Online detection system for life paper post-processing based on AI vision

Abstract

This invention belongs to the field of paper manufacturing inspection technology, and relates to an online inspection system for post-processing of household paper based on AI vision. It includes an environment-deformation perception module, an imaging calibration module, a speed acquisition module, a threshold calculation module, a defect statistics module, a sensitivity calculation module, an AI vision inspection core module, and a production linkage module, with each module establishing a communication connection. The environment-deformation perception module collects temperature, humidity, and paper contour data and outputs normalized deviation and deformation. The imaging calibration module outputs imaging calibration coefficients based on multi-parameter quantization calculations. The threshold calculation module and the sensitivity calculation module, respectively, output dynamic detection thresholds and process detection sensitivity coefficients based on pre-set parameters. This invention improves the parameter matching and process coordination of post-processing inspection of household paper.

Description

Technical Field

[0001] This invention belongs to the field of paper manufacturing inspection technology, specifically relating to an AI vision-based online inspection system applied to the post-processing stage of household paper. Background Technology

[0002] AI-powered online visual inspection in the post-processing of tissue paper is a crucial step in ensuring finished product quality. Current AI visual inspection systems for tissue paper post-processing only collect dynamic paper deformation data for correction during the imaging calibration stage. They fail to consider the non-linear impact of environmental temperature and humidity changes on industrial camera imaging, nor do they perform quantified calculations to coordinate temperature and humidity deviations with paper deformation. They rely solely on fixed calibration coefficients to correct the paper image. Because tissue paper is soft and easily deformed, temperature and humidity fluctuations in the production workshop directly cause pixel shifts in industrial camera imaging. Single deformation calibration cannot compensate for the imaging distortion caused by temperature and humidity, leading to biases in feature extraction during subsequent defect identification and frequent false positives and false negatives. Furthermore, existing inspection systems use fixed values ​​for defect detection thresholds and process detection sensitivity, failing to dynamically adjust based on imaging calibration results, production line speed fluctuations, and upstream process defects. This disconnect between detection parameters and actual production conditions makes them unsuitable for the full-process inspection requirements of tissue paper post-processing.

[0003] Based on the above problems, there is an urgent need for an AI vision online detection system that can achieve multi-parameter quantification of imaging calibration and dynamic linkage between detection threshold and sensitivity, so as to solve the problem of insufficient detection accuracy caused by single imaging calibration in the existing technology. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an online inspection system for post-processing of household paper based on AI vision. This system includes an environment-deformation sensing module, an imaging calibration module, a speed acquisition module, a threshold calculation module, a defect statistics module, a sensitivity calculation module, an AI vision inspection core module, and a production linkage module. Communication connections are established between these modules. The environment-deformation sensing module collects actual temperature and humidity data of the production environment and dynamic paper contour data, and calculates and outputs temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation. The imaging calibration module receives the temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation output by the environment-deformation sensing module, performs multi-parameter quantification calculations, and outputs imaging calibration coefficients. The speed acquisition module collects instantaneous speed data of the production line and calculates and outputs the speed deviation rate. The threshold calculation module... The calculation module receives the imaging calibration coefficients output by the imaging calibration module and the speed deviation rate output by the speed acquisition module, performs dual-parameter quantization calculations, and outputs the dynamic detection threshold. The defect statistics module collects defect detection data from upstream processes and calculates and outputs the probability of defect occurrence. The sensitivity calculation module receives the dynamic detection threshold output by the threshold calculation module and the probability of defect occurrence output by the defect statistics module, performs dual-parameter quantization calculations, and outputs the process detection sensitivity coefficient. The AI ​​visual inspection core module receives all data output by the imaging calibration module, threshold calculation module, and sensitivity calculation module, performs defect recognition operations on the paper images of each process in the post-processing of tissue paper, and outputs the defect recognition results. The production linkage module receives the defect recognition results output by the AI ​​visual inspection core module and outputs equipment parameter adjustment instructions to the tissue paper post-processing equipment.

[0005] Preferably, the environment-deformation sensing module includes a temperature and humidity sensor, a high-speed profile camera, and a first data processing unit. The temperature and humidity sensor and the high-speed profile camera are both electrically connected to the first data processing unit. The temperature and humidity sensor is used to collect actual temperature data and actual humidity data of the production environment and transmit them to the first data processing unit. The high-speed profile camera is used to collect dynamic profile data of paper in each post-processing step of the tissue paper and transmit it to the first data processing unit. The first data processing unit is used to receive all the data transmitted by the temperature and humidity sensor and the high-speed profile camera, and calculate and output the temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation.

[0006] More preferably, the imaging calibration module includes an FPGA image processing unit and a first reference parameter storage unit, with the FPGA image processing unit and the first reference parameter storage unit electrically connected. The first reference parameter storage unit stores the imaging basic calibration coefficients, temperature influence weighting coefficients, and humidity influence weighting coefficients and transmits them to the FPGA image processing unit. The FPGA image processing unit receives the temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation output by the environment-deformation sensing module, receives all reference parameters transmitted by the first reference parameter storage unit, performs quantization calculations, and outputs the imaging calibration coefficients.

[0007] In a further preferred embodiment, the core AI visual inspection module includes a high-definition industrial camera group, a ring-shaped shadowless light source, a GPU inference unit, and a defect feature library storage unit. All three components are electrically connected to the GPU inference unit. The high-definition industrial camera group consists of a hyperspectral camera and a high-speed linear array camera. It is used to acquire raw images of paper from various post-processing stages of household paper and transmit them to the GPU inference unit. The ring-shaped shadowless light source provides illumination for the raw paper image acquisition by the high-definition industrial camera group. The defect feature library storage unit stores a multi-dimensional defect feature library and an improved YOLO neural network model and transmits them to the GPU inference unit. The GPU inference unit receives the imaging calibration coefficients output by the imaging calibration module, performs image calibration on the raw paper image, receives the dynamic detection threshold output by the threshold calculation module and the process detection sensitivity coefficient output by the sensitivity calculation module, calls the multi-dimensional defect feature library and the improved YOLO neural network model to perform defect recognition on the calibrated paper image, and outputs the defect recognition result.

[0008] More preferably, the quantization calculation performed by the FPGA image processing unit adopts the imaging calibration coefficient calculation formula. The imaging calibration coefficient calculation formula is constructed with imaging basic calibration coefficient, temperature influence weight coefficient, temperature normalization deviation, humidity influence weight coefficient, humidity normalization deviation, and paper dynamic deformation as core calculation parameters, and is specifically used to calculate the imaging calibration coefficient.

[0009] Further preferably, the threshold calculation module includes a second reference parameter storage unit and a first embedded computing chip, with the second reference parameter storage unit and the first embedded computing chip electrically connected; the second reference parameter storage unit is used to store the basic pixel threshold of the defect feature and transmit it to the first embedded computing chip; the first embedded computing chip is used to receive the imaging calibration coefficient output by the imaging calibration module and the velocity deviation rate output by the velocity acquisition module, receive the basic pixel threshold of the defect feature transmitted by the second reference parameter storage unit, perform quantization calculation, and output the dynamic detection threshold.

