Digital point inspection and online reinspection system and method for visual inspection equipment on industrial production line

Abstract

The invention discloses a digital point inspection and online reinspection system and method for visual inspection equipment on an industrial production line. The system comprises a digital point inspection module, an intelligent reinspection module and a data management and tracing module, the digital point inspection module carries out high-frequency state verification and algorithm validity verification on visual inspection equipment at each station of the packaging machine through cooperation of an automatic execution mechanism and an intelligent algorithm; the intelligent reinspection module performs secondary analysis and judgment on the suspected defective products identified by the detection equipment; and the data management and tracing module is used for recording point inspection and reinspection related data, realizing data visualization and retrieval, and performing centralized management and optimization updating on the detection model. The system realizes real-time data interaction with the PLC control cabinet and the visual inspection equipment at each station through an industrial Ethernet wired or wireless transmission network.

Description

Technical Field

[0001] This invention belongs to the field of industrial production line quality inspection technology, specifically relating to a digital inspection and online re-inspection system and method for visual inspection equipment on industrial production lines. It is particularly suitable for digital inspection and online defect re-inspection, and realizes digital inspection, online re-inspection, model lifecycle management and inspection data traceability through an integrated management platform. Background Technology

[0002] In automated industrial production lines, vision inspection equipment is a crucial link in ensuring product quality; its inspection accuracy and stability directly affect the product pass rate. Existing technologies suffer from the following problems: 1. Low efficiency of manual inspection: Traditional inspection relies on manual labor, which has shortcomings such as insufficient coverage, inconsistent standards, low frequency (generally once per shift), need to stop the machine for operation, and untimely detection of equipment abnormalities. In addition, the types of defective samples manufactured manually are limited, and there is also the risk of them being mixed into the normal production process.

[0003] 2. Limited accuracy of defect identification: Due to the constraints of workstation space and real-time detection requirements, some detection equipment cannot be configured with high-performance computing resources and can only run lightweight algorithms, resulting in frequent missed detection of minor defects or false detection of similar features, which affects the quality control effect.

[0004] 3. Lack of integrated management: The detection data, models and equipment status lack centralized management, resulting in delayed model updates, scattered and difficult-to-trace data, and low utilization of computing resources.

[0005] Therefore, a system is needed that can perform high-frequency automatic inspections, precise and intelligent re-inspections, and has centralized model management and full-process traceability functions to improve the quality control level and production efficiency of industrial production lines. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention proposes a digital inspection and online re-inspection system and method for visual inspection equipment on industrial production lines, which solves problems such as low efficiency of manual inspection, insufficient accuracy of re-inspection, and lack of integrated management.

[0007] To achieve the above objective 1, the present invention adopts the following technical solution: A digital inspection and online re-inspection system for visual inspection equipment on an industrial production line includes a digital inspection module, an intelligent re-inspection module, and a data management and traceability module; The digital inspection module, through the collaboration of automated actuators and intelligent algorithms, performs high-frequency status verification and algorithm effectiveness validation of the visual inspection equipment at each station of the packaging machine. The intelligent re-inspection module performs secondary analysis and judgment on suspected defective products identified by the testing equipment; The data management and traceability module is used to record data related to spot checks and re-inspections, realize data visualization and retrieval, and centrally manage and optimize the detection model.

[0008] Furthermore, the digital inspection module includes a sensing and actuator mechanism, an image acquisition unit, a PLC control unit, a performance verification unit, and a defect database management unit; The sensing and actuator integrates the original vision equipment, adds position sensors, light sensors, temperature and humidity sensors, and is equipped with cylinders, solenoid valves, and motors to collect equipment operating parameters and simulate defect scenarios. The image acquisition unit acquires images of the detection area using an industrial camera and uploads them to the central system when the inspection is triggered. The PLC control unit schedules inspection tasks based on a preset strategy, controls the actuators and data acquisition devices, judges the feedback signals from the detection devices, and dynamically adjusts the inspection frequency. The performance verification unit inputs the verification image to the detection equipment through the industrial network channel for algorithm performance testing, and evaluates the algorithm performance status of the equipment based on multi-period statistical indicators, confidence drift trend or image quality parameters. The defect database management unit collects and integrates historical defect images, optimizes and constructs a standardized defect image database using data augmentation technology, and supports automatic invocation of verification algorithm effectiveness during non-production periods.

