RFID label production line defect real-time detection system and compensation method based on machine vision

Abstract

The invention discloses an RFID tag production line defect real-time detection system and compensation method based on machine vision, and belongs to the technical field of intelligent manufacturing and quality control. The system comprises an image acquisition module, a defect detection module, a compensation execution module and a feedback control module, based on defect types and positions, a grading compensation strategy is triggered, the grading compensation strategy comprises real-time printing parameter adjustment, laser accurate repair of a defect area or automatic rejection of a mechanical arm, and a process parameter library is synchronously updated through an industrial Internet of Things platform. The innovation points are as follows: (1) a multi-modal defect data fusion method is provided, and the detection robustness under complex working conditions is improved; (2) a lightweight double-branch detection network is designed, and the detection speed and precision are both considered; and (3) developing a compensation instruction cooperative control algorithm to realize seamless connection between defect repair and production line rhythm. According to the scheme, the defect omission ratio can be reduced to 0.5% or below, the compensation response time is shortened to be within 200 ms, and the production line yield and the production efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent manufacturing and quality control, and particularly to a real-time defect detection system and compensation method for RFID tag production lines based on machine vision. Background Art

[0002] As a key component of the Internet of Things, the production process of radio frequency identification (RFID) tags involves multiple precision processes such as printing, chip mounting, and antenna etching. Current mainstream detection methods rely on manual visual inspection or traditional machine vision technology, suffering from problems such as low efficiency, high missed detection rate (about 3%-5%), and lag in defect compensation. Especially in the scenario of high-speed production lines (≥200 tags / minute), existing technologies are difficult to capture micron-level defects (such as antenna breakage and chip offset) in real time, and there is a lack of dynamic compensation mechanism after defect discovery, resulting in a high scrap rate (about 4%-8%), seriously restricting production yield and cost control.

[0003] Referring to the Chinese patent "A Defect Detection Method and System for RFID Tags", it proposes a defect detection solution based on multi-spectral imaging and machine learning algorithms, improving the recognition ability of surface and internal defects through the fusion imaging of visible light and infrared light. However, this solution has significant limitations: 1) The detection model is not optimized for high-speed production lines, and the processing time for a single tag is ≥500 ms, making it difficult to meet the real-time requirements; 2) It does not integrate a defect compensation mechanism, and the detection results are only used for offline sorting, unable to dynamically adjust production parameters or repair defects, resulting in defect compensation lagging behind the production line beat; 3) The multi-spectral data fusion method is single, and the false detection rate is relatively high (about 2.5%) under complex lighting or material differences.

[0004] Based on the existing technology, the current defect control of RFID tag production lines faces the following core problems: 1) Disconnection between detection and compensation: Traditional systems only achieve defect recognition, without forming a "perception - decision - execution" closed loop, and the lag in compensation leads to defect accumulation; 2) Poor dynamic adaptability: Fixed threshold or rule-driven compensation strategies cannot adapt to multiple types of defects (such as blurred printing and antenna breakage requiring different treatments); 3) Low production line coordination efficiency: There is a lack of timing synchronization algorithms between the detection module and the execution mechanism (such as laser repair and robotic arm), easily causing production line beat disorders. The present invention aims to break through the above bottlenecks and construct an integrated solution for real-time detection and intelligent compensation. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the present invention provides a real-time defect detection system and compensation method for RFID tag production lines based on machine vision, which solves the problems of disconnection between detection and compensation, poor dynamic adaptability, and low production line coordination efficiency.

[0007] To solve the above technical problems, the present invention provides the following technical solutions:

[0008] In a first aspect, the present invention provides a real-time defect detection system for RFID tag production lines based on machine vision, including:

[0009] An image acquisition module: A multi-spectral imaging unit is used to collect surface and internal structure images of RFID tags transmitted on the production line. The imaging spectral range covers visible light (400 - 700nm) and near-infrared (800 - 1200nm), and the spatial resolution is not less than 20μm / pixel;

[0010] A defect detection module: A lightweight dual-branch deep learning network is equipped. The first branch uses an improved YOLOv5 model for defect localization, and the second branch uses ResNet-18 for defect classification, outputting defect types and coordinate information;

[0011] A compensation execution module: Triggers a hierarchical compensation strategy according to the defect type, including a printing parameter dynamic adjustment unit, a laser repair unit, and a robotic arm rejection unit;

[0012] A feedback control module: Feeds back defect data and compensation results to the production line PLC through an industrial Internet of Things platform, and updates the process parameter library in real time.