[0010] More preferably, the sensitivity calculation module includes a third reference parameter storage unit and a second embedded computing chip, with the third reference parameter storage unit and the second embedded computing chip electrically connected. The third reference parameter storage unit stores the basic detection sensitivity coefficient and the influence weight of upstream defects on downstream processes and transmits them to the second embedded computing chip. The second embedded computing chip receives the dynamic detection threshold output by the threshold calculation module and the defect occurrence probability output by the defect statistics module, receives the basic detection sensitivity coefficient and the influence weight of upstream defects on downstream processes transmitted by the third reference parameter storage unit, performs quantitative calculations, and outputs the process detection sensitivity coefficient.

[0011] In a further preferred embodiment, the production linkage module includes a PLC communication module, a pneumatic control component, and a second data processing unit. Both the PLC communication module and the pneumatic control component are electrically connected to the second data processing unit. The PLC communication module establishes a communication connection with the tissue paper post-processing equipment, transmits equipment parameter adjustment commands to the tissue paper post-processing equipment, collects process operation data of the tissue paper post-processing equipment, and feeds it back to the defect statistics module. The second data processing unit receives defect identification results output by the AI ​​visual inspection core module, generates equipment parameter adjustment commands, and transmits them to the PLC communication module and the pneumatic control component. The pneumatic control component receives the equipment parameter adjustment commands transmitted by the second data processing unit and executes the parameter adjustment operation of the tissue paper post-processing equipment.

[0012] A further preferred embodiment includes a defect intelligent marking module, which establishes communication connections with the AI ​​visual inspection core module, the imaging calibration module, and the threshold calculation module. The defect intelligent marking module includes a pneumatic defect marking device, a paper weight sensor, and a third data processing unit. Both the pneumatic defect marking device and the paper weight sensor are electrically connected to the third data processing unit. The paper weight sensor collects the actual weight data of the tissue paper and transmits it to the third data processing unit. The third data processing unit receives the imaging calibration coefficient output by the imaging calibration module, the dynamic detection threshold output by the threshold calculation module, and the actual weight data transmitted by the paper weight sensor. It then generates marking force parameters and marking position parameters and transmits them to the pneumatic defect marking device. The pneumatic defect marking device receives the marking force parameters and marking position parameters transmitted by the third data processing unit, receives the defect identification results output by the AI ​​visual inspection core module, and performs defect marking operations on the defective tissue paper.

[0013] A further preferred embodiment includes a defective product rejection module, which establishes a communication connection with the defective intelligent marking module. The defective product rejection module is located at the operating position of the post-processing and packaging process of tissue paper. The defective product rejection module includes a visual recognition unit, a mechanical execution unit, and a fourth data processing unit. The visual recognition unit and the mechanical execution unit are both electrically connected to the fourth data processing unit. The visual recognition unit is used to collect surface images of the tissue paper before the packaging process and transmit them to the fourth data processing unit, identify defective markings on the surface of the tissue paper, and transmit the marking recognition signal to the fourth data processing unit. The fourth data processing unit is used to receive the marking recognition signal transmitted by the visual recognition unit, generate a rejection execution command, and transmit it to the mechanical execution unit. The mechanical execution unit is used to receive the rejection execution command transmitted by the fourth data processing unit and perform a defective product rejection operation on the tissue paper with defective markings. The mechanical execution unit is a pneumatic execution structure.

[0014] The technical effects achieved by the above embodiments include:

[0015] The core inventive technology of this invention lies in the multi-parameter quantification calculation of the imaging calibration module. It incorporates temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation into the calculation dimension of imaging calibration, solving the imaging distortion problem caused by single imaging calibration in the prior art. Simultaneously, through threshold calculation and sensitivity calculation modules, it achieves step-by-step linkage quantification of detection parameters, allowing dynamic detection thresholds, process detection sensitivity coefficients, imaging calibration results, and production conditions to be matched in real time. This forms a data closed loop between modules, achieving dynamic adaptation of detection parameters. The AI ​​visual inspection core module performs defect identification based on the aforementioned quantified parameters, and the production linkage module adjusts production equipment according to the identification results, enabling deep collaboration between inspection and production. This fundamentally reduces false positives and false negatives caused by parameter misalignment, adapting to the full-process inspection needs of post-processing of household paper. Attached Figure Description

[0016] Figure 1 This is a connection diagram of the main modules of the AI ​​vision-based online inspection system for post-processing of household paper products in this application.

[0017] Figure 2 This is a diagram showing the internal structure and connection of the deformation sensing module in the environment of this application.

[0018] Figure 3 This is a diagram showing the internal structure and connection of the threshold and sensitivity calculation module in this application;

[0019] Figure 4 This is a diagram showing the internal structure and connection of the production linkage module in this application;

[0020] Figure 5 This is a diagram showing the internal linkage of the defect marking and non-conforming product rejection module in this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] Traditional AI visual inspection systems for post-processing of household paper rely solely on single paper morphology parameters for imaging calibration. The detection threshold and process detection sensitivity are fixed values, making it impossible to make adaptive adjustments based on the production environment, production line status, and upstream process conditions. This results in a severe disconnect between the detection parameters and actual production conditions, making it difficult to match the full-process inspection requirements.

[0023] Based on this, please refer to Figures 1-5 This embodiment provides an online inspection system for post-processing of household paper products based on AI vision, including: an environment-deformation sensing module, an imaging calibration module, a speed acquisition module, a threshold calculation module, a defect statistics module, a sensitivity calculation module, an AI vision inspection core module, and a production linkage module. These modules establish communication connections to achieve real-time data interaction and transmission. The environment-deformation sensing module collects actual temperature and humidity data from the production environment and dynamic paper contour data, and calculates and outputs temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation, providing basic quantitative data for imaging calibration. The imaging calibration module receives the three types of quantitative data output by the environment-deformation sensing module, performs multi-parameter quantification calculations, and outputs imaging calibration coefficients, providing a quantitative basis for image calibration. The speed acquisition module collects instantaneous speed data from the production line and calculates and outputs a speed deviation rate, reflecting the deviation between the actual speed and the reference speed. The threshold calculation module receives data from the imaging calibration module. The imaging calibration coefficient and the speed deviation rate of the speed acquisition module are used to calculate and output a dynamic detection threshold through dual-parameter quantization, which serves as the core pixel judgment standard for defect identification. The defect statistics module collects defect detection data from upstream processes and outputs the probability of defect occurrence through statistical calculation, reflecting the upstream production quality. The sensitivity calculation module receives the dynamic detection threshold from the threshold calculation module and the probability of defect occurrence from the defect statistics module, and outputs a process detection sensitivity coefficient through dual-parameter quantization, which is used to adjust the accuracy of defect identification. The AI ​​visual inspection core module receives all the data output by the imaging calibration module, the threshold calculation module, and the sensitivity calculation module, and performs defect identification operations on the paper images of each process based on the built-in recognition model and feature library, outputting the defect identification results. The production linkage module receives the defect identification results from the AI ​​visual inspection core module and outputs equipment parameter adjustment instructions to the tissue paper post-processing equipment, realizing the transformation of detection results into production control.