[0009] Furthermore, the intelligent re-inspection module includes an image transmission unit, a re-inspection calculation unit, and a decision and rule application unit; When the defect confidence level is within 20% of the set threshold, the image transmission unit sends the image and metadata of the suspected defective product to the central system in real time. The re-inspection computing unit uses GPU resources to run a high-precision deep learning model, dynamically selects the model according to the inspection station to perform secondary inspection, and sends the re-inspection results to the production line execution equipment. The judgment and rule application unit combines the defect processing rule base and historical statistical model to optimize the judgment threshold, supports manual review and error sample labeling, and uses the labeled data for model optimization.

[0010] Furthermore, the data management and traceability module includes a data recording unit, a visualization and retrieval unit, and a model management unit; The data recording unit records the time, parameters, images, detection results, and model version information of the inspection and re-inspection process throughout the entire process. The visualization and retrieval unit retrieves data in multiple dimensions through a web interface and presents it in the form of charts, curves, and image comparisons. The model management unit centrally manages the model versions and training datasets of each device, supports batch updates and performance monitoring, and the central system optimizes the model and pushes updates based on the feedback results.

[0011] To achieve the above objective 2, the present invention adopts the following technical solution: A digital inspection and online re-inspection method for a visual inspection device on an industrial production line, employing a digital inspection and online re-inspection system for a visual inspection device on an industrial production line as described above, including a digital inspection step, an intelligent re-inspection step, and a data management and traceability step; The digital inspection step, through the collaboration of automated actuators and intelligent algorithms, performs high-frequency status verification and algorithm effectiveness validation of the visual inspection equipment at each station of the packaging machine. The intelligent re-inspection step performs secondary analysis and judgment on suspected defective products identified by the detection equipment. The data management and traceability steps are used to collect, store, and visualize inspection and re-inspection related data, and to realize centralized management, optimization and updating of the detection model and two-way traceability of product quality.

[0012] Furthermore, the digital inspection step includes: 1) The central system is powered on and started. The digital inspection module is initialized and loads the preset inspection strategy containing inspection trigger conditions and time period rules from the XML configuration file. 2) Based on the current time, production counter, and historical data from the equipment health status database, prioritize workstation inspections and generate inspection task instructions; 3) Determine the system time period. During production hours, the PLC controls the actuator to move the defect test block to the detection area, triggering the industrial camera to acquire and upload images. During non-production hours, randomly select labeled defect images from the defect library. 4) The central system sends the image to the visual inspection device under test, the device runs the local algorithm and returns the detection result; 5) The performance verification unit evaluates the imaging status and algorithm stability of the detection equipment based on the quality parameters of the verification image, the algorithm confidence fluctuation, and multi-period statistical characteristics, and automatically generates a structured equipment health report. 6) Use Bayesian algorithm to analyze historical inspection data and dynamically update the inspection strategy to adaptively optimize the inspection interval; 7) The images acquired in step 3) and enhanced in step 4) are added to the defect database. The defect database is expanded and optimized by data augmentation techniques such as contrast-limited adaptive histogram equalization, wavelet denoising and morphological operations.

[0013] Furthermore, the intelligent re-inspection step includes: 1) The visual inspection equipment outputs the defect confidence score. If it falls within the 20% of the suspected defect range below the set threshold, the reporting process is triggered. 2) The inspection equipment encapsulates the industrial images of the product and metadata containing defect categories and confidence levels into a standard format data packet, and transmits it asynchronously to the central re-inspection system via industrial Ethernet; 3) After parsing the data packet, the central system obtains basic information, assigns the re-inspection task to the idle GPU computing nodes in the GPU computing cluster, and dynamically selects a deep learning model for secondary detection based on the basic information to generate the re-inspection judgment result and the corresponding new confidence level. 4) Receive the re-inspection judgment result and corresponding new confidence level generated in step 3), combine it with the preset defect handling rule base and the statistical model trained with historical data, output the final quality judgment result, and send it to the PLC control cabinet to execute the product rejection operation; Quality engineers use a visual interface to obtain the re-inspection decision criteria, including defect heatmaps, re-inspection confidence levels, rule matching clauses, and feature importance analysis of statistical models, for manual review. They manually annotate and correct cases where the system misjudged the decision. The correct results of the manual annotations are transmitted to the model training database in real time for periodic fine-tuning and optimization of the re-inspection deep learning model and statistical model, forming a continuous improvement loop.

