Air conditioner filter screen weaving defect online detection alarm method

Abstract

This application discloses an online detection and alarm method for defects in air conditioning filter fabric, belonging to the field of online inspection technology in textile weaving. The steps are as follows: dynamically adjusting the frame rate of the visual sensor according to the loom running speed to maintain a constant physical resolution; preprocessing the acquired image, extracting the dynamic region of interest, extracting multi-dimensional grid feature parameters and comparing them item by item with the pre-stored standard template; calculating the deviation rate weighted comprehensive difference degree for initial screening, and conducting a secondary verification of suspected areas; after confirming the defect, identifying the defect type and classifying it into three levels: minor, general, and severe, triggering a graded audible and visual alarm. When a level three alarm is triggered, the defect is physically marked first and then the loom is stopped. The data can be stored and uploaded, and it has threshold self-learning optimization driven by false alarms and missed alarms, and emergency response functions triggered by abnormal detection frequency. It can realize real-time online detection, intelligent graded alarm and self-optimization of various fabric defects in air conditioning filter fabric, thereby improving product quality.

Description

Technical Field

[0001] This application relates to the field of online inspection technology for textile machinery, and more specifically, to an online detection and alarm method for defects in the weaving of air conditioning filters. Background Technology

[0002] As a core component in air conditioning systems that intercepts dust, particulate matter, and microorganisms from the air, the quality of air conditioning filters directly determines the filtration efficiency, air resistance characteristics, and service life of the air conditioner. Current production primarily uses polypropylene, polyamide, and polyester fibers, woven on Titan series rapier looms or air-jet looms. During the weaving process, factors such as warp tension fluctuations, abnormal weft yarn introduction, mechanical failures, and raw material quality fluctuations can easily lead to various weaving defects, including warp breakage, weft breakage, uneven mesh size, uneven density, holes, and skipped threads. Among these, structural defects such as broken threads and holes significantly weaken the filter's filtration accuracy and mechanical strength, and must be detected promptly during production.

[0003] Currently, the detection of defects in air conditioner filter weaving mainly relies on manual inspection. Operators visually inspect the fabric surface while the loom is running, stopping the machine to address any abnormalities. This method has significant drawbacks: a single operator needs to monitor multiple high-speed looms simultaneously, making continuous and comprehensive inspection difficult, resulting in excessive labor intensity; defects are often only discovered after a considerable length of fabric has been woven, causing a large amount of woven fabric to become waste; for small defects on the high-speed running fabric surface, such as a single broken filament or slight unevenness in mesh size, the human eye has limited ability to detect them, leading to a high rate of missed detections, and the inspection results are greatly affected by the inspector's experience and condition.

[0004] While some looms are equipped with mechanical or photoelectric yarn breakage sensors, their functionality is extremely limited. These sensors can only detect yarn breaks and cannot identify visual defects such as mesh misalignment, uneven density, or holes. Furthermore, the sensors require specific interface design with the loom's control system, making installation and debugging complex, and compatibility between different loom models is poor. More importantly, for air conditioner filters with regular, periodic mesh patterns, the sensors struggle to distinguish between normal mesh gaps and abnormal defects, leading to frequent false alarms and missed alarms.

[0005] In recent years, automatic fabric defect detection technology based on machine vision has developed, with various algorithms emerging such as texture analysis, Gabor filters, low-rank decomposition, and deep learning methods. While some industrial systems can identify certain defects, these systems are mostly general-purpose designs and lack adaptation to the specific mounting structure of Titan looms. Different Titan loom models have significant differences in the position of the feed rollers, the fabric feeding space, and the machine structure. General-purpose detection equipment is prone to physical interference with the loom's transmission mechanism, heddle frame, or fabric winding mechanism during installation, making stable installation and imaging impossible. Furthermore, existing visual inspection algorithms are not optimized for the unique periodic grid texture of air conditioner filters, resulting in poor performance in detecting characteristic defects such as mesh misalignment and uneven density, making it difficult to simultaneously meet the integrated requirements of broken yarn detection and mesh appearance defect detection.

