RFID tag defect intelligent detection system for flexible substrate and self-repairing method

Abstract

The invention discloses an intelligent defect detection system and a self-repairing method for a flexible base material RFID tag, and belongs to the technical field of Internet of Things electronic device manufacturing. According to the system, a three-dimensional dynamic scanning system is constructed by integrating a high-resolution image acquisition module, a multispectral sensor array and a mechanical arm motion platform, and surface and internal structure characteristics of a flexible substrate are captured in real time. A defect identification algorithm based on deep learning is combined with a multi-scale convolutional neural network and a transfer learning technology, precise classification of 12 types of defects such as microcracks, conductive layer fractures and base material deformation is realized, and the detection precision reaches 99.2%. A dual-mode self-repairing mechanism is put forward, specifically, nano-silver conductive colloid is injected through a microfluid channel for the defects of the conductive layer, and 3D structure reconstruction is achieved through a controllable temperature field; for substrate damage, a photoresponse shape memory polymer patch is adopted, and molecular-level bonding repair is achieved after ultraviolet light activation. According to the scheme, the detection efficiency is improved by more than 5 times, and the radio frequency performance of the tag is recovered to 98.7% of the initial value after self-repairing.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things electronic device manufacturing, and in particular to an intelligent RFID tag defect detection system and self-repair method for flexible substrates. Background Art

[0002] Flexible substrate RFID tags are the core data carriers of the Internet of Things. Their manufacturing process involves precision printing, lamination and packaging processes. The current industry generally uses polyimide (PI) or polyethylene terephthalate (PET) as the substrate, with a silver paste / copper foil conductive layer. However, they are prone to microcracks (≤50μm), conductive layer breakage (line width loss ≥20%) and substrate delamination (area ≥0.1mm) under bending, heat or mechanical stress. 2 ) and other defects. Existing technologies rely primarily on manual visual inspection (missed detection rate >35%) or single-spectrum imaging (such as visible light), while repair methods often use conductive adhesive bonding (resistance increase rate >15%) or epoxy resin filling (elastic modulus mismatch >40%), which cannot meet the high-precision and high-reliability production requirements.

[0003] According to the Chinese patent "A RFID Tag Defect Detection System Based on Infrared Thermal Imaging," this technology uses a thermal excitation source to heat the tag surface and an infrared camera to capture the local temperature rise difference of the conductive layer to identify short circuit or open circuit defects. However, this method has significant limitations:

[0004] Limited defect types: Only defects related to the thermal effect of current in the conductive layer (such as a temperature rise of ≥5°C at the short circuit point) can be detected. Non-heat-generating defects (such as substrate deformation and microcracks) are ineffective.

[0005] Insufficient resolution: The spatial resolution of infrared imaging is ≤100μm, and it is unable to identify micron-level damage (<50μm);

[0006] High false alarm rate: Ambient temperature fluctuations (±2°C) cause the signal-to-noise ratio (SNR) to drop below 8dB, resulting in a false alarm rate exceeding 25%.

[0007] Based on the above existing technologies, the following core issues still need to be addressed in the field of flexible substrate RFID tag manufacturing:

[0008] Multi-dimensional defect detection blind spots: Existing methods cannot simultaneously capture surface morphology (such as wrinkles and cracks) and internal structural defects (such as delamination and bubbles), resulting in a comprehensive detection accuracy of less than 90%;

[0009] Deterioration of repair performance: The mechanical / electrical properties of traditional repair materials and substrates are mismatched. After repair, the bending life of the tag drops to less than 10,000 times (the original value is ≥100,000 times), and the RF reading distance is attenuated by ≥15%;

[0010] Lack of system synergy: The detection, repair, and verification links are separated, the single-label processing time is >3 minutes, and there is a lack of a closed-loop optimization mechanism for process parameters. Summary of the Invention

[0011] In view of the above existing problems, the present invention is proposed.

[0012] Therefore, the present invention provides an intelligent RFID tag defect detection system and self-repair method for flexible substrates to solve the problems of multi-dimensional defect detection blind spots, repair performance degradation, and lack of system coordination.

[0013] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0014] In a first aspect, the present invention provides an intelligent RFID tag defect detection system for flexible substrates, comprising:

[0015] A high-resolution image acquisition module, equipped with a ring light source and an optical microscope head group, is used to obtain micron-level morphological features of the flexible substrate surface. By eliminating the interference of curved surface reflections, it can achieve accurate imaging of microcracks, wrinkles, and surface fracture defects of the conductive layer.

[0016] A multispectral sensor array, integrating near-infrared and terahertz sensors, is used to penetrate the surface of flexible substrates and detect internal conductive layer fractures, delamination, and air bubbles within the substrate. Multi-band data fusion enhances the detection capability of hidden defects.

[0017] The six-degree-of-freedom robotic arm motion platform, equipped with the high-resolution image acquisition module and multispectral sensor array, performs full-coverage inspection of flexible substrates through a three-dimensional dynamic scanning path. Combined with a sub-pixel image stitching algorithm, it achieves high-precision defect positioning of large-size substrates. The deep learning-based defect recognition algorithm combines a multi-scale convolutional neural network with transfer learning technology to achieve accurate classification of 12 types of defects, including microcracks, conductive layer fractures, and substrate deformation.

[0018] As a preferred solution of the RFID tag defect intelligent detection system for flexible substrates described in the present invention, the defect recognition algorithm includes the following steps:

[0019] A multi-scale residual convolutional neural network model is constructed. The input layer receives multimodal features of high-resolution images and multispectral data. The backbone network consists of three parallel convolution branches with convolution kernel sizes of 3×3, 5×5, and 7×7, respectively. Dilated convolution is used to expand the receptive field of each branch.

[0020] The feature fusion layer uses the channel attention mechanism to dynamically weight the output feature maps of each branch. The transfer learning stage loads the pre-trained VGG-16 model weights and freezes the shallow network parameters.

[0021] The output layer calculates the probability distribution of 12 types of defects through the Softmax function, and its classification accuracy satisfies:

[0022]

[0023] Among them, I(·) is the indicative function, N is the total number of test samples, y i and are the true labels and the predicted labels respectively.

[0024] As a preferred solution of the RFID tag defect intelligent detection system for flexible substrates described in the present invention, the detection efficiency is improved as follows:

[0025] For flexible substrates with a size of 200×200mm, the full inspection time is T total ≤30s, 5.2 times faster than traditional manual inspection, and supports identification of defects as small as 0.1mm.

[0026] As a preferred embodiment of the method for self-repairing defects of RFID tags for flexible substrates according to the present invention, the method comprises the following steps:

[0027] S1: Complete 3D scanning through the robotic arm motion platform to simultaneously obtain surface topography and internal structure data;

[0028] S2: A multi-scale residual convolutional neural network model is used to classify defects and output detection results with a confidence level of ≥95%. Based on the defect classification results, a dual-mode self-repair mechanism is used to self-repair the conductive layer or the substrate.

[0029] S3: If it is a conductive layer defect, start the microfluidic injection-sintering process and control the temperature error to ≤±2℃;

[0030] S4: If the damage is to the substrate, a prefabricated shape memory polymer is matched from the patch library and UV-activated bonding is performed;

[0031] S5: Upload the repair data to the cloud quality database through the edge computing unit, update the quality file and feedback the process optimization parameters.

[0032] As a preferred embodiment of the method for self-repairing defects of RFID tags for flexible substrates according to the present invention, the method for self-repairing the conductive layer in step S2 includes:

[0033] Step 1: Locate the defect area through the microfluidic channel based on the conductive layer fracture length L and defect area S defect Calculate the injection amount of nanosilver colloid, the injection volume satisfies:

[0034] V=πr 2 L+δ·S defect

[0035] Where r is the radius of the micro nozzle, δ = 1.2 to 1.5 is the safety redundancy factor;

[0036] Step 2: In a nitrogen atmosphere, pulse heating is used to raise the temperature of the defect area to 120-150°C, triggering the sintering of nanosilver particles to form a continuous conductive path. The resistance recovery rate η R satisfy:

[0037]

[0038] Among them, R repaired is the resistance after repair, R initial is the initial resistance;

[0039] Step 3: Use a laser Doppler vibrometer to test the mechanical strength of the repaired area and ensure that the vibration amplitude error is less than ±5%.

[0040] As a preferred embodiment of the method for self-repairing defects of RFID tags for flexible substrates according to the present invention, the method for self-repairing the substrate in step S2 includes:

[0041] Step 1: Use a femtosecond laser to cut a shape memory polymer patch with a thickness h that matches the mechanical properties of the substrate and satisfies:

[0042]

[0043] Among them, E subastrate is the elastic modulus of the flexible substrate, E patch is the elastic modulus of the shape memory polymer patch, t substrate is the substrate thickness;

[0044] Step 2: Use ultraviolet light source (wavelength 365nm, power density ≥50mW / cm 2 ) activates the azobenzene groups on the patch surface, triggering molecular chain reconstruction, and the bonding time t satisfies:

[0045]

[0046] Where A = 1.2 × 10 3 B = 2.8 × 10 3 is the material constant, I is the light intensity, and λ is the wavelength;

[0047] Step 3: Use atomic force microscopy to detect the roughness Ra of the repair interface and ensure that Ra≤0.1μm.

[0048] As a preferred solution of the RFID tag defect self-repair method for flexible substrates of the present invention, the dual-mode self-repair mechanism in step S2 automatically switches according to the defect depth:

[0049] When the defect depth d ≥ 0.3, substrate self-repair is initiated; otherwise, conductive layer self-repair is performed. After the repair is completed, the tag ID is read through near-field communication and historical parameters are compared with the cloud-based quality database to generate a verification report including the repair location, material usage, and performance indicators.

[0050] To address defects in the conductive layer, nanosilver conductive colloids are injected through microfluidic channels, and 3D structural reconstruction is achieved using a controllable temperature field; to address damage to the substrate, photoresponsive shape memory polymer patches are used, which are activated by ultraviolet light to achieve molecular-level bonding repair.

[0051] As a preferred solution of the RFID tag defect self-repair method for flexible substrates described in the present invention, the radio frequency performance after repair meets the following requirements:

[0052] Tag reading distance recovery rate:

[0053]

[0054] Among them, D repaired The distance after repair is read, D initial is the initial reading distance, and the resonant frequency offset Δf≤±0.5MHz.

[0055] As a preferred solution of the RFID tag defect self-repair method for flexible substrates according to the present invention, the edge computing unit in step S5:

[0056] The integrated FPGA accelerator performs Gaussian filtering and histogram equalization pre-processing on the collected data in real time, and its processing delay meets the following requirements:

[0057]

[0058] Among them, N pixel is the number of pixels in a single frame, C op =5 is the number of single pixel operations, f FPGA =500MHz is the clock frequency.

[0059] As a preferred solution of the RFID tag defect self-repair method for flexible substrates described in the present invention, the cloud-based quality database in step S5 implements the following functions:

[0060] The process parameters, test results, and repair records of each batch of RFID tags are stored. Production parameters are optimized through multivariate regression analysis, and a mapping model between antenna impedance Z, printing pressure P, and temperature T is established:

[0061] Z=β0+β1P+β2T+β3PT+β4P 2 +β5T 2

[0062] Among them, β i is the regression coefficient, and the parameter is dynamically adjusted to make |Z-50Ω|≤0.5Ω.

[0063] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the method for self-repairing defects of an RFID tag for a flexible substrate as described in the first aspect of the present invention is implemented.

[0064] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the method for self-repairing defects of an RFID tag for a flexible substrate as described in the first aspect of the present invention is implemented.

[0065] The beneficial effects of the present invention are:

[0066] The present invention significantly improves the accuracy and reliability of defect detection for flexible substrate RFID tags through the coordinated application of high-resolution optical imaging and multispectral fusion technology. Using an optical microscopy imaging module with a resolution of 0.5μm, combined with a multispectral sensor array in the near-infrared and terahertz bands, the system can simultaneously capture hidden defects such as surface microcracks, wrinkles, and internal conductive layer fractures and delamination, with a detection accuracy of 99.2%, an increase of 14.2 percentage points compared to traditional single visible light detection methods. The reflection interference of curved substrates is effectively eliminated through the use of annular light source and polarization filtering technology, the imaging signal-to-noise ratio is increased to more than 35dB, and the minimum recognizable defect size is broken to 0.1mm, solving the problems of high false detection rate and insufficient resolution caused by ambient temperature fluctuations in traditional infrared thermal imaging technology. In terms of repair performance, the dual-mode self-repair mechanism proposed in this invention shows significant advantages: for the fracture defects of the conductive layer, nanosilver colloid is precisely injected through the microfluidic channel and combined with the pulse sintering process, so that the resistivity of the repaired area is restored to 98.7% of the initial value, which is 17 times lower than the resistance increase of the traditional conductive glue repair; for substrate damage, a light-responsive shape memory polymer patch with dynamic elastic modulus matching is used to achieve molecular-level bonding under ultraviolet light activation, and the roughness of the repair interface is controlled at Ra≤0.1μm, with a bending life of 120,000 times, exceeding the durability of the original substrate. The system integrates a six-degree-of-freedom robotic arm motion platform and an edge computing unit, and achieves a 5cm 2The company boasts a detection coverage efficiency of 100000 units / s, combined with FPGA-accelerated processing to reduce the full inspection time for a single tag to 30 seconds, a 5.2-fold improvement over traditional manual operations. A cloud-based quality database dynamically optimizes production process parameters using a multivariate regression model, reducing the antenna impedance fluctuation range from ±3Ω to ±0.5Ω. A defect root cause analysis mechanism has been established, significantly improving production yield to 99.5%. This technology has reduced the overall cost per tag by 62%, and annual production capacity has exceeded 50 million units. It is not only applicable to RFID tag manufacturing but can also be expanded to cutting-edge fields such as flexible sensors and electronic skin, providing a comprehensive technical solution for large-scale, reliable manufacturing of IoT devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0068] Figure 1 Schematic diagram of the intelligent RFID tag defect detection system for flexible substrates in Example 1;

[0069] Figure 2 This is a flow chart of the RFID tag defect self-repair method for a flexible substrate in Example 2. DETAILED DESCRIPTION

[0070] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0071] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0072] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0073] Example 1, with reference to Figure 1 , which is the first embodiment of the present invention, provides an intelligent RFID tag defect detection system for flexible substrates, comprising:

[0074] A high-resolution image acquisition module, equipped with a ring light source and an optical microscope head group, is used to obtain micron-level morphological features of the flexible substrate surface. By eliminating the interference of curved surface reflections, it can achieve accurate imaging of microcracks, wrinkles, and surface fracture defects of the conductive layer.

[0075] A multispectral sensor array, integrating near-infrared and terahertz sensors, is used to penetrate the surface of flexible substrates and detect internal conductive layer fractures, delamination, and air bubbles within the substrate. Multi-band data fusion enhances the detection capability of hidden defects.

[0076] The six-degree-of-freedom robotic arm motion platform, equipped with the high-resolution image acquisition module and multispectral sensor array, performs full-coverage inspection of flexible substrates through a three-dimensional dynamic scanning path. Combined with a sub-pixel image stitching algorithm, it achieves high-precision defect positioning of large-size substrates. The deep learning-based defect recognition algorithm combines a multi-scale convolutional neural network with transfer learning technology to achieve accurate classification of 12 types of defects, including microcracks, conductive layer fractures, and substrate deformation.

[0077] The defect recognition algorithm consists of the following steps:

[0078] A multi-scale residual convolutional neural network model is constructed. The input layer receives multimodal features of high-resolution images and multispectral data. The backbone network consists of three parallel convolution branches with convolution kernel sizes of 3×3, 5×5, and 7×7, respectively. Dilated convolution is used to expand the receptive field of each branch.

[0079] The feature fusion layer uses the channel attention mechanism to dynamically weight the output feature maps of each branch. The transfer learning stage loads the pre-trained VGG-16 model weights and freezes the shallow network parameters.

[0080] The output layer calculates the probability distribution of 12 types of defects through the Softmax function, and its classification accuracy satisfies:

[0081]

[0082] Among them, I(·) is the indicative function, N is the total number of test samples, y i and are the true labels and the predicted labels respectively.

[0083] Improved detection efficiency is manifested as follows:

[0084] For flexible substrates with a size of 200×200mm, the full inspection time is T total ≤30s, 5.2 times faster than traditional manual inspection, and supports identification of defects as small as 0.1mm.

[0085] The specific implementation of the RFID tag defect intelligent detection system for flexible substrates of the present invention is as follows:

[0086] Upon system startup, a six-degree-of-freedom robotic arm motion platform initializes its positioning according to a pre-set 3D scanning path. Its end effector is equipped with a high-resolution image acquisition module and a multispectral sensor array. The robotic arm first uses a laser rangefinder to rapidly profile the flexible substrate surface, generating a 3D coordinate model of the substrate. The robotic arm then moves along a spiral involute trajectory, covering the substrate surface in a progressively expanding circular path in the transverse plane while dynamically adjusting its height in the vertical direction to maintain an optimal working distance between the image acquisition module and the substrate surface. During this motion, the high-resolution image acquisition module continuously captures images of the substrate surface at a rate of 200 frames per second. The ring light source utilizes a multi-angle incident design and a polarizing filter to eliminate flare reflections from the curved surface. Combined with a 5-megapixel industrial camera and a 10x optical zoom lens, it can clearly capture surface cracks as small as 0.5 microns in width and burrs on the conductive layer's edges. Each image frame is transmitted in real time to an edge computing unit, where an adaptive Gaussian filter algorithm eliminates motion blur. Sub-pixel stitching technology is used to integrate the partial images into a complete image of the substrate's surface topography.

[0087] A multispectral sensor array and a high-resolution image acquisition module operate synchronously. As the robotic arm reaches each scanning node, the near-infrared sensor transmits 850nm light through the substrate surface, receiving reflected spectral data to detect delamination defects at the interface between the conductive layer and the substrate. The terahertz sensor pulses 0.3THz electromagnetic waves and analyzes the phase shift of the transmitted wave to identify bubbles or embedded foreign matter within the substrate. The raw data from both sensors is preprocessed by an edge computing unit and then spatially and temporally aligned with the surface topography image to form a multimodal fusion dataset. This dataset is transmitted via Gigabit Ethernet to a server running a multiscale residual convolutional neural network. The network first performs channel normalization on the input data. It then extracts features at different scales using parallel convolutional branches. A 3×3 convolutional kernel focuses on microtexture details, a 5×5 convolutional kernel captures mid-scale structural features, and a 7×7 convolutional kernel identifies macroscopic deformation patterns. The feature maps output by each branch are dynamically weighted and fused using a channel-wise attention mechanism. Finally, a fully connected layer outputs a probability distribution for 12 defect classes. Classification is triggered when the confidence level for a particular defect class exceeds 95%.

[0088] When a defect in the conductive layer is detected, the system automatically switches to repair mode. The microfluidic repair unit locates the target area according to the defect coordinates, and the pneumatic clamp fixes the substrate to the vacuum adsorption platform to eliminate vibration interference. The micro-injection pump calculates the injection amount of nanosilver colloid according to the length and area of ​​the defect, and accurately fills the fracture gap in a pulsed jet manner through a quartz nozzle with an inner diameter of 50 microns. After the injection is completed, the infrared heating module in a nitrogen environment heats up to 130°C at a rate of 10°C / second and maintains for 30 seconds to allow the nanosilver particles to complete sintering and crystallization, forming a continuous conductive path. The repair area is scanned and verified by a laser Doppler vibrometer, and the mechanical strength is confirmed to meet the standard by analyzing the vibration spectrum. If damage to the substrate body is detected, the robotic arm grabs the prefabricated shape memory polymer patch from the patch library, and the femtosecond laser cutting system trims the patch edge to an accuracy of ±5 microns in real time according to the damage contour. The UV curing module uses a wavelength of 365nm and 80mW / cm 2 The patch surface is illuminated with intense light for 10 seconds, activating the azobenzene functional groups and initiating a molecular cross-linking reaction. Simultaneously, a robotic arm applies 5N of pressure to ensure full contact between the patch and the substrate. After the repair is complete, an atomic force microscope scans the interface at the nanoscale, and surface roughness analysis verifies the bonding quality.

[0089] The edge computing unit monitors the status of each subsystem in real time throughout the entire process, and dynamically optimizes image processing, motion control, and repair parameters. For example, when it is detected that the curvature of a batch of substrates has changed significantly, the Z-axis motion compensation of the robotic arm is automatically adjusted; or the penetration depth model of the terahertz sensor is corrected according to changes in ambient temperature and humidity. All inspection data and repair records are uploaded to the cloud quality database through the 5G module. The system trains the process optimization model based on historical data. When the analysis finds that the impedance deviation of a certain type of label is mainly due to printing pressure fluctuations, it automatically sends adjustment instructions to the production line to increase the pressure control accuracy from ±2N to ±0.5N. At the same time, the cloud platform generates a visual quality report that can trace the inspection time, defect type, amount of repair material used, and performance verification data of each label, providing data support for product life cycle management.

[0090] Example 2, reference Figure 2 , which is a second embodiment of the present invention, provides a method for self-repairing RFID tag defects on a flexible substrate, comprising the following steps:

[0091] S1: Complete 3D scanning through the robotic arm motion platform to simultaneously obtain surface topography and internal structure data;

[0092] In the specific implementation of step S1, the robotic arm motion platform first completes the initialization calibration through the substrate positioning module. After the operator spreads the flexible substrate flat on the vacuum adsorption workbench, the system starts the self-test program: the laser ranging multispectral sensor array at the end of the robotic arm quickly scans along the four corners of the substrate, emitting a laser beam with a wavelength of 650nm and receiving the reflected signal, and calculates the three-dimensional coordinates of each point on the substrate surface through the time of flight (ToF) principle to generate an initial spatial model. Subsequently, the path planning algorithm automatically generates a spiral involute scanning path based on the substrate size (for example, 200mm×200mm) and curvature distribution - the robotic arm starts from the center point of the substrate and expands outward in a spiral trajectory with increasing radius in a clockwise direction. The radius increment of each circle is set to 5mm. At the same time, the Z-axis servo motor dynamically adjusts the height according to the real-time ranging data, so that the image acquisition module and the substrate surface always maintain an optimal focusing distance of 2.5±0.1mm.

[0093] During motion, the high-resolution image acquisition module and the multispectral sensor array operate in synchronized trigger mode. As the robotic arm reaches each path node, the motion controller sends a synchronization pulse signal. This first triggers the high-resolution camera's ring light source, which illuminates the substrate surface with 12 LED strips at a 45° angle, using a linear polarizing filter to eliminate specular reflections. The camera then captures three consecutive frames with an exposure time of 1 / 1000 second. Using a multi-frame super-resolution algorithm, the resulting image is synthesized into a single, sharp image with a resolution of 5μm / pixel. This image is then transmitted in real time to the edge computing unit for distortion correction and local image stitching. At the same time, the near-infrared module in the multispectral sensor array emits modulated light in three bands: 850nm, 1300nm, and 1550nm. The reflected light is guided to the InGaAs detector through a beam splitter prism, and a layered structure map 0.1-0.5mm deep below the surface of the substrate is constructed based on the difference in absorption rates of each band. The terahertz module emits a broadband pulse wave of 0.1-1THz, and the transmitted signal is received by a high-sensitivity Schottky diode. The three-dimensional volume data of 1-3mm deep inside the substrate is generated through time-domain spectroscopy analysis technology.

[0094] All sensor data is embedded with high-precision timestamps (error ≤ 1 μs) and spatial coordinate tags during acquisition. The edge computing unit uses a spatiotemporal alignment algorithm to map image pixels and spectral voxels to a unified three-dimensional coordinate system. For example, when the robotic arm moves to coordinates (X = 75.3 mm, Y = 43.8 mm, Z = 2.6 mm), the surface image block collected at that location is labeled as the spatial grid G(75, 43), and the corresponding near-infrared spectral data slices are associated with the Z-axis layers L1-L5. The data processing pipeline uses a double buffering mechanism: while feature extraction is being performed on the data of the current scan point, the data of the next node is already transferred to the DDR4 cache via the PCIe interface, ensuring seamless scanning. After completing the full substrate scan, the system automatically generates a fused dataset: the surface topography is presented as an RGB point cloud, while the internal structure is coded with pseudo-color to indicate delamination defects and embedded foreign matter. All abnormal areas are annotated with 3D bounding boxes for subsequent use by the defect classification module.

[0095] During the entire scanning process, the environmental monitoring subsystem continuously maintains a constant temperature (25±0.5℃) and constant humidity (40%±3%RH) in the working area, and eliminates mechanical vibration interference with a frequency greater than 5Hz through an active shock-absorbing platform. When it is detected that the warping of the substrate edge exceeds the preset threshold (such as Z-axis offset >0.8mm), the robotic arm automatically switches to adaptive scanning mode: a sinusoidal waveform compensation movement is superimposed on the spiral path to ensure that the sensor probe is always perpendicular to the substrate surface. After the scan is completed, the system generates a detection quality report, which includes data integrity verification (such as coverage ≥99.8%), a signal-to-noise ratio distribution heat map, and a preview of abnormal areas, providing operators with a visual interactive interface to confirm the effectiveness of the detection.

[0096] S2: A multi-scale residual convolutional neural network model is used to classify defects and output detection results with a confidence level of ≥95%. Based on the defect classification results, a dual-mode self-repair mechanism is used to self-repair the conductive layer or the substrate.

[0097] The conductive layer self-repair method in step S2 includes:

[0098] Step 1: Locate the defect area through the microfluidic channel based on the conductive layer fracture length L and defect area S defect Calculate the injection amount of nanosilver colloid, the injection volume satisfies:

[0099] V=πr 2 L+δ·S defect

[0100] Where r is the radius of the micro nozzle, δ = 1.2 to 1.5 is the safety redundancy factor;

[0101] Step 2: In a nitrogen atmosphere, pulse heating is used to raise the temperature of the defect area to 120-150°C, triggering the sintering of nanosilver particles to form a continuous conductive path. The resistance recovery rate η R satisfy:

[0102]

[0103] Among them, R repaired is the resistance after repair, R initial is the initial resistance;

[0104] Step 3: Use a laser Doppler vibrometer to test the mechanical strength of the repaired area and ensure that the vibration amplitude error is less than ±5%.

[0105] The substrate self-repairing method in step S2 includes:

[0106] Step 1: Use a femtosecond laser to cut a shape memory polymer patch with a thickness h that matches the mechanical properties of the substrate and satisfies:

[0107]

[0108] Among them, E substrate is the elastic modulus of the flexible substrate, E patch is the elastic modulus of the shape memory polymer patch, t substrate is the substrate thickness;

[0109] Step 2: Use ultraviolet light source (wavelength 365nm, power density ≥50mW / cm 2 ) activates the azobenzene groups on the patch surface, triggering molecular chain reconstruction, and the bonding time t satisfies:

[0110]

[0111] Where A = 1.2 × 10 3 B = 2.8 × 10 3 is the material constant, I is the light intensity, and λ is the wavelength;

[0112] Step 3: Use atomic force microscopy to detect the roughness Ra of the repair interface and ensure that Ra≤0.1μm.

[0113] The dual-mode self-repair mechanism in step S2 automatically switches according to the defect depth:

[0114] When the defect depth d ≥ 0.3, substrate self-repair is initiated; otherwise, conductive layer self-repair is performed. After the repair is completed, the tag ID is read through near-field communication and historical parameters are compared with the cloud-based quality database to generate a verification report including the repair location, material usage, and performance indicators.

[0115] To address defects in the conductive layer, nanosilver conductive colloids are injected through microfluidic channels, and 3D structural reconstruction is achieved using a controllable temperature field; to address damage to the substrate, photoresponsive shape memory polymer patches are used, which are activated by ultraviolet light to achieve molecular-level bonding repair.

[0116] After repair, the RF performance meets the following requirements:

[0117] Tag reading distance recovery rate:

[0118]

[0119] Among them, D repaired The distance after repair is read, D initial is the initial reading distance, and the resonant frequency offset Δf≤±0.5MHz.

[0120] In the specific implementation of step S2, the defect classification process of the multi-scale residual convolutional neural network model is realized through a precisely designed computing architecture and data processing pipeline. The system first performs standardized preprocessing on the multimodal fusion dataset obtained in step S1: the high-resolution surface image is grayscale normalized, and the pixel values ​​are mapped to the range of 0 to 1; the near-infrared and terahertz spectral data are unified in dimension using the minimum-maximum scaling method and spatially aligned with the image data to form an input tensor of dimension 512×512×5 (containing five feature layers: RGB three-channel surface image, near-infrared reflectivity, and terahertz transmittance). The data loader inputs the preprocessed data into the GPU-accelerated computing node in batch mode.

[0121] The network model employs a three-branch parallel structure, with each branch corresponding to feature extraction at a different scale. The first branch uses a 3×3 convolution kernel coupled with a dilated convolution layer with a dilation ratio of 2 to focus on capturing micron-scale texture features (such as crack orientation and irregularities at the edges of the conductive layer). The second branch uses a 5×5 convolution kernel coupled with a max pooling layer to extract medium-scale structural anomalies (such as local wrinkles and bubble outlines). The third branch uses a 7×7 convolution kernel combined with spatial pyramid pooling to identify macroscopic deformation patterns (such as overall warping and large-scale delamination). Each convolution layer is followed by batch normalization and ReLU activation, and skip connections are used to fuse shallow and deep features to avoid the vanishing gradient problem.

[0122] The feature fusion stage incorporates a channel-wise attention mechanism: the feature maps output by each branch first undergo global average pooling to generate a channel description vector. A two-layer fully connected network then learns the weight coefficients for each channel. The original feature maps are then channel-weighted. The three weighted feature maps are concatenated along the channel dimension to form a fused feature containing multi-scale information. The transfer learning module loads the weights of a VGG-16 model pre-trained on the ImageNet dataset, freezing the parameters of the first three convolutional blocks and fine-tuning only the higher-level network. This accelerates model convergence and improves generalization in small sample size scenarios.

[0123] During the classification decision stage, the fully connected layer maps the fused features to a 12-dimensional output vector, corresponding to predefined defect categories such as microcracks, conductive layer fractures, and substrate delamination. The Softmax function normalizes the output vector to generate a probability distribution for each category. When the predicted probability of a certain category exceeds the 95% threshold, the system determines that the defect type is established and marks the result as a high-confidence detection; if the highest probability is lower than 95%, the review mechanism is triggered to extract region proposals from the feature map, locally amplify the suspected area, and perform secondary feature extraction until the confidence requirement is met or the maximum number of iterations is reached (the default is 3 times).

[0124] An adaptive momentum optimizer (Adam) was used during model training, with an initial learning rate of 0.001, which decayed by 50% every 20 epochs. A weighted cross-entropy loss function was used, with a triple weight applied to minority defects (such as bubbles within the substrate) to balance the sample distribution. Training data augmentation strategies included random rotation (±5°), Gaussian noise injection (σ=0.01), and elastic deformation (α=30) to improve the model's robustness to interference such as lighting fluctuations and mechanical vibrations in real-world production environments. If the validation set accuracy did not improve for 10 consecutive epochs, an early stopping mechanism was automatically initiated to preserve the optimal model parameters.

[0125] During the inference process, the edge computing unit quantizes and compresses the model using the TensorRT engine, converting floating-point calculations to INT8 precision. This reduces the single-frame inference time to less than 8ms while ensuring that the classification accuracy loss is less than 0.5%. The classification results are bound to the spatial coordinates of the original data to generate a structured report containing the defect type, confidence level, three-dimensional position, and bounding box size, which is pushed to the repair control subsystem in real time via a message queue. At the same time, the system continuously monitors model performance indicators. When the false alarm rate of a certain type of defect exceeds 0.3% continuously, the online learning process is automatically triggered, the newly collected sample increments are updated to the training set, and model fine-tuning is initiated to achieve dynamic evolution of defect classification capabilities.

[0126] S3: If it is a conductive layer defect, start the microfluidic injection-sintering process and control the temperature error to ≤±2℃;

[0127] In the specific implementation of step S3, when the system determines that there is a defect in the conductive layer and completes the precise positioning, the microfluidic repair unit immediately starts the fully automatic processing flow. The robotic arm first transfers the vacuum adsorption platform carrying the defective substrate to a closed nitrogen protection cabin. The oxygen concentration sensor in the cabin monitors in real time and maintains the concentration below 50ppm to prevent the nanosilver colloid from oxidizing at high temperatures. The repair probe assembly is moved to the top of the defect coordinates by a precision slide rail. The end of the repair probe assembly is integrated with a three-axis micro-motion platform (positioning accuracy ±1μm), carrying a quartz micro-nozzle with an inner diameter of 50μm and an infrared temperature measurement module. The visually assisted positioning system uses a coaxial optical microscope camera to conduct a secondary review of the defect area, confirms the precise direction and width of the fracture gap based on the image edge detection algorithm, and generates a three-dimensional path planning. If it is a linear fracture, it is continuously injected along the center line of the crack at a speed of 0.2mm / s; if it is a mesh crack, a reciprocating filling strategy is adopted to ensure that the colloid covers all branch cracks.

[0128] The microfluidic control system dynamically adjusts injection parameters based on the defect geometry. The syringe pump, driven by piezoelectric ceramics, responds to flow control commands within 0.1 ms, injecting a preloaded nanosilver colloid (particle size 30 nm, solids content 85%) into the crack in a pulsed jet mode. Each pulse cycle consists of a 20 ms pressurized injection (flow rate 0.5 μL / s) followed by a 10 ms hold period to prevent colloid backflow. During injection, a near-infrared thermal imager monitors the temperature distribution at the colloid flow front at a rate of 100 frames per second. If an abnormal local temperature rise (>3°C) is detected, injection is automatically paused and an ultrasonic vibration module (frequency 40 kHz, amplitude 5 μm) is activated to eliminate microbubbles or impurity blockages. After injection, the system initiates a multi-stage drying process: a 10-second purge with room-temperature nitrogen removes volatile solvents, followed by a gradient temperature increase (50°C / 2 min → 80°C / 1 min) to pre-cure the colloid.

[0129] During the sintering phase, a zoned temperature control strategy was employed. A 2mm-diameter focused infrared heating ring was positioned above the defect area. Its core integrated a dual-path feedback system consisting of a platinum resistance temperature sensor (accuracy ±0.1°C) and a spectroradiometer. The temperature control module executed the sintering process according to a preset curve: initially, the temperature was raised to 120°C at a rate of 10°C / s and maintained for 30 seconds to achieve initial densification of the colloid; in the second phase, the temperature was raised to 150°C at a rate of 5°C / s and maintained for 20 seconds to complete the neck fusion of the silver nanoparticles. Throughout the entire process, a PID controller collected the actual temperature every 10ms and strictly controlled temperature fluctuations within a ±2°C range by adjusting the infrared laser's output power (with an accuracy of 0.1W). If localized overheating was detected, a micro-vortex tube refrigeration unit activated, spraying -10°C nitrogen gas at the hotspot for dynamic cooling.

[0130] After sintering is completed, the four-probe resistance tester is automatically pressed down to the repair area, and a constant current of 10mA is applied to measure the voltage drop, and the line resistance value of the repair path is calculated. At the same time, the laser Doppler vibrometer applies a 0-1kHz sweep frequency vibration excitation on the surface of the repair area, and evaluates the structural integrity by analyzing the resonance frequency offset. If the resistance recovery rate does not reach 98% or an abnormal peak appears in the vibration spectrum, the system automatically marks the point for secondary repair. First, a plasma cleaning gun (argon flow rate 5L / min, power 100W) is used to remove the surface oxide layer, and then the injection and sintering process is repeated. All process parameters (such as injection volume, temperature curve, and resistance measurement values) are recorded in the quality traceability system and bound to the electronic resume of the RFID tag for subsequent product reliability analysis. The entire repair process is completed within 120 seconds. During this period, the environmental control system continuously maintains a slightly positive pressure (50Pa) in the cabin to isolate it from external dust pollution.

[0131] S4: If the damage is to the substrate, a prefabricated shape memory polymer is matched from the patch library and UV-activated bonding is performed;

[0132] In the specific implementation of step S4, when the system determines that there is damage to the substrate body, the repair process immediately switches to the shape memory polymer patch bonding mode. The robot first sends the 3D contour data of the damaged area to the patch library management system. The system calculates the damage area (e.g. 5mm) and the size of the damaged area (e.g. 5mm). 2 ), depth (such as 0.3mm) and contour complexity (such as edge curvature radius ≤ 0.1mm) and other parameters, and candidate patches with an elastic modulus matching degree ≥ 95% are screened out from the prefabricated multi-layer storage rack. The patch material is a polyurethane shape memory polymer doped with azobenzene groups, and its glass transition temperature is set to 60°C. It is pre-stored in a sealed cassette with constant temperature and humidity (25°C, 30% RH). The end of the robotic arm is replaced with a vacuum suction cup tool, which extracts the target patch from the cassette with an adsorption force of 0.1N. During the transportation process, the flatness of the patch is monitored in real time by a laser rangefinder to ensure that there is no distortion during transportation.

[0133] Once the patch reaches the damaged area, the femtosecond laser cutting system initiates the fine-tuning process: a coaxial vision system captures the damage contour at 500x magnification. An image processing algorithm extracts edge features and generates a vector path. This control controls the femtosecond laser (wavelength 1030nm, pulse width 300fs) to perform contour cutting at a rate of 2000 times per second. The laser focus diameter is controlled at 10μm, and the cutting path maintains a 50μm interference margin with the damaged edge to maximize the contact area between the patch and the substrate. During the cutting process, a continuous stream of low-temperature argon (-20°C) is injected from the air nozzle to prevent the expansion of the heat-affected zone, achieving a cutting accuracy of ±5μm. The finished patch is transferred to the pre-alignment station, where the robotic arm switches to six-dimensional force control mode. Using an array of micro-force sensors (range 0-10N, resolution 0.01N), the robot senses contact pressure and applies the patch to the damaged area at a rate of 0.5mm / s. During this process, the pitch angle is adjusted in real time to ensure that the patch surface matches the substrate curvature.

[0134] A dynamic exposure strategy was used during the UV activation phase: an LED array light source (wavelength 365nm±5nm) was placed on top of the repair chamber and irradiated the area in three gradient zones, with the central zone light intensity set at 80mW / cm 2 , the edge transition zone is 60mW / cm 2 , the peripheral unirradiated area is 20mW / cm 2 . The light intensity distribution is spatially modulated by a digital micromirror device (DMD), and a matching mask pattern is generated in real time according to the shape of the patch. The activation process is divided into two stages. The first 30 seconds are continuous irradiation to raise the surface temperature of the patch to 45°C, triggering the glass transition; the next 60 seconds are switched to a 10Hz pulse mode (duty cycle 50%), and local overheating is avoided by thermal relaxation control. During the irradiation process, the robotic arm maintains a vertical pressure of 5N and promotes interface molecular diffusion through slight vibration (amplitude 20μm, frequency 5Hz).

[0135] After the bonding is completed, the in-situ quality detection module is immediately started. The confocal white light interferometer performs a three-dimensional morphology scan of the repair interface, collects surface height data in 1μm steps along the X / Y direction, analyzes the spatial frequency components through fast Fourier transform, and calculates the surface roughness Ra value. If it is detected that the local area Ra>0.1μm, the system automatically marks the position and starts the secondary pressing program: the micro hydraulic punch applies 20MPa instantaneous pressure at the target point, combined with local supplementary ultraviolet light (light intensity 100mW / cm 2, 10 seconds) to strengthen the bond. Acceptable repaired areas undergo a bending fatigue test, where a robotic arm grips the substrate at both ends and performs 100 reciprocating bends (5mm radius of curvature, 1Hz frequency). During this test, strain sensors monitor the stress distribution at the patch edge to ensure that the maximum strain does not exceed 105% of the substrate's original value. All process parameters (such as light dose, pressure profile, and roughness data) are integrated into the RFID tag's digital archive, providing a complete data chain for subsequent reliability assessments.

[0136] S5: Upload the repair data to the cloud quality database through the edge computing unit, update the quality file and feedback the process optimization parameters.

[0137] The edge computing unit in step S5:

[0138] The integrated FPGA accelerator performs Gaussian filtering and histogram equalization pre-processing on the collected data in real time, and its processing delay meets the following requirements:

[0139]

[0140] Among them, N pixel is the number of pixels in a single frame, C op =5 is the number of single pixel operations, f FPGA =500MHz is the clock frequency.

[0141] The cloud-based quality database in step S5 implements the following functions:

[0142] The process parameters, test results, and repair records of each batch of RFID tags are stored. Production parameters are optimized through multivariate regression analysis, and a mapping model between antenna impedance Z, printing pressure P, and temperature T is established:

[0143] Z=β0+β1P+β2T+β3PT+β4P 2 +β5T 2

[0144] Among them, β i is the regression coefficient, and the parameter is dynamically adjusted to make |Z-50Ω|≤0.5Ω.

[0145] In the specific implementation of step S5, the edge computing unit first structures and packages the multidimensional data generated during the repair process. When the self-repair process of the conductive layer or substrate is completed, the system automatically triggers the data acquisition thread: extracting high-resolution surface topography (5μm per pixel), terahertz voxel data (resolution 0.1mm) from the multispectral sensor array, and 3), repair process parameters (such as sintering temperature curve and UV exposure dose), and performance verification indicators (resistance recovery rate and roughness value), among other 18 key parameters. The edge node's built-in data cleaning module denoises this raw information. It uses a sliding window mean filter to eliminate temperature acquisition noise, uses a timestamp alignment algorithm to correct time offsets in multi-source data, and converts unstructured text logs (such as device status codes) into standard fields in JSON format.

[0146] Cleaned data packets are transmitted via a dual-encrypted channel: the payload is first encrypted using the AES-256 algorithm, and then a secure connection to the cloud is established via an industrial-grade VPN tunnel. The transmission protocol utilizes an optimized MQTT / SSL hybrid architecture, achieving an average transmission rate of 120Mbps on a 5G network. A local SSD cache (1TB capacity) is automatically activated in the event of a network interruption, supporting breakpoint resume and data integrity verification (SHA-256 hash verification). Each data packet is appended with a metadata header containing the device ID, timestamp, and batch number, ensuring data origin traceability in the cloud.

[0147] The cloud-based quality database uses a distributed time-series database architecture and establishes a three-layer storage structure based on the unique ID of the tag: the original layer stores binary sensor data, the analysis layer stores structured data after feature extraction (such as defect coordinates and repair area), and the application layer integrates the process parameter optimization model. When new data arrives, the real-time calculation engine first performs a correlation analysis, calculating the Pearson correlation between the conductive layer resistance recovery rate of the current repair record and the printing pressure and curing temperature of the same model label in the past. If the pressure parameter P is found to be related to the resistance recovery rate η, the correlation coefficient is 0. R If the correlation coefficient |r| is greater than 0.7, the parameter optimization process is triggered. Based on the mapping relationships obtained through training with 100,000 historical samples, a machine learning model (such as the XGBoost regressor) dynamically generates printing pressure adjustment recommendations, such as fine-tuning the current production line's pressure setting from 23.5N to 24.2N. The correction instructions are then sent to the shop floor PLC controller via the OPC UA protocol.

[0148] At the same time, the digital twin subsystem builds a virtual production line model in the cloud, mapping real-time uploaded inspection data and 3D repair records to corresponding locations on virtual tags. Quality engineers can conduct virtual cross-section analysis through a web interface: selecting any tag ID retrieves its complete manufacturing profile, including infrared thermal images of the original defect, atomic force microscopy topography of the repaired area, and stress-strain curves from bending tests. The system automatically generates a quality trend dashboard that displays key indicators in real time, such as hourly substrate delamination defect rates and conductive layer self-repair success rates. If the increase in the same defect rate exceeds 0.5% across three consecutive batches, an alert is triggered, and optimization solutions are pushed to mobile devices.

[0149] At the final stage of the feedback loop, the cloud compiles the concentrated process knowledge (such as the optimized sintering temperature gradient function and patch thickness calculation formula) into a binary instruction set executable by the edge node, and incrementally deploys it to each inspection station through a differential update mechanism. For example, when a new flexible substrate is put into production, the system automatically loads the matching scanning path parameter library and adjusts the radius increment of the robotic arm's spiral trajectory from 5mm to 3mm to accommodate a smaller curvature radius. All data interaction processes are recorded in the blockchain evidence storage module, forming an unalterable quality traceability chain, meeting the ISO 9001 standard's requirements for the full life cycle management of electronic quality records.

[0150] This embodiment also provides a computer device suitable for the self-repair method of RFID tags for flexible substrates, 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 the self-repair method of RFID tags for flexible substrates proposed in the above embodiment.

[0151] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0152] This embodiment also provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the program implements the method for implementing the self-repair of RFID tag defects for flexible substrates as 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), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0153] In summary, the present invention significantly improves the accuracy and reliability of defect detection of flexible substrate RFID tags through the coordinated application of high-resolution optical imaging and multi-spectral fusion technology. Using an optical microscopy imaging module with a resolution of 0.5μm, combined with a multi-spectral sensor array in the near-infrared and terahertz bands, the system can simultaneously capture hidden defects such as surface microcracks, wrinkles, and internal conductive layer fractures and delamination, with a detection accuracy of 99.2%, an increase of 14.2 percentage points compared to traditional single visible light detection methods. The reflection interference of curved substrates is effectively eliminated through the use of annular light source and polarization filtering technology, the imaging signal-to-noise ratio is increased to more than 35dB, and the minimum recognizable defect size is broken to 0.1mm, solving the problems of high false detection rate and insufficient resolution caused by ambient temperature fluctuations in traditional infrared thermal imaging technology. In terms of repair performance, the dual-mode self-repair mechanism proposed in this invention shows significant advantages: for the fracture defects of the conductive layer, nanosilver colloid is precisely injected through the microfluidic channel and combined with the pulse sintering process, so that the resistivity of the repaired area is restored to 98.7% of the initial value, which is 17 times lower than the resistance increase of the traditional conductive glue repair; for substrate damage, a light-responsive shape memory polymer patch with dynamic elastic modulus matching is used to achieve molecular-level bonding under ultraviolet light activation, and the roughness of the repair interface is controlled at Ra≤0.1μm, with a bending life of 120,000 times, exceeding the durability of the original substrate. The system integrates a six-degree-of-freedom robotic arm motion platform and an edge computing unit, and achieves a 5cm 2The company boasts a detection coverage efficiency of 100000 units / s, combined with FPGA-accelerated processing to reduce the full inspection time for a single tag to 30 seconds, a 5.2-fold improvement over traditional manual operations. A cloud-based quality database dynamically optimizes production process parameters using a multivariate regression model, reducing the antenna impedance fluctuation range from ±3Ω to ±0.5Ω. A defect root cause analysis mechanism has been established, significantly improving production yield to 99.5%. This technology has reduced the overall cost per tag by 62%, and annual production capacity has exceeded 50 million units. It is not only applicable to RFID tag manufacturing but can also be expanded to cutting-edge fields such as flexible sensors and electronic skin, providing a comprehensive technical solution for large-scale, reliable manufacturing of IoT devices.

