Paper product packaging defect real-time detection system based on edge calculation

Abstract

The invention discloses a paper product packaging defect real-time detection system based on edge calculation, and relates to the technical field of packaging defect detection. Comprising a multi-modal data acquisition module for acquiring packaging multi-dimensional image data by using a camera and a sensor; the edge intelligent processing unit and the heterogeneous architecture are combined with dynamic scheduling to realize rapid data processing; the defect feature enhancement network is used for improving an algorithm to accurately extract defect features; the edge decision optimization engine is used for reinforcing learning to optimize a detection strategy; the industrial-grade communication module is used for guaranteeing data transmission through 5G and TSN hybrid; the closed-loop repair execution system is used for realizing automatic repair of defects; in addition, auxiliary modules such as a traceability module and an energy optimization module are further arranged, and all the modules cooperate to achieve efficient detection and repair of packaging defects. The paper product packaging detection level is greatly improved, the speed is improved by 2.5 times, energy consumption is reduced, detection and closed loop repairing are achieved, intelligent upgrading of a production line is promoted, and the labor cost and the product reject ratio are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of packaging defect detection, and in particular to a real-time detection system for paper product packaging defects based on edge computing. Background Art

[0002] In the field of paper packaging production, quality control directly affects product market competitiveness, and defect detection is a key link. Traditional manual inspection methods rely on visual observation by quality inspectors, which is extremely inefficient. In a high-speed production line environment, it is difficult to meet the inspection needs of over 100 products per minute. In addition, due to factors such as worker fatigue and subjective judgment differences, the missed detection rate is as high as 12%, and the false detection rate is as high as 8%, seriously affecting product quality stability and production efficiency. Even if some companies use traditional machine vision inspection systems, their algorithms based on fixed thresholds and simple feature extraction are insufficient to identify defects such as fine scratches less than 0.1mm in width and tiny wrinkles with height differences less than 0.15mm when faced with complex textured packaging, transparent film materials, and changing lighting conditions. The detection accuracy fluctuates greatly, making it difficult to guarantee inspection quality.

[0003] With the development of intelligent manufacturing, while edge computing and deep learning technologies are beginning to be applied to packaging defect detection, significant shortcomings remain. Existing systems often transmit data to the cloud for processing, resulting in high latency in data transmission and an inability to meet the real-time detection needs of production lines. Deep learning models have large parameters and place heavy demands on the computing resources of edge devices, resulting in slow inference speeds and difficulty adapting to the high-speed pace of production. Furthermore, most inspection systems only identify defects and lack the ability to perform subsequent defect processing, preventing a closed-loop process from detection to repair and resulting in poor integration between production links. Furthermore, system energy consumption remains high, and data from each production line is isolated, making cross-production line collaborative optimization impossible, thus limiting the progress of enterprises' intelligent transformation. Summary of the Invention

[0004] The present invention proposes a real-time detection system for paper product packaging defects based on edge computing to solve the problems mentioned in the above-mentioned prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a real-time detection system for paper product packaging defects based on edge computing, comprising: Data acquisition module: Using polarization-structured light fusion imaging technology, the integrated linear array CCD camera and polarization sensor synchronously acquire RGB images, depth maps, and polarization characteristics of the packaging surface. The degree of polarization (DoP) is calculated using the Stokes vector to enhance transparent packaging defect detection. , where DoP stands for degree of polarization, which is used to quantify the degree of polarization of light; I is the total light intensity; Q, U, and V are Stokes parameters that describe the polarization state of light, corresponding to polarization components in different directions respectively; the influence of surface microstructure on polarized light is quantified by the calculation formula; Edge processing unit: built-in heterogeneous computing architecture CPU+GPU+FPGA, using dynamic task scheduling algorithm, according to the task complexity index , computing-intensive tasks are assigned to FPGA, while logical judgment tasks are accelerated by GPU, realizing parallel processing and real-time fusion of data; where C is the task complexity index; I is the input data volume, representing the scale of data required to be processed by the task; O is the output data volume, that is, the amount of data output after task processing; P is the processing priority, which is used to identify the urgency or importance of the task; the formula is used to evaluate task characteristics to achieve task allocation; Defect Feature Enhancement Network: We designed the dilated attention module (MSCAM) to capture multi-scale features through three sets of dilated convolutions with different dilation rates of 3, 5, and 7. We also introduced a channel attention mechanism to dynamically adjust feature weights to identify irregularly shaped defects. Communication module: Using the TSN time-sensitive network and edge cache collaboration mechanism, VLAN channels are configured for defect image data. Historical data is pre-fetched through the edge cache algorithm to reduce bandwidth usage, achieving zero-packet-loss real-time communication within a distance of 100m. Closed-loop repair execution system: Integrates a high-speed piezoelectric-driven micro-positioning platform and a laser repair unit. Based on the three-dimensional reconstruction data of the defect, it repairs micro defects by fitting the repair path with a non-uniform rational B-spline (NURBS) curve.