[0024] The system comprises an organic detection system consisting of eight core modules. These modules are interconnected to form an uninterrupted data transmission link. The environment-deformation perception module, acting as the data acquisition source, transforms raw acquired data into standardized quantitative data, providing a foundation for subsequent calculations. The imaging calibration module calculates imaging calibration coefficients based on multi-dimensional quantitative data, solving the imaging distortion problem caused by traditional single calibration. The speed acquisition module and the defect statistics module quantify data from the perspectives of production line status and upstream production quality, providing working condition data for dynamic calculation of detection parameters. The threshold calculation module and the sensitivity calculation module perform linked quantization based on data from the front-end modules, replacing traditional fixed parameters and allowing detection parameters to match imaging quality and production conditions in real time. The AI ​​vision inspection core module integrates all quantitative parameters to complete defect identification, ensuring that the identification operation aligns with actual inspection needs. The production linkage module transforms the identification results into production control commands, achieving seamless integration between inspection and production, forming a closed-loop system. During the operation of the entire system, the functions of each module are interconnected and progressively advanced. Data is transmitted level by level and quantified at each step. The output of each module serves as the calculation input for the subsequent modules. There are no independent operating modules. This system realizes the quantification and dynamic adaptation of parameters throughout the entire process, from data acquisition and imaging calibration to defect identification and production linkage. This allows the detection parameters to match the actual production conditions in real time, greatly improving the synergy between the detection system and each process of post-processing of tissue paper, making the detection results more accurate and the production control more timely.

[0025] Traditional environmental deformation data acquisition suffers from the problem of separation between data acquisition and quantification. Various acquisition devices work independently without a unified data processing terminal. The raw data formats are inconsistent and need to be manually converted before they can be used for subsequent calculations, which is inefficient and prone to data errors.

[0026] Based on this, the environment-deformation sensing module is equipped with a temperature and humidity sensor, a high-speed profile camera, and a first data processing unit. The temperature and humidity sensor and the high-speed profile camera are electrically connected to the first data processing unit to form an integrated acquisition and processing system. The temperature and humidity sensor collects the actual temperature and humidity data of the production environment and transmits the raw environmental data to the first data processing unit in real time. The high-speed profile camera collects the dynamic profile data of the paper in each post-processing step of the tissue paper and transmits the raw morphological data to the first data processing unit in real time. The first data processing unit receives all the raw data transmitted by the two types of devices, processes it through built-in quantization calculation logic, and outputs the temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation.

[0027] The temperature and humidity sensor and the high-speed contour camera cover the entire process of slitting, embossing, laminating, rewinding, and folding, ensuring that the collected data reflects the actual working conditions of each process and has no blind spots in data collection; the stable electrical connection between the two types of equipment and the first data processing unit enables real-time data transmission without delay, avoiding processing delays caused by data backlog.

[0028] The first data processing unit transforms raw data in different formats into standardized dimensionless quantized data, allowing various types of data to be directly used for calculations in subsequent modules without additional format conversion. This module achieves integrated acquisition and standardized quantization of environmental and paper morphology data, creating a seamless connection between data acquisition and quantization. This provides standardized basic data for subsequent imaging calibration, improves data acquisition and transmission efficiency, and avoids calculation errors caused by inconsistent data formats, resulting in more accurate calculations in subsequent modules.

[0029] Traditional imaging calibration processes lack dedicated reference parameter storage units. Reference parameters are stored in a scattered manner, resulting in low retrieval efficiency and difficulty in updating. Multi-parameter calculations lack dedicated processing terminals and share hardware resources with other operations, leading to low computational efficiency and susceptibility to interference.

[0030] Based on this, the imaging calibration module is equipped with an FPGA image processing unit and a first reference parameter storage unit. The FPGA image processing unit is electrically connected to the first reference parameter storage unit to realize dedicated storage and real-time retrieval of reference parameters. The first reference parameter storage unit serves as a dedicated storage carrier, specifically storing three types of core reference parameters: imaging basic calibration coefficients, temperature influence weighting coefficients, and humidity influence weighting coefficients. It can be flexibly updated according to production needs and transmits all reference parameters to the FPGA image processing unit in real time. The FPGA image processing unit serves as a dedicated computing core, specifically responsible for multi-parameter quantization calculations for imaging calibration. It receives the temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation output by the environment-deformation sensing module, and simultaneously receives all reference parameters from the first reference parameter storage unit. It uses both types of data as calculation inputs, quantizes them using a preset algorithm, and outputs the imaging calibration coefficients.

[0031] The separate setup of the first reference parameter storage unit allows for independent storage of reference parameters, separating them from other system data storage, thus improving retrieval efficiency. It also allows for flexible updates based on changes in the type of tissue paper and processing technology, ensuring the reference parameters always match production needs. The FPGA image processing unit is solely responsible for multi-parameter quantization calculations for imaging calibration, without handling other data processing operations. This avoids computational interference caused by hardware resource sharing. Simultaneously, the FPGA's high-speed parallel computing capabilities enable rapid multi-parameter calculations, guaranteeing efficient output of imaging calibration coefficients and ensuring that imaging calibration matches the production line's operating speed. This module achieves independent storage of imaging calibration reference parameters and dedicated quantization calculations for multiple parameters, making the imaging calibration calculation process more standardized and efficient. It avoids errors caused by reference parameter retrieval delays and computational interference, resulting in output imaging calibration coefficients that better reflect actual production conditions and providing accurate quantitative data for subsequent image calibration.

[0032] Traditional AI visual defect recognition uses a single imaging acquisition device, which makes it difficult to fully capture defect features. Without dedicated lighting equipment, uneven lighting can easily lead to imaging deviations. Furthermore, defect recognition models and feature libraries are stored in a scattered manner, resulting in low retrieval efficiency.

[0033] Based on this, the AI ​​visual inspection core module is equipped with a high-definition industrial camera group, a ring-shaped shadowless light source, a GPU inference unit, and a defect feature library storage unit. The high-definition industrial camera group, the ring-shaped shadowless light source, and the defect feature library storage unit are all electrically connected to the GPU inference unit, forming an integrated defect identification system encompassing acquisition, illumination, storage, and processing. The high-definition industrial camera group consists of a hyperspectral camera and a high-speed linear array camera, acquiring original images of paper from each process and transmitting the image data to the GPU inference unit in real time. The ring-shaped shadowless light source provides uniform and stable illumination for image acquisition, eliminating light shadows and reflections. The defect feature library storage unit is dedicated to storing a multi-dimensional defect feature library and an improved YOLO neural network model, and transmits the stored content to the GPU inference unit in real time. The GPU inference unit receives the imaging calibration coefficients from the imaging calibration module, performs calibration operations on the original paper image to correct imaging distortion, and simultaneously receives the dynamic detection threshold from the threshold calculation module and the process detection sensitivity coefficient from the sensitivity calculation module. It calls the multi-dimensional defect feature library and the improved YOLO neural network model, integrates the dynamic detection threshold and the process detection sensitivity coefficient into the recognition process, performs defect recognition on the calibrated image, and outputs the defect recognition result.

[0034] The high-definition industrial camera group combines the advantages of two types of cameras: a hyperspectral camera captures the spectral characteristics of paper to identify hidden defects, and a high-speed linear array camera achieves high-definition imaging under high-speed motion, matching the production line speed. Together, they achieve multi-dimensional acquisition of defect features. The uniform illumination of the ring-shaped shadowless light source avoids imaging deviations, making the original image clearer. The dedicated settings of the defect feature library storage unit improve the retrieval efficiency of the model and feature library and facilitate updates and optimizations based on changes in product categories. The powerful parallel computing capabilities of the GPU inference unit can quickly complete image calibration and defect recognition, while integrating various quantization parameters into the recognition process, allowing the recognition operation to match the imaging quality and production conditions in real time. This module achieves multi-dimensional acquisition and precise calibration of paper images, enabling defect recognition to be performed based on standardized calibration images and dynamic detection parameters, effectively improving the accuracy of defect recognition and reducing false detections and missed detections.

[0035] Traditional imaging calibration quantitative calculations lack dedicated mathematical formulas, and multi-parameter calculations lack clear logical basis, relying solely on empirical values, resulting in poor scientific validity and repeatability of the calculation results.

[0036] Based on this, the quantization calculation of the FPGA image processing unit adopts the imaging calibration coefficient calculation formula. This formula is constructed based on six core parameters: imaging basic calibration coefficient, temperature influence weighting coefficient, temperature normalization deviation, humidity influence weighting coefficient, humidity normalization deviation, and paper dynamic deformation. It is specifically used to calculate the imaging calibration coefficient, providing clear mathematical logic support for imaging calibration. The construction of this formula relies entirely on the two types of input data of the imaging calibration module: the three types of reference parameters in the first reference parameter storage unit and the three types of quantization data in the environment-deformation sensing module. All six parameters are direct input parameters of the module, with no redundant parameters. The formula's unique output is the imaging calibration coefficient, achieving a precise correspondence between input and output.

[0037] In actual system operation, the formula for calculating the imaging calibration coefficient is expressed as follows:

[0038] ;

[0039] All parameters in this formula are dimensionless quantities, with a dimension of 1.

[0040] It is worth mentioning that the logical derivation of this formula is based on the characteristics of post-processing technology of tissue paper and the imaging principle of industrial cameras. Tissue paper is soft, and fluctuations in temperature and humidity in the production environment can cause slight deformation of the industrial camera lens, resulting in pixel shift. Moreover, experimental data shows that the effect of temperature and humidity deviation on imaging is non-linear. The larger the deviation, the greater the increase in the magnitude of pixel shift. Therefore, the formula uses the form of multiplying the temperature influence weighting coefficient by the square of the normalized temperature deviation and the humidity influence weighting coefficient by the square of the normalized humidity deviation to reflect the non-linear effect of temperature and humidity deviation. At the same time, the greater the dynamic deformation of the paper, the more blurred the surface texture features. The correction effect of imaging calibration decreases exponentially with the increase of deformation. Therefore, the formula adopts the form of a natural exponential function to reflect the decaying effect of the dynamic deformation of paper on imaging calibration.

[0041] This imaging calibration coefficient formula is a multi-parameter quantification calculation formula specifically designed for the AI ​​visual inspection imaging calibration stage of post-processing of tissue paper. Its core function is to accurately quantify the coupled influence of temperature and humidity deviations in the production environment and the dynamic deformation of the paper, transforming them into specific imaging calibration coefficients. This provides a scientific and accurate quantitative basis for the calibration of the original paper image, overcoming the shortcomings of traditional calibration relying on empirical values. The formula contains... The final output imaging calibration coefficient is the core quantitative indicator for subsequent image calibration. It is calculated by the FPGA image processing unit, is dimensionless, and has a value range of greater than 0 and less than or equal to 1. The value directly reflects the degree of correction of imaging calibration under the current production conditions. The larger the value, the better the imaging quality and the smaller the calibration correction range. The smaller the value, the higher the degree of imaging distortion and the larger the calibration correction range. The imaging baseline calibration coefficient is the imaging calibration benchmark value of tissue paper under the benchmark production conditions. It is pre-stored in the first benchmark parameter storage unit and is obtained by machine learning training from a large amount of historical detection data. It is dimensionless and its value is less than or equal to 1. It is the basis for the calculation of the entire formula and provides a benchmark reference for multi-parameter adjustment. It can be flexibly updated according to the type of tissue paper and changes in processing technology. The temperature influence weighting coefficient is a quantitative indicator of the degree of influence of temperature changes on industrial camera imaging. It is also pre-stored in the first reference parameter storage unit and is obtained by fitting a large amount of experimental data on temperature and humidity-imaging deviation. It is dimensionless and has a value less than 1. It is specifically used to measure the influence weight of temperature normalization deviation on imaging calibration. The larger the value, the greater the influence of temperature changes on imaging and the higher its proportion in the calculation. It can be updated and optimized according to the industrial camera model and service life. The temperature normalization deviation is the relative deviation between the actual production environment temperature and the reference environment temperature. It is calculated from the actual temperature data collected by the environment-deformation sensing module. It is dimensionless and ranges from -0.5 to 0.5. The sign of the value represents the relative temperature of the actual temperature to the reference temperature, and the absolute value represents the degree of deviation. The quadratic calculation reflects the nonlinear effect of temperature deviation on imaging, which is consistent with the relationship between deviation and imaging pixel offset obtained from experiments. The humidity influence weighting coefficient, designed in the same logic as the temperature influence weighting coefficient, is a quantitative indicator of the degree of influence of humidity changes on industrial camera imaging. It is pre-stored in the first reference parameter storage unit and is obtained by fitting temperature and humidity-imaging deviation experimental data. It is dimensionless and has a value less than 1. It is used to measure the influence weight of humidity normalization deviation on imaging calibration. The larger the value, the greater the influence of humidity changes on imaging. It can be flexibly adjusted according to the humidity control range of the production workshop. Humidity normalization deviation is the relative deviation between the actual production environment humidity and the reference environment humidity. It is calculated from the actual humidity data of the environment-deformation sensing module. It is dimensionless and has a value range of [-0.5, 0.5]. The positive or negative value represents the level of the actual humidity relative to the reference humidity, and the absolute value represents the degree of deviation. It is also calculated by quadratic power to reflect the nonlinear effect, which is consistent with the calculation form of temperature deviation, making the calculation logic of the formula more consistent. The paper's dynamic deformation is the ratio of its actual deformation to a reference dimension, representing the paper's strain. It is calculated from the contour data collected by the high-speed contour camera in the environment-deformation sensing module. It is dimensionless and ranges from 0 to 1, with the value representing the degree of paper deformation. This is determined using a natural exponential function. This reflects the effect of exponential decay, because the greater the paper deformation, the more blurred the surface texture, and the faster the calibration correction effect decays. After the deformation exceeds the threshold, the calibration effect approaches 0. This calculation method accurately matches the actual relationship between paper deformation and calibration effect.

[0042] The formula is implemented by embedding a dedicated algorithm program into the FPGA image processing unit, embedding the calculation logic in the hardware. During system operation, the FPGA image processing unit retrieves three types of reference parameters from the first reference parameter storage unit in real time, and simultaneously receives three types of quantized data from the environment-deformation sensing module. The six parameters are then calculated sequentially according to the formula's logic: first, the two weighting coefficients are multiplied by the square of the corresponding normalized deviation; then, the result is added to 1; subsequently, it is multiplied by the imaging baseline calibration coefficient; and finally, it is multiplied by the natural exponential function. The calculation results are multiplied to obtain the final imaging calibration coefficient, which is output in real time. The core innovation of this formula lies in breaking the limitation of traditional imaging calibration that only considers paper deformation. For the first time, it couples and quantifies temperature and humidity deviation with paper dynamic deformation, fully considering the nonlinear effects of temperature and humidity and the exponential decay effect of paper deformation. This allows the calculated imaging calibration coefficient to accurately reflect the actual imaging situation, rather than relying on empirical values ​​for estimation. At the same time, all parameters in the formula come from real-time system acquisition and preset benchmarks. The calculation process has a high degree of scientificity and repeatability, making imaging calibration operations more accurate and more in line with actual production needs.

[0043] Traditional detection threshold calculation lacks a dedicated reference parameter storage unit. The basic pixel threshold of defect features is stored together with other data, resulting in low retrieval efficiency. Furthermore, there is no dedicated computing chip to perform quantization calculations. Instead, it shares hardware resources with other operations, leading to low computational efficiency and susceptibility to interference.

[0044] Based on this, the threshold calculation module is equipped with a second reference parameter storage unit and a first embedded computing chip. The second reference parameter storage unit is electrically connected to the first embedded computing chip to realize dedicated storage and dedicated calculation of the reference threshold. The second reference parameter storage unit serves as a dedicated storage carrier, specifically storing the basic pixel threshold of defect features. It can be flexibly updated according to changes in the type of household paper and detection standards, and transmits the parameter to the first embedded computing chip in real time. The first embedded computing chip serves as a dedicated computing terminal, specifically responsible for the dual-parameter quantization calculation of the dynamic detection threshold. It receives the imaging calibration coefficient from the imaging calibration module and the speed deviation rate from the speed acquisition module, and simultaneously receives the basic pixel threshold of defect features from the second reference parameter storage unit. It uses the three types of parameters as calculation inputs, and outputs the dynamic detection threshold after quantization calculation by a preset algorithm.

[0045] The separate configuration of the second reference parameter storage unit allows for independent storage of the defect feature's basic pixel threshold, improving retrieval efficiency and facilitating flexible updates based on detection needs, ensuring the reference threshold always matches the detection standard. The first embedded computing chip is solely responsible for dynamic detection threshold quantization calculation, featuring small size, high calculation speed, and strong anti-interference capabilities. It can quickly complete dual-parameter calculations, guaranteeing the output efficiency of the dynamic detection threshold and ensuring the detection threshold matches the production line's operating status in real time. In actual system operation, the quantization calculation mathematical formula used by the first embedded computing chip is:

[0046] ;

[0047] All parameters are dimensionless, with a uniform dimension of 1, conforming to the principle of dimensional homogeneity and ensuring the scientific validity of the calculation logic. This module implements independent storage of the detection threshold benchmark parameters and dedicated quantization calculation of two parameters, allowing dynamic detection thresholds to match imaging quality and production speed in real time. It replaces the traditional fixed detection thresholds, making the pixel judgment standard for defect identification more closely aligned with actual production conditions.

[0048] This dynamic detection threshold formula is a dual-parameter quantification calculation formula designed for the defect detection threshold stage of AI vision inspection in the post-processing of tissue paper. Its core function is to accurately quantify the influence of imaging calibration coefficients and production line speed deviation rates, transforming them into dynamic detection thresholds. This replaces traditional fixed thresholds, allowing defect feature extraction standards to adapt to actual production conditions in real time, thus solving the problem of false detections and missed detections caused by the disconnect between fixed thresholds and actual operating conditions. The formula contains... The final output dynamic detection threshold is the core pixel standard for judging whether paper has defects in defect identification. It is calculated by the first embedded computing chip, is dimensionless and has a value range of greater than 0 and less than or equal to 1. The larger the value, the higher the defect feature extraction standard. Only when the proportion of defect pixels exceeds this value is it judged as a defect. The smaller the value, the lower the extraction standard, which can identify more subtle defects. The real-time change of this parameter allows the defect judgment standard to accurately match the imaging quality and production line speed. The baseline pixel threshold for defect features is the minimum pixel percentage threshold for defect features in tissue paper under the baseline detection conditions. It is pre-stored in the second baseline parameter storage unit and is obtained by machine learning training from a multi-dimensional defect feature library. It is dimensionless and its value is less than or equal to 1. It is the basis for the calculation of the formula and provides a baseline reference for dynamic detection threshold adjustment. It can be flexibly updated according to the type of tissue paper and the detection accuracy requirements to ensure that the baseline threshold meets the actual detection standards. The imaging calibration coefficient, output by the FPGA image processing unit of the imaging calibration module, is dimensionless and ranges from 0 to 1. It serves as the core adjustment parameter in the formula, and its value directly reflects the current imaging quality; the better the imaging quality, the higher the accuracy. The higher the value, the closer the dynamic detection threshold is to the basic pixel threshold, and the worse the image quality. The smaller the value, the lower the dynamic detection threshold, effectively avoiding missed defects caused by imaging distortion, and ensuring that the defect judgment standard matches the imaging quality. The speed deviation rate is the relative deviation between the actual instantaneous speed of the production line and the reference production speed. It is calculated from the instantaneous speed data of the speed acquisition module, is dimensionless, and ranges from -0.3 to 0.3. The positive or negative value represents how fast the actual speed is relative to the reference speed, and the absolute value represents the degree of deviation. The calculation method reflects the non-linear effect of speed deviation on the detection threshold. This derivation is based on the matching relationship between production line speed and image acquisition: when the speed deviation is positive, the paper moves too fast, the matching degree between the camera acquisition frame rate and the paper movement speed decreases, and the image is prone to ghosting. When the value is positive, the dynamic detection threshold is increased accordingly to avoid false detections caused by motion blur; when the speed deviation is negative, the paper moves too slowly, and the image features are clearer. When the value is negative, the dynamic detection threshold decreases accordingly, allowing for the identification of more subtle defects. The formula uses... Normalizing the impact of speed deviation allows the adjustment range of the dynamic detection threshold to match actual detection needs, avoiding excessive or insufficient adjustment.

[0049] The formula is implemented by embedding the corresponding algorithm program into the first embedded computing chip, solidifying the calculation logic. During system operation, the first embedded computing chip retrieves the defect feature base pixel threshold from the second reference parameter storage unit in real time, and simultaneously receives the imaging calibration coefficient from the imaging calibration module and the speed deviation rate from the speed acquisition module. The three types of parameters are then calculated sequentially according to the formula logic: first, the speed deviation rate is multiplied by its absolute value, then added to 1, and subsequently multiplied by the imaging calibration coefficient and the defect feature base pixel threshold in sequence, ultimately obtaining the dynamic detection threshold, which is then output to the AI ​​vision inspection core module in real time. The core innovation of this formula lies in integrating two key operating condition parameters—imaging quality and production line speed fluctuation—into the detection threshold calculation, allowing the detection threshold to be dynamically adjusted in real time based on the imaging calibration coefficient and the speed deviation rate. This changes the traditional fixed threshold mode and solves the problem of the fixed threshold being disconnected from the operating conditions; simultaneously, through… It reflects the nonlinear effect of speed deviation, making the detection threshold adjustment more in line with the actual operation of the production line, making the defect feature extraction standard more accurate, and matching the dynamic production needs of tissue paper post-processing.

[0050] Traditional process inspection sensitivity calculation lacks a dedicated benchmark parameter storage unit. Benchmark parameters such as basic inspection sensitivity coefficients and the influence weight of upstream defects on downstream processes are stored in a scattered manner, resulting in low retrieval efficiency and difficulty in updating. There is no dedicated computing chip to perform multi-parameter quantification calculations, making the calculation process susceptible to interference and the results not matching the actual process requirements.

[0051] Based on this, the sensitivity calculation module is equipped with a third reference parameter storage unit and a second embedded computing chip. The third reference parameter storage unit is electrically connected to the second embedded computing chip to realize dedicated storage and dedicated calculation of reference parameters. The third reference parameter storage unit serves as a dedicated storage carrier, specifically storing two types of core reference parameters: the basic detection sensitivity coefficient and the influence weight of upstream defects on downstream processes. It can be flexibly updated according to changes in processing technology and process connection relationships, and transmits all reference parameters to the second embedded computing chip in real time. The second embedded computing chip serves as a dedicated computing terminal, specifically responsible for the multi-parameter quantization calculation of the process detection sensitivity coefficient. It receives the dynamic detection threshold from the threshold calculation module and the defect occurrence probability from the defect statistics module, and simultaneously receives the two types of reference parameters from the third reference parameter storage unit. It uses the four types of parameters as calculation inputs, and outputs the process detection sensitivity coefficient after quantization calculation by a preset algorithm.

[0052] The separate configuration of the third reference parameter storage unit allows for independent storage of reference parameters, improving retrieval efficiency and enabling flexible updates based on changes in the production process, ensuring that the reference parameters match the actual connection relationships between processes. The second embedded computing chip is solely responsible for the quantization calculation of the process detection sensitivity coefficient, possessing high-speed and anti-interference computing characteristics. It can quickly complete multi-parameter calculations, ensuring coefficient output efficiency and allowing the detection sensitivity to match the detection threshold and upstream process conditions in real time. In actual system operation, the quantization calculation mathematical formula used by the second embedded computing chip is:

[0053] ;

[0054] All parameters are dimensionless quantities with a uniform dimension of 1, conforming to the principle of dimensional homogeneity and ensuring the scientific and rational nature of the calculation logic. This module realizes independent storage of sensitivity benchmark parameters and dedicated quantization calculation of multiple parameters, enabling real-time linkage between process detection sensitivity and dynamic detection thresholds and upstream defect conditions. This replaces the traditional fixed sensitivity setting, making defect identification accuracy more aligned with the actual production needs of each process.

[0055] The sensitivity coefficient formula for this process is a multi-parameter quantification formula designed for the AI ​​visual inspection process in the post-processing of tissue paper. Its core function is to accurately quantify the coupled influence of dynamic detection thresholds, the probability of defects occurring in upstream processes, and the weighting of the impact of upstream defects on downstream processes, transforming this into a process detection sensitivity coefficient. This allows the detection sensitivity of each process to adapt in real time to the actual detection thresholds and upstream production conditions, solving the problem of missed hidden defects caused by traditional fixed sensitivity and improving the comprehensiveness of defect identification. The formula contains... The final output process detection sensitivity coefficient is the core quantitative indicator for adjusting the confidence level of the defect identification model. It is calculated by the second embedded computing chip, is dimensionless, and has a value range of greater than 0 and less than or equal to 1. The larger the value, the higher the detection sensitivity and the lower the confidence threshold of the defect identification model, which can identify more subtle hidden defects. The smaller the value, the lower the detection sensitivity and the more obvious visible defects can be identified. The real-time change of this parameter allows the defect identification accuracy of each process to match the actual working conditions. The basic detection sensitivity coefficient is the benchmark value of the process detection sensitivity of tissue paper under the benchmark detection conditions. It is pre-stored in the third benchmark parameter storage unit and is obtained by machine learning training from process defect correlation data. It is dimensionless and its value is less than or equal to 1. It is the basis for the calculation of the formula and provides a benchmark reference for coefficient adjustment. It can be flexibly updated according to the type of tissue paper and the process detection accuracy requirements to ensure that the benchmark sensitivity meets the actual detection needs. The dynamic detection threshold is output by the first embedded computing chip of the threshold calculation module. It is dimensionless and its value ranges from 0 to 1. It serves as one of the core adjustment parameters of the formula. The value reflects the current defect feature extraction standard. The higher the extraction standard, the better. The higher the value, the higher the sensitivity coefficient of the process detection, avoiding missed detections due to excessively high extraction standards; the lower the extraction standard, the better. The smaller the value, the lower the sensitivity coefficient of the process detection, which reduces false detections caused by excessively low extraction standards and achieves real-time linkage between detection sensitivity and detection threshold. The probability of upstream process defects is the ratio of the number of defects occurring in the upstream process per unit time to the total number of inspections. It is calculated by the defect statistics module in real time by collecting upstream defect data. It is dimensionless and its value ranges from [0,1]. The value directly reflects the upstream production quality. The larger the value, the more upstream defects there are, and the higher the possibility of downstream derivative defects. The real-time update of this parameter allows the downstream detection sensitivity to match the upstream production quality. The weight of the upstream defect on the downstream is a quantitative indicator of the degree to which the upstream defect affects the downstream derivative defect. It is pre-stored in the third benchmark parameter storage unit and is obtained by training the process defect coupling model with a large amount of experimental data. It is dimensionless and the value is less than or equal to 1. The larger the value, the greater the impact of the upstream defect on the downstream and the higher the probability of the downstream generating derivative defects. It can be flexibly updated according to the processing technology and the tightness of the process connection to ensure accurate reflection of the actual defect coupling relationship between processes.

[0056] In the formula By coupling the probability of upstream defect occurrence with its impact weight, the actual impact of upstream defects on downstream processes can be accurately quantified. Normalize this effect so that the adjustment range of the process inspection sensitivity coefficient matches the actual inspection requirements: when the probability of upstream defects occurring in downstream processes is high and the impact weight is large, As the calculated value increases, the sensitivity coefficient of the process inspection also increases, and the defect identification model can accurately identify downstream derivative latent defects caused by upstream defects, avoiding omissions; when the probability of upstream defects occurring is low and their influence weight is small, The process inspection sensitivity coefficient is close to 1, meaning it is close to the product of the basic inspection sensitivity coefficient and the dynamic inspection threshold, maintaining conventional inspection accuracy and reducing unnecessary false detections. This formula is implemented by embedding the corresponding algorithm program into the second embedded computing chip, solidifying the calculation logic. During system operation, the second embedded computing chip retrieves two types of benchmark parameters from the third benchmark parameter storage unit in real time, and simultaneously receives the dynamic inspection threshold from the threshold calculation module and the upstream defect occurrence probability from the defect statistics module. The four types of parameters are then calculated sequentially according to the formula logic: first, the upstream defect occurrence probability is multiplied by its influence weight, then added to 1, and subsequently multiplied sequentially by the dynamic inspection threshold and the basic inspection sensitivity coefficient. Finally, the process inspection sensitivity coefficient is obtained and output to the AI ​​visual inspection core module in real time. The core innovation of this formula lies in its full consideration of the defect coupling characteristics between various processes in the post-processing of tissue paper. It integrates three key parameters—dynamic detection threshold, probability of upstream defect occurrence, and weight of the impact of upstream defects on downstream processes—into the calculation of detection sensitivity. This allows the detection sensitivity of each process to be dynamically adjusted in real time based on the detection threshold and upstream production conditions, changing the traditional fixed sensitivity model and solving the problem that fixed sensitivity cannot identify latent derivative defects caused by upstream defects. At the same time, it achieves the linkage adjustment of detection sensitivity and detection threshold, ensuring that the accuracy of defect identification fully matches the actual production conditions of each process, thereby improving the defect identification capability of the entire detection system.

[0057] Traditional production processes lack dedicated communication modules to interface with post-processing equipment for household paper, resulting in low data transmission efficiency and susceptibility to interference. There is no dedicated data processing unit to generate equipment parameter adjustment instructions, and equipment adjustments are made manually based on test results, leading to delayed and inaccurate control.

[0058] Based on this, the production linkage module is equipped with a PLC communication module, a pneumatic control component, and a second data processing unit. The PLC communication module and the pneumatic control component are electrically connected to the second data processing unit, forming an integrated production control system for instruction generation, transmission, and execution. The PLC communication module, as a dedicated communication module, establishes a stable communication connection with the post-processing equipment for household paper, transmitting the equipment parameter adjustment instructions generated by the second data processing unit to the production equipment in real time. Simultaneously, it collects process operation data from the production equipment and feeds the data back to the defect statistics module in real time, providing data support for calculating the upstream defect probability. The second data processing unit, as the core data processing terminal, receives the defect identification results from the AI ​​visual inspection core module and converts the identification results into specific executable equipment parameter adjustment instructions according to preset control logic. These instructions are then transmitted in real time to the PLC communication module and the pneumatic control component. The pneumatic control component, as a dedicated execution component, receives the equipment parameter adjustment instructions from the second data processing unit and directly executes the parameter adjustment operations of the production equipment according to the instructions, achieving precise and rapid control of the production equipment.

[0059] The PLC communication module features strong anti-interference capabilities and stable data transmission, enabling bidirectional data transmission between the detection system and production equipment. This ensures both rapid and accurate transmission of control commands and real-time acquisition of equipment operating data, allowing the detection system to monitor equipment status and forming a data closed loop. The second data processing unit replaces manual command generation, improving command generation efficiency and accuracy and avoiding errors from manual operation. The pneumatic control component boasts fast response speed and high execution precision, enabling rapid and accurate execution of control commands. This allows production equipment to adjust promptly based on detection results, preventing the continuous production of substandard products. This module achieves bidirectional data linkage and precise parameter adjustment between detection results and production equipment, creating a closed-loop collaboration between detection and production. It achieves seamless transformation of detection results into production control, reducing the production of substandard batches and improving the production quality of post-processing of household paper.

[0060] Traditional post-processing inspection systems for household paper only identify defects and lack a dedicated defect marking module. Defective paper cannot be marked accurately, manual marking is prone to errors, and the marking parameters cannot match the inspection parameters, which can easily cause secondary damage to the paper.

[0061] Based on this, the detection system also includes a defect intelligent marking module. This module establishes communication connections with the AI ​​visual detection core module, the imaging calibration module, and the threshold calculation module to achieve real-time linkage between marking parameters and detection parameters. The defect intelligent marking module includes a pneumatic defect marking device, a paper weight sensor, and a third data processing unit. The pneumatic defect marking device and the paper weight sensor are electrically connected to the third data processing unit, forming an integrated defect marking system encompassing parameter acquisition, generation, and execution. The paper weight sensor collects the actual weight data of the tissue paper and transmits the data to the third data processing unit in real time. The first data processing unit provides a basis for setting the marking intensity based on paper characteristics. The second data processing unit receives the imaging calibration coefficient from the imaging calibration module, the dynamic detection threshold from the threshold calculation module, and the actual weight data from the paper weight sensor. It integrates and processes the three types of data according to a preset marking logic to generate marking intensity parameters and marking position parameters that match the paper characteristics and defect conditions. The parameters are then transmitted to the pneumatic defect marking device in real time. The pneumatic defect marking device receives the marking parameters from the third data processing unit and the defect recognition results from the AI ​​visual detection core module. Based on the marking parameters and recognition results, it performs precise defect marking operations on the defective paper.

[0062] The communication connection between the intelligent defect marking module and the core detection module allows for real-time linkage between marking parameters, imaging calibration coefficients, and dynamic detection thresholds. This ensures precise matching of the marking position to the defect location, and the marking force is adjusted according to the paper weight to avoid secondary damage from excessive force or unclear marking from insufficient force. The paper weight sensor ensures that the marking force setting fully considers paper characteristics, improving the adaptability of the marking. The pneumatic defect marking device features high precision and fast response, replacing manual marking methods and improving marking efficiency and accuracy. This module achieves real-time matching of defect marking parameters with detection parameters and paper characteristics, making defect marking more accurate and avoiding secondary paper damage. It also provides clear and accurate marking criteria for subsequent rejection of defective products.

[0063] Traditional post-processing inspection systems for household paper lack a dedicated module for rejecting defective products. Paper with defect markings must be manually removed, which is inefficient and prone to missing defects. Without dedicated visual recognition and execution components, accurate rejection is impossible.

[0064] Based on this, the detection system also includes a defective product rejection module. This module establishes a communication connection with the intelligent defect marking module, achieving seamless integration of marking recognition and rejection operations. This module is located at the packaging process location, completing rejection before finished product packaging to prevent defective products from entering the consumer market. The defective product rejection module includes a visual recognition unit, a mechanical execution unit, and a fourth data processing unit. The visual recognition unit and the mechanical execution unit are electrically connected to the fourth data processing unit, forming an integrated rejection system encompassing recognition, instruction generation, and execution. The visual recognition unit collects packaging data... The surface image of the paper before the process is transmitted to the fourth data processing unit in real time. Simultaneously, defect marks on the paper surface are identified using built-in recognition logic, and the mark recognition signals are transmitted to the fourth data processing unit in real time. The fourth data processing unit receives the mark recognition signals from the visual recognition unit and converts the signals into specific rejection execution instructions according to preset rejection logic. These instructions are then transmitted to the mechanical execution unit in real time. The mechanical execution unit is a pneumatic actuator that receives the rejection execution instructions from the fourth data processing unit and performs precise rejection of defective paper with defect marks according to the instructions.

[0065] The defective product rejection module, as the final quality control step, completes rejection before packaging, effectively ensuring the quality of the finished product. The visual recognition unit accurately identifies defect marks, providing accurate basis for rejection operations and avoiding missed or incorrect rejections. The fourth data processing unit automates the connection between recognition and rejection, improving rejection efficiency. The pneumatic actuator's mechanical execution unit has a fast response speed, high execution accuracy, and no contact damage, enabling rapid and accurate rejection operations while avoiding damage to qualified paper. This module achieves accurate identification and automated rejection of defective products, replacing manual rejection methods, significantly improving rejection efficiency and accuracy, preventing defective products from entering the consumer market, ensuring the quality of finished household paper products, and seamlessly integrating with the intelligent defect marking module to form a complete quality control system.

[0066] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An online inspection system for post-processing of household paper products based on AI vision, characterized in that, The system includes an environment-deformation sensing module, an imaging calibration module, a speed acquisition module, a threshold calculation module, a defect statistics module, a sensitivity calculation module, an AI visual inspection core module, and a production linkage module. These modules establish communication connections with each other. The environment-deformation sensing module collects actual temperature and humidity data from the production environment and dynamic paper contour data, and calculates and outputs temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation. The imaging calibration module receives the temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation output from the environment-deformation sensing module, performs multi-parameter quantization calculations, and outputs imaging calibration coefficients. The speed acquisition module collects instantaneous speed data from the production line and calculates and outputs the speed deviation rate. The threshold calculation module receives the imaging calibration coefficients output from the imaging calibration module and the speed deviation rate output from the speed acquisition module, performs dual-parameter quantization calculations, and outputs a dynamic detection threshold. The defect statistics module is used to collect defect detection data from upstream processes and calculate and output the probability of defect occurrence. The sensitivity calculation module receives the dynamic detection threshold output by the threshold calculation module and the defect occurrence probability output by the defect statistics module, performs dual-parameter quantization calculation, and outputs the process detection sensitivity coefficient. The AI ​​visual inspection core module receives all the data output by the imaging calibration module, threshold calculation module, and sensitivity calculation module, performs defect recognition operations on the paper images of each process in the post-processing of tissue paper, and outputs the defect recognition results. The production linkage module receives the defect recognition results output by the AI ​​visual inspection core module and outputs equipment parameter adjustment instructions to the tissue paper post-processing equipment.