[0014] Furthermore, the data management and traceability steps include: 1) Collect multi-source data for digital inspection, intelligent re-inspection, and model management through a distributed data acquisition gateway based on the MQTT protocol; 2) After standardizing the collected data, a hybrid storage architecture is adopted, in which structured data is stored in a time-series database and unstructured data is stored in object storage and indexed and associated with the time-series database. 3) Build a front-end and back-end separated web visualization and retrieval platform to support multi-dimensional retrieval, multi-format data visualization, and role-based fine-grained permission management; 4) The platform centrally manages model versions and training datasets, supports batch updates, performance monitoring and unified optimization push, and automatically triggers alarms when key indicators are abnormal. 5) Install barcode scanning devices at the product packaging station to assign a unique identifier to the product or batch, and bind the identifier to the data of the entire process of production, inspection, and re-inspection to achieve forward tracing of the production and testing information corresponding to the product, and reverse tracing of the product batch and flow affected by specific defects or failures. Furthermore, the data augmentation technology process is as follows: first, the local details of the image are enhanced and noise is suppressed by limiting the contrast adaptive histogram equalization algorithm; then, the image is denoised and reconstructed using wavelet denoising method; finally, morphological operations are performed based on the defect morphological characteristics to improve the saliency of the defect area.

[0015] The beneficial technical effects of this invention are as follows: The digital inspection and intelligent re-inspection system and method of the visual inspection equipment adopted in this invention can realize high-frequency automatic inspection, accurate intelligent re-inspection, centralized model management and full-process data traceability in industrial production lines. It effectively solves the problems of low efficiency of manual inspection, poor re-inspection accuracy and low resource utilization, improves the accuracy of defect detection and production efficiency, and is applicable to various types of packaging equipment. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the digital inspection and online re-inspection system architecture of a visual inspection device on an industrial production line according to the present invention; Figure 2 This is a flowchart of the digital inspection steps in this invention; Figure 3 This is a flowchart of the intelligent re-inspection steps in this invention; Figure 4 This is a schematic diagram of data management and traceability in this invention; Detailed Implementation

[0017] The specific embodiments of the present invention will be further described below with reference to specific examples: A digital inspection and online re-inspection system for visual inspection equipment on an industrial production line, such as Figure 1 As shown, the system includes a digital inspection module, an intelligent re-inspection module, and a data management and traceability module. The system achieves real-time data interaction with the PLC control cabinet and the visual inspection equipment at each workstation through an industrial Ethernet wired or wireless transmission network. The digital inspection module, through the collaboration of automated actuators and intelligent algorithms, performs high-frequency status verification and algorithm validity checks on the visual inspection equipment at each station of the packaging machine. The intelligent re-inspection module performs secondary analysis and judgment on suspected defective products identified by the testing equipment; The data management and traceability module is used to record data related to spot checks and re-inspections, realize data visualization and retrieval, and centrally manage and optimize the detection model.

[0018] Specifically, the digital inspection module includes a sensing and actuator mechanism, an image acquisition unit, a PLC control unit, an algorithm verification unit, and a defect library management unit; The sensing and actuator integrates the original vision equipment and adds position sensors, light sensors, and temperature and humidity sensors to collect equipment operating parameters; it is also equipped with cylinders, solenoid valves, and motors to move the detected object to a preset position during inspection, simulating defect scenarios such as positional offset and missing images. When the inspection is triggered, the image acquisition unit acquires images of the inspection area using an industrial camera and uploads them to the central system; The PLC control unit schedules inspection tasks based on preset frequencies, trigger conditions, and other strategies, controls the actuators and data acquisition devices, judges whether the feedback signals from the detection devices are correct, and can dynamically adjust the inspection frequency according to the historical false alarm rate and failure rate of the equipment. The performance verification unit inputs the verification image to the detection equipment through the industrial network channel for algorithm performance testing, and evaluates the performance status of the equipment algorithm based on multi-period statistical indicators, confidence drift trends or image quality parameters. The defect library management unit collects and integrates historical defect images from various visual inspection devices, expands and optimizes them using data augmentation technology, constructs a standardized defect image library, supports automatic retrieval of defect images during non-production periods, and directly distributes them to inspection devices, completing regular verification and performance evaluation of the algorithm's effectiveness.