[0006] In summary, existing technologies face four key problems: manual inspection is labor-intensive, slow, and has a high rate of missed detections; traditional broken yarn sensors have limited functionality and cannot cover defects in the mesh appearance; machine vision systems lack specific adaptations for the Titan loom's installation structure, making installation difficult and prone to interference with machine components; and there is a lack of dedicated detection algorithms for the periodic mesh texture of air conditioning filters, resulting in the inability to coordinate broken yarn detection and mesh defect detection.

[0007] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0008] The purpose of this application is to provide an online detection and alarm method for defects in the fabrication of air conditioning filters. This method can detect various fabrication defects in the fabrication process of air conditioning filters in real time and accurately, realize automatic graded alarm and physical marking, reduce the labor intensity and missed detection rate of manual inspection, and improve product quality and production efficiency.

[0009] This application provides an online detection and alarm method for fabric defects in air conditioning filters, the technical solution of which is as follows:

[0010] Includes the following steps:

[0011] Step 1: Image Acquisition and Speed ​​Synchronization: The loom drives the air conditioner filter fabric forward at a running speed V. A backlight located below the filter fabric illuminates the fabric surface evenly. A vision sensor located above the filter fabric acquires a full-width grayscale image of the fabric surface in real time, while simultaneously acquiring the current running speed signal of the loom. The acquisition frame rate of the vision sensor is dynamically adjusted according to the running speed signal to maintain a constant physical resolution of the acquired image in the direction of fabric travel.

[0012] Step 2, Image Preprocessing: The intelligent control host sequentially performs median filtering for noise reduction and contrast-limited adaptive histogram equalization on the acquired raw grayscale image to enhance the grayscale contrast between the fabric texture and the background, resulting in the preprocessed image.

[0013] Step 3: Dynamic Region of Interest Extraction: The intelligent control host automatically locates and extracts the effective detection area in the preprocessed image based on the edge position of the fabric and the periodic texture features of the grid, eliminating interference from non-detection areas on both sides of the fabric.

[0014] Step 4: Extraction of Mesh Feature Parameters: The intelligent control host performs latitude and longitude edge detection on the image within the effective detection area and extracts the following feature parameters of the filter mesh: longitude spacing, latitude spacing, single mesh area, longitude gray-level gradient change rate, latitude gray-level gradient change rate, and number of mesh intersections.

[0015] Step 5: Multi-template standard comparison: The intelligent control host calls the pre-stored standard mesh texture database, which stores the standard feature parameter range and standard mesh topology structure corresponding to the current production variety; the real-time feature parameters extracted in Step 4 are compared with the standard feature parameters item by item, and the deviation rate of each item is calculated.

[0016] Step 6: Calculation and initial screening of comprehensive difference: The intelligent control host calculates the weighted sum of various deviation rates to obtain the comprehensive difference D. When D is greater than the preset first dynamic threshold T1, it is determined that there is a suspected defect in the area, and the location coordinates, frame number and local image block of the suspected defect are recorded.

[0017] Step 7, Local Secondary Verification: The intelligent control host controls the vision sensor to perform optical zoom or digital cropping magnification on the area where the suspected defect is located, and acquires a local high-resolution verification image; Steps 2 to 6 are repeated on the verification image. If the comprehensive verification difference D' is less than the preset second dynamic threshold T2, it is determined to be a false alarm, the suspected signal is ignored and the detection continues; if D' ≥ T2, it is confirmed as a real defect.

[0018] Step 8: Intelligent Defect Type Identification: For areas confirmed as genuine defects, the intelligent control host classifies the type based on local image features: If a grayscale abrupt change occurs along the meridian direction and the latitude continuity is interrupted, it is determined to be a broken meridian wire; if a grayscale abrupt change occurs along the latitude direction and the meridian continuity is interrupted, it is determined to be a broken latitude wire; if the area deviation rate of adjacent grids exceeds the third threshold T3 and is continuously distributed, it is determined to be uneven mesh size; if the grid area abruptly changes to several times the normal value and the boundary is irregularly torn, it is determined to be a hole; if the spacing between meridians or latitudes shows periodic width changes, it is determined to be a sparse-dense path; if a free thread shadow appears locally that is not in the meridian or latitude direction, it is determined to be a skipped thread; if none of the above criteria can be matched, it is determined to be an unclassified structural defect.