2. The online inspection system for post-processing of household paper products based on AI vision according to claim 1, characterized in that, The environment-deformation sensing module includes a temperature and humidity sensor, a high-speed profile camera, and a first data processing unit. Both the temperature and humidity sensor and the high-speed profile camera are electrically connected to the first data processing unit. The temperature and humidity sensor is used to collect actual temperature data and actual humidity data of the production environment and transmit them to the first data processing unit. The high-speed profile camera is used to collect dynamic profile data of paper in each post-processing step of the tissue paper and transmit it to the first data processing unit. The first data processing unit is used to receive all the data transmitted by the temperature and humidity sensor and the high-speed profile camera, and calculate and output the normalized temperature deviation, normalized humidity deviation, and dynamic deformation of the paper.

3. The online inspection system for post-processing of household paper products based on AI vision according to claim 1, characterized in that, The imaging calibration module includes an FPGA image processing unit and a first reference parameter storage unit. The FPGA image processing unit and the first reference parameter storage unit are electrically connected. The first reference parameter storage unit is used to store the imaging basic calibration coefficients, temperature influence weighting coefficients, and humidity influence weighting coefficients and transmit them to the FPGA image processing unit. The FPGA image processing unit is used to receive the temperature normalization deviation, humidity normalization deviation, and paper dynamic deformation output by the environment-deformation sensing module, receive all reference parameters transmitted by the first reference parameter storage unit, perform quantization calculations, and output the imaging calibration coefficients.

4. The online inspection system for post-processing of household paper products based on AI vision according to claim 1, characterized in that, The core module of AI visual inspection includes a high-definition industrial camera group, a ring-shaped shadowless light source, a GPU inference unit, and a defect feature library storage unit. The high-definition industrial camera group, the ring-shaped shadowless light source, and the defect feature library storage unit are all electrically connected to the GPU inference unit. The high-definition industrial camera group consists of a hyperspectral camera and a high-speed linear array camera. The high-definition industrial camera group is used to acquire raw images of paper at each stage of paper processing and transmit them to the GPU inference unit. The ring-shaped shadowless light source is used to provide illumination for the acquisition of raw paper images by the high-definition industrial camera group. The defect feature library storage unit is used to store the multi-dimensional defect feature library and the improved YOLO neural network model and transmit them to the GPU inference unit. The GPU inference unit is used to receive the imaging calibration coefficients output by the imaging calibration module, perform image calibration operations on the original paper image, receive the dynamic detection threshold output by the threshold calculation module and the process detection sensitivity coefficient output by the sensitivity calculation module, call the multi-dimensional defect feature library and the improved YOLO neural network model to perform defect recognition operations on the calibrated paper image, and output the defect recognition results.

5. The online inspection system for post-processing of household paper products based on AI vision according to claim 3, characterized in that, The quantization calculation performed by the FPGA image processing unit adopts the imaging calibration coefficient calculation formula. The imaging calibration coefficient calculation formula is constructed with imaging basic calibration coefficient, temperature influence weight coefficient, temperature normalization deviation, humidity influence weight coefficient, humidity normalization deviation, and paper dynamic deformation as core calculation parameters, and is specifically used to calculate the imaging calibration coefficient.

6. The online inspection system for post-processing of household paper products based on AI vision according to claim 1, characterized in that, The threshold calculation module includes a second reference parameter storage unit and a first embedded computing chip, and the second reference parameter storage unit is electrically connected to the first embedded computing chip; The second reference parameter storage unit is used to store the basic pixel threshold of the defect feature and transmit it to the first embedded computing chip; The first embedded computing chip is used to receive the imaging calibration coefficients output by the imaging calibration module and the speed deviation rate output by the speed acquisition module, and to receive the defect feature basic pixel threshold transmitted by the second reference parameter storage unit, perform quantization calculations and output the dynamic detection threshold.

7. The online inspection system for post-processing of household paper products based on AI vision according to claim 1, characterized in that, The sensitivity calculation module includes a third reference parameter storage unit and a second embedded computing chip, and the third reference parameter storage unit and the second embedded computing chip are electrically connected. The third reference parameter storage unit is used to store the basic detection sensitivity coefficient and the influence weight of upstream defects on downstream and transmit them to the second embedded computing chip. The second embedded computing chip is used to receive the dynamic detection threshold output by the threshold calculation module, the defect occurrence probability output by the defect statistics module, the basic detection sensitivity coefficient transmitted by the third reference parameter storage unit, and the influence weight of upstream defects on downstream, perform quantitative calculations, and output the process detection sensitivity coefficient.

8. The online inspection system for post-processing of household paper products based on AI vision according to claim 1, characterized in that, The production linkage module includes a PLC communication module, a pneumatic control component, and a second data processing unit. Both the PLC communication module and the pneumatic control component are electrically connected to the second data processing unit. The PLC communication module establishes a communication connection with the tissue paper post-processing equipment, transmits equipment parameter adjustment commands to the equipment, collects process operation data from the equipment, and feeds it back to the defect statistics module. The second data processing unit receives defect identification results from the AI ​​visual inspection core module, generates equipment parameter adjustment commands, and transmits them to the PLC communication module and the pneumatic control component. The pneumatic control component receives the equipment parameter adjustment commands transmitted by the second data processing unit and executes the parameter adjustment operations on the tissue paper post-processing equipment.

9. The online inspection system for post-processing of household paper products based on AI vision according to claim 1, characterized in that, It also includes a defect intelligent marking module, which establishes communication connections with the AI ​​visual inspection core module, imaging calibration module, and threshold calculation module. The defect intelligent marking module includes a pneumatic defect marking device, a paper weight sensor, and a third data processing unit. The pneumatic defect marking device and the paper weight sensor are electrically connected to the third data processing unit. The paper weight sensor is used to collect the actual weight data of the tissue paper and transmit it to the third data processing unit. The third data processing unit is used to receive the imaging calibration coefficient output by the imaging calibration module, the dynamic detection threshold output by the threshold calculation module, and the actual weight data transmitted by the paper weight sensor, generate marking force parameters and marking position parameters, and transmit them to the pneumatic defect marking device. The pneumatic defect marking device is used to receive the marking force parameters and marking position parameters transmitted by the third data processing unit, receive the defect identification results output by the AI ​​visual inspection core module, and perform defect marking operations on the tissue paper with defects.

10. The online inspection system for post-processing of household paper products based on AI vision according to claim 9, characterized in that, It also includes a defective product rejection module, which establishes a communication connection with the defective intelligent marking module. The defective product rejection module is located at the working position of the post-processing and packaging process of tissue paper. The defective product rejection module includes a visual recognition unit, a mechanical execution unit, and a fourth data processing unit. The visual recognition unit and the mechanical execution unit are both electrically connected to the fourth data processing unit. The visual recognition unit is used to collect surface images of tissue paper before the packaging process and transmit them to the fourth data processing unit, identify defective marks on the surface of tissue paper, and transmit the mark recognition signal to the fourth data processing unit. The fourth data processing unit is used to receive the mark recognition signal transmitted by the visual recognition unit, generate rejection execution instructions, and transmit them to the mechanical execution unit. The mechanical execution unit is used to receive the rejection execution instructions transmitted by the fourth data processing unit and perform defective product rejection operations on tissue paper with defective marks. The mechanical execution unit is a pneumatic execution structure.

Images