[0019] Specifically, the intelligent re-inspection module includes an image transmission unit, a re-inspection calculation unit, and a decision and rule application unit; When the defect confidence level is within 20% of the set threshold, the detection equipment will send the image and metadata (defect type, confidence level, workstation number, time, etc.) of the suspected defective product to the central system in real time. The re-inspection unit uses GPU resources to run a high-precision deep learning model, dynamically selects the model for secondary inspection based on the inspection station, and sends the re-inspection results (whether to reject) to the production line execution equipment in a timely manner to reject defective products online. The judgment and rule application unit combines the defect handling rule base and historical statistical model to optimize the judgment threshold, supports manual review and error sample labeling, and uses the labeled data for model optimization.

[0020] Specifically, the data management and traceability module includes a data recording unit, a visualization and retrieval unit, and a model management unit; the data recording unit records the time, parameters, images, test results, model version, and other information of the inspection and re-inspection process to ensure traceability; The visualization and retrieval unit retrieves data through a web interface by multiple dimensions such as time, device, batch, and defect type, and presents it in the form of charts, curves, and image comparisons. The model management unit centrally manages the model versions and training datasets of each device, supports batch updates and performance monitoring, and the central system optimizes the model and pushes updates based on feedback results to ensure continuous improvement in overall detection performance.

[0021] A digital inspection and online re-inspection method for a visual inspection device on an industrial production line, employing a digital inspection and online re-inspection system for a visual inspection device on an industrial production line as described above, including a digital inspection step, an intelligent re-inspection step, and a data management and traceability step; The digital inspection process, through the collaboration of automated actuators and intelligent algorithms, performs high-frequency status verification and algorithm effectiveness validation of the visual inspection equipment at each station of the packaging machine. like Figure 2 As shown, the digital inspection steps include: 1) The central system is powered on and started. The digital inspection module is initialized and loads the preset inspection strategy containing inspection trigger conditions and time period rules from the XML configuration file. These strategies are stored in XML form, including inspection trigger conditions and production / non-production time period rules. 2) Start the inspection task. Based on the strategy loaded in step 1), sort the inspection priorities of the workstations according to the current time, production counter and historical data of the equipment health status database, and generate inspection task instructions. 3) Determine the system time period. During production hours, the PLC controls the actuator to move the defect test block to the detection area, triggering the industrial camera to acquire and upload images. During non-production hours, randomly select labeled defect images from the defect library. Specifically: To accurately determine whether the system is currently in a production or non-production period, if it is determined to be in a production period, to prevent disruption to normal production, the inspection process must be rapid and non-interfering with product flow. The central system sends instructions to the PLC. The PLC controls the servo slide or cylinder to move a specific defect test block (such as a calibration block with a preset 0.5mm position offset) on the inspection fixture quickly and accurately to the inspection area directly below the vision camera, simulating a real defect scenario. Simultaneously, the PLC triggers the industrial camera to capture an image of the current test block. The industrial camera uploads the acquired raw defect image to the central system's image processing unit via the network. If it is determined to be in a non-production period, a "soft" inspection mode is used. The central system directly sends instructions to the defect database management unit. The defect database management unit randomly retrieves a data-enhanced, annotated defect image from its database based on the type of equipment under test and the verification purpose.

[0022] 4) The central system sends the image to the visual inspection device under test, the device runs its local algorithm and returns the detection result; specifically: The central system transmits the image file obtained in step 3) directly to the vision inspection device under test via industrial Ethernet, simulating real-time camera triggering. Upon receiving the image, the vision device treats it as a real inspection frame, invokes its locally deployed detection algorithm for real-time analysis, and generates a detection result (such as "NG - Position Offset"). This result is then returned to the central system.

[0023] 5) The performance verification unit evaluates the imaging status and algorithm stability of the detection equipment based on the quality parameters of the verification image, algorithm confidence fluctuations, and multi-period statistical characteristics, and automatically generates a structured equipment health report; specifically: The performance verification unit performs confidence analysis, category stability assessment, and deviation detection from historical statistical models on the results returned by the testing equipment to determine the working status of the equipment algorithm in the current environment. When the test results show abnormal confidence, abnormal category fluctuation, or significant deviation from the previous stable range, the system automatically marks them as "performance abnormal"; when the test results are within the stable range and the image quality, illumination consistency, and other indicators meet the requirements, they are marked as "performance normal". All verification records (time, workstation, image quality indicators, confidence distribution, deviation, stability score, etc.) are automatically summarized to generate a structured equipment performance health report (PDF and Excel formats). 6) Use Bayesian algorithm to analyze historical inspection data and dynamically update the inspection strategy to adaptively optimize the inspection interval; specifically: The system analyzes long-term historical data in inspection reports and dynamically updates strategies using a Bayesian algorithm. It evaluates the reward function (including equipment operational stability and inspection cost) at different inspection frequencies and continuously iterates to find the inspection strategy with the highest long-term expected return. For equipment with unstable recent inspection results, the system automatically shortens the inspection interval; while for equipment with stable operating conditions, the inspection interval is extended accordingly, achieving adaptive and optimized allocation of inspection resources.

[0024] 7) The acquired and processed images are added to the defect database. The defect database is expanded and optimized using data augmentation techniques such as contrast-limited adaptive histogram equalization, wavelet denoising, and morphological operations. Specifically: During each inspection, high-quality images acquired in step 3) and enhanced in step 4) are added to the defect database after administrator confirmation. Simultaneously, historical defect images from various visual inspection devices are collected and integrated, expanded and optimized using data augmentation techniques. The processing flow includes: First, a contrast-limited adaptive histogram equalization algorithm is used to divide the image into multiple sub-blocks. The histogram of each sub-block is calculated and cropped to control the local contrast enhancement. Then, histogram equalization is performed on each sub-block, and bilinear interpolation is used to fuse the sub-block boundaries, effectively eliminating artifacts and enhancing local image details while suppressing noise amplification. Second, wavelet denoising is applied, performing multi-scale wavelet transform on the image to decompose it into sub-band components of different frequencies. For high-frequency detail coefficients representing noise, soft or hard thresholding functions are used to shrink or reduce the noise coefficients to zero. Finally, the image is reconstructed through inverse wavelet transform to obtain a clear, denoised result. Finally, based on the morphological characteristics of the defects (such as cracks, holes, protrusions, etc.), appropriate morphological operations (such as dilation, erosion, opening operations, closing operations) are selected to enhance or suppress defects of specific shapes, so as to improve the salience of the defect area or reduce background interference.

[0025] This allows for the construction of a standardized defect image library, which is then expanded through self-learning to promote the continuous growth of the number and diversity of samples in the defect library over time, providing richer test cases for spot checks during non-production periods.

[0026] The intelligent re-inspection step performs secondary analysis and judgment on suspected defective products identified by the testing equipment; like Figure 3 As shown, the intelligent re-inspection steps include: 1) The visual inspection equipment outputs a defect confidence score. If the score falls within the 20% suspected defect range below a set threshold, a reporting process is triggered; specifically: When any vision inspection device on the production line performs real-time inspection of a product, its built-in algorithm model outputs a defect confidence score. The system then compares this confidence score with a preset judgment threshold. When the confidence score falls into the "suspected defect range" (for example, within 20% below the set threshold), the reporting process is automatically triggered, thereby ensuring that all difficult-to-determine boundary cases are captured.

[0027] 2) The inspection equipment encapsulates the product's industrial image and metadata, including defect categories and confidence levels, into a standard format data packet, which is then asynchronously transmitted to the central re-inspection system via industrial Ethernet; specifically: Upon triggering, the vision inspection equipment performs the following operations: It encapsulates the captured industrial image of the product and its corresponding inspection metadata into a data packet. The metadata includes the defect category prediction value, confidence level, equipment station number, manufacturer brand, and timestamp. The data packet is then formatted into a predefined standard format, encapsulated using an "image file (e.g., PNG, PEG) + JSON description file" structure. The JSON file clearly stores all metadata in a key-value pair structure for easy subsequent parsing. Finally, the encapsulated data packet is transmitted asynchronously to the central re-inspection system in real time via high-speed industrial Ethernet, ensuring that the continuous operating cycle of the production line's vision inspection equipment is not affected.

[0028] 3) After parsing the data packet, the central system obtains basic information, assigns the re-inspection task to idle GPU computing nodes in the GPU computing cluster, and dynamically selects a deep learning model for secondary detection based on the basic information to generate a re-inspection judgment result; specifically: After receiving the data packet from the detection device, the central system executes the following re-inspection process: The central management system parses the JSON file to obtain basic information, and then schedules and allocates tasks, assigning the re-inspection task to an idle GPU computing node in the GPU computing cluster. Based on the obtained basic information, the management system automatically assigns a suitable deep learning detection model to the current task. After the re-inspection is completed, the unit generates a re-inspection judgment result ("qualified" or "defective" and its specific type) and the corresponding new confidence level.