[0013] As a preferred solution of the real-time defect detection system for RFID tag production lines based on machine vision according to the present invention, wherein: The imaging parameters of the multi-spectral imaging unit satisfy:

[0014] The surface imaging bands are 450nm, 630nm, and 850nm, used to detect printing layer defects;

[0015] The internal imaging band is 1050nm, and the penetration depth is ≥200μm, used to detect antenna breaks and chip misalignments; The imaging frame rate is synchronized with the production line speed, satisfying the relationship:

[0016]

[0017] Wherein, F is the imaging frame rate (fps), v is the production line transmission speed (mm / s), d is the minimum defect size of the tag (mm), and η = 1.2 - 1.5 is a redundancy coefficient.

[0018] As a preferred solution of the real-time defect detection system for RFID tag production lines based on machine vision according to the present invention, wherein: The loss function of the lightweight dual-branch deep learning network is defined as:

[0019] L total = αL loc + βL cls + γL reg

[0020] Among them, L loc is the positioning loss (CIoU loss), and L cls is the classification loss (focal loss), and L reg is the model complexity regularization term, and the weight coefficients satisfy α:β:γ = 3:2:0.5.

[0021] As a preferred solution of the real-time defect detection system for RFID tag production lines based on machine vision according to the present invention, wherein: the triggering condition of the hierarchical compensation strategy is:

[0022] First-level compensation: When the proportion of the defect area S defect / S tag ≤ 5%, start the laser repair unit, and the repair path planning satisfies:

[0023]

[0024] Second-level compensation: When 5% < S defect / S tag ≤ 15%, adjust the printing parameters (temperature T ± 3°C, pressure P ± 5%);

[0025] Third-level compensation: When S defect / S tag > 15%, the robotic arm rejection unit removes the defective label within t ≤ 200 ms;

[0026] Among them, S defect is the defect area, and S tag is the label area.

[0027] As a preferred solution of the real-time defect detection system for RFID tag production lines based on machine vision according to the present invention, wherein: the optimization of the network structure of the improved YOLOv5 model includes:

[0028] Replace the C3 module in the backbone network with a Ghost module, reducing the number of parameters by 40%;

[0029] Add a coordinate attention mechanism (CA), and the feature map weight calculation is:

[0030]

[0031] Among them, F c is the channel feature, and δ(·) is the Sigmoid function.

[0032] As a preferred solution of the real-time defect detection system for RFID tag production lines based on machine vision according to the present invention, wherein: the power adjustment formula of the laser repair unit is:

[0033]

[0034] Among them, A defect is the defect area, A0 = 0.1mm 2 , k = 12W / mm 2 , b = 5W.

[0035] As a preferred solution of the real-time defect detection system for RFID tag production lines based on machine vision according to the present invention, wherein: the feedback control module dynamically optimizes the production line parameters through the PID algorithm:

[0036]

[0037] Among them, e(t) is the defect rate deviation, and the proportionality coefficient K p = 0.8, the integral coefficient K i = 0.2, the differential coefficient K d = 0.1.

[0038] As a preferred solution of the defect compensation method for RFID tag production lines based on machine vision according to the present invention, wherein: the following steps are included:

[0039] Step S1: The multi-spectral imaging unit synchronously acquires the surface and internal images of the tag;

[0040] Step S2: The dual-branch deep learning network outputs the defect type and location. If the confidence conf≥0.9, it is determined as a defect;

[0041] Step S3: Select the compensation method according to the defect level, and after execution, re-inspect through an industrial camera. If there are still defects, trigger secondary compensation or rejection.

[0042] As a preferred solution of the defect compensation method for RFID tag production lines based on machine vision according to the present invention, wherein: in step S2, the confidence calculation uses an improved Softmax function:

[0043]

[0044] Among them, ρ = 0.5 is the temperature coefficient, which is used to enhance the discrimination of the classification confidence, and z i represents the original output value of the i-th category (i.e., the unnormalized logits), and adjusts the smoothness of the confidence distribution through the temperature coefficient ρ.

[0045] As a preferred solution of the defect compensation method for RFID tag production lines based on machine vision according to the present invention, wherein: the response time t response of the compensation instruction in step S3 satisfies:

[0046]

[0047] Among them, L tag is the label length, v line is the production line speed, t process ≤50ms is the data processing time delay.

[0048] In a second aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program is executed by the processor, any step of a method for compensating defects in an RFID tag production line based on machine vision as described in the first aspect of the present invention is implemented.

[0049] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and: when the computer program is executed by the processor, any step of a method for compensating defects in an RFID tag production line based on machine vision as described in the first aspect of the present invention is implemented.

[0050] The beneficial effects of the present invention are as follows:

[0051] 1. High-precision real-time detection and low missed detection rate

[0052] By fusing visible light and the near-infrared band (400 - 1200nm) through multi-spectral imaging technology and combining a lightweight dual-branch deep learning network (improved YOLOv5 + ResNet-18), synchronous detection of surface printing defects and internal structure defects (such as antenna breakage, chip misalignment, etc.) is achieved. The defect recognition accuracy reaches over 99.5%, the missed detection rate is reduced to below 0.5%, and the single-tag detection time ≤ 50ms, meeting the real-time requirements of high-speed production lines (≥ 300 tags / minute).

[0053] 2. Dynamic closed-loop compensation and efficient repair

[0054] A hierarchical compensation strategy (parameter adjustment, laser repair, automatic rejection) is proposed. Combining PID feedback control and laser power adjustment algorithms, the compensation response time is shortened to within 200ms, the defect repair success rate ≥ 95%, and the scrap rate is reduced from 4% - 8% of the traditional method to below 1%.

[0055] 3. Intelligent production line coordination and process optimization

[0056] Based on the feedback control module of the industrial Internet of Things platform, defect data and production parameters (temperature, pressure, speed) are synchronized in real time. Through PID dynamic adjustment (K p = 0.8, K i = 0.2, K d= 0.1) realizes the self - optimization of the production line, reduces the manual intervention frequency by more than 80%, and supports the continuous iterative update of the process parameter library to meet the flexible production requirements of multiple types of RFID tags (high - frequency, ultra - high - frequency).

[0057] 4. Cost Savings and Scalability

[0058] Adopts a lightweight network (replacing the C3 module with the Ghost module, reducing the number of parameters by 40%) and a multi - modal data fusion method, reducing the hardware resource occupancy by 35%. At the same time, it is compatible with the upgrade and transformation of existing production line equipment, and the deployment cost is only 60% - 70% of the traditional solution, suitable for large - scale mass production scenarios.

[0059] In summary, through the full - process closed - loop management of detection - compensation - feedback, the present invention significantly improves the yield, efficiency and economy of RFID tag production, providing a highly reliable quality control solution for intelligent manufacturing. Brief Description of the Drawings

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0061] Figure 1 It is a schematic diagram of a real - time defect detection system for an RFID tag production line based on machine vision in Embodiment 1;

[0062] Figure 2 It is a flowchart of a defect compensation method for an RFID tag production line based on machine vision in Embodiment 2. Detailed Embodiments

[0063] To make the above - mentioned objects, features and advantages of the present invention more obvious and understandable, the following will give a detailed description of the specific embodiments of the present invention with reference to the drawings of the specification.

[0064] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described here. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0065] Secondly, the so - called "one embodiment" or "embodiment" refers to a specific feature, structure or characteristic that can be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it an individual or alternative embodiment that mutually excludes other embodiments.

[0066] Example 1, referring to Figure 1 , which is the first embodiment of the present invention. This embodiment provides a real-time defect detection system for RFID tag production lines based on machine vision, including:

[0067] Image acquisition module: The multi-spectral imaging unit 101 is used to collect surface and internal structure images of RFID tags transmitted on the production line. The imaging spectral range covers visible light (400 - 700nm) and near-infrared (800 - 1200nm), and the spatial resolution is not less than 20μm / pixel;

[0068] Defect detection module: The lightweight dual-branch deep learning network 201 is equipped. The first branch uses an improved YOLOv5 model for defect localization, and the second branch uses ResNet-18 for defect classification, outputting defect type and coordinate information;

[0069] Compensation execution module: Trigger a hierarchical compensation strategy according to the defect type, including a printing parameter dynamic adjustment unit 301, a laser repair unit 302, and a robotic arm rejection unit 303;

[0070] Feedback control module: Feedback defect data and compensation results to the production line PLC through the industrial Internet of Things platform 401, and update the process parameter library in real time.

[0071] The imaging parameters of the multi-spectral imaging unit 101 meet the following:

[0072] The surface imaging bands are 450nm, 630nm, and 850nm, which are used to detect printing layer defects;

[0073] The internal imaging band is 1050nm, and the penetration depth ≥ 200μm, which is used to detect antenna breakage and chip misalignment; The imaging frame rate is synchronized with the production line speed and satisfies the relationship:

[0074]

[0075] where F is the imaging frame rate (fps), v is the production line transmission speed (mm / s), d is the minimum defect size of the tag (mm), and η = 1.2 - 1.5 is the redundancy coefficient.

[0076] The loss function of the lightweight dual-branch deep learning network 201 is defined as:

[0077] L total = αL loc + βL cls + γL reg

[0078] where L loc is the localization loss (using CIoU loss), Lcls is the classification loss (focal loss), L reg is the model complexity regularization term, and the weight coefficients satisfy α:β:γ = 3:2:0.5.

[0079] The triggering conditions for the hierarchical compensation strategy are as follows:

[0080] First-level compensation: When the proportion of the defect area S defect / S tag ≤ 5%, start the laser repair unit 302, and the repair path planning satisfies:

[0081]

[0082] Second-level compensation: When 5% < S defect / S tag ≤ 15%, adjust the printing parameters (temperature T ± 3°C, pressure P ± 5%);

[0083] Third-level compensation: When S defect / S tag > 15%, the robotic arm rejection unit 303 removes the defect label within t ≤ 200 ms;

[0084] Among them, S defect is the defect area, and S tag is the label area.

[0085] The network structure optimization of the improved YOLOv5 model includes:

[0086] Replace the C3 module in the backbone network with a Ghost module, reducing the number of parameters by 40%;

[0087] Add a coordinate attention mechanism (CA), and the feature map weight calculation is as follows:

[0088]

[0089] Among them, F c is the channel feature, and δ(·) is the Sigmoid function.

[0090] The power adjustment formula of the laser repair unit 302 is:

[0091]

[0092] Among them, A defect is the defect area, A0 = 0.1 mm 2 , k = 12 W / mm 2 , b = 5 W.

[0093] The feedback control module 401 dynamically optimizes the production line parameters through the PID algorithm:

[0094]

[0095] Among them, e(t) is the defect rate deviation, and the proportionality coefficient K p = 0.8, the integral coefficient K i = 0.2, the differential coefficient K d = 0.1.

[0096] Example 2, referring to Figure 2 , which is the second embodiment of the present invention. This embodiment provides a method for compensating defects in an RFID tag production line based on machine vision, including the following steps:

[0097] Step S1: The multispectral imaging unit 101 synchronously acquires images of the tag surface and inside;

[0098] 1. Hardware configuration and light source design

[0099] The multispectral imaging unit consists of two groups of optical modules:

[0100] Surface imaging module: Adopt a high-resolution visible light camera (resolution ≥ 20μm / pixel), equipped with a three-band LED array light source of 450nm (blue light), 630nm (red light), and 850nm (near-infrared), covering the detection requirements of the printing layer color and surface topography.

[0101] Internal imaging module: Adopt a near-infrared camera (wavelength 1050nm, penetration depth ≥ 200μm) and a coaxial light source, and shield stray light through an optical filter, which is specifically used to capture the internal structure of the antenna copper foil layer and the chip package.

[0102] Synchronous trigger mechanism: Real-time monitor the production line transmission position through a photoelectric sensor (or encoder). When the RFID tag enters the imaging area, trigger the two modules to synchronously expose (time deviation ≤ 1ms) to ensure the spatio-temporal alignment of the surface and internal images.

[0103] 2. Dynamic adaptation of imaging parameters

[0104] Frame rate matching with the production line speed: According to the production line transmission speed v (unit: mm / s) and the minimum defect size d of the tag (unit: mm), dynamically adjust the imaging frame rate:

[0105]

[0106] For example, when v = 300mm / s and d = 0.2mm, calculate F = 1800fps to ensure that the overlapping rate of adjacent frames

[0107] ≥ 20%, to avoid missed detection.

[0108] Adaptive Exposure Control: Based on surface reflectivity and material light transmission characteristics, the light source intensity and camera exposure time (range: 10 μs - 10 ms) are adjusted in real time through the PID algorithm to ensure that the imaging signal-to-noise ratio (SNR) of different material labels (such as PET, paper) is ≥ 40 dB.

[0109] 3. Image Preprocessing and Calibration

[0110] Spatial Alignment: A checkerboard calibration plate is used to geometrically calibrate the fields of view of the surface and internal cameras, and pixel-level alignment (error ≤ 2 μm) is achieved through an affine transformation matrix.

[0111] Multispectral Fusion: The three-band visible light images are weighted and fused (weight coefficients: w 450 = 0.4, w 630 = 0.3, w 850 = 0.3) to enhance the contrast of printed patterns; the internal near-infrared images are sharpened by Unsharp Mask filtering to enhance the antenna edge details.

[0112] Data Encapsulation: The synchronously acquired multispectral images (surface RGB + internal NIR) are bound with timestamps and production line speed information, and transmitted to the defect detection module 201 through Gigabit Ethernet, with a single-frame transmission delay ≤ 5 ms.

[0113] 4. Anti-Interference and Stability Guarantee

[0114] Enclosed Imaging Cavity: Light-shielding materials are used to isolate ambient light, and nitrogen is filled inside to avoid mirror condensation.

[0115] Vibration Compensation: In the high-speed production line scenario, mechanical vibrations are monitored in real time through inertial sensors, and the position of the optical platform is adjusted using a servo motor (compensation accuracy ±5 μm) to eliminate motion blur.

[0116] Self-Cleaning System: An air curtain device and periodic ultraviolet lamp irradiation are integrated to prevent dust from adhering to the lens and ensure long-term imaging stability.

[0117] It should be noted that through the above implementation method, in step S1, surface and internal images can be synchronously obtained on a high-speed production line of 300 labels per minute with a spatial resolution of ≤ 0.1 mm, providing high-precision and low-noise input data for subsequent defect detection and compensation, while ensuring strict synchronization with the production line rhythm and avoiding missed detections or false triggers caused by imaging delays.

[0118] Step S2: The dual-branch deep learning network 201 outputs the defect type and location. If the confidence conf ≥ 0.9, it is determined as a defect;

[0119] In step S2, the confidence calculation uses an improved Softmax function:

[0120]

[0121] Among them, ρ = 0.5 is the temperature coefficient, which is used to enhance the discrimination of classification confidence, and z i represents the original output value of the i-th category (i.e., the unnormalized logits), and adjusts the smoothness of the confidence distribution through the temperature coefficient ρ.

[0122] 1. Dual-branch network architecture design

[0123] Input data:

[0124] Receive the multi-spectral image (RGB three channels) of the label surface and the internal near-infrared image (NIR single channel) synchronously collected in step S1, and splice them into a four-channel input (size: 640×640×4).

[0125] Branch structure:

[0126] Localization branch (improved YOLOv5):

[0127] Backbone network: Replace the C3 module of YOLOv5 with a Ghost module (parameter quantity reduced by 40%), and embed the Coordinate Attention (CA) mechanism to enhance the sensitivity to tiny defects. Output layer: Output the coordinates (x, y, w, h) of the defect bounding box and the localization confidence conf loc .

[0128] Classification branch (lightweight ResNet-18):

[0129] Feature extraction: Receive the candidate regions (ROIs) output by the localization branch, adjust them to a size of 224×224 through adaptive pooling, and input them into ResNet-18 for feature extraction.

[0130] Classification head: Output the defect category probability p i (including 6 categories in total such as blurred printing, antenna breakage, chip misalignment, etc.) and the classification confidence conf cls .

[0131] 2. Confidence fusion and defect determination

[0132] Joint confidence calculation:

[0133] Based on the comprehensive localization and classification results, calculate the final confidence: conf = conf loc ×conf cls

[0134] Among them:

[0135] conf locThe target existence probability output by the localization branch (after Sigmoid activation);

[0136] conf cls Calculated by the classification branch using an improved Softmax function:

[0137]

[0138] where z i is the logits value of the i-th class, and ρ reduces the smoothness of the classification probability distribution and enhances the distinguishability of high-confidence classes.

[0139] Decision logic:

[0140] If conf ≥ 0.9, it is determined as a valid defect, and the type and coordinates are output;

[0141] If 0.7 ≤ conf < 0.9, a review mechanism is triggered (such as secondary detection of local images);

[0142] If conf < 0.7, it is determined as a normal label.

[0143] 3. Network training and optimization

[0144] Loss function:

[0145] The total loss L total includes the localization loss L loc , the classification loss L cls and the regularization term L reg :

[0146] L total = 3L loc + 2L cls + 0.5L reg

[0147] L loc : The CIoU loss is adopted to measure the overlap degree and the center point distance between the predicted box and the ground truth box;

[0148] L cls : The Focal Loss is adopted to alleviate the problem of class imbalance;

[0149] L reg : L2 regularization is used to constrain the model complexity.

[0150] Training data:

[0151] The dataset contains 100,000 annotated RFID tag images (with the surface and the interior aligned), covering 6 types of defects and normal samples;

[0152] Data augmentation: Random multi-spectral channel perturbation, motion blur simulation, Gaussian noise injection.

[0153] Training strategy:

[0154] Pre-training: The localization branch is pre-trained on the COCO dataset, and the classification branch is pre-trained on ImageNet;

[0155] Joint fine-tuning: Freeze the backbone of the localization branch, only train the CA module and the classification branch, and the learning rate

[0156] 1×10 -4 ;

[0157] End-to-end optimization: Unfreeze the entire network, use the AdamW optimizer, and the learning rate is 5×10 -5 .

[0158] 4. Real-time guarantee

[0159] Hardware acceleration:

[0160] Quantize the dual-branch network (FP16 precision) through the TensorRT engine and deploy it on NVIDIA Jetson AGX Xavier, with the single-frame inference time ≤ 15ms.

[0161] Dynamic resource allocation:

[0162] Adjust the number of ROIs according to the production line speed (v):

[0163]

[0164] where t cls is the classification time for a single ROI (about 1.2ms), ensuring that the total processing time ≤ 50ms.

[0165] It should be noted that the present invention has the advantages of high-precision determination, low false detection rate and real-time response. The confidence threshold of 0.9 is verified by the ROC curve, balancing the precision (Precision≥98%) and recall (Recall≥96%); in complex lighting and material variation scenarios, the false detection rate ≤ 0.3%; the end-to-end delay from image input to defect determination ≤ 30ms, supporting the high-speed production line demand of 300 labels / minute.

[0166] Step S3: Select the compensation method according to the defect level, and perform a re-inspection through an industrial camera after execution. If there are still defects, trigger secondary compensation or rejection.

[0167] The response time t of the compensation instruction in step S3 response satisfies:

[0168]

[0169] where, L tagis the label length, v line is the production line speed, t process ≤50ms is the data processing delay.

[0170] 1. Defect level classification and compensation strategy matching

[0171] Level classification criteria:

[0172] Based on the defect type, area ratio (S defect / S tag ) and position sensitivity (e.g., higher defect weight in the chip area) output by step S2, dynamically assign defect levels:

[0173] First-level defect (minor):

[0174] Condition: S defect / S tag ≤5% and located in a non-critical area (such as the antenna edge).

[0175] Compensation method: The laser repair unit 302 performs local repair.

[0176] Second-level defect (moderate):

[0177] Condition: 5% < S defect / S tag ≤15% or located in a critical area (such as the chip mounting area).

[0178] Compensation method: Dynamically adjust printing parameters (temperature T±3°C, pressure P±5%).

[0179] Third-level defect (severe):

[0180] Condition: S defect / S tag >15% or functional failure (such as antenna breakage).

[0181] Compensation method: The robotic arm rejection unit 303 removes the defective label and marks it as a scrap.

[0182] Dynamic priority scheduling:

[0183] If a label has multiple defects, perform compensation in the order of "third level > second level > first level" to ensure that critical defects are processed first.

[0184] 2. Compensation execution and parameter adaptation

[0185] Laser repair unit 302:

[0186] Path planning: Generate the optimal laser path based on the defect shape to minimize the repair time:

[0187]

[0188] Power adjustment: Dynamically adjust the laser power according to the defect area Adefect

[0189]

[0190] Ensure that the repair depth matches the material properties (such as PET substrate or copper foil) to avoid overburning or insufficient repair.

[0191] Parameter adjustment unit 301:

[0192] Real-time adjust the printing parameters through the PID algorithm:

[0193]

[0194] where e(t) is the deviation between the current defect rate and the target value, and the parameter adjustment period ≤ 100 ms.

[0195] Robotic arm rejection unit 303:

[0196] Adopt a Delta robotic arm with a response time ≤ 200 ms, and the motion trajectory is calculated based on the production line speed v:

[0197]

[0198] where t delay is the compensation delay, and d tag is the label spacing to ensure accurate grasping.

[0199] 3. Re-inspection process and secondary compensation logic

[0200] Re-inspection trigger condition:

[0201] After completing the primary or secondary compensation, trigger the re-inspection industrial camera (the same model as the main imaging unit) to image the repaired / adjusted area again.

[0202] The re-inspection image is transmitted to the dual-branch network (201), and only the original defect area is locally detected, with a time consumption ≤ 10 ms.

[0203] Re-inspection judgment rule:

[0204] Pass: If the re-inspection confidence level conf recheck < 0.7, it is determined that the compensation is successful, and the label flows into the next process.

[0205] Fail: If conf recheck ≥ 0.7, trigger according to the residual defect level:

[0206] Primary residual defect: Initiate secondary laser repair with a 10% - 20% power increase (such as P laser × 1.15).

[0207] Secondary residual defects: Adjust the printing parameters twice (such as temperature T ± 5°C, pressure P ± 8%), and extend the PID integral time to 200 ms.

[0208] Tertiary residual defects or failed secondary repair: Force rejection and record in the process parameter library for subsequent production optimization.

[0209] Number of attempts limit:

[0210] Each single label allows a maximum of two compensation attempts. If it still does not meet the standard, force rejection and give an alarm prompt.

[0211] 4. Timing synchronization and production line coordination

[0212] Beat matching:

[0213] The total time for compensation and re-inspection needs to meet the production line transmission constraint:

[0214]

[0215] Among them, t comp is the compensation time, t recheck is the re-inspection time, L tag is the label length, v line is the production line speed, t process ≤ 50 ms is the data processing time delay.

[0216] Exception handling:

[0217] If the compensation times out or the robotic arm fails, immediately activate the emergency stop protocol, synchronously notify the MES system, and switch to the redundant execution unit.

[0218] It should be noted that the present invention has the advantages of high repair success rate, low mis-rejection rate, production line compatibility, and process self-optimization. The first repair success rate of primary defects ≥ 95%, and reaches 99% after secondary repair; through the re-inspection mechanism, the mis-rejection rate ≤ 0.2%; supports a production line speed ≤ 400 labels / minute, and the compensation and re-inspection processes are seamlessly embedded in the production beat; the cumulative defect data is fed back to the parameter library to achieve continuous iterative improvement of the printing and mounting processes.

[0219] This embodiment also provides a computer device, applicable to a situation of a machine vision-based RFID label production line defect compensation method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement a machine vision-based RFID label production line defect compensation method as proposed in the above embodiment.

[0220] The computer device may be a terminal, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0221] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the method for compensating defects in an RFID tag production line based on machine vision proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM for short), Electrically Erasable Programmable Read-Only Memory (EEPROM for short), Erasable Programmable Read Only Memory (EPROM for short), Programmable Red-Only Memory (PROM for short), Read-Only Memory (ROM for short), magnetic memory, flash memory, a magnetic disk, or an optical disc.

[0222] The present invention combines visible light and the near-infrared band (400 - 1200 nm) through multi-spectral imaging technology, and combines a lightweight dual-branch deep learning network (improved YOLOv5 + ResNet-18) to achieve synchronous detection of surface printing defects and internal structure defects (such as antenna breakage, chip misalignment, etc.). The defect recognition accuracy reaches over 99.5%, the missed detection rate is reduced to below 0.5%, and the single-label detection time is ≤50 ms, meeting the real-time requirements of high-speed production lines (≥300 labels / minute). A hierarchical compensation strategy (parameter adjustment, laser repair, automatic rejection) is proposed, combined with a PID feedback control and a laser power adjustment algorithm, and the compensation response time is shortened to within 200 ms, and the defect repair success rate is ≥95%. The scrap rate is reduced from 4% - 8% of the traditional method to below 1%. Based on the feedback control module of the industrial Internet of Things platform, defect data and production parameters (temperature, pressure, speed) are synchronized in real time, and through PID dynamic adjustment (K p = 0.8, K i = 0.2, K d = 0.1), the production line is self-optimized, the frequency of manual intervention is reduced by more than 80%, and continuous iterative updates of the process parameter library are supported to meet the flexible production requirements of multiple types of RFID tags (high frequency, ultra-high frequency). By using a lightweight network (replacing the C3 module with a Ghost module, reducing the number of parameters by 40%) and a multi-modal data fusion method, the hardware resource occupancy is reduced by 35%, and at the same time, it is compatible with the upgrade and transformation of existing production line equipment, and the deployment cost is only 60% - 70% of the traditional solution, suitable for large-scale mass production scenarios. Through the full-process closed-loop management of detection - compensation - feedback, the present invention significantly improves the yield, efficiency and economy of RFID tag production, providing a highly reliable quality control solution for intelligent manufacturing.