[0006] Furthermore, it also includes: Edge decision optimization engine: Build a deep reinforcement learning (DRL) decision model, take the defect missed detection rate (FNR), false positive rate (FPR), and computing energy consumption (E) as comprehensive optimization objectives, and design a weighted reward function: , where R is the reward value, which is used to evaluate the quality of the decision-making strategy; w1, w2, and w3 are the weight coefficients corresponding to the defect missed detection rate, false detection rate, and computing energy consumption, respectively, reflecting the importance of different indicators in the optimization goal; FNR is the defect missed detection rate, which is the proportion of undetected defects to actual defects; FPR is the false detection rate, which is the proportion of normal samples that are mistakenly detected as defects; E is the current computing energy consumption; E max is the maximum allowable energy consumption of the system; the reward value is calculated through the formula to guide the model to optimize the decision strategy; Furthermore, the data acquisition module also includes an ultraviolet fluorescence imaging unit, which uses a 365nm wavelength ultraviolet LED array to excite the fluorescent substance on the packaging surface, and captures the wavelength fluorescence signal through a narrow-band filter to detect invisible ink printing defects.

[0007] Furthermore, the edge processing unit also includes an adaptive noise suppression module, which dynamically adjusts the denoising threshold based on the analysis of each resolution of wavelet transform and the Bayesian shrinkage algorithm according to the statistical characteristics of the noise to achieve robust denoising in a production environment.

[0008] Furthermore, the defect feature enhancement network also includes a feature pyramid fusion layer, which adopts a bidirectional feature pyramid network BiFPN structure and realizes adaptive aggregation of features of different scales through a weighted mechanism to recall micro-target defect detection.

[0009] Furthermore, the edge decision optimization engine also includes a knowledge distillation acceleration module, which compresses the neural network knowledge trained in the cloud to the edge model, and improves the edge reasoning speed by minimizing the output distribution difference between the teacher model and the student model.

[0010] Furthermore, the communication module also includes an edge cache scheduling unit, which uses a dual-queue scheduling algorithm DQS to distinguish real-time control data from historical analysis data, and ensures zero packet loss rate of control instructions by dynamically adjusting queue length and service priority.

[0011] Furthermore, the closed-loop repair execution system also includes a laser energy adaptive control module, which dynamically adjusts the laser repair energy based on the nonlinear mapping relationship between defect depth and area to ensure that the bonding strength between the repair area and the substrate is consistent with the original strength.

[0012] Furthermore, it also includes: Defect traceability management module: Establishes a defect-process parameter association database, identifies the degree of correlation between defect types and process parameters temperature T, pressure P, and speed v through a correlation analysis algorithm, and triggers the process parameter adaptive adjustment mechanism when the correlation exceeds the threshold.

[0013] Energy optimization management module: Using deep reinforcement learning (DRL) algorithm, it dynamically adjusts the hardware operating frequency according to the real-time load status of the system to reduce system energy consumption.

[0014] Federated learning update module: Using a secure aggregation protocol, edge nodes train model parameters locally and then generate a global model through encrypted gradient aggregation, ensuring data privacy while enabling continuous model evolution.

[0015] Compared with the existing technology, the beneficial effects of the present invention are: In terms of detection accuracy, the system uses multimodal data acquisition and deep feature enhancement networks to collect and integrate multi-dimensional information such as RGB images, polarization characteristics, and three-dimensional morphology on the packaging surface, accurately capturing tiny defects and increasing the average detection accuracy from 94.3% of traditional machine vision to 99.2%. The detection rate of tiny defects is improved, greatly reducing the outflow of unqualified products.

[0016] In terms of inspection efficiency and production collaboration, the application of edge computing architecture and lightweight deep learning models reduces system response latency and achieves an inspection speed of 200 pieces per minute, meeting the demands of high-speed production. The closed-loop repair execution system automatically repairs defects with a 98% success rate, achieving a seamless integration between inspection and production, significantly improving production efficiency and product qualification rates.

[0017] The system also boasts exceptional energy-saving and data collaboration capabilities. The energy optimization management module dynamically adjusts hardware operating frequency, reducing energy consumption per unit product and saving production costs. The federated learning update module enables cross-production line data sharing and model optimization, protecting data privacy while driving continuous system performance evolution and providing strong support for enterprise intelligent upgrades. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic block diagram of the real-time detection system for paper product packaging defects based on edge computing proposed in the present invention; Figure 2 This is a schematic diagram comparing the defect detection rates of different detection technologies in the real-time detection system for paper product packaging defects based on edge computing proposed in the present invention; Figure 3 This is a schematic diagram showing how the detection speed of the real-time detection system for paper product packaging defects based on edge computing proposed in the present invention changes with the production line speed. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0021] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below with reference to the accompanying drawings.

