A deep learning-based screen printing defect detection system

Abstract

The present application belongs to the technical field of screen printing quality control, and specifically relates to a screen printing defect detection system based on deep learning, which comprises: multi-spectral image acquisition, multi-band imaging and mechanical linkage control logic configuration, and multi-type printing product image synchronous acquisition according to a preset viewing angle; a cloud-edge collaborative module, which comprises: an edge computing unit deploying a lightweight model to realize single-frame image inference delay and meet the high-speed real-time detection requirements of a production line; a cloud platform end constructing a defect sample intelligent management system, integrating a federal learning mechanism to realize privacy-protected sharing and model joint training of multi-production-line defect data; and real-time signal linkage with a screen printing device, an automatic sorting device and an MES system, so that when serious defects are detected, the printing device is automatically triggered to stop and give a warning, the sorting device is synchronously controlled to sort defective products to a specified area, and quality data in the MES system is updated in real time.

Description

Technical Field

[0001] This invention belongs to the field of screen printing quality control technology, specifically a screen printing defect detection system based on deep learning. Background Technology

[0002] Screen printing, as a high-precision transfer process for precision graphics, is widely used in high-end electronic manufacturing fields such as PCB circuit boards, photovoltaic cell silver electrodes, microelectronic packaging devices, and ceramic precision components.

[0003] Printing defects such as broken lines, missing ink, pinholes, misregistration, ink splatter, color difference, and pattern distortion directly determine the electrical conductivity reliability, appearance quality, and service life of components. They are the core and critical link in the quality control of the entire printing process. With the rapid development of mass production in the industry and the continuous improvement of production line cycle time, the demand for online, high-precision, and fully automated detection of screen printing defects throughout the entire process is becoming increasingly stringent.

[0004] However, existing detection systems suffer from interference from the coupling of various defect types, insufficient accuracy in defect classification, and a lack of an integrated closed-loop system architecture encompassing image acquisition and preprocessing, deep learning-based intelligent judgment, defect classification and alarm, data archiving and tracing, and coordinated control of printing equipment. Furthermore, edge-embedded lightweight deployment is difficult, and the iteration cost for non-standard small-batch product models is high. These systems struggle to simultaneously meet the demands of high-precision recognition, high-speed real-time detection, and industrial-scale adaptation, severely hindering the intelligent manufacturing and quality improvement of screen printing. Therefore, this invention provides a deep learning-based screen printing defect detection system. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: a deep learning-based screen printing defect detection system, comprising: Multispectral image acquisition, configured with multi-band imaging and mechanical linkage control logic, synchronously acquires images of multiple types of printed products from a preset perspective; The adaptive preprocessing module is configured with dynamic compensation and texture suppression control logic to perform normalization fusion and pixel-level defect pre-annotation on multispectral images. The dual-path lightweight deep learning detection model module includes: a) The backbone feature extraction network is based on the lightweight ShuffleNetV3 and adopts a convolutional structure that combines grouped convolution and channel shuffling; b) Introduce a dynamic attention feature pyramid network and a bidirectional path aggregation network to achieve multi-scale feature adaptive fusion and enhance the feature extraction capability of target defects; c) Set up a defect localization branch and a defect classification branch in parallel. The localization branch uses a simplified CenterNet algorithm to achieve high-precision regression of the defect center point and size. The classification branch uses a multi-label classifier combined with a lightweight Focal Loss loss function to achieve defect classification during screen printing and output the defect level classification results. The cloud-edge collaboration module includes: a) The edge computing unit deploys a lightweight model to reduce single-frame image inference latency and meet the high-speed real-time detection requirements of the production line; b) Build an intelligent management system for defect samples on the cloud platform, integrate a federated learning mechanism, and realize privacy-preserving sharing of defect data from multiple production lines and joint training of models; c) Real-time signal linkage with screen printing equipment, automatic sorting equipment, and MES system. When a serious defect is detected, the printing equipment will be automatically triggered to stop the machine. The sorting equipment will be controlled to sort the defective products to the designated area and the quality data in the MES system will be updated in real time. The adaptive model iterative optimization module includes: Configure meta-learning adaptation, performance monitoring, and knowledge transfer control logic to dynamically update detection model parameters.