[0223] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A real-time defect detection system for RFID tag production lines based on machine vision, characterized in that Including: Image acquisition module: A multi-spectral imaging unit (101) is used to acquire surface and internal structure images of RFID tags transmitted on the production line. The imaging spectral range covers visible light (400 - 700nm) and near-infrared (800 - 1200nm), and the spatial resolution is not less than 20μm / pixel; Defect detection module: A lightweight dual-branch deep learning network (201) is carried. The first branch uses an improved YOLOv5 model for defect localization, and the second branch uses ResNet-18 for defect classification, outputting defect type and coordinate information; Compensation execution module: Trigger a hierarchical compensation strategy according to the defect type, including a printing parameter dynamic adjustment unit (301), a laser repair unit (302), and a robotic arm rejection unit (303); Feedback control module: Through the industrial Internet of Things platform (401), defect data and compensation results are fed back to the production line PLC to update the process parameter library in real time.

2. The real-time defect detection system for RFID tag production lines based on machine vision according to claim 1, wherein The imaging parameters of the multi-spectral imaging unit (101) satisfy: The surface imaging bands are 450nm, 630nm, and 850nm, which are used to detect printing layer defects; The internal imaging band is 1050nm, and the penetration depth is ≥200μm, which is used to detect antenna breakage and chip misalignment; The imaging frame rate is synchronized with the production line speed, and satisfies the relationship: Where, F is the imaging frame rate (fps), v is the production line transmission speed (mm / s), d is the minimum defect size of the tag (mm), and η = 1.2 - 1.5 is a redundancy coefficient.

3. The real-time defect detection system for RFID tag production lines based on machine vision according to claim 1, characterized in that, The loss function of the lightweight dual-branch deep learning network (201) is defined as: L total = αL loc + βL cls + γL reg Among them, L loc is the localization loss (using CIoU loss), and L cls is the classification loss (focal loss), and L reg is the model complexity regularization term, and the weight coefficients satisfy α:β:γ = 3:2:0.

5.

4. A real-time defect detection system for RFID tag production lines based on machine vision as claimed in claim 1, characterized in that, The trigger condition of the hierarchical compensation strategy is: Primary compensation: When the proportion of the defect area S defect / S tag ≤ 5%, start the laser repair unit (302), and the repair path planning satisfies: Secondary compensation: When 5% < S defect / S tag ≤ 15%, adjust the printing parameters (temperature T ± 3°C, pressure P ± 5%); Three - level compensation: S defect / S tag When it is > 15%, the manipulator rejection unit (303) removes the defective label within t ≤ 200 ms; Among them, S defect is the defect area, and S tag is the label area.

5. The real-time defect detection system for RFID tag production lines based on machine vision according to claim 1, wherein The network structure optimization of the improved YOLOv5 model includes: Replace the C3 module in the backbone network with a Ghost module, reducing the number of parameters by 40%; Add a coordinate attention mechanism (CA), and the feature map weight calculation is: Among them, F c is the channel feature, and δ(·) is the Sigmoid function.

6. The real-time defect detection system for RFID tag production line based on machine vision according to claim 1, characterized in that, The power adjustment formula of the laser repair unit (302) is: Among them, A defect is the defect area, A0 = 0.1 mm 2 , k = 12 W / mm 2 , b = 5 W.

7. The real-time defect detection system for RFID tag production lines based on machine vision according to claim 1, wherein, The feedback control module (401) dynamically optimizes the production line parameters through the PID algorithm: Among them, e(t) is the defect rate deviation, and the proportionality coefficient K p = 0.8, the integral coefficient K i = 0.2, the differential coefficient K d = 0.

1.

8. A method for compensating defects in the RFID tag production line based on machine vision, which is implemented based on a real-time defect detection system for the RFID tag production line based on machine vision as described in any one of claims 1 to 7, and is characterized in that, Including the following steps: Step S1: The multi-spectral imaging unit (101) synchronously acquires the surface and internal images of the tag; Step S2: The dual-branch deep learning network (201) outputs the defect type and location. If the confidence conf≥0.9, it is determined as a defect; Step S3: Select a compensation method according to the defect level. After execution, re-inspect through an industrial camera. If there are still defects, trigger secondary compensation or rejection.

9. A method for compensating defects in an RFID tag production line based on machine vision according to claim 8, characterized in that, In step S2, the confidence calculation uses an improved Softmax function: Among them, ρ = 0.5 is the temperature coefficient, which is used to enhance the discrimination of classification confidence, and z i represents the original output value of the i-th category (i.e., the unnormalized logits), and the smoothness of the confidence distribution is adjusted by the temperature coefficient ρ.

10. A method for compensating defects in an RFID tag production line based on machine vision according to claim 8, characterized in that, The response time t of the compensation instruction in step S3 response Satisfies: Among them, L tag is the label length, v line is the production line speed, and t process ≤ 50 ms is the data processing time delay.

Images