[0022] Reference Figure 1 and Figure 3 :A specific implementation of a real-time detection system for paper product packaging defects based on edge computing 1. Multimodal data acquisition module The system of this application adopts polarization-structured light fusion imaging technology to build a multimodal data acquisition module. Its specific implementation structure is as follows: Figure 1 As shown. The linear array CCD camera uses the German Baslerace2 series product, with a resolution of 5μm, and is combined with an 8K high-speed image sensor to achieve continuous acquisition at 12,000 frames per second. The multi-channel polarization sensor uses the Sony IMX250MZR, which has a built-in four-directional polarization filter array (0°, 45°, 90°, and 135°) and a measurement angle accuracy of 0.1°. This sophisticated hardware configuration ensures that the system can capture characteristic information on the packaging surface from multiple dimensions, providing a rich data foundation for subsequent defect detection. During the transparent packaging defect detection process, the system calculates the degree of polarization (DoP) using the Stokes vector: .

[0023] When light strikes a packaging surface, different surface structures affect the polarization state of the light to varying degrees. A smooth surface maintains relatively stable polarization characteristics for the reflected light, while defects such as scratches and wrinkles alter the polarization direction and intensity distribution of the reflected light. By calculating the DoP value, the system quantifies these variations, enabling the detection of subtle defects that are imperceptible to the naked eye. The Stokes parameters I, Q, U, and V in the formula represent the polarization components in different directions, where I represents the total light intensity, Q represents the linear polarization component in the horizontal-vertical direction, U represents the linear polarization component in the ±45° direction, and V represents the circular polarization component (in this system, V = 0 is assumed to be circularly polarized). Benefits: This polarization-based inspection method is particularly suitable for inspecting transparent packaging materials, addressing the shortcomings of traditional visual inspection methods for such materials. Experimental data shows that when the DoP value varies by more than 0.05, it can effectively identify subtle scratches with a width ≥ 0.08 mm and wrinkles with a height difference ≥ 0.15 mm, improving detection accuracy by 20% and reducing false detection rates by 35% compared to traditional methods. The UV fluorescence imaging unit uses a 365nm wavelength UVLED array, evenly distributed in a ring light source, to excite the fluorescent material on the packaging surface. The narrowband filter uses Semrock FF01-450 / 50-25 with a bandwidth of 10nm, and is combined with the Hamamatsu S11155-01 avalanche photodiode (APD) detector to achieve 0.05mg / cm 2 Invisible ink defect detection sensitivity.

[0024] Some packaging materials may use invisible inks for anti-counterfeiting marking or product information labeling. These inks cannot be detected under visible light, but they will emit fluorescence when excited by ultraviolet light of a specific wavelength. The system precisely controls the wavelength and intensity of ultraviolet light to excite the fluorescent substance to emit light of a specific wavelength, and then filters out interfering light of other wavelengths through a narrow-band filter, allowing only the fluorescent signal to pass through. Finally, it is captured by a highly sensitive APD detector and converted into an electrical signal for analysis. Beneficial effects: This detection method can effectively identify invisible ink defects on packaging, such as incomplete printing, uneven ink concentration, etc. The detection sensitivity is 50% higher than that of traditional fluorescence detection methods, and the minimum detectable ink concentration is from 0.1mg / cm 2 Reduced to 0.05 mg / cm 2 , providing strong guarantee for the quality control of high-end packaging products.

[0025] 2. Edge Intelligent Processing Unit The edge intelligent processing unit deploys a heterogeneous computing architecture, and the specific hardware configuration is an Intel Xeon E-2274G CPU, an NVIDIA Jetson AGX Xavier GPU, and a Xilinx UltraScale+ FPGA. The design of this heterogeneous architecture fully considers the characteristics of different types of computing tasks. The CPU is good at processing logic control tasks, the GPU has strong advantages in parallel computing, and the FPGA is suitable for computing-intensive tasks with extremely high real-time requirements. The dynamic task scheduling algorithm is based on the task complexity index for task allocation.

[0026] This formula allocates tasks to the most suitable computing unit by comprehensively considering the data volume and priority of the tasks, thereby achieving the optimal utilization of system resources. Among them, I is the input data volume (MB), representing the data scale that the task needs to process; O is the output data volume (MB), reflecting the data volume generated after the task is processed; P is the processing priority (levels 1-5), representing the importance of the task. When C>10, it means that the task has a large data volume and high computational complexity. At this time, the task is allocated to the FPGA for processing to utilize its hardware parallelism to improve the processing speed; when 3<C≤10, it is accelerated by the GPU; when C≤3, the task is relatively simple and can be processed by the CPU.