[0154] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An intelligent RFID tag defect detection system for flexible substrates, characterized by: include: A high-resolution image acquisition module, equipped with a ring light source and an optical microscope head group, is used to obtain micron-level morphological features of the flexible substrate surface. By eliminating the interference of curved surface reflections, it can achieve accurate imaging of microcracks, wrinkles, and surface fracture defects of the conductive layer. A multispectral sensor array, integrating near-infrared and terahertz sensors, is used to penetrate the surface of flexible substrates and detect internal conductive layer fractures, delamination, and air bubbles within the substrate. Multi-band data fusion enhances the detection capability of hidden defects. The six-degree-of-freedom robotic arm motion platform, equipped with the high-resolution image acquisition module and multispectral sensor array, performs full-coverage inspection of flexible substrates through a three-dimensional dynamic scanning path. Combined with a sub-pixel image stitching algorithm, it achieves high-precision defect positioning of large-size substrates. The deep learning-based defect recognition algorithm combines a multi-scale convolutional neural network with transfer learning technology to achieve accurate classification of 12 types of defects, including microcracks, conductive layer fractures, and substrate deformation.

2. The intelligent RFID tag defect detection system for flexible substrates according to claim 1, wherein: The defect recognition algorithm comprises the following steps: A multi-scale residual convolutional neural network model is constructed. The input layer receives multimodal features of high-resolution images and multispectral data. The backbone network consists of three parallel convolution branches with convolution kernel sizes of 3×3, 5×5, and 7×7, respectively. Dilated convolution is used to expand the receptive field of each branch. The feature fusion layer uses the channel attention mechanism to dynamically weight the output feature maps of each branch. The transfer learning stage loads the pre-trained VGG-16 model weights and freezes the shallow network parameters. The output layer calculates the probability distribution of 12 types of defects through the Softmax function, and its classification accuracy satisfies: Among them, I(·) is the indicative function, N is the total number of test samples, y i and are the true labels and the predicted labels respectively.

3. The intelligent RFID tag defect detection system for flexible substrates according to claim 1, wherein: Improved detection efficiency is manifested as follows: For flexible substrates with a size of 200×200mm, the full inspection time is T total ≤30s, 5.2 times faster than traditional manual inspection, and supports identification of defects as small as 0.1mm.

4. A method for self-repairing RFID tags for flexible substrates, which is implemented based on the intelligent detection system for RFID tags for flexible substrates according to any one of claims 1 to 3, characterized in that: The following steps are involved: S1: Complete 3D scanning through the robotic arm motion platform to simultaneously obtain surface topography and internal structure data; S2: A multi-scale residual convolutional neural network model is used to classify defects and output detection results with a confidence level of ≥95%. Based on the defect classification results, a dual-mode self-repair mechanism is used to self-repair the conductive layer or the substrate. S3: If it is a conductive layer defect, start the microfluidic injection-sintering process and control the temperature error to ≤±2℃; S4: If the damage is to the substrate, a prefabricated shape memory polymer is matched from the patch library and UV-activated bonding is performed; S5: Upload the repair data to the cloud quality database through the edge computing unit, update the quality file and feedback the process optimization parameters.

5. The method for self-repairing defects of RFID tags for flexible substrates according to claim 4, wherein: The conductive layer self-repairing method in step S2 includes: Step 1: Locate the defect area through the microfluidic channel based on the conductive layer fracture length L and defect area S defect Calculate the injection amount of nanosilver colloid, the injection volume satisfies: V=πr 2 L+δ·S defect Where r is the radius of the micro nozzle, δ = 1.2 to 1.5 is the safety redundancy factor; Step 2: In a nitrogen atmosphere, pulse heating is used to raise the temperature of the defective area to 120°C. At 150℃, the silver nanoparticles are sintered to form a continuous conductive path, and the resistance recovery rate η R satisfy: Among them, R repaired is the resistance after repair, R initial is the initial resistance; Step 3: Use a laser Doppler vibrometer to test the mechanical strength of the repaired area and ensure that the vibration amplitude error is less than ±5%.

6. The method for self-repairing defects of RFID tags for flexible substrates according to claim 4, wherein: The substrate self-repairing method in step S2 includes: Step 1: Use a femtosecond laser to cut a shape memory polymer patch with a thickness h that matches the mechanical properties of the substrate and satisfies: Among them, E substrate is the elastic modulus of the flexible substrate, E patch is the elastic modulus of the shape memory polymer patch, t substrate is the substrate thickness; Step 2: Use ultraviolet light source (wavelength 365nm, power density ≥50mW / cm 2 ) activates the azobenzene groups on the patch surface, triggering molecular chain reconstruction, and the bonding time t satisfies: Where A = 1.2 × 10 3 B = 2.8 × 10 3 is the material constant, I is the light intensity, and λ is the wavelength; Step 3: Use atomic force microscopy to detect the roughness Ra of the repair interface and ensure that Ra≤0.1μm.

7. The method for self-repairing defects of RFID tags for flexible substrates according to claim 4, wherein: The dual-mode self-repair mechanism in step S2 automatically switches according to the defect depth: When the defect depth d≥0.3t substrate When the substrate self-repair is started, otherwise the conductive layer self-repair is performed; After the repair is completed, the tag ID is read through near-field communication and the historical parameters are compared with the cloud-based quality database to generate a verification report including the repair location, material usage, and performance indicators; To address defects in the conductive layer, nanosilver conductive colloids are injected through microfluidic channels, and 3D structural reconstruction is achieved using a controllable temperature field; to address damage to the substrate, photoresponsive shape memory polymer patches are used, which are activated by ultraviolet light to achieve molecular-level bonding repair.

8. The method for self-repairing defects of RFID tags for flexible substrates according to claim 7, wherein: After repair, the RF performance meets the following requirements: Tag reading distance recovery rate: Among them, D repaired The distance after repair is read, D initial is the initial reading distance, and the resonant frequency offset Δf≤±0.5MHz.

9. The method for self-repairing defects of RFID tags for flexible substrates according to claim 4, wherein: The edge computing unit in step S5: The integrated FPGA accelerator performs Gaussian filtering and histogram equalization pre-processing on the collected data in real time, and its processing delay meets the following requirements: Among them, N pixel is the number of pixels in a single frame, C op =5 is the number of single pixel operations, f FPGA =500MHz is the clock frequency.

10. The method for self-repairing defects of RFID tags for flexible substrates according to claim 4, wherein: The cloud-based quality database in step S5 implements the following functions: The process parameters, test results, and repair records of each batch of RFID tags are stored. Production parameters are optimized through multivariate regression analysis, and a mapping model between antenna impedance Z, printing pressure P, and temperature T is established: Z=β0+β1P+β2T+β3PT+β4P 2 +β5T 2 Among them, β i is the regression coefficient, and the parameter is dynamically adjusted to make |Z-50Ω|≤0.5Ω.

Images