[0007] Preferably, the multispectral image acquisition includes three imaging bands: visible light, near-infrared, and ultraviolet. It is equipped with a ring-shaped shadowless light source array and a six-axis adjustable mechanical lens bracket. Through mechanical linkage control logic, it can adapt to screen printing products of different sizes and substrates to achieve multi-view, blind-angle-free synchronous image acquisition.

[0008] Preferably, the adaptive preprocessing module incorporates a dynamic illumination compensation algorithm and a multi-scale substrate texture suppression algorithm to offset industrial ambient light fluctuations, printing substrate reflections, and substrate texture interference in real time. It performs registration and fusion, noise reduction and enhancement, and grayscale normalization on multi-band images, and automatically completes pixel-level pre-annotation of suspected defect areas, providing standardized image input for deep learning models.

[0009] Preferably, the lightweight ShuffleNetV3 backbone feature extraction network achieves cross-group information interaction by splitting channel features through grouped convolution and channel shuffling, and reduces redundant convolutional layers and fully connected layers, compressing the number of model parameters to less than 1 / 10 of that of traditional convolutional networks.

[0010] Preferably, the dynamic attention feature pyramid network has a built-in pixel-level attention weight allocation mechanism that adaptively adjusts feature weights according to the size of the defect target, enhances the features of micron-level fine pinholes and short line defects, and suppresses interference from invalid background features.

[0011] Preferably, the simplified CenterNet algorithm removes the redundant upsampling module of the traditional CenterNet, retains the core center point prediction and size offset prediction branches, optimizes the coordinate regression loss function, controls the positioning error of minor defects within a suitable range, and adjusts the positioning accuracy of minor printing defects.

[0012] Preferably, a three-level defect classification threshold of minor, moderate, and severe is preset, and the defect level result is automatically output in combination with defect area, shape, and location parameters.

[0013] Preferably, the edge computing unit has a built-in local data cache and fault tolerance mechanism, and is configured with breakpoint continuation detection, abnormal frame filtering, and dynamic computing power scheduling logic. The edge computing unit locally builds a lightweight temporary defect database, caches on-duty detection images, defect data and model inference logs, and continues to independently complete the detection work under network outages and temporary equipment fluctuations. After the network is restored, the offline data is automatically synchronized to the cloud platform.

[0014] Preferably, the cloud platform federated learning mechanism adopts a training mode in which local data is not uploaded and model parameters are encrypted and interacted with. Under the premise of protecting the privacy of production data of each production line, it aggregates the defect feature parameters of multiple sites to complete the joint iteration of the model.

[0015] Preferably, the adaptive model iterative optimization module is based on a meta-learning algorithm to achieve rapid adaptation to small samples. For non-standard small-batch printing products, model fine-tuning can be completed with a small number of labeled samples. At the same time, knowledge distillation technology is used to transfer the feature learning experience of large models to lightweight models, and the false negative rate and false positive rate are monitored in real time to automatically trigger incremental model updates.

[0016] The beneficial effects of this invention are as follows: 1. The deep learning-based screen printing defect detection system described in this invention employs a multispectral, multi-view image acquisition method, combined with adaptive illumination compensation and texture suppression preprocessing techniques. This effectively counteracts interference from industrial lighting fluctuations, substrate reflection, and substrate texture. Through a dynamic attention feature fusion mechanism and a lightweight Focal Loss function, it focuses on enhancing the expression of micron-level minute defect features, balancing the long-tail distribution of defect samples, and significantly reducing the false negative and false positive rates of difficult-to-detect defects such as fine pinholes, short lines, and light color differences. This significantly improves the accuracy and stability of printing defect detection under complex industrial conditions.

[0017] 2. The deep learning-based screen printing defect detection system of the present invention significantly reduces the number of model parameters and computational load by using a lightweight backbone network based on ShuffleNetV3 and a simplified CenterNet localization structure. Combined with edge-end optimized deployment strategies, it effectively reduces inference latency while ensuring accurate multi-scale defect localization and hierarchical classification capabilities. It can adapt to high-speed production line cycles, has strong hardware adaptability, low deployment cost, and supports long-term uninterrupted stable operation. It solves the shortcomings of traditional deep learning models, such as high computational cost, difficulty in implementation, and inability to adapt to high-speed industrial production lines.

[0018] 3. The deep learning-based screen printing defect detection system described in this invention, through a cloud-edge collaborative federated learning mechanism, achieves joint iterative optimization of multiple production line models while ensuring the privacy and security of the factory's original production data. Combined with meta-learning small-sample rapid adaptation and knowledge distillation transfer technology, it can quickly adapt to non-standard, small-batch, and new variety printing product detection scenarios, effectively solving the problems of long iteration and adaptation cycles and weak generalization ability of traditional models. At the same time, relying on the closed-loop management of production line equipment and MES system, it realizes real-time defect early warning, automatic sorting, and quality data traceability, significantly improving the intelligent quality control level of the screen printing process. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a schematic diagram of the overall working process of this invention; Figure 2 This is a flowchart illustrating the internal computation of the dual-path lightweight deep learning detection model in this invention. Figure 3 This is a schematic diagram of the cloud-edge collaboration and adaptive model iterative optimization process in this invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0022] like Figure 1-3 As shown in the embodiment of the present invention, a screen printing defect detection system based on deep learning includes: Multispectral image acquisition, configured with multi-band imaging and mechanical linkage control logic, synchronously acquires images of multiple types of printed products from a preset perspective; The adaptive preprocessing module is configured with dynamic compensation and texture suppression control logic to perform normalization fusion and pixel-level defect pre-annotation on multispectral images. The dual-path lightweight deep learning detection model module includes: a) The backbone feature extraction network is based on the lightweight ShuffleNetV3 and adopts a convolutional structure that combines grouped convolution and channel shuffling; b) Introduce a dynamic attention feature pyramid network and a bidirectional path aggregation network to achieve multi-scale feature adaptive fusion and enhance the feature extraction capability of target defects; c) Set up a defect localization branch and a defect classification branch in parallel. The localization branch uses a simplified CenterNet algorithm to achieve high-precision regression of the defect center point and size. The classification branch uses a multi-label classifier combined with a lightweight Focal Loss loss function to achieve defect classification during screen printing and output the defect level classification results. The cloud-edge collaboration module includes: a) The edge computing unit deploys a lightweight model to reduce single-frame image inference latency and meet the high-speed real-time detection requirements of the production line; b) Build an intelligent management system for defect samples on the cloud platform, integrate a federated learning mechanism, and realize privacy-preserving sharing of defect data from multiple production lines and joint training of models; c) Real-time signal linkage with screen printing equipment, automatic sorting equipment, and MES system. When a serious defect is detected, the printing equipment will be automatically triggered to stop the machine. The sorting equipment will be controlled to sort the defective products to the designated area and the quality data in the MES system will be updated in real time. The adaptive model iterative optimization module includes: Configure meta-learning adaptation, performance monitoring, and knowledge transfer control logic to dynamically update detection model parameters; The system consists of five core functional modules: multispectral image acquisition, adaptive preprocessing, dual-path lightweight deep learning detection model, cloud-edge collaboration, and adaptive model iteration optimization. These modules work in a hierarchical, closed-loop manner, and the overall system workflow is as follows: First, multispectral image acquisition equipment, combined with a mechanical linkage structure, synchronously acquires multi-band images of various screen printing products from multiple angles. Then, an adaptive preprocessing module completes image illumination correction, texture noise reduction, multispectral fusion, and defect pre-annotation, outputting standardized image data. Subsequently, the data is fed into a dual-path lightweight deep learning detection model. The lightweight backbone network extracts deep defect features, and a multi-scale feature fusion mechanism is used to enhance subtle defect features. In parallel, the model completes accurate defect localization and classification of multiple types and levels of defects. The detection process relies on a cloud-edge collaborative architecture to achieve high-speed real-time detection at the edge and privacy-preserving model iteration in the cloud. At the same time, it links production line equipment and the MES system to complete early warning, sorting, and data archiving. Finally, through an adaptive model iteration optimization module, based on meta-learning and knowledge transfer technology, the model parameters are continuously and dynamically updated to achieve adaptive adaptation of the system to multiple categories and small batches of non-standard printing products. This solves the technical pain points of traditional detection systems, such as low accuracy, poor real-time performance, weak generalization ability, and inability to share and iterate data.

[0023] like Figure 1-3 As shown, the multispectral image acquisition includes three imaging bands: visible light, near infrared, and ultraviolet. It is equipped with a ring-shaped shadowless light source array and a six-axis adjustable mechanical lens bracket. Through mechanical linkage control logic, it can adapt to screen printing products of different sizes and substrates to achieve multi-view, blind-angle synchronous image acquisition. The multispectral image acquisition module is equipped with three core imaging bands: visible light, near-infrared, and ultraviolet. The visible light band is fixed at 400nm to 780nm and is mainly used to acquire visible macroscopic defects such as color difference, registration misalignment, pattern distortion, and large-area missing print in printed patterns. The near-infrared band is fixed at 780nm to 900nm and can penetrate the weak ink blockage on the surface to identify hidden defects such as hidden breaks, substrate bulges, and uneven ink layer thickness. The ultraviolet band is fixed at the commonly used industrial 365nm ultraviolet band, which is highly sensitive to micron-level pinholes, microcracks, and tiny ink delamination defects, and can capture minute flaws that cannot be identified by visible light imaging. On the hardware side, it is equipped with a high-brightness ring-shaped shadowless light source array. The light source adopts a zoned and independently controllable design, supporting stepless adjustment of brightness from 0 to 100%. It can adaptively adjust the brightness and activate the zone according to the different reflective characteristics of PCB substrates, ceramic substrates, photovoltaic silicon wafer substrates, and thin film substrates, completely eliminating interference from board surface reflection, shadows, and uneven local exposure. It is also equipped with a high-precision six-axis adjustable mechanical lens bracket, supporting X-axis, Y-axis, and Z-axis translation adjustment and three-axis rotation fine adjustment. The shooting distance adjustment range is 50mm to 800mm, and the angle adjustment accuracy can reach 0.1°. It can be fully adapted to printed products of different specifications and sizes, including micro precision components, conventional circuit boards, and large-size photovoltaic printed boards. The module has a built-in dedicated mechanical linkage control logic that can be linked with the production line's sensor signals. When a product arrival signal is detected, it automatically calls the preset shooting parameters, viewing angle parameters, and light source parameters corresponding to the product to complete multi-view, multi-band synchronous capture. This enables full-coverage image acquisition of the product's upper surface and side printing area without blind spots, reducing defect omissions and under-collection issues from the source, and providing complete and rich multi-dimensional image data support for subsequent high-precision inspection.

[0024] like Figure 1-3 As shown, the adaptive preprocessing module incorporates a dynamic illumination compensation algorithm and a multi-scale substrate texture suppression algorithm to counteract industrial ambient light fluctuations, printing substrate reflections, and substrate texture interference in real time. It performs registration and fusion, noise reduction and enhancement, and grayscale normalization on multi-band images, and automatically completes pixel-level pre-annotation of suspected defect areas, providing standardized image input for deep learning models. First, pixel-level noise removal is performed on the original multispectral image. A Gaussian adaptive noise reduction algorithm is used to filter out salt-and-pepper noise, Gaussian noise, and equipment circuit noise during the shooting process. Then, dynamic illumination compensation is activated, and the global grayscale mean and local grayscale variance of the image are statistically analyzed. Overexposed and underexposed areas are automatically corrected, and the overall image brightness is balanced. Then, a multi-scale texture suppression algorithm is used to extract the fixed texture features of the substrate and perform masking to retain ink printing features and defect and abnormal features, and to completely separate background interference from effective detection targets. Finally, multi-band image registration, alignment, and weighted fusion are completed to output a standardized image with uniform size, normalized grayscale, and clear features. Among them, the multispectral image weighted fusion adopts an adaptive dynamic weight formula, which can adjust the weight of each band in real time according to the on-site working conditions, and has strong adaptability. In the formula: For the final fused output image; , , These are the original images acquired in visible light, near-infrared, and ultraviolet light, respectively. , , The adaptive weighting coefficients satisfy the constraints. When the ambient light is strong and the substrate is highly reflective, the system automatically increases the near-infrared weight. When it is necessary to detect minute pinholes or microcracks, the weight of the ultraviolet band is automatically increased. Equalizing visible light weights in conventional detection scenarios To maximize the retention of effective features of various defects; Meanwhile, the module automatically completes pixel-level pre-annotation of suspected defect areas at the end of preprocessing. Through gray-level gradient abrupt change detection and texture feature difference comparison, it locks down abnormal areas in the image and generates pixel-level masks, eliminating large areas of defect-free background areas. This effectively reduces the computational load of subsequent models, improves model inference speed and detection accuracy, and provides high-quality, high signal-to-noise ratio, and standardized image input for deep learning models.

[0025] like Figure 1-3 As shown, the lightweight ShuffleNetV3 backbone feature extraction network achieves cross-group information interaction by splitting channel features through grouped convolution and channel shuffling, and reduces redundant convolutional layers and fully connected layers, compressing the number of model parameters to less than 1 / 10 of that of traditional convolutional networks. This achieves extreme lightweighting of the network while retaining feature extraction capabilities. The lightweight feature extraction network is specifically optimized for high-speed detection scenarios of screen printing. Its core purpose is to compress the model size and reduce the amount of computation to the maximum extent without losing the ability to extract defect features, thus adapting to the high-speed inference requirements of edge devices. The network front end adopts grouped convolution operation, which evenly divides the input feature channels into 8 independent branches. Each branch performs convolution operation independently. Compared with traditional full-channel convolution, it can directly reduce the convolution computation by 7 / 8. After completing the grouped convolution, the channel shuffling operation is started. The order of each group of channels is shuffled by matrix transposition and channel rearrangement to realize cross-group feature information interaction, which solves the problems of channel information fragmentation and low feature utilization in traditional grouped convolution. Meanwhile, this solution specifically prunes the native ShuffleNetV3 network for industrial applications, removing deep redundant convolutional layers, ineffective dimensionality reduction layers, and excess fully connected layers, while retaining the core feature extraction layer adapted for printing defect detection. Furthermore, by combining model pruning and 8-bit quantization compression techniques, the network weight parameters are simplified and compressed. After optimization, the overall parameter count of this backbone network can be compressed to less than 1 / 10 of the traditional ResNet50 network, significantly reducing floating-point computation. While effectively preserving shallow printing edge details and deep defect semantic features, this solution achieves extreme network lightweighting, fully meeting the computing power requirements for real-time inference at the edge of high-speed production lines.

[0026] like Figure 1-3 As shown, the dynamic attention feature pyramid network has a built-in pixel-level attention weight allocation mechanism, which adaptively adjusts the feature weights according to the size of the defect target, enhances the features of micron-level fine pinholes and short line defects, suppresses background invalid feature interference, and adapts to the feature extraction needs of multi-scale printing defects. The module has a built-in pixel-level attention weight allocation mechanism that can determine the effectiveness of features pixel by pixel, adaptively adjust the weight ratio of different pixels and different feature layers, automatically enhance defect features and suppress invalid background features. Its core attention weight calculation formula is as follows: In the formula: Image pixel coordinates The corresponding attention weights; This is a convolutional feature transformation function used to extract local feature differences between pixels; Using the Sigmoid activation function, the weights are normalized to the 0-1 range. Defective regions exhibit significant pixel grayscale abrupt changes and high feature diversity, resulting in weight values ​​approaching 1, thus enhancing features. Meanwhile, the substrate background and uniform ink regions show stable features, with weight values ​​approaching 0, achieving feature suppression. Simultaneously, a bidirectional path aggregation network is combined to construct a top-down and bottom-up bidirectional feature fusion pathway, fully integrating the network's shallow high-resolution detail features with deep high-semantic features. This preserves the contour details of minute defects while possessing semantic discrimination capabilities for defect types, perfectly adapting to the feature extraction needs of multi-scale and multi-morphological defects in screen printing, significantly reducing the false negative and false positive rates for minor defects.

[0027] like Figure 1-3 As shown, the simplified CenterNet algorithm removes the redundant upsampling module of the traditional CenterNet, retains the core center point prediction and size offset prediction branches, optimizes the coordinate regression loss function, controls the positioning error of minor defects within a suitable range, and adjusts the positioning accuracy of minor printing defects. First, redundant triple upsampling modules, offset compensation auxiliary branches, and boundary redundant regression branches are removed from the network, leaving only the core defect center point prediction branch and defect size offset prediction branch. This significantly simplifies the network structure and reduces the number of parameters and computational load. Second, the coordinate regression loss function is specifically improved. The traditional L2 loss function is not sensitive to small offsets, and an improved CIoU loss function is adopted. At the same time, the defect center coordinates, defect width and height dimensions, and boundary overlap are constrained to accurately constrain the coordinate offset error of small defects. Through extensive iterative training and parameter optimization using numerous industrial samples, this simplified CenterNet algorithm can stably control the positioning error of micro-pinholes, minor breaks, and local ink shortages larger than 5μm within a 5μm range. The accuracy of defect boundary fitting is significantly improved, solving the problems of inaccurate positioning of minor defects, boundary drift, and missed positioning of small targets in traditional algorithms. This provides accurate coordinate data support for subsequent defect size calculation and defect level determination.

[0028] like Figure 1-3 As shown, the lightweight Focal Loss loss function solves the long-tail distribution problem of screen printing defect samples by reducing the weight of simple samples and increasing the weight of scarce small defect samples. It presets three levels of defect classification thresholds: minor, general, and severe, and automatically outputs the defect level result by combining defect area, shape, and location parameters. By employing a lightweight Focal Loss function in the classification branch and balancing sample distribution differences through a dynamic weight adjustment mechanism, the focus is on learning features with scarce and minor defects. The core calculation formula for lightweight Focal Loss is as follows: In the formula: This represents the model's predicted probability of the true defect category. To balance the weights across categories, high weights are assigned to rare and minor defects, while low weights are assigned to common defects. The optimal value for the modulation coefficient is 2.0, which can effectively reduce the loss weight of simple normal samples and conventional large defect samples, force the model to focus on learning the feature patterns of small and scarce defects, and completely improve the model bias problem caused by the long-tail distribution of samples. Meanwhile, the system incorporates standardized defect classification logic, presets three defect classification thresholds: minor, moderate, and severe. It performs a weighted comprehensive evaluation based on four dimensions: actual defect area, outline shape, defect location, and defect quantity. The defect location is defined as both critical and non-critical printing areas. The system automatically outputs standardized defect level results. Minor defects are extremely small, superficial flaws that do not affect electrical performance; moderate defects are medium-sized printing flaws that can be reworked and repaired; and severe defects are fatal defects such as broken lines, large-area missing prints, and pinholes in critical locations that affect the product's conductivity and lifespan. This provides accurate data for subsequent production line classification and quality traceability.

[0029] like Figure 1-3 As shown, the edge computing unit has a built-in local data cache and fault tolerance mechanism, and is configured with breakpoint resumption inspection, abnormal frame filtering, and dynamic computing power scheduling logic. The edge computing unit locally builds a lightweight temporary defect database, caches the on-duty inspection images, defect data and model inference logs in real time, and continues to independently complete the inspection work under network outage and temporary equipment fluctuation conditions. After the network is restored, the offline data is automatically synchronized to the cloud platform to ensure that the inspection process is not interrupted and the data is not lost under complex industrial conditions. The system has built-in abnormal frame filtering logic, which can automatically identify invalid abnormal images such as overexposed shots, stuttering images, transmission interference, and empty frames where products have not arrived. These images are directly filtered out and do not participate in model inference, reducing unnecessary computing power consumption and error detection result output. At the same time, it is configured with a dynamic computing power scheduling mechanism, which can adaptively allocate computing power resources for model inference, data caching, and log storage based on the real-time computing power utilization rate of edge devices. This reduces inference stuttering, latency spikes, and program crashes under high load conditions, ensuring the continuous operation of high-speed production lines. The edge computing unit has complete breakpoint resume inspection and offline working capabilities. In the event of workshop network interruption, cloud signal fluctuation, or temporary abnormal equipment fluctuation, it can automatically cut off the cloud data interaction link and switch to local independent operation mode to continuously complete the entire process of inspection, including image acquisition, model inference, defect judgment, and local storage of results. There will be no production line shutdown, inspection interruption, or missed inspection. All data generated during offline operation, such as original images, defect coordinates, defect types, defect levels, inference logs, and equipment operating parameters, are encrypted and cached in a local temporary database. Once network communication returns to normal, the edge computing unit automatically identifies the network status, initiates a batch data synchronization mechanism, and uploads the offline cached data to the cloud platform and MES system in a complete manner according to the time sequence. This completes data supplementation and archiving, ensuring uninterrupted testing processes, no loss of quality data, and no missing production records. It greatly adapts to the complex and ever-changing industrial production conditions in factories, improving the overall reliability of the system and the stability of its industrial implementation.

[0030] like Figure 1-3 As shown, the cloud platform federated learning mechanism adopts a training mode in which local data is not uploaded and model parameters are encrypted and interacted. Under the premise of protecting the privacy of production data of each production line, it aggregates defect feature parameters of multiple sites to complete the joint iteration of the model and improves the generalization ability of the model. The specific process is as follows: the edge computing devices at each production site only complete the defect sample collection, model forward inference and gradient parameter calculation locally, and do not transmit the original printing images, product models and core privacy data of the production process to the outside world. Each edge device only uploads the encrypted model gradient parameters and weight update parameters to the cloud platform. The cloud uses a secure parameter aggregation algorithm to perform weighted averaging and fusion iteration on the encrypted parameters of multiple factories and multiple production lines to generate the globally optimal model parameters. Finally, the cloud distributes the updated global model parameters to each edge device to complete the synchronous upgrade of all production line models. This model can aggregate massive amounts of defect feature data from multiple scenarios and processes while absolutely protecting the privacy of production data in each factory area. It makes up for the shortcomings of single production line defect sample types and scenarios, continuously improves the model's ability to identify printing defects of different processes, equipment and working conditions, greatly improves the model's generalization ability, and reduces the problems of overfitting and poor adaptability of single production line models.

[0031] like Figure 1-3 As shown, the adaptive model iterative optimization module achieves rapid adaptation to small samples based on the meta-learning algorithm. For non-standard small-batch printing products, model fine-tuning can be completed with a small number of labeled samples. At the same time, knowledge distillation technology is combined to transfer the feature learning experience of large models to lightweight models, and the false negative rate and false positive rate indicators are monitored in real time, automatically triggering incremental model updates. The adaptive model iteration optimization module is the core module for achieving intelligent and long-term upgrades. It integrates four core control logics: rapid adaptation of meta-learning small samples, real-time monitoring of model performance, intelligent optimization of knowledge transfer, and automatic incremental training. It solves the technical shortcomings of traditional deep learning models, such as reliance on massive samples, long adaptation cycles for non-standard products, and the inability to upgrade models independently. First, the module builds a small-sample rapid adaptation mechanism based on meta-learning algorithm. Relying on the prior knowledge of general printing defects learned in the early stage, for non-standard customized, small-batch, and new variety screen printing products, only 50 to 100 manually labeled samples are needed to complete the rapid fine-tuning of the model. The overall fine-tuning cycle can be controlled within 8 hours. Compared with the traditional model that requires thousands of samples and several days of training cycle, it greatly reduces the adaptation cost of non-standard products and improves the flexible production adaptation capability of the production line. Secondly, the module introduces knowledge distillation technology to construct a teacher-student transfer architecture consisting of a high-precision large model (teacher model) and a lightweight deployment model (student model). This transfers the massive defect feature patterns and subtle defect discrimination experience learned by the high-precision large model to the lightweight model of this system. Without increasing edge-end inference time or reducing detection speed, this further improves the detection accuracy and discrimination stability of the lightweight model. The knowledge distillation loss fusion formula is as follows: In the formula: The hard classification loss corresponding to the true label ensures basic detection accuracy; Output the supervisory loss corresponding to the soft label for the teacher model, which is used to transfer higher-order feature experience; To balance the weighting coefficients, the optimal value is 0.5, which takes into account both the constraints of the actual labels and the effect of knowledge transfer. Finally, the module monitors the four core performance indicators of the model in real time: false negative rate, false positive rate, detection accuracy, and inference time. It presets standard threshold ranges. When the production line process is adjusted, ink parameters change, or equipment wear causes the defect morphology to change, causing the model performance indicators to fall below the preset threshold, the system automatically triggers incremental sample collection and incremental model training, dynamically updates the model parameters, realizes unattended autonomous iterative upgrades of the model, and ensures the long-term detection accuracy and stability of the system.

[0032] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based screen printing defect detection system, characterized in that, include: Multispectral image acquisition, configured with multi-band imaging and mechanical linkage control logic, synchronously acquires images of multiple types of printed products from a preset perspective; The adaptive preprocessing module is configured with dynamic compensation and texture suppression control logic to perform normalization fusion and pixel-level defect pre-annotation on multispectral images. The dual-path lightweight deep learning detection model module includes: a) The backbone feature extraction network is based on the lightweight ShuffleNetV3 and adopts a convolutional structure that combines grouped convolution and channel shuffling; b) Introduce a dynamic attention feature pyramid network and a bidirectional path aggregation network to achieve multi-scale feature adaptive fusion and enhance the feature extraction capability of target defects; c) Set up a defect localization branch and a defect classification branch in parallel. The localization branch uses a simplified CenterNet algorithm to achieve high-precision regression of the defect center point and size. The classification branch uses a multi-label classifier combined with a lightweight Focal Loss loss function to achieve defect classification during screen printing and output the defect level classification results. The cloud-edge collaboration module includes: a) The edge computing unit deploys a lightweight model to reduce single-frame image inference latency and meet the high-speed real-time detection requirements of the production line; b) Build an intelligent management system for defect samples on the cloud platform, integrate a federated learning mechanism, and realize privacy-preserving sharing of defect data from multiple production lines and joint training of models; c) Real-time signal linkage with screen printing equipment, automatic sorting equipment, and MES system. When a serious defect is detected, the printing equipment will be automatically triggered to stop the machine. The sorting equipment will be controlled to sort the defective products to the designated area and the quality data in the MES system will be updated in real time. The adaptive model iterative optimization module includes: Configure meta-learning adaptation, performance monitoring, and knowledge transfer control logic to dynamically update detection model parameters.

2. The screen printing defect detection system based on deep learning according to claim 1, characterized in that: The multispectral image acquisition includes three imaging bands: visible light, near-infrared, and ultraviolet. It is equipped with a ring-shaped shadowless light source array and a six-axis adjustable mechanical lens bracket. Through mechanical linkage control logic, it can adapt to screen printing products of different sizes and substrates, and realize multi-view synchronous image acquisition without blind spots.

3. The screen printing defect detection system based on deep learning according to claim 1, characterized in that: The adaptive preprocessing module incorporates a dynamic illumination compensation algorithm and a multi-scale substrate texture suppression algorithm to counteract industrial ambient light fluctuations, printing substrate reflections, and substrate texture interference in real time. It performs registration and fusion, noise reduction and enhancement, and grayscale normalization on multi-band images, and automatically completes pixel-level pre-annotation of suspected defect areas, providing standardized image input for deep learning models.

4. The screen printing defect detection system based on deep learning according to claim 1, characterized in that: The lightweight ShuffleNetV3 backbone feature extraction network achieves cross-group information interaction by splitting channel features through grouped convolution and channel shuffling, and reduces redundant convolutional layers and fully connected layers, compressing the number of model parameters to less than 1 / 10 of that of traditional convolutional networks.

5. The screen printing defect detection system based on deep learning according to claim 1, characterized in that: The dynamic attention feature pyramid network incorporates a pixel-level attention weight allocation mechanism, which adaptively adjusts feature weights according to the size of the defect target, enhances features for micron-level pinholes and short line defects, and suppresses interference from invalid background features.

6. The screen printing defect detection system based on deep learning according to claim 1, characterized in that: The simplified CenterNet algorithm removes the redundant upsampling module of the traditional CenterNet, retains the core center point prediction and size offset prediction branches, optimizes the coordinate regression loss function, controls the positioning error of minor defects within a suitable range, and adjusts the positioning accuracy of minor printing defects.

7. The screen printing defect detection system based on deep learning according to claim 1, characterized in that: The system presets three defect classification thresholds: minor, moderate, and severe. It then automatically outputs the defect classification result based on the defect area, shape, and location parameters.

8. The screen printing defect detection system based on deep learning according to claim 1, characterized in that: The edge computing unit has a built-in local data cache and fault tolerance mechanism, and is configured with breakpoint continuation inspection, abnormal frame filtering and dynamic computing power scheduling logic. The edge computing unit locally builds a lightweight temporary defect database, caches the on-duty inspection images, defect data and model inference logs, and continues to independently complete the inspection work under network outage and temporary equipment fluctuation conditions. After the network is restored, the offline data is automatically synchronized to the cloud platform.

9. The screen printing defect detection system based on deep learning according to claim 1, characterized in that: The cloud platform's federated learning mechanism adopts a training mode that does not upload local data and uses encrypted interaction of model parameters. Under the premise of protecting the privacy of production data from each production line, it aggregates defect feature parameters from multiple sites to complete the joint iteration of the model.

10. The screen printing defect detection system based on deep learning according to claim 1, characterized in that: The adaptive model iterative optimization module is based on meta-learning algorithm to achieve rapid adaptation to small samples. For non-standard small-batch printing products, model fine-tuning can be completed with a small number of labeled samples. At the same time, knowledge distillation technology is combined to transfer the feature learning experience of large models to lightweight models, and the false negative rate and false positive rate are monitored in real time, automatically triggering incremental model updates.

Images