[0027] This dynamic task scheduling algorithm increases the system throughput to 2000 frames per second, the processing delay of the FPGA is stable within 8 ms, the processing efficiency is increased by 40% compared with the traditional fixed allocation method, and the resource utilization rate is increased by 35%. It can better meet the strict real-time requirements of high-speed production lines. The adaptive noise suppression module is based on the multi-resolution analysis of wavelet transform and uses the Daubechies db4 wavelet basis function for 4-layer decomposition. The Bayesian shrinkage algorithm dynamically adjusts the threshold according to the local variance of the noise: .

[0028] The core of wavelet transform is to decompose the signal into different frequency sub-bands and process them separately in each sub-band. For image signals, the low-frequency sub-band contains the main information of the image (such as contours, textures), while the high-frequency sub-bands mainly contain noise and detail information. The Bayesian shrinkage algorithm is based on the method of statistical inference and determines the optimal threshold by estimating the variances of the signal and noise. When the absolute value of the wavelet coefficient is greater than the threshold, it is considered that the coefficient mainly contains signal information and is appropriately shrunk; when it is less than the threshold, it is considered that the coefficient is mainly noise and is set to zero. This can remove noise while maximizing the retention of the detail information of the signal.

[0029] In actual tests, this module improved the system SNR (signal-to-noise ratio) by 12dB, effectively suppressing random noise and pulse interference in the production line environment, reducing the image noise level from 15% of traditional methods to below 5%, providing cleaner and more reliable data for subsequent defect feature extraction, and improving defect detection accuracy by 8%.

[0030] 3. Defect Feature Enhancement Network The defect feature enhancement network adopts the multi-scale void attention module (MSCAM), the specific structure is as follows Figure 2 As shown in the figure, the dilation rates of the three dilated convolutional layers are 3, 5, and 7, respectively, and the convolution kernel size is 3×3. This design enables the network to capture defect features at different scales. Dilated convolutions with larger dilation rates can capture a wider range of contextual information, while dilated convolutions with smaller dilation rates can retain more detailed information. The channel attention mechanism generates a weight vector through global average pooling and fully connected layers: .

[0031] The core idea of ​​the channel attention mechanism is to automatically learn the importance of each channel, assign higher weights to important channels, and suppress responses to unimportant channels. Specifically, the feature map of each channel is first compressed into a scalar through global average pooling (GAP), resulting in a 1×1×C feature vector, where C is the number of channels. This feature vector is then nonlinearly transformed through two fully connected layers (W1 and W2). The first fully connected layer reduces the feature vector dimension to the original (r is the compression ratio, typically 16), and a ReLU activation function is applied to increase the model's nonlinear expressiveness. The second fully connected layer restores the dimension to C and applies a Sigmoid activation function to limit the output value to between 0 and 1, thereby obtaining the weight coefficient for each channel. Finally, the original feature map is multiplied by these weight coefficients to achieve channel weighting.

[0032] This channel attention mechanism enables the network to focus more on features relevant to defect detection and suppress interference from irrelevant background information. Experiments show that after introducing the MSCAM module, the network's recognition accuracy for irregularly shaped defects increased from 95.7% to 99.2%, and the recall rate for small defects increased by 12%. Meanwhile, the number of model parameters increased by only 2.3%, effectively balancing detection accuracy and computational complexity. The feature pyramid fusion layer uses a BiFPN structure, which achieves adaptive aggregation of features at different scales through a learnable weighting mechanism: .

[0033] In traditional feature pyramid networks, features of different scales are simply added or concatenated, which does not take into account the importance of different features to the final detection results. BiFPN, by introducing learnable weights, can dynamically adjust the fusion method according to the importance of different features. Specifically, for each input feature P i , the network learns a corresponding weight w i , and then normalize these weights through the Softmax function to get the contribution ratio of each feature, and finally perform weighted fusion of each feature according to these ratios. is a small constant used to prevent the denominator from being zero.

[0034] This adaptive feature fusion method enables the network to better utilize feature information of different scales and improve the detection ability of defects of different sizes. Especially for defects with a size of ≤0.15mm 2 The recall rate of small target defects is improved from 85.0% of traditional FPN to 97.8%, and the detection accuracy is improved by 15.1%. At the same time, the inference speed of the model is only reduced by 3.2%, maintaining a high efficiency while ensuring detection performance.

[0035] 4. Edge decision optimization engine The edge decision optimization engine builds a deep reinforcement learning model and uses the PPO algorithm for training. The state space contains 12 dimensions, including defect detection metrics (FNR, FPR), computing resource utilization (CPU / GPU / FPGA), and energy consumption. The action space is a three-dimensional continuous vector that controls the camera exposure time (0-100ms), light source intensity (0-100%), and algorithm processing accuracy (low / medium / high). The reward function is designed as follows: .

[0036] The design idea of ​​this reward function is to minimize the false positive rate and energy consumption while ensuring the accuracy of detection. The weight coefficients w1=0.6, w2=0.3, and w3=0.1 reflect the importance of each indicator. The detection accuracy (1-FNR) is the most critical indicator, so it is given the highest weight; the false positive rate (FPR) affects the subsequent processing flow, so it is also given a higher weight; energy consumption Optimization is performed while ensuring the first two indicators. The deep reinforcement learning model continuously interacts with the environment, tries different action combinations, and learns the optimal decision-making strategy based on the reward signals it receives. Specifically, it automatically adjusts camera exposure time, light intensity, and algorithm processing accuracy under different production environments and packaging types to achieve optimal detection results and resource utilization efficiency.

[0037] In this way, the deep reinforcement learning model can learn the optimal decision-making strategy, balancing detection performance and resource consumption. Experiments show that compared with traditional fixed-parameter detection methods, this engine reduces the system's FNR (missed detection rate) from 4.8% to 0.8%, and the FPR (false positive rate) from 6.2% to 1.5%. At the same time, the system's energy consumption is reduced by 15%, achieving a dual improvement in detection accuracy and energy efficiency. The knowledge distillation acceleration module compresses the cloud-trained ResNet-50 teacher model (with 25.6M parameters) into a MobileNetV3 student model (with 2.9M parameters), achieving knowledge transfer by minimizing the KL divergence: .

[0038] The core idea of ​​knowledge distillation is to transfer the knowledge learned by the large model (teacher model) to the small model (student model). During the training process, the student model must not only learn the true labels, but also learn the output probability distribution of the teacher model. The KL divergence KL(p t ||p s ) measures the softened probability distribution p of the teacher model t And the softened probability distribution p of the student model s The softening probability distribution is achieved by introducing a temperature parameter T, that is, when performing a softmax operation on the original output of the model, the logits are divided by T, making the probability distribution smoother, thereby exposing more implicit knowledge learned by the teacher model. This ensures that the student model can still learn the true label information. By adjusting the values ​​of α and T, the degree to which the student model learns the true label and the teacher model knowledge can be balanced.

[0039] Experimental results show that the distilled model's inference speed increased by 4.2 times, from 22ms / frame to 5.2ms / frame, while maintaining 98.7% detection accuracy. This represents a 3.5% improvement in accuracy compared to traditional model compression methods with the same number of parameters. This enables the model to run efficiently on edge devices, meeting the real-time requirements of high-speed production lines while reducing hardware costs and energy consumption.

[0040] 5.Industrial-grade communication module The industrial communication module uses a TSN switch (Moxa EDS-408A) to build a time-sensitive network. A dedicated VLAN channel is configured for defect image data, with a priority level of 7. The edge cache scheduling unit uses a dual-queue scheduling algorithm (DQS), with a strict priority queue for real-time control data (P1) and a weighted fair queue for historical analysis data (P2). The cache prefetch algorithm is based on a Markov prediction model: .

[0041] The algorithm analyzes historical request sequences to predict possible future request contents and caches these contents in edge devices in advance. Specifically, the algorithm maintains a state transition table that records the number of transitions from one state (request content) to another state C(st,st{t+1}). For the current state , the probability of the next state st{t+1} can be calculated using the above formula. Based on these probabilities, the system can predict the most likely requested content and cache it in advance, thereby reducing data transmission delays.

[0042] In actual tests, this algorithm achieved a cache hit rate of 85.3%. At a transmission distance of 100m, the P1 data transmission latency jitter was controlled within 50μs, achieving zero packet loss communication. Compared to the traditional FIFO scheduling algorithm, latency was reduced by 75% and throughput was increased by 30%. This efficient communication mechanism ensures the system can acquire and process data in real time, providing a strong guarantee for high-speed operation of production lines. It is particularly suitable for industrial automation scenarios with extremely strict real-time requirements.

[0043] 6. Closed-loop repair execution system The closed-loop repair execution system integrates PI's P-621.ZCD piezoelectric drive micro-positioning platform, with a positioning accuracy of ±0.03mm and a maximum acceleration of 20m / s. 2 The laser repair unit uses IPGYLR-50 fiber laser with an adjustable power density range of 10-100W / cm 2 The laser energy adaptive control module is based on the nonlinear mapping relationship between defect depth d and area A: .

[0044] This formula dynamically adjusts the laser energy according to the depth and area of ​​the defect. For deeper and larger defects, higher laser energy is provided to ensure complete repair; for smaller and shallower defects, lower energy is used to avoid damage to surrounding materials. Among them, k0=5 is the basic energy value, k1=0.8 is the energy coefficient, k2=1.2 and k3=0.6 are exponential parameters. These parameters are obtained by fitting a large amount of experimental data and can accurately reflect the relationship between defect characteristics and required laser energy. The repair path is fitted by NURBS curve, and the number of control points is dynamically adjusted according to the complexity of the defect (usually 5-15) to ensure that the repair path is smooth and accurately covers the defect area.

[0045] Experiments have shown that the system has a 96.7% success rate for repairing tiny defects as small as 0.2mm, with the bond strength of the repaired area reaching 95.2% of the substrate's strength. This represents a 20% improvement in success rate and a 15% improvement in repair quality compared to traditional fixed-energy repair methods. In actual production, this system can effectively reduce scrap rates, improve production efficiency, and lower production costs, making it particularly suitable for defect repair in high-value packaging products.

[0046] 7. Defect traceability management module The defect traceability management module establishes a defect-process parameter correlation database containing more than 100,000 samples and uses the Pearson correlation coefficient to analyze the correlation between defect types and process parameters: .

[0047] The Pearson correlation coefficient measures the degree of linear correlation between two variables, and its value ranges from -1 to 1. When the correlation coefficient is close to 1, it indicates that the two variables are positively correlated, that is, when one variable increases, the other variable tends to increase; when it is close to -1, it indicates a negative correlation, that is, when one variable increases, the other variable tends to decrease; when it is close to 0, it indicates that there is no linear relationship between the two variables. In this system, by calculating the correlation coefficient between the defect type and each process parameter, the process parameter that has the greatest impact on the defect can be found. For example, when the correlation coefficient between the fold mark defect and the temperature parameter is calculated to be 0.82, it means that temperature has a strong positive correlation with the generation of fold mark defects, and an increase in temperature may lead to an increase in fold mark defects.

[0048] When the correlation exceeds a threshold of 0.7, the system automatically triggers process parameter adjustments. In actual applications, this module has increased process parameter adjustment efficiency by 75% and reduced defect recurrence by 68%. By continuously optimizing process parameters, defects are eliminated at the source. Compared with traditional manual process parameter adjustment methods, this system can more quickly and accurately identify the root cause of problems, improve production efficiency and product quality, and reduce production costs.

[0049] 8.Energy optimization management module The energy optimization management module uses a deep reinforcement learning algorithm. The state space includes eight dimensions, including system load, temperature, and power consumption of each hardware. The action space is the adjustment amount of each hardware operating frequency. The reward function is designed as: .

[0050] The design goal of this reward function is to reduce energy consumption while ensuring that the system temperature does not exceed the safety threshold. Indicates the ratio of current energy consumption to baseline energy consumption, represents the ratio of the temperature change to the temperature threshold. α = 0.7 and β = 0.3 are weight coefficients reflecting the relative importance of energy consumption and temperature. The deep reinforcement learning model interacts with the environment, tries different hardware frequency adjustment strategies, and learns the optimal energy management strategy based on the reward signal obtained. For example, when the system load is low, the model reduces the hardware operating frequency to reduce energy consumption. When the temperature approaches the threshold, the model appropriately increases the frequency to enhance heat dissipation while maintaining system performance.

[0051] In actual testing, this module reduced average system energy consumption by 25.4%, from 387W to 289W, while also ensuring that hardware temperatures did not exceed 75°C, a 50% reduction in temperature fluctuation compared to traditional fixed-frequency operation. This not only saves energy costs, but also extends the lifespan of hardware equipment, reducing the frequency of equipment maintenance and replacement, bringing significant economic benefits to the enterprise.

[0052] 9. Federated Learning Update Module The federated learning update module uses a secure aggregation protocol. The edge node first trains the model parameters locally and then encrypts the gradients using homomorphic encryption technology: .

[0053] Homomorphic encryption is a special encryption technology that allows specific computations to be performed on ciphertext without first decrypting it. In this system, each edge node uses its own private key k i The locally calculated gradient g i Encrypt and get the encrypted gradient After these encrypted gradients are sent to the cloud server, the server can aggregate them (such as adding them) in the ciphertext state to obtain the aggregated encrypted gradients Due to the properties of homomorphic encryption, the result of this aggregation operation is equivalent to first aggregating the original gradients and then encrypting them. Finally, the server uses the master key to decrypt the aggregated encrypted gradients to obtain the global gradient G. This approach ensures that the original gradient data remains encrypted throughout the entire process, without leaking any private information.

[0054] Experiments have shown that this solution, while ensuring data privacy, improves model accuracy by 3.2%, shortens the model update cycle to 2 hours per update, and enables continuous model evolution and optimization. Compared with traditional centralized learning methods, federated learning avoids the risk of data leakage during transmission and storage, meeting the stringent requirements of enterprises for data privacy and security. It also enables joint training using data from multiple edge nodes, improving the model's generalization and performance.

