level is higher than the preset threshold, the detection result is determined to be normal. If the feature codes are inconsistent, the component to be detected is not detected, or the confidence level is lower than the preset threshold, the detection result is determined to be abnormal.

7. The AI ​​vision-based intelligent error prevention system for automotive parts according to claim 1, characterized in that: The data recording unit is used to automatically save complete process data for each test. The complete process data includes the original image, the processed image, the AI ​​recognition result, the comparison record, the VIN, and the timestamp, to generate a traceable electronic quality archive. PLC communication module: It adopts the snap7 communication library and communicates with the production line PLC in real time via the Siemens S7 protocol based on TCP / IP to obtain the VIN number data stored in the DB block; Database query module: Connects to the enterprise MES production database through Oracle Instant Client and oracledb library, and executes standard query process: Query table UDA_EXTEND_WORK_ORDER_V to get WIP_ENTITY_NAME_W based on VIN number, and then query table ERP_ORDER_BOM_V to get PART_NUM standard part number; Feature extraction module: Extracts feature codes from standard part numbers using preset regular expression rules, including: Main rule: r'-ED(\d{2})-', used to match part numbers in the format XXXX-ED12-XXXX; Alternative rule: r'ED(\d{2})', used to match part numbers in the format ED12 or XXXXED12XXXX; If extraction fails, the system will trigger an enterprise WeChat robot interface to push a system abnormality alarm, notifying relevant personnel that "feature code extraction failed, please check the VIN number or database connection", and prompting the operator to intervene in the human-computer interaction interface.

8. A method for intelligent error prevention of automotive parts based on AI vision and real-time data interaction, applied to the system described in any one of claims 1-7, characterized in that, Includes the following steps: Step 1: System initialization and configuration, start the core processing software on the industrial control computer, and load the pre-trained YOLO object detection and classification model; Configure system parameters via the graphical user interface (GUI): Camera parameters: Select camera type, set screen rotation angle, define ROI (Region of Interest); PLC communication parameters: Set the PLC's IP address, rack number, slot number, DB block number, and VIN data offset address; Database connection parameters: Set the database address, port, service name, username, and password; Warning parameters: Set the three-color light control port, the Enterprise WeChat robot Webhook URL, and the list of alarm recipients; The system automatically tests the connection status of each hardware component, including the camera, PLC, database, and serial port LEDs. If the test fails, a specific fault message is displayed. Automatic retry is supported. If the retry fails, manual repair guidance is provided. After initialization, "System ready" is displayed on the GUI interface. Step 2: VIN Triggering and Image Acquisition. The system continuously monitors the VIN number data area at offset address 486 in the specified DB block (DB23) of the PLC through the PLC interface. When it detects that the currently read VIN number is inconsistent with the last valid VIN number recorded by the system, and the newly read VIN number is not an empty string, a complete error prevention detection process is automatically triggered. The image acquisition unit immediately controls the camera to capture a high-definition image of the current workstation (resolution not less than 1280×720 pixels, clear image, and identifiable feature areas of the component to be inspected). If the acquisition fails or the image quality does not meet the set threshold, the system will automatically retry up to 3 times. If it fails three times in a row, the system will record the abnormality and send an "image acquisition failure" alarm notification through the enterprise WeChat robot interface, while prompting the operator to intervene on the human-machine interface. Step 3: AI recognition and database query, rotate the acquired image, supporting four fixed angle rotations of 0°, 90°, 180° and 270° to ensure that the parts maintain a positive posture in the image, and then crop the image according to the defined ROI area; The preprocessed image is input into a detection model optimized based on the YOLOv11 architecture. The model outputs detection boxes with confidence scores. An exact bounding box is defined as a rectangular region that accurately matches the target location with at least a 50% Intersection over Union (IoU) and has a detection confidence score higher than a set threshold (default 0.5). The component regions within this bounding box are extracted, and the YOLO-optimized classification model outputs visual classification results and corresponding confidence scores. Simultaneously, using the current VIN number, the core processing unit connects to the database interface of the data interaction unit to perform a VIN number query, obtaining the standard component part number corresponding to the vehicle. From the returned part number, the system extracts its feature code according to preset regular expression rules. The regular expressions used in this system mainly adapt to the "ED + two digits" format, such as the pattern r'-ED(\d{2})-' or r'ED(\d{2})', which can accommodate most standard part numbers. If the part number format changes, the system supports updating the regular expression rules through configuration to adapt to different encoding structures. If feature code extraction fails, the system will record the exception, push an alarm, skip the subsequent comparison process, and mark this detection as "ED code extraction failed". At the same time, it supports manual intervention by the operator in the graphical interface. Step 4: Intelligent comparison, The intelligent comparison unit acquires the visual category results and confidence level output by AI recognition, as well as the feature codes extracted from the database query. The system compares whether the feature codes mapped from the visual results are completely consistent with the feature codes extracted from the part number. If they are completely consistent and the confidence level is higher than the set threshold, the detection is deemed to have passed. If the comparison is inconsistent, the system will immediately trigger the enterprise WeChat robot interface to push a serious alarm, notifying relevant personnel that "parts do not match or assembly is incorrect", and send an emergency stop command to the assembly line control system through the PLC interface to prevent the vehicle from leaving the current workstation. Step 5: Early Warning and Data Recording Under normal circumstances, the core processing unit sends a "detection passed" signal to the human-machine interface unit, the GUI interface displays "detection passed", the green indicator light stays on, and the detection record is saved to the specified folder in local storage; When an abnormal situation occurs: the core processing unit immediately sends an abnormal signal to the multi-level early warning unit, executes multi-level early warning, and automatically saves the abnormal data packet to a designated folder in local storage, generates a traceable text report, and the save path is simultaneously displayed in the result area of ​​the GUI interface; the abnormal data packet includes at least the original image acquired, the image after the marked detection box, the VIN number, the visual classification result and confidence level, the queried part number and the extracted ED code, the comparison result, the timestamp, and the description of the abnormal type; Step 6: Manual Handling and Reset. Authorized personnel, after username and password verification, will proceed to the workstation to check and handle any abnormalities based on the lights, WeChat alarm prompts, and the elevator's shutdown status. After handling, the authorized personnel will click the "Reset" button in the GUI interface. The system will clear the alarm status, the tri-color lights will stop flashing and return to standby mode. The system will continue to monitor the PLC's VIN number changes, preparing for the next test. During initialization and operation, the system will continuously or periodically test the connection status of various hardware components (including cameras, PLCs, databases, and tri-color lights). If a test fails, the system will clearly display the faulty module in the GUI status bar and result text box. For automatically recoverable faults, the system will automatically attempt reconnection according to a preset strategy. If automatic retry fails, the system will provide detailed diagnostic information in the GUI interface (such as specific error information for database connection errors). The system allows manual intervention: operators can manually trigger reconnection tests for specific modules in the "Camera Settings" or "System Control" areas. Only when all necessary hardware connections are restored to normal will the system status be updated to "Ready," allowing the testing task to continue.

Images