[0019] Step Nine, Defect Grading: The intelligent control host classifies defects into three levels—minor defects, general defects, and severe defects—according to preset grading rules based on defect type, defect area S, number of defects N per unit length, and defect distribution density ρ. When the defect type is a broken warp thread or a hole, the grade is at least general defect. When the defect type is an unclassified structural defect, it is directly classified as severe defect. When the hole area exceeds the second area threshold S2, it is directly classified as severe defect.

[0020] Step 10: Graded Audible and Visual Alarms and Shutdown Control: The intelligent control host sends corresponding control signals to the audible and visual alarms according to the defect level. Minor defects trigger a Level 1 alarm and are recorded in the log. General defects trigger a Level 2 audible and visual alert. Severe defects trigger a Level 3 audible and visual alarm. When a Level 3 audible and visual alarm is triggered, the intelligent control host first sends a marking execution signal. After the physical marking of the defect in Step 11 is completed, a shutdown signal is sent to the loom control system.

[0021] Step 11, Physical Marking of Defects: The intelligent control host controls the inkjet marking device located behind the vision sensor to spray an identifiable mark at the edge of the fabric corresponding to the defect, while simultaneously recording the mark position and the longitudinal distance of the defect on the fabric surface; after receiving the marking execution signal, the inkjet marking device immediately performs the marking operation, and after marking is completed, it sends a marking completion signal back to the intelligent control host;

[0022] Step 12: Data storage and upload: Before triggering the stop signal, the intelligent control host first writes the image, type, grade, occurrence time, loom number and fabric speed information of the current defect into the non-volatile memory, and uploads it to the workshop-level quality management system via industrial Ethernet.

[0023] Step Thirteen: Self-learning Optimization and Emergency Response: The intelligent control host periodically calculates the false alarm rate and the missed alarm rate. When the false alarm rate exceeds the preset threshold, it automatically raises the first dynamic threshold T1 or the second dynamic threshold T2. When the missed alarm rate exceeds the preset threshold, it automatically lowers T1 or T2 to achieve adaptive optimization of the judgment threshold. When the defect detection frequency per unit time exceeds a preset multiple of the historical average, the system temporarily lowers T1 and T2 and triggers a process anomaly warning.

[0024] Furthermore, this application also proposes that, in step one, the brightness of the backlight is automatically adjusted based on real-time feedback from the ambient light sensor. The adjustment method is as follows: when the ambient light intensity increases, the backlight brightness is automatically increased; when the ambient light intensity decreases, the backlight brightness is automatically decreased, so that the average gray value of the fabric image collected by the visual sensor is always maintained within the preset gray target range.

[0025] Furthermore, this application also proposes that, in step three, the method for extracting the dynamic region of interest is as follows: performing row grayscale projection and column grayscale projection on the preprocessed image, determining the left and right boundaries of the fabric surface based on the abrupt change points at the edges of the projection curves, and further determining the upper and lower boundaries based on the periodic characteristics of the latitude and longitude lines, thus forming a rectangular effective detection window that dynamically changes with the position of the fabric surface.

[0026] Furthermore, this application also proposes that, in step four, the latitude and longitude line edge detection adopts the Canny edge detection operator, and after extracting the edges, the main straight line clusters in the longitude and latitude directions are identified by Hough line transformation, the pixel distance between adjacent parallel lines is calculated and converted into physical size, and the longitude and latitude line spacing is obtained; the single grid area is calculated by the closed quadrilateral formed by the intersection points of the detected longitude and latitude lines and converted into physical area.

[0027] Furthermore, this application also proposes that, in step six, the formula for calculating the comprehensive difference degree D is:

[0028] D = α·Δd_warp + β·Δd_weft + γ·ΔA + δ·ΔG_warp + ε·ΔG_weft +ζ·ΔP,

[0029] Where Δd_warp is the deviation rate of meridian spacing, Δd_weft is the deviation rate of parallel spacing, ΔA is the deviation rate of single grid area, ΔG_warp is the deviation rate of meridian gray-level gradient change rate, ΔG_weft is the deviation rate of parallel gray-level gradient change rate, and ΔP is the deviation rate of grid intersection number; α, β, γ, δ, ε, and ζ are the weight coefficients of the corresponding terms, and α+β+γ+δ+ε+ζ=1.

[0030] Furthermore, this application also proposes that in step six, the first dynamic threshold T1 and the second dