Intelligent identification and correction system for printing packaging process defects based on AI vision detection

Abstract

The application relates to the field of printed packaging production quality control and discloses a printed packaging process defect intelligent identification and correction system based on AI visual detection, which comprises the following: a visual acquisition module that acquires image data of products on a production line in real time; an AI intelligent identification module that processes the image data based on a pre-trained AI detection model and outputs defect judgment results containing defect types, grades and positions; a data management module that adopts a distributed architecture of edge storage and cloud hierarchical storage and is used for storing and managing image data, defect identification results and production data; a defect tracing module that associates full-link data based on production task numbers or product batch numbers and adopts a hybrid architecture combining a rule engine, a machine learning model and an explainability analysis tool to perform defect root cause analysis; and an intelligent correction module that is used for generating correction instructions according to the defect judgment results and the root cause analysis results. The application improves detection accuracy and real-time performance.

Description

Technical Field

[0001] This invention relates to the field of quality control technology in printing and packaging production, specifically to an intelligent identification and correction system for defects in printing and packaging processes based on AI visual inspection. Background Technology

[0002] The printing and packaging production process covers many key steps, including plate making, printing, surface treatment, paper mounting, die cutting, box gluing, book making, binding, cutting, and finished product inspection. The entire process must strictly follow the incoming material standards, sampling standards, and process requirements of each step (such as the ZZ / WI-22-005 standard), and product quality is ensured through manual sampling or full inspection.

[0003] The existing technology has the following prominent problems:

[0004] (1) Conflict between detection accuracy and real-time performance: Traditional equipment such as spectrophotometers cannot meet the real-time detection requirements of high-speed production of 400m / min. Manual detection is affected by subjective experience and fatigue, resulting in a high rate of missed detection and false detection of defects such as local color difference and small ink droplets. Moreover, the detection speed cannot match the rhythm of the production line.

[0005] (2) Poor adaptability of defect detection technology: The traditional CIEDE2000 algorithm has fixed correction parameters and low sensitivity to local color defects; the template matching difference method is easily affected by artifact interference, leading to misjudgment; the deep learning model relies on a large number of samples, while defect samples are scarce in the small batch production scenario of printing and packaging, and the model's generalization ability is insufficient.

[0006] (3) Defect tracing and correction are lagging: Defect data is scattered and it is difficult to link information on the entire chain such as raw materials, equipment and processes, resulting in low efficiency in root cause location; correction instructions need to be transmitted manually and lack real-time linkage with production equipment, which can easily lead to batches of unqualified products and increase production costs.

[0007] (4) High system compatibility and maintenance costs: The communication protocols of production equipment of different brands and new and old models are very different, and the traditional system has poor adaptability; multi-process detection requires independent models, which have high training, storage and update costs and are prone to "catastrophic forgetting";

[0008] (5) Insufficient data management and interactive experience: The storage volume of full-process data (images, production parameters, test results, etc.) is large, and traditional storage architectures are difficult to balance real-time access and long-term storage costs; the monitoring interface is prone to data delays in multi-line concurrent scenarios, affecting operational decisions.

[0009] These problems result in poor quality stability, low production efficiency, and high labor and material costs in printing and packaging production, necessitating an integrated, high-precision, and highly adaptable intelligent detection and correction solution. Summary of the Invention

[0010] This invention aims to overcome the shortcomings of existing technologies and provide an intelligent identification and correction system for printing and packaging process defects based on AI visual inspection. It achieves integrated detection of color and shape defects, improving detection accuracy and real-time performance, and adapting to high-speed production rhythms of 400m / min. It solves the problem of training with small sample defect data, optimizes multi-process adaptability, and reduces model maintenance costs. It establishes a full-chain defect traceability mechanism, enabling intelligent root cause analysis and improving problem-solving efficiency. It constructs an intelligent correction closed loop that is equipment-compatible and has a fast response time, reducing the risk of batches of non-conforming products.

[0011] To achieve the above objectives, the following technical solution is adopted:

[0012] This invention provides an AI-based intelligent identification and correction system for defects in printing and packaging processes, comprising: a visual acquisition module deployed at multiple workstations on a printing and packaging production line to collect image data of products on the production line in real time; an AI intelligent identification module to receive the image data and process it based on a pre-trained AI detection model, outputting a defect judgment result including defect type, level, and location; wherein the AI ​​detection model includes a color defect detection model and a shape defect detection model; a data management module employing a distributed architecture of edge storage and cloud-based layered storage to store and manage image data, defect identification results, and production data; a defect traceability module to associate full-chain data based on production task number or product batch number, and to perform root cause analysis of defects using a hybrid architecture combining a rule engine, machine learning model, and interpretable analysis tools to obtain root cause analysis results; and an intelligent correction module to generate correction instructions based on the defect judgment results and root cause analysis results, wherein the intelligent correction module includes a device adaptation layer to be compatible with production equipment using different communication protocols and to execute a graded correction strategy.

[0013] Furthermore, the vision acquisition module includes multiple vision acquisition subsystems, which are deployed at different workstations in the printing and packaging production process. Among them, a high-speed linear array camera for acquiring color information is deployed in the printing process, a 3D structured light camera for acquiring three-dimensional morphological information is deployed in the surface treatment process, and a multispectral camera or high-precision dimension measurement sensor for acquiring geometric dimension information is deployed in the die-cutting or lamination process.

[0014] Furthermore, the color defect detection model is built based on the improved CIEDE2000 color difference algorithm. By introducing Gamma correction, physical brightness is converted into psychological brightness, and the correction parameters in the CIEDE2000 formula are dynamically adjusted according to the characteristics of the brightness L, saturation C, and hue H of the image to be detected in the LAB color space.

[0015] Furthermore, the specific construction process of the color defect detection model includes: converting the acquired RGB image to the LAB color space, calculating luminance L, saturation C, and hue H; introducing Gamma correction to convert the physical luminance of the image into the psychological luminance perceived by the human eye, wherein the Gamma correction coefficient is 2.2; and dynamically determining the luminance correction parameters of the CIEDE2000 formula based on the preset value ranges of luminance L, saturation C, and hue H. Saturation correction parameters and tone correction parameters The value.

[0016] Furthermore, the shape defect detection model adopts a structure combining a shared feature extraction network and a process-specific classification head. Its construction process includes: using a deep convolutional generative adversarial network with self-attention mechanism and spectral normalization to augment the shape defect samples and expand the training dataset; using a pre-trained convolutional neural network as the shared backbone network, and training independent process-specific output heads for different processes in printing and packaging production; when it is necessary to adapt to a new process, using elastic weight merging technology to incrementally learn the image classification model, and achieving rapid adaptation by fine-tuning the output head parameters corresponding to the new process.

[0017] Furthermore, the AI ​​intelligent recognition module is used to receive the image data and process the image data based on a pre-trained AI detection model, outputting a defect judgment result including defect type, level, and location. This includes: preprocessing the received image data by denoising and enhancing it, and converting it into RGB and LAB dual-color space data; loading the corresponding process-specific output head according to the process identifier associated with the image data; calculating the color difference ΔE based on the color defect detection model in the LAB color space, and judging color defects by combining dynamically adjusted correction parameters; obtaining candidate regions for shape defects through an improved template matching difference method, and inputting the image of this region into the shape defect detection model to identify the type, level, and location of the shape defect; and comprehensively judging the product's conformity by combining the color and shape defect recognition results.

[0018] Furthermore, the data management module adopts a distributed architecture combining edge storage nodes and cloud-based tiered storage services. The edge storage nodes are deployed on the production line side to store real-time access hot data, which includes at least real-time acquired image data and defect identification results. The cloud-based tiered storage services include object storage services for storing non-real-time access warm data, and low-cost archive storage services for storing long-term archived cold data. The system has a built-in data lifecycle management strategy to automatically migrate the hot data to the object storage service or the archive storage service based on the data's access frequency and storage duration.

[0019] Furthermore, in the defect tracing module: the rule engine has a predefined defect root cause rule library built-in, used to match potential explicit causes according to defect type; the machine learning model is built based on the random forest algorithm, and its input includes defect features, equipment parameters, and raw material attributes, used to predict implicit root causes; the interpretability analysis tool is used to calculate and output the contribution of each input feature to the prediction result of the machine learning model; the defect tracing module is also used to integrate the matching results of the rule engine, the prediction results of the machine learning model, and the contribution output of the interpretability analysis tool to generate a tracing report containing root causes and key influencing factors.

[0020] Furthermore, the intelligent correction module includes: a device adaptation layer for compatibility with production equipment using different communication protocols, comprising at least a protocol parsing module and an instruction standardization module to convert system-generated standardized correction instructions into instruction formats recognizable by the corresponding device; a correction strategy execution unit for executing a graded correction strategy based on the defect level in the defect determination result; and a correction verification unit for controlling the visual acquisition module to re-acquire images of the product at the original defect location after the correction action is executed, and re-detecting the re-acquired images through the AI ​​intelligent recognition module to verify the effectiveness of the correction.

[0021] Furthermore, the system also includes an interactive display module for providing a visual operation interface. The operation interface includes: a real-time monitoring view, which establishes a long connection with the backend server based on WebSocket technology to receive real-time data pushes and adopts a hierarchical view loading mechanism, prioritizing the rendering of key production line indicator views; a defect details view, which displays defect images, annotation information, and associated production batch data; a traceability and correction status view, which centrally displays the traceability reports generated by the defect traceability module and the instruction execution status of the intelligent correction module; and a system configuration view, which manages detection parameters, equipment communication protocols, and user permissions.

[0022] Compared with the prior art, the present invention achieves the following beneficial effects:

[0023] 1. Excellent detection accuracy and real-time performance: Adopting the improved CIEDE2000 algorithm, the color detection accuracy is effectively improved compared with the traditional algorithm, and the sensitivity of local color difference recognition is significantly improved; the quadratic difference strategy effectively removes artifacts, and the shape defect detection accuracy is ≥98.3%; the processing time of a single channel high-definition image is ≤83.63ms. In high-concurrency scenarios, the latency is extended through edge nodes to ensure that the latency is controlled within the acceptable range of industry (≤200ms), which is fully adapted to high-speed production lines.

[0024] 2. Strong adaptability to small samples and multiple processes: The improved DCGAN introduces a self-attention mechanism and spectral normalization to generate high-quality defect samples, solving the problem of insufficient training with small samples; the VGG16 classification network adopts a "shared backbone + process-specific output head" design, combined with EWC incremental learning technology to avoid catastrophic forgetting, and only ≥50 new process samples are needed for rapid adaptation, reducing the model training and maintenance costs.

[0025] 3. Intelligent Defect Tracing: Adopting a hybrid architecture of "rule engine + random forest + SHAP tool", the root cause prediction accuracy is ≥92%, and the feature contribution can be quantified. The traceability report generation time is ≤3 seconds, realizing a rapid closed loop from defect discovery to root cause location, and reducing the recurrence of similar defects.

[0026] 4. Highly efficient and reliable correction response: The equipment adaptation layer is compatible with industrial protocols such as OPC UA and Modbus TCP, supporting adaptation to mainstream brands and older equipment; the hierarchical correction strategy combines fuzzy PID algorithm and incremental adjustment, with automatic correction response time of ≤50ms for minor defects, timely shutdown warning for serious defects, and the effectiveness of correction is verified through precise positioning, reducing material loss.

[0027] 5. Convenient data management and interaction: The "edge + cloud layered storage" architecture ensures hot data query response time of ≤50ms and low-cost archiving of cold data, balancing performance and cost; the interactive display module adopts WebSocket + view hierarchical loading technology, with a data refresh frequency of ≤1 second, supports multi-permission management and multi-channel alarms, and is convenient and intuitive to operate.

[0028] 6. Significant overall benefits: After the system is in operation, labor costs are significantly reduced, order delivery cycles are effectively shortened, and it is adaptable to different types of printing and packaging products and production scenarios. It has strong versatility and important industrial application value.

[0029] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0030] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0031] Figure 1This is a schematic diagram of a module of an AI-based visual inspection-based intelligent identification and correction system for printing and packaging process defects according to an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of the overall system architecture according to an embodiment of the present invention;

[0033] Figure 3 This is a flowchart of the AI ​​intelligent recognition module according to an embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of the interactive display module providing a visual operation interface according to an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0037] Figure 1 This is a schematic diagram of a module of an AI-based visual inspection-based intelligent identification and correction system for printing and packaging process defects according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall system architecture according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, an AI-based visual inspection intelligent identification and correction system 100 for printing and packaging process defects is applicable to the integrated online detection, intelligent traceability, and rapid correction of color and shape defects throughout the entire process of printing and packaging products such as flexible packaging, cartons, and color boxes. It can adapt to the real-time quality control needs of high-speed production lines. The system includes a visual acquisition module 110, an AI intelligent identification module 120, a data management module 130, a defect traceability module 140, an intelligent correction module 150, and an interactive display module (optional). These modules work collaboratively to cover the entire quality control process of printing and packaging production. The specific technical solution is as follows:

[0038] The visual acquisition module 110 is deployed at multiple workstations on the printing and packaging production line to collect image data of products on the production line in real time.

[0039] S1. Constructing the visual acquisition module 110

[0040] The visual acquisition module 110 is deployed at key process nodes in printing and packaging production (including incoming material inspection, printing, surface treatment, mounting, die-cutting, box gluing, book making, binding / gluing, cutting, and finished product inspection). The corresponding visual acquisition equipment is configured according to the inspection requirements of each process, adapting to the high-speed production rhythm of 400m / min.

[0041] S11 Incoming Material Inspection Process

[0042] Equipped with a high-definition industrial camera (resolution ≥ 30 million pixels) and a spectral sensor, it performs image acquisition and material analysis on raw materials such as cardboard boxes and colored paper based on preset incoming material standards and sampling standards. It collects visual and physical information such as the size, color, flatness, and material composition of the raw materials. The image resolution supports multiple specifications such as 3320×1440 and 1996×3324.

[0043] S12, Process Inspection Steps

[0044] For different processes such as printing, surface treatment, mounting, die-cutting, and box assembly, high-speed linear scan cameras (with a frame rate ≥ 60fps), 3D structured light cameras, and multispectral cameras are configured to acquire surface images, 3D morphology, and color information of the products in real time. Preferably, a high-speed linear scan camera is deployed in the printing stage to acquire color information, and a 3D structured light camera is deployed in the surface treatment stage to acquire 3D morphology information. A multispectral camera or high-precision dimensional measurement sensor is deployed in the die-cutting or box assembly stage to acquire geometric dimension information. Specifically:

[0045] (1) Printing process: Focus on collecting image information such as the registration accuracy, color difference, text clarity, and pattern integrity of the printed pattern to provide RGB raw data for color defect detection.

[0046] (2) Surface treatment process: Collect image information on surface gloss, flatness, scratches, bubbles, etc. after polishing / oiling / UV treatment / hot stamping, to meet the needs of shape defect detection.

[0047] (3) Paper mounting / die cutting / box pasting process: collect image information such as paper mounting fit, air bubbles, die cutting size accuracy, and box pasting alignment to ensure the integrity of geometric features and surface condition data.

[0048] (4) Book making / binding / perfect binding process: collect image information such as binding firmness, page margin consistency, and no glue overflow in perfect binding.

[0049] S13, Finished Product Inspection Process

[0050] Equipped with a panoramic camera and a high-precision dimensional measurement sensor (measurement accuracy ≤ ±0.01mm), the final product is subjected to full-size, full-surface image acquisition to verify whether it meets the finished product quality standards.

[0051] The visual acquisition module has a built-in calibration unit that automatically calibrates camera parameters and light source intensity periodically (calibration cycle is configurable, defaulting to once every 2 hours). It uses Gaussian filtering to remove image noise and histogram equalization to enhance image gradient features, ensuring the clarity, consistency, and accuracy of the acquired images. At the same time, it supports flexible adjustment of acquisition parameters (exposure time, resolution, shooting angle, etc.) according to the process requirements of different processes (such as process conditions and quality assurance requirements in the ZZ / WI-22-005 standard), with a parameter adjustment response time ≤10ms.

[0052] S14, Hardware and Deployment Configuration

[0053] The visual acquisition equipment adopts an industrial-grade protection design (IP65 protection level) and supports PoE power supply; it is recommended to configure an edge computing gateway (such as NVIDIA Jetson AGX Xavier) to process the acquired images locally and reduce transmission latency; the network adopts gigabit industrial Ethernet, requiring end-to-end transmission latency ≤20ms to ensure real-time image data upload; for scenarios with multiple concurrent image streams, it supports horizontal expansion deployment of multiple edge nodes, with each edge node responsible for the acquisition and preprocessing of ≤8 high-definition image streams (30 million pixels), avoiding single-point performance bottlenecks.

[0054] AI intelligent recognition module 120 is used to receive the image data, process the image data based on a pre-trained AI detection model, and output a defect judgment result including defect type, level, and location; wherein, the AI ​​detection model includes a color defect detection model and a shape defect detection model.

[0055] like Figure 3 The diagram shows the workflow of the AI ​​intelligent recognition module according to an embodiment of the present invention. In actual operation, after receiving image data from the visual acquisition module 110, the AI ​​intelligent recognition module 120 first calls its preprocessing unit to perform Gaussian filtering for noise reduction and histogram equalization enhancement on the image, and then converts it into RGB and LAB dual color space data. Subsequently, the model scheduling unit loads the corresponding process-specific output head according to the process identifier attached to the image. In the LAB color space, the color recognition unit calculates the color difference ΔE based on the color defect detection model, and combines it with dynamically adjusted correction parameters ( , , The color defect is determined. Simultaneously, the shape recognition unit obtains candidate defect regions using an improved template matching difference method and inputs the image of these regions into the shape defect detection model for recognition. Finally, the comprehensive judgment unit summarizes the recognition results of the two types of defects and determines the product's conformity based on preset process standards.

[0056] S2, Constructing an AI intelligent recognition module 120

[0057] The AI ​​intelligent recognition module 120 is the core of the system, including a model training unit and a real-time recognition unit. It enables integrated detection of color and shape defects, supports multi-process adaptation, and balances detection accuracy with deployment and maintenance efficiency.

[0058] S21, Constructing Model Training Units

[0059] S211, Sample Library Construction

[0060] A printing and packaging defect sample library has been built, covering color defects (local color difference, global color difference, color deviation, etc.) and shape defects (ink splatter, breakage, ink dragging, missing print, etc.). The samples are defect images collected from the actual production process of printing and packaging manufacturers, including 800 shape defect samples and 400 color defect samples (including 200 local color defect samples and 200 global color defect samples), covering different image sizes, material types and process scenarios; it supports automatic sample annotation auxiliary tools to reduce manual annotation costs.

[0061] S212, Color Defect Detection Model Training

[0062] A color defect detection model is constructed based on the improved CIEDE2000 color difference algorithm. The training process is as follows:

[0063] (1) Data preprocessing: Convert the acquired RGB image into LAB color space and calculate the three elements of brightness (L), saturation (C), and hue (H).

[0064] (2) Parameter optimization: Introducing the Gamma value ( The physical brightness is converted into psychological brightness that conforms to human visual perception, using the following formula:

[0065]

[0066] Psychological brightness, which is the image brightness value that conforms to human visual perception after Gamma correction; : Gamma correction coefficient, fixed at 2.2, used to establish the mapping relationship between physical brightness and psychological brightness; 1.36: Brightness normalization coefficient, used to calibrate the psychological brightness value to the standard range; R: Red color component of the image, with a value range of 0-255; A: The green color component of the image, with a value range of 0-255; B: The blue color component of the image, with a value range of 0-255. , , RGB component weighting coefficients correspond to the human eye's sensitivity ratios for red, green, and blue colors.

[0067] (3) Determination of adaptive correction parameters: Based on the distribution range of brightness, saturation, and hue, the correction parameters of the CIEDE2000 formula are dynamically adjusted. , , The specific mapping relationship is as follows:

[0068]

[0069] The brightness correction parameters of the CIEDE2000 color difference formula are dynamically adjusted according to the actual brightness L of the image to optimize the detection sensitivity of local brightness defects. The saturation correction parameters of the CIEDE2000 color difference formula are dynamically adjusted according to the actual saturation C of the image to optimize the detection accuracy of saturation abnormality defects. The color tone correction parameters of the CIEDE2000 color difference formula are dynamically adjusted according to the actual color tone H of the image to optimize the recognition effect of color tone deviation defects. The actual brightness value of the image is calculated based on the LAB color space and is used to adjust the brightness correction parameters. The basis; The actual saturation value of the image is calculated based on the LAB color space and is used to adjust the saturation correction parameters. The basis; The actual tonal value of the image, calculated based on the LAB color space, is used to adjust the tone correction parameters. The basis for this is: 67.84 and 82.92: brightness division thresholds, used to divide brightness values ​​into three intervals, corresponding to different... Values: 10.51, 15.76: Saturation threshold values, used to divide saturation values ​​into three intervals, corresponding to different... Values: 0.46, 0.57: Hue segmentation thresholds, used to divide hue values ​​into three intervals, corresponding to different... Values.

[0070] The thresholds for brightness (L), saturation (C), and hue (H) were determined by clustering analysis (such as the K-means algorithm) on the feature distribution of a large number of standard printed image samples in the LAB color space and combining it with experiments on Just Noticeable Difference (JND). The aim is to divide the color space into typical regions with different sensitivities to defects, thereby achieving refined adaptive matching of correction parameters.

[0071] (4) Threshold setting: based on As a threshold for judging color defects, when the calculated color difference... It was determined that there was a color defect.

[0072] in, : CIEDE2000 color difference calculation result, used to quantify the degree of color difference between the image to be detected and the standard image; 1.0: Color defect judgment threshold, when When this happens, the image is determined to have color defects (including local color difference, global color difference, color cast, etc.).

[0073] S213, Training of Shape Defect Detection Model

[0074] S2131, Data Augmentation

[0075] The shape defect detection model adopts a structure combining a shared feature extraction network and a process-specific classification head. Its construction process includes: using a deep convolutional generative adversarial network (DCGAN) with self-attention mechanism and spectral normalization to augment the shape defect samples and expand the training dataset; using a pre-trained convolutional neural network as the shared backbone network, and training independent process-specific output heads for different processes in printing and packaging production; when it is necessary to adapt to a new process, using elastic weight merging technology to incrementally learn the image classification model, and achieving rapid adaptation by fine-tuning the output head parameters corresponding to the new process.

[0076] Specifically, the improved deep convolutional generative adversarial network (DCGAN) architecture is as follows:

[0077] The generator takes a 1×100 random noise vector as input and contains 5 deconvolution layers. Self-attention (SA) is introduced after the 3rd and 4th deconvolution layers. Each layer is configured with spectral normalization (SN) and batch normalization (BN). The activation function is LeakyReLU, and the activation function of the output layer is Tanh.

[0078] Discriminator: The input is a 64×64×3 image, containing 4 convolutional layers. A self-attention mechanism (SA) is introduced after the 3rd and 4th convolutional layers. Each layer is configured with spectral normalization (SN) and batch normalization (BN). The activation function is LeakyReLU, and the output layer activation function is Sigmoid. Training parameters: The Adam optimizer is used, with a learning rate of 0.003, a batch size of 16, and 300 iterations. The generated discrete defect images are fused with normal images through rotation, translation, mirroring, etc., to construct a diverse training dataset.

[0079] S2132, Classification Model Training

[0080] A lightweight design of "shared backbone network + process-specific output head" is adopted. VGG16 is used as the shared backbone network. After initial training, the backbone network parameters are frozen. Independent output heads (containing 2-3 fully connected layers) are trained only for the defect features of different processes (printing, die-cutting, mounting, etc.), reducing model storage and training overhead. Online incremental learning is supported. When adding a process, elastic weight merging (EWC) technology is used to protect the model knowledge of the original process and avoid catastrophic forgetting. The output head parameters are fine-tuned by ≥50 new process samples to quickly adapt to the requirements of the new process without reconstructing the entire model.

[0081] S214, Model Fusion and Deployment

[0082] The color defect detection model and the shape defect sub-models of each process are encapsulated into an integrated detection engine, which supports automatic switching of the appropriate output head according to the process type; TensorRT is used to accelerate the model and ensure inference efficiency; the model storage adopts version management, supports rollback and incremental updates, and reduces maintenance complexity.

[0083] S22, Construct a real-time recognition unit

[0084] S221, Image Preprocessing

[0085] The image data received from the vision acquisition module is first subjected to Gaussian filtering for noise reduction and histogram equalization for enhancement, and then converted into RGB and LAB dual color space data. This preprocessing process is time-consuming. .

[0086] S222, Process Adaptation and Model Selection

[0087] Based on the process identifiers attached to the images (synchronized by the production MES system), the shared backbone network and the output heads of the corresponding processes are automatically loaded, eliminating the need to load the full model and reducing loading overhead.

[0088] S223, Color Defect Recognition

[0089] Based on the improved CIEDE2000 algorithm after training, calculate the color difference in the LAB space. Combined with adaptive correction parameters , , It can determine whether there are local or global color defects, and its color detection accuracy is improved compared with traditional algorithms. Single frame processing time .

[0090] S224. Shape Defect Identification

[0091] An improved template matching differential method is used for preliminary detection of shape defects. The procedure is as follows:

[0092] 1. Initial Differentiation: The image to be detected is directly differentiated from the standard template image of this process to obtain an initial difference map;

[0093] 2. Mask generation: The edges in the initial difference map are extracted using the Canny edge detection algorithm. Morphological dilation is performed on the detected edge contours. The structuring element is a circle with a radius of 2 pixels, thereby generating a mask image to cover the artifact region.

[0094] 3. Secondary Difference: The generated binarized mask image (white for artifact regions, pixel value 255; black for background, pixel value 0) is bitwise NOT to obtain an inverse mask. Then, the initial difference map and the inverse mask are bitwise ANDed to set the pixel values ​​of the corresponding artifact regions in the initial difference map to zero, resulting in a secondary difference map containing only the real defect candidate regions.

[0095] 4. Classification and Recognition: The processed difference map is input into the trained lightweight classification model to identify the defect type, level (minor, moderate, severe), and location coordinates. The accuracy of shape defect detection is [not specified]. Single frame processing time .

[0096] S225, Acceptance Criteria

[0097] By combining the process standards of each step (such as allowable consumption values ​​and quality assurance requirements) and integrating the results of color and shape defect recognition, the system automatically determines whether the product is qualified and the time required. Mark defect information (type, grade, location) for non-conforming products. (Value, SSIM value) and trigger subsequent processing flow.

[0098] S23, Performance Specifications

[0099] "Full-process detection time ≤ 83.63ms" is defined as the end-to-end processing time (including preprocessing, recognition, and judgment) of a single high-definition image (30 million pixels) under ideal conditions (no network congestion, no concurrent task queuing). When processing multiple concurrent image streams, load balancing is achieved by horizontally expanding the number of edge nodes. The system can linearly expand its processing capacity to ensure that the single-channel image processing latency is still within the acceptable range for industrial production (usually ≤ 200ms) in high-concurrency scenarios.

[0100] S24 Hardware Configuration Requirements

[0101] The AI ​​inference unit is recommended to be configured with a GPU (NVIDIA A100 or equivalent industrial-grade GPU) to support parallel processing of multiple image data streams; the edge computing node adopts a CPU+GPU heterogeneous architecture, and a single node can support the simultaneous processing of ≥8 high-definition image streams (30 million pixels).

[0102] The data management module 130 adopts a distributed architecture of edge storage and cloud-layered storage to store and manage image data, defect identification results and production data;

[0103] S3, Constructing the Data Management Module 130

[0104] Furthermore, the data management module 130 adopts a distributed architecture combining edge storage nodes and cloud-based tiered storage services, balancing real-time access performance with long-term storage costs. The edge storage nodes are deployed on the production line side to store real-time access hot data, which includes at least real-time acquired image data and defect identification results. The cloud-based tiered storage services include object storage services for storing non-real-time access warm data, and low-cost archive storage services for storing long-term archived cold data. The system has a built-in data lifecycle management strategy to automatically migrate hot data to the object storage service or the archive storage service based on the data access frequency and storage duration. Specifically:

[0105] S31, Edge Storage

[0106] Configure ≥1TB SSD industrial-grade storage to store recent (default 30 days) hot data, including real-time acquired raw images, pre-processed images, recognition results, device operating parameters, etc., and support high-speed random read and write (read rate ≥500MB / s) to meet the needs of real-time detection and local query.

[0107] S32, Cloud Tiered Storage

[0108] Warm data storage: Use object storage services (such as AWS S3, Alibaba Cloud OSS) to store warm data for 30 days to 1 year, including non-real-time access image data, production reports, model training logs, etc., and dynamically adjust storage redundancy according to access frequency to reduce storage costs.

[0109] Cold data storage: Use low-cost archive storage (such as AWS Glacier, Alibaba Cloud Archive Storage) to store more than one year of cold data, including historical defect samples, annual production data, etc., and support batch archiving and on-demand recovery (recovery time ≤ 24 hours).

[0110] S33, Data Lifecycle Management

[0111] It has a built-in automated data migration strategy that automatically migrates hot data to warm data storage and cold data storage according to preset rules (such as access frequency and storage duration), and automatically cleans up expired and worthless data (such as defect-free redundant images). Manual configuration of retention rules is also supported.

[0112] The preset rules of the automated data migration strategy can be configured by the user. For example, a configurable rule is: the system automatically marks the data and migrates it to cloud object storage when both conditions are met simultaneously: "data generation time > 30 days" and "number of active queries and accesses in the last 7 calendar days < 5 times".

[0113] S34, Storage Performance Guarantee

[0114] Edge storage and the cloud are transmitted via dedicated lines or 5G industrial internet, supporting asynchronous data synchronization and breakpoint resumption; the cloud uses distributed caching (such as Redis) to accelerate warm data queries, with complex query response times ≤1 second; for real-time queries of hot data, the response time is ≤50ms; for historical analysis queries of warm / cold data, asynchronous query and result notification mechanisms are supported, allowing users to continue other operations after initiating a query, and receive notifications via system messages or emails upon query completion, avoiding long waiting times; data sharding storage is supported, and cross-process and cross-time period query performance is improved by 50% compared to traditional architectures.

[0115] The stored data includes:

[0116] (1) Basic data: incoming material standards, sampling standards, process parameters of each process, quality standards, defect type dictionary, CIEDE2000 calibration parameter mapping table, equipment communication protocol configuration information, etc.

[0117] (2) Data Acquisition: Original images captured by the visual acquisition module, pre-processed images after noise reduction, LAB spatial transformation images, etc.;

[0118] (3) Identification data: Defect identification results (defect type, level, location, timestamp, color difference) Values ​​(SSIM value, FID value), product qualification status;

[0119] (4) Production data: production task sheet information, order information, process flow records, equipment operating parameters, model training logs, correction instruction execution records, etc.;

[0120] (5) System data: equipment calibration records, network status logs, module operation status information, etc.

[0121] The data management module 130 supports standardized data formats (compatible with common industrial formats such as JSON, XML, and CSV), provides standard API interfaces for easy integration with enterprise ERP and MES systems to achieve information sharing, and supports encrypted data transmission and storage (using the AES-256 encryption algorithm) to ensure data security.

[0122] The defect tracing module 140 is used to associate full-chain data based on production task order number or product batch number, and to perform defect root cause analysis using a hybrid architecture that combines rule engine, machine learning model and interpretable analysis tool to obtain root cause analysis results.

[0123] S4. Construct defect tracing module 140:

[0124] The defect tracing module 140, based on the associated data in the data management module, adopts a hybrid architecture of "rule engine + machine learning model + interpretability tool" to achieve full-link defect tracing and intelligent root cause analysis, balancing accuracy and interpretability.

[0125] In the defect tracing module 140: the rule engine has a predefined defect root cause rule library built-in, which is used to match potential explicit causes according to defect type; the machine learning model is built based on the random forest algorithm, and its input includes defect features, equipment parameters and raw material attributes, which is used to predict implicit root causes; the interpretability analysis tool is used to calculate and output the contribution of each input feature to the prediction result of the machine learning model; the defect tracing module 140 is also used to integrate the matching results of the rule engine, the prediction results of the machine learning model and the contribution output of the interpretability analysis tool to generate a tracing report containing root causes and key influencing factors.

[0126] S41, Data Association

[0127] Using production task order number and product batch number as indexes, the system automatically associates the entire data chain of defective products, including raw material batch, production equipment model and operating parameters, operator information, process flow records, inspection time, camera acquisition parameters, and model recognition parameters, with an association time of ≤100ms.

[0128] S42, Root Cause Analysis

[0129] Rule Engine Layer: Built-in rule library of common defects in the printing and packaging industry (such as "printing color difference - ink ratio deviation", "die-cutting burrs - die wear" and more than 200 other rules) to quickly match explicit root causes.

[0130] Machine Learning Layer: A root cause prediction model is constructed using the Random Forest algorithm. Inputting defect features (type, grade, location), equipment parameters, raw material properties, environmental parameters, and other data, it predicts latent root causes (such as "overprinting deviation caused by long-term temperature fluctuations") with a prediction accuracy of ≥92%. The SHAP (SHapley Additive exPlanations) interpretability tool is introduced to calculate the contribution (weight value) of each feature to the root cause prediction results and identify key influencing factors.

[0131] Report generation: Integrates rule matching results, model prediction results, and SHAP contribution analysis to generate a defect traceability report, which clarifies the root cause of the defect, the responsible process, the scope of impact, and key influencing factors (with contribution percentages marked). The report generation time is ≤3 seconds and supports exporting to PDF / Excel format.

[0132] S43, Statistical Analysis

[0133] It supports statistical analysis by multiple dimensions such as defect type (color / shape), production batch, process, time, equipment, and operator, and generates visual charts such as defect distribution heatmaps, trend change curves, and equipment failure frequency rankings, providing data support for production optimization.

[0134] The intelligent correction module 150 is used to generate correction instructions based on the defect judgment results and root cause analysis results. The intelligent correction module includes a device adaptation layer for compatibility with production equipment with different communication protocols and to execute a hierarchical correction strategy.

[0135] S5, Construct Intelligent Correction Module 150

[0136] The intelligent correction module 150 adopts a design of "device adaptation layer + protocol integration + incremental adjustment + closed-loop verification" to achieve rapid defect correction while ensuring compatibility and reliability, forming a closed-loop control of "detection-identification-traceability-correction-verification". The device adaptation layer is designed to be compatible with production equipment using different communication protocols. It includes at least a protocol parsing module and an instruction standardization module to convert the standardized correction instructions generated by the system into an instruction format recognizable by the corresponding equipment. The correction strategy execution unit executes a graded correction strategy based on the defect level in the defect determination result. The correction verification unit, after the correction action is executed, controls the visual acquisition module to re-acquire images of the product at the original defect location and uses the AI ​​intelligent recognition module to re-inspect the re-acquired images to verify the effectiveness of the correction. Details are as follows:

[0137] S51, Equipment Adaptation Layer Design

[0138] S511, Adaptor Layer Architecture

[0139] The built-in device adaptation layer supports protocol conversion, command mapping, and permission adaptation, and is compatible with production equipment of different brands and models. The adaptation layer includes a device driver library (covering mainstream brands such as Heidelberg, Komori, and Roland), a protocol parsing module (compatible with industrial general protocols such as OPC UA, Modbus TCP, and Profinet), and a command standardization module (mapping system-wide commands into device-specific command formats).

[0140] S512, Driver Library Maintenance and Updates

[0141] It provides a driver development framework and graphical configuration tools to lower the barrier to user customization and adaptation. Users can add new device drivers visually through the configuration tools or develop custom drivers based on the development framework. It also supports online updates of the driver library and automatically pushes official driver update packages through the system background. Users can choose to install manually or automatically.

[0142] S513, Permissions and Security Adaptation

[0143] Supports device security authentication mechanisms (such as username and password, digital certificate), automatically adapting to the authentication requirements of different devices; provides manual correction guidance for scenarios where older devices do not support remote parameter adjustment, generating standardized operating steps for operators to execute; clarifies that the "automatic correction" function requires the device to have remote control interface permissions, and automatically downgrades to the "early warning + manual correction" mode if such permissions are not available.

[0144] S52, Device Interoperability and Communication Mechanism

[0145] S521, Interface Configuration

[0146] It supports configuring device communication parameters (baud rate, IP address, port number, timeout, etc.) through a visual interface, and automatically detects device connection status and supported protocol types.

[0147] S522, communication guarantee

[0148] A heartbeat mechanism (heartbeat interval ≤ 1 second) is used to monitor the device connection status. After disconnection, it will automatically reconnect (reconnection count ≤ 5 times, each time interval 2 seconds). If reconnection fails, an alarm will be triggered. The correction command adopts a three-stage transmission mechanism of "send-confirm-feedback". The command is encrypted (using RSA encryption algorithm) and accompanied by a verification code to avoid command loss or tampering.

[0149] S523, Security Lockout Adaptor

[0150] For scenarios where devices have security locking mechanisms, it supports configuring an "automatic unlock application - manual confirmation - unlock execution" process, or directly triggering a manual unlock prompt to ensure compliant operation.

[0151] S53, Graded Correction Strategy

[0152] S531, Minor Defect Correction

[0153] Parameter calculation: Based on defect tracing results and historical equipment adjustment data, an incremental adjustment strategy is adopted to calculate correction parameters (e.g., ink ratio adjustment increment ≤ ±5% / time, die-cutting die position adjustment increment ≤ ±0.02mm / time) to avoid over-correction; considering the nonlinear and hysteretic mapping relationship between ink ratio, die position and defects, fuzzy PID method is used for parameter adjustment, or PID parameters are optimized offline based on historical correction effects to improve control accuracy.

[0154] Command issuance: Automatically generate parameter adjustment commands, convert them into a device-recognizable format through the device adaptation layer, and then issue them. The command issuance response time is ≤50ms.

[0155] Anti-shake mechanism: Set adjustment interval (≥3 seconds), the same parameter will not be adjusted repeatedly during the anti-shake period to prevent frequent fluctuations of the equipment.

[0156] S532, General Defect Correction

[0157] Warning notification: Warning information is sent through multiple channels, including audible and visual alarms, pop-up windows on workshop management terminals, and push notifications to operators' mobile apps. The warning information includes the type of defect, its location, and suggested correction solutions.

[0158] Human interaction: Operators can confirm and correct operations or modify parameters on the terminal, and the system records the operation process and parameter change logs.

[0159] S533, Critical Defect Correction

[0160] Stop command: Automatically triggers the production line stop command and locks equipment operation permissions to prevent accidental start-up.

[0161] Troubleshooting: Send detailed defect reports and troubleshooting guidelines to operators. The shutdown status can only be unlocked after the operators have investigated the problem (such as replacing ink or die-cutting molds, adjusting equipment precision) and confirmed it in the system.

[0162] S54, Verification of Correction Effect

[0163] After correction, the system calculates the estimated arrival time of the defective product at the re-collection station based on the product serial number and defect pixel coordinates relative to the product's starting position recorded in the defect judgment result, combined with the real-time encoder signal of the production line conveyor belt. The visual acquisition module triggers acquisition at a predetermined time and uses a QR code reader or feature matching algorithm installed at the re-collection station to finally confirm the product's arrival, thus ensuring that the re-collected image corresponds to the original defect location. The AI ​​intelligent recognition module then re-inspects according to the same standards.

[0164] If color defect If the shape defect detection result is qualified, the correction is deemed effective, and the system resumes normal production process.

[0165] If the defect is not eliminated ( If the shape defect still exists, the correction process will be repeated: for minor defects, the parameter adjustment increment will be halved and recalculated and issued; for general or serious defects, the system will trigger the corresponding warning or shutdown process again according to the current level of the defect.

[0166] Optionally, in some embodiments, the AI-based visual inspection-based intelligent identification and correction system 100 for printing and packaging process defects of the present invention further includes an interactive display module for providing a visual operation interface. The operation interface includes: a real-time monitoring view, which establishes a long connection with the backend server based on WebSocket technology to receive real-time data pushes and adopts a view hierarchical loading mechanism, prioritizing the rendering of key production line indicator views; a defect details view, used to display defect images, annotation information, and associated production batch data; a traceability and correction status view, used to centrally display the traceability report generated by the defect traceability module and the instruction execution status of the intelligent correction module; and a system configuration view, used to manage detection parameters, equipment communication protocols, and user permissions.

[0167] S6. Construct the interactive display module:

[0168] like Figure 4As shown, the interactive display module provides a visual operation interface (supporting access from both web and local clients) and features hierarchical permission management (administrators, operators, quality inspectors, etc.). Different permissions correspond to different operation permissions, balancing real-time requirements with adaptability to large-scale monitoring scenarios.

[0169] (1) Real-time monitoring interface: Real-time data push is achieved using "WebSocket + data aggregation" technology, and the data refresh frequency is set to default. The time limit is 2 seconds, which can be dynamically adjusted according to the monitoring scale (e.g., 2 seconds / time when monitoring multiple production lines); it supports hierarchical loading of views, prioritizing the rendering of key indicators (pass rate, total number of defects, alarm information), and loading detailed data (defect details of single process, equipment parameters) as needed, reducing the front-end rendering pressure.

[0170] (2) Defect details interface: Displays defect images and annotation information (defect location selection, color difference value, defect level identifier, inspection time, associated batch), supports zooming in and out of defect images, and supports quick retrieval by defect type and time range.

[0171] (3) Traceability and Correction Interface: Displays defect traceability reports (including SHAP contribution analysis), correction instruction execution status (pending execution / executed / execution failed), equipment parameter adjustment records, and correction effect verification results. It supports exporting and printing traceability reports.

[0172] (4) System configuration interface: Supports manual setting of detection parameters (such as color difference threshold) The system includes features such as defect level assessment criteria, quality standard modification, device communication protocol configuration, user permission management, and system log viewing. Configuration operations support version management and one-click rollback.

[0173] (5) Alarm function: Supports sound alarm (alarm volume is adjustable), pop-up alarm, and mobile APP push alarm. Alarm information can be manually confirmed or automatically confirmed (after defect correction). Alarm logs are permanently stored and traceable. Statistical analysis can be performed by alarm type, time, and device.

[0174] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0175] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A smart identification and correction system for printing and packaging process defects based on AI visual inspection, characterized in that, include: The vision acquisition module is deployed at multiple workstations on the printing and packaging production line to collect image data of products on the production line in real time. The AI ​​intelligent recognition module is used to receive the image data and process the image data based on a pre-trained AI detection model, and output a defect judgment result including defect type, level and location; wherein, the AI ​​detection model includes a color defect detection model and a shape defect detection model; The data management module adopts a distributed architecture of edge storage and cloud-layered storage to store and manage image data, defect identification results, and production data. The defect tracing module is used to associate full-chain data based on production task order number or product batch number, and to perform defect root cause analysis using a hybrid architecture that combines rule engine, machine learning model and interpretable analysis tools to obtain root cause analysis results. The intelligent correction module is used to generate correction instructions based on the defect determination results and root cause analysis results. The intelligent correction module includes a device adaptation layer for compatibility with production equipment with different communication protocols and to execute a hierarchical correction strategy.

2. The intelligent identification and correction system for printing and packaging process defects based on AI visual inspection according to claim 1, characterized in that, The vision acquisition module includes multiple vision acquisition subsystems, which are deployed at different workstations in the printing and packaging production process. Among them, a high-speed linear array camera for acquiring color information is deployed in the printing process, a 3D structured light camera for acquiring three-dimensional morphological information is deployed in the surface treatment process, and a multispectral camera or high-precision dimension measurement sensor for acquiring geometric dimension information is deployed in the die-cutting or lamination process.

3. The intelligent identification and correction system for printing and packaging process defects based on AI visual inspection according to claim 1, characterized in that, The color defect detection model is built on the improved CIEDE2000 color difference algorithm. It converts physical brightness into psychological brightness by introducing Gamma correction, and dynamically adjusts the correction parameters in the CIEDE2000 formula according to the characteristics of the brightness L, saturation C and hue H of the image to be detected in the LAB color space.

4. The intelligent identification and correction system for printing and packaging process defects based on AI visual inspection according to claim 3, characterized in that, The specific construction process of the color defect detection model includes: The acquired RGB image is converted to the LAB color space, and the brightness (L), saturation (C), and hue (H) are calculated. Gamma correction is introduced to convert the physical brightness of the image into the psychological brightness perceived by the human eye, where the Gamma correction coefficient is 2.2; Based on the preset value ranges of luminance L, saturation C, and hue H, the luminance correction parameters of the CIEDE2000 formula are dynamically determined. Saturation correction parameters and tone correction parameters The value.

5. The intelligent identification and correction system for printing and packaging process defects based on AI visual inspection according to claim 1, characterized in that, The shape defect detection model adopts a structure that combines a shared feature extraction network with a process-specific classification head. Its construction process includes: A deep convolutional generative adversarial network with self-attention mechanism and spectral normalization is used to augment the shape defect samples to expand the training dataset; A pre-trained convolutional neural network is used as the shared backbone network, and independent process-specific output heads are trained for different processes in printing and packaging production. When it is necessary to adapt to a new process, the image classification model is incrementally learned using the elastic weight merging technique, and the output header parameters corresponding to the new process are fine-tuned to achieve rapid adaptation.

6. The intelligent identification and correction system for printing and packaging process defects based on AI visual inspection according to claim 5, characterized in that, The AI ​​intelligent recognition module is used to receive the image data, process the image data based on a pre-trained AI detection model, and output a defect judgment result including defect type, level, and location, including: Perform preprocessing of denoising and enhancement on the received image data, and convert it into RGB and LAB dual color space data; Load the corresponding process-specific output header according to the process identifier associated with the image data; In the LAB color space, calculate the color difference ΔE based on the color defect detection model, and judge the color defect in combination with dynamically adjusted correction parameters; Obtain the candidate area of the shape defect through the improved template matching difference method, and input the image of this area into the shape defect detection model to identify the type, grade and position of the shape defect; Comprehensively judge the qualification of the product based on the recognition results of color and shape defects.

7. The intelligent identification and correction system for printing and packaging process defects based on AI visual inspection according to claim 1, characterized in that, The data management module adopts a distributed architecture that combines edge storage nodes and cloud hierarchical storage services; Among them, the edge storage nodes are deployed on the production line side and are used to store hot data for real-time access. The hot data at least includes image data collected in real time and defect recognition results; The cloud hierarchical storage service includes an object storage service for storing warm data for non-real-time access, and a low-cost archival storage service for storing cold data for long-term archiving; The system has a built-in data life cycle management strategy, which is used to automatically migrate the hot data to the object storage service or the archival storage service according to the access frequency and storage duration of the data.

8. The intelligent identification and correction system for printing and packaging process defects based on AI visual inspection according to claim 1, characterized in that, In the defect traceability module: The rule engine has a pre-defined defect root cause rule library built in, which is used to match potential obvious causes according to the defect type; The machine learning model is constructed based on the random forest algorithm, and its inputs include defect features, equipment parameters and raw material attributes, which are used to predict the hidden root cause; The interpretability analysis tool is used to calculate and output the contribution degree of each input feature to the prediction result of the machine learning model; The defect traceability module is also used to integrate the matching result of the rule engine, the prediction result of the machine learning model and the contribution degree output by the interpretability analysis tool, and generate a traceability report including the root cause and key influencing factors.

9. The intelligent identification and correction system for printing and packaging process defects based on AI visual inspection according to claim 1, characterized in that, The intelligent correction module includes: The device adaptation layer is used to be compatible with production devices with different communication protocols, and at least includes a protocol parsing module and an instruction standardization module to convert the standardized correction instructions generated by the system into an instruction format recognizable by the corresponding device; The correction strategy execution unit is used to execute the hierarchical correction strategy according to the grade of the defect in the defect determination result; The correction verification unit is used to control the vision acquisition module to re-acquire the image of the product at the original defect position after the correction action is executed, and re-detect the re-acquired image through the AI intelligent recognition module to verify whether the correction is effective.

10. The intelligent identification and correction system for printing and packaging process defects based on AI visual inspection according to claim 9, characterized in that, The system also includes an interactive display module, which is used to provide a visual operation interface. The operation interface includes: A real-time monitoring view, which establishes a long connection with the backend server based on WebSocket technology to receive real-time data push, and adopts a view hierarchical loading mechanism to preferentially render the key index views of the production line; A defect details view, which is used to display defect images, annotation information and associated production batch data; The traceability and correction status view is used to centrally display the traceability report generated by the defect traceability module and the instruction execution status of the intelligent correction module; The system configuration view is used to manage testing parameters, device communication protocols, and user permissions.

Images