[0029] 4) Receive the re-inspection judgment result and corresponding new confidence level generated in step 3), combine it with the statistical model trained on the preset defect handling rule base and historical data, output the final quality judgment result, and send it to the PLC control cabinet to execute the product rejection operation. Specifically: The system receives the results from the intelligent re-inspection analysis (3) and processes them as follows: It accesses a pre-defined defect handling rule base. This rule base contains business rules defined based on product standards, process requirements, and customer needs (e.g., "Zero tolerance for Class A defects, immediate removal," "Class B defects with an area less than 0.1% are acceptable"). Simultaneously, the system introduces a statistical model trained on historical data. This model dynamically optimizes the release threshold based on historical risk data (such as the actual complaint rate of a certain type of minor defect). The final decision is the result of the combined effect of the rule base logic and the statistical model recommendation, and this result is immediately sent to the PLC control cabinet for online removal of products confirmed as defective.

[0030] All review and decision criteria, including defect heatmaps, review confidence levels, matching rules and clauses, and feature importance analysis of statistical models, are presented to quality engineers through a visual interface. Engineers manually review these criteria and correct any erroneous decisions (whether false negatives or false negatives). The correct results of these manual corrections are then uploaded to the system's model training database in real time. This newly added sample data is used to periodically fine-tune and optimize the review deep learning and statistical models, forming a continuous improvement loop of "data collection -> model training -> deployment and application -> error annotation -> retraining," making the system increasingly accurate and reliable over time.

[0031] The data management and traceability steps are used to collect, store, and visualize inspection and re-inspection related data, and to realize centralized management, optimization and updating of the testing model and two-way traceability of product quality.

[0032] like Figure 4 As shown, the data management and traceability steps include: 1) Collect multi-source data for digital inspection, intelligent re-inspection, and model management through a distributed data acquisition gateway using the MQTT protocol; specifically: A distributed data acquisition gateway is deployed on the production line side to collect and aggregate multi-source data in real time via a predefined interface protocol (MQTT). The collection scope includes: sensor readings (light source brightness, temperature and humidity) from the digital inspection module, inspection metadata (time, trigger conditions), image data (original image, enhanced image), and algorithm verification results; edge detection results (defect type, coordinates, confidence level) from the intelligent re-inspection module, central re-inspection images, and manual review labels; and model version information, training metadata, and performance metrics (accuracy, false negative rate) from the model management module.

[0033] 2) After standardizing the collected data, a hybrid storage architecture is adopted, storing structured data in a time-series database and unstructured data in object storage, with an index established between the object storage and the time-series database; specifically: After the collected data is standardized (unified timestamp, device code, defect classification dictionary), a hybrid storage architecture is adopted: all time-series structured data (such as time, device ID, and results) is stored in a time-series database, supporting high-speed query and aggregation analysis. All unstructured data (such as raw images and model files) is stored in object storage and associated with records in the time-series database through unique indexes to ensure data integrity and access efficiency.

[0034] 3) Build a front-end and back-end separated web visualization and retrieval platform that supports multi-dimensional retrieval, multi-format data visualization, and role-based fine-grained permission management; specifically: A front-end and back-end separation architecture is adopted to build a web visualization and retrieval platform, supporting multi-dimensional retrieval and data visualization. It supports queries by time (accurate to the second), equipment / workstation, product batch, and defect type, and supports filtering by multiple conditions such as defect severity, model version, and processing results (e.g., release / rejection). Search results are presented intuitively in various formats, such as defect type distribution pie charts, daily inspection trend line charts, and equipment status heatmaps. Edge inspection images and central re-inspection images can be displayed side-by-side. A refined permission system is built based on a role-based access control model: for example, operators can only view the inspection data and search results of their assigned workstation. Quality inspectors have permissions to query re-inspection data, perform manual review, and correct results. Administrators have full data access, model management, rule configuration, and user permission management capabilities.

[0035] 4) The platform centrally manages model versions and training datasets, supporting batch updates, performance monitoring, and unified optimization pushes, automatically triggering alarms when key metrics are abnormal. Specifically: The platform built using step 3) centrally manages information such as model versions and training datasets for each device, supports batch updates and performance monitoring, and performs unified optimization based on feedback results. The center optimizes models uniformly based on feedback results and pushes updates to ensure continuous improvement in overall detection performance. A model performance monitoring dashboard is established to display the real-time running status of models on each device. When key indicators (such as accuracy and false negative rate) continuously exceed thresholds, the system automatically triggers an alarm.

[0036] 5) Install barcode scanning devices at the product packaging station to assign a unique identifier (such as a QR code) to each product or batch, and link this identifier to data from all stages of production, inspection, and re-inspection. This two-way linking enables forward traceability (tracing the root cause): Product identifiers allow for quick lookup of corresponding production time, equipment used, model version, inspection images, and results, accurately pinpointing the root cause of the problem. Reverse traceability (tracing the scope of impact): By using specific defect types or equipment failure records, all affected product batches and their distribution can be traced backwards, providing decision support for quality recalls.

[0037] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A digital inspection and online re-inspection system for visual inspection equipment on an industrial production line, characterized in that, It includes a digital inspection module, an intelligent re-inspection module, and a data management and traceability module; The digital inspection module, through the collaboration of automated actuators and intelligent algorithms, performs high-frequency status verification and algorithm effectiveness validation of the visual inspection equipment at each station of the packaging machine. The intelligent re-inspection module performs secondary analysis and judgment on suspected defective products identified by the testing equipment; The data management and traceability module is used to record data related to spot checks and re-inspections, realize data visualization and retrieval, and centrally manage and optimize the detection model.

2. The digital inspection and online re-inspection system for visual inspection equipment on an industrial production line according to claim 1, characterized in that, The digital inspection module includes a sensing and actuator, an image acquisition unit, a PLC control unit, a performance verification unit, and a defect database management unit. The sensing and actuator integrates the original vision equipment, adds position sensors, light sensors, temperature and humidity sensors, and is equipped with cylinders, solenoid valves, and motors to collect equipment operating parameters and simulate defect scenarios. The image acquisition unit acquires images of the detection area using an industrial camera and uploads them to the central system when the inspection is triggered. The PLC control unit schedules inspection tasks based on a preset strategy, controls the actuators and data acquisition devices, judges the feedback signals from the detection devices, and dynamically adjusts the inspection frequency. The performance verification unit inputs the verification image to the detection equipment through the industrial network channel for algorithm performance testing, and evaluates the algorithm performance status of the equipment based on multi-period statistical indicators, confidence drift trend or image quality parameters. The defect database management unit collects and integrates historical defect images, optimizes and constructs a standardized defect image database using data augmentation technology, and supports automatic invocation of verification algorithm effectiveness during non-production periods.

3. The digital inspection and online re-inspection system for visual inspection equipment on an industrial production line according to claim 1, characterized in that, The intelligent re-inspection module includes an image transmission unit, a re-inspection calculation unit, and a decision and rule application unit; When the defect confidence level is within 20% of the set threshold, the image transmission unit sends the image and metadata of the suspected defective product to the central system in real time. The re-inspection computing unit uses GPU resources to run a high-precision deep learning model, dynamically selects the model according to the inspection station to perform secondary inspection, and sends the re-inspection results to the production line execution equipment. The judgment and rule application unit combines the defect processing rule base and historical statistical model to optimize the judgment threshold, supports manual review and error sample labeling, and uses the labeled data for model optimization.

4. The digital inspection and online re-inspection system for visual inspection equipment on an industrial production line according to claim 1, characterized in that, The data management and traceability module includes a data recording unit, a visualization and retrieval unit, and a model management unit; The data recording unit records the time, parameters, images, detection results, and model version information of the inspection and re-inspection process throughout the entire process. The visualization and retrieval unit retrieves data in multiple dimensions through a web interface and presents it in the form of charts, curves, and image comparisons. The model management unit centrally manages the model versions and training datasets of each device, supports batch updates and performance monitoring, and the central system optimizes the model and pushes updates based on the feedback results.

5. A digital inspection and online re-inspection method for visual inspection equipment on an industrial production line, characterized in that, A digital inspection and online re-inspection system using a visual inspection device on an industrial production line as described in any one of claims 1-4, comprising a digital inspection step, an intelligent re-inspection step, and a data management and traceability step; The digital inspection step, through the collaboration of automated actuators and intelligent algorithms, performs high-frequency status verification and algorithm effectiveness validation of the visual inspection equipment at each station of the packaging machine. The intelligent re-inspection step performs secondary analysis and judgment on suspected defective products identified by the detection equipment. The data management and traceability steps are used to collect, store, and visualize inspection and re-inspection related data, and to realize centralized management, optimization and updating of the detection model and two-way traceability of product quality.