[0055] Beneficial effect data characterization and analysis

[0056] It can be seen from the above data that the system of this application is significantly superior to the traditional system in many key performance indicators. The improvement in detection accuracy is mainly due to the synergy between multimodal data acquisition and feature enhancement network. The multimodal data acquisition module obtains information on the packaging surface from different angles, including RGB images, depth maps, polarization characteristics and fluorescence signals, etc., providing a rich data source for defect detection. The MSCAM module and BiFPN structure in the feature enhancement network can effectively extract and fuse features of different scales and types, thereby improving the ability to identify tiny defects and irregularly shaped defects.

[0057] The significant reduction in missed and false positive rates is attributed to the deep reinforcement learning model within the edge decision optimization engine. This model dynamically adjusts detection strategies by comprehensively considering factors such as FNR, FPR, and energy consumption, minimizing missed and false positives while maintaining high detection accuracy. In particular, the application of knowledge distillation technology significantly improves the model's inference speed without compromising performance, enabling the system to accurately detect a wide range of defects in real time on high-speed production lines.

[0058] The significant improvement in processing latency is due to the heterogeneous computing architecture and dynamic task scheduling algorithm. By assigning different types of tasks to the most appropriate computing units, the respective strengths of the CPU, GPU, and FPGA are fully utilized, achieving optimal utilization of computing resources. The FPGA's accelerated processing of compute-intensive tasks and the GPU's efficient execution of logical judgment tasks have increased overall system throughput by nearly fourfold, reducing processing latency from 52ms to 7.8ms, meeting the stringent real-time requirements of high-speed production lines.

[0059] The energy optimization management module dynamically adjusts hardware operating frequencies through deep reinforcement learning, achieving a 25.4% reduction in energy consumption while maintaining detection performance. This not only saves energy costs but also helps reduce carbon emissions, aligning with the development trend of green manufacturing. Furthermore, effective system temperature control extends the lifespan of hardware equipment and reduces maintenance costs.

[0060] The extended filter cloth lifespan and reduced number of failures are due to the closed-loop repair execution system's timely repair of minor defects and the defect traceability management module's optimized adjustment of process parameters. The closed-loop repair system accurately locates and repairs minor defects, preventing them from expanding and damaging the filter cloth. The defect traceability module analyzes the correlation between defects and process parameters, adjusting them promptly and reducing defects at the source. These comprehensive performance improvements make the system significantly more economical and competitive in practical applications, potentially bringing higher production efficiency and lower costs to paper packaging manufacturers.

[0061] Tabular data representation corresponding to beneficial effects

[0062] These data fully demonstrate the significant advantages of the system of this application in multiple key technical indicators. The improvement in polarization detection sensitivity enables the system to detect more subtle defects, which is especially important for the detection of high-quality packaging products. The improvement in fluorescence detection accuracy enhances the system's ability to detect special marks such as invisible ink, which is helpful for anti-counterfeiting and quality control. The improvement in small target recall rate means that the system can more accurately identify and locate tiny defects, reducing the possibility of missed detection. The significant improvement in model inference speed enables the system to process large amounts of image data in real time on high-speed production lines, meeting the actual needs of industrial production. The improvement in process adjustment efficiency reduces defects caused by inappropriate process parameters, improves production efficiency and product quality. The reduction in communication delay jitter ensures reliable communication between the various components of the system, and ensures the stability and accuracy of the entire detection process. In summary, through the application of a series of innovative technologies, the system of this application has achieved a comprehensive improvement in performance in the field of paper product packaging defect detection, and has broad application prospects and market value.

[0063] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A real-time detection system for paper product packaging defects based on edge computing, characterized in that: include: Data acquisition module: Using polarization-structured light fusion imaging technology, the integrated linear array CCD camera and polarization sensor synchronously acquire RGB images, depth maps, and polarization characteristics of the packaging surface. The degree of polarization (DoP) is calculated using the Stokes vector to enhance transparent packaging defect detection. , where DoP stands for degree of polarization, which is used to quantify the degree of polarization of light; I is the total light intensity; Q, U, and V are Stokes parameters that describe the polarization state of light, corresponding to polarization components in different directions respectively; the influence of surface microstructure on polarized light is quantified by the calculation formula; Edge processing unit: built-in heterogeneous computing architecture CPU+GPU+FPGA, using dynamic task scheduling algorithm, according to the task complexity index , computing-intensive tasks are assigned to FPGA, while logical judgment tasks are accelerated by GPU, realizing parallel processing and real-time fusion of data; where C is the task complexity index; I is the input data volume, representing the scale of data required to be processed by the task; O is the output data volume, that is, the amount of data output after task processing; P is the processing priority, which is used to identify the urgency or importance of the task; the formula is used to evaluate task characteristics to achieve task allocation; Defect Feature Enhancement Network: We designed the dilated attention module (MSCAM) to capture multi-scale features through three sets of dilated convolutions with different dilation rates of 3, 5, and 7. We also introduced a channel attention mechanism to dynamically adjust feature weights to identify irregularly shaped defects. Communication module: Using the TSN time-sensitive network and edge cache collaboration mechanism, VLAN channels are configured for defect image data. Historical data is pre-fetched through the edge cache algorithm to reduce bandwidth usage, achieving zero-packet-loss real-time communication within a distance of 100m. Closed-loop repair execution system: Integrates a high-speed piezoelectric-driven micro-positioning platform and a laser repair unit. Based on the three-dimensional reconstruction data of the defect, it repairs micro defects by fitting the repair path with a non-uniform rational B-spline (NURBS) curve.

2. The real-time detection system for paper product packaging defects based on edge computing according to claim 1 is characterized in that: Also includes: Edge decision optimization engine: Build a deep reinforcement learning (DRL) decision model, take the defect missed detection rate (FNR), false positive rate (FPR), and computing energy consumption (E) as comprehensive optimization objectives, and design a weighted reward function: , where R is the reward value, which is used to evaluate the quality of the decision-making strategy; w1, w2, and w3 are the weight coefficients corresponding to the defect missed detection rate, false detection rate, and computing energy consumption, respectively, reflecting the importance of different indicators in the optimization goal; FNR is the defect missed detection rate, which is the proportion of undetected defects to actual defects; FPR is the false detection rate, which is the proportion of normal samples that are mistakenly detected as defects; E is the current computing energy consumption; E max is the maximum allowable energy consumption of the system; Calculate reward values ​​through formulas to guide the model to optimize decision strategies; The data acquisition module also includes an ultraviolet fluorescence imaging unit, which uses a 365nm wavelength ultraviolet LED array to excite the fluorescent substance on the packaging surface and captures the wavelength fluorescence signal through a narrow-band filter to detect invisible ink printing defects.

3. The real-time detection system for paper product packaging defects based on edge computing according to claim 1 is characterized in that: The edge processing unit also includes an adaptive noise suppression module, which dynamically adjusts the denoising threshold based on the statistical characteristics of noise based on the analysis of each resolution of wavelet transform combined with the Bayesian shrinkage algorithm to achieve robust denoising in a production environment.

4. The real-time detection system for paper product packaging defects based on edge computing according to claim 1 is characterized in that: The defect feature enhancement network also includes a feature pyramid fusion layer, which adopts a bidirectional feature pyramid network BiFPN structure and realizes adaptive aggregation of features of different scales through a weighted mechanism to recall micro-target defect detection.

5. The real-time detection system for paper product packaging defects based on edge computing according to claim 2 is characterized in that: The edge decision optimization engine also includes a knowledge distillation acceleration module, which compresses the neural network knowledge trained in the cloud to the edge model, and improves the edge reasoning speed by minimizing the output distribution difference between the teacher model and the student model.

6. The real-time detection system for paper product packaging defects based on edge computing according to claim 1 is characterized in that: The communication module also includes an edge cache scheduling unit, which uses a dual-queue scheduling algorithm DQS to distinguish real-time control data from historical analysis data, and ensures zero packet loss rate of control instructions by dynamically adjusting queue length and service priority.

7. The real-time detection system for paper product packaging defects based on edge computing according to claim 1 is characterized in that: The closed-loop repair execution system also includes a laser energy adaptive control module, which dynamically adjusts the laser repair energy based on the nonlinear mapping relationship between defect depth and area to ensure that the bonding strength between the repair area and the substrate is consistent with the original strength.

8. The real-time detection system for paper product packaging defects based on edge computing according to claim 1 is characterized in that: Also includes: Defect traceability management module: Establishes a defect-process parameter association database, identifies the degree of correlation between defect types and process parameters temperature T, pressure P, and speed v through a correlation analysis algorithm, and triggers the process parameter adaptive adjustment mechanism when the correlation exceeds the threshold.

9. The real-time detection system for paper product packaging defects based on edge computing according to claim 1 is characterized in that: Also includes: Energy optimization management module: Using deep reinforcement learning (DRL) algorithm, it dynamically adjusts the hardware operating frequency according to the real-time load status of the system to reduce system energy consumption.

10. The real-time detection system for paper product packaging defects based on edge computing according to claim 1 is characterized in that: Also includes: Federated learning update module: Using a secure aggregation protocol, edge nodes train model parameters locally and then generate a global model through encrypted gradient aggregation, ensuring data privacy while enabling continuous model evolution.

Images