6. The digital inspection and online re-inspection method for a visual inspection device on an industrial production line according to claim 5, characterized in that, The digital inspection steps include: 1) The central system is powered on and started. The digital inspection module is initialized and loads the preset inspection strategy containing inspection trigger conditions and time period rules from the XML configuration file. 2) Based on the current time, production counter, and historical data from the equipment health status database, prioritize workstation inspections and generate inspection task instructions; 3) Determine the system time period. During production hours, the PLC controls the actuator to move the defect test block to the detection area, triggering the industrial camera to acquire and upload images. During non-production hours, randomly select labeled defect images from the defect library. 4) The central system sends the image to the visual inspection device under test, the device runs the local algorithm and returns the detection result; 5) The performance verification unit evaluates the imaging status and algorithm stability of the detection equipment based on the quality parameters of the verification image, the algorithm confidence fluctuation, and multi-period statistical characteristics, and automatically generates a structured equipment health report. 6) Use Bayesian algorithm to analyze historical inspection data and dynamically update the inspection strategy to adaptively optimize the inspection interval; 7) The images acquired in step 3) and enhanced in step 4) are added to the defect database. The defect database is expanded and optimized by data augmentation techniques such as contrast-limited adaptive histogram equalization, wavelet denoising and morphological operations.

7. The digital inspection and online re-inspection method for a visual inspection device on an industrial production line according to claim 5, characterized in that, The intelligent re-inspection steps include: 1) The visual inspection equipment outputs the defect confidence score. If it falls within the 20% of the suspected defect range below the set threshold, the reporting process is triggered. 2) The inspection equipment encapsulates the industrial images of the product and metadata containing defect categories and confidence levels into a standard format data packet, and transmits it asynchronously to the central re-inspection system via industrial Ethernet; 3) After parsing the data packet, the central system obtains basic information, assigns the re-inspection task to the idle GPU computing nodes in the GPU computing cluster, and dynamically selects a deep learning model for secondary detection based on the basic information to generate the re-inspection judgment result and the corresponding new confidence level. 4) Receive the re-inspection judgment result and corresponding new confidence level generated in step 3), combine it with the preset defect handling rule base and the statistical model trained with historical data, output the final quality judgment result, and send it to the PLC control cabinet to execute the product rejection operation; Quality engineers use a visual interface to obtain the re-inspection decision criteria, including defect heatmaps, re-inspection confidence levels, rule matching clauses, and feature importance analysis of statistical models, for manual review. They manually annotate and correct cases where the system misjudged the decision. The correct results of the manual annotations are transmitted to the model training database in real time for periodic fine-tuning and optimization of the re-inspection deep learning model and statistical model, forming a continuous improvement loop.

8. The digital inspection and online re-inspection method for a visual inspection device on an industrial production line according to claim 5, characterized in that, The data management and traceability steps include: 1) Collect multi-source data for digital inspection, intelligent re-inspection, and model management through a distributed data acquisition gateway based on the MQTT protocol; 2) After standardizing the collected data, a hybrid storage architecture is adopted, in which structured data is stored in a time-series database and unstructured data is stored in object storage and indexed and associated with the time-series database. 3) Build a front-end and back-end separated web visualization and retrieval platform to support multi-dimensional retrieval, multi-format data visualization, and role-based fine-grained permission management; 4) The platform centrally manages model versions and training datasets, supports batch updates, performance monitoring and unified optimization push, and automatically triggers alarms when key indicators are abnormal. 5) Install barcode scanning devices at the product packaging station to assign a unique identifier to the product or batch, and bind the identifier to the data of the entire process of production, inspection, and re-inspection to achieve forward tracing of the production and testing information corresponding to the product, and reverse tracing of the product batch and flow affected by specific defects or failures.

9. The digital inspection and online re-inspection method for a visual inspection device on an industrial production line according to claim 6, characterized in that, The data augmentation technology process is as follows: first, the local details of the image are enhanced and noise is suppressed by limiting the contrast adaptive histogram equalization algorithm; then, the image is de-denoised and reconstructed by wavelet denoising method; finally, morphological operations are performed according to the defect morphological characteristics to improve the saliency of the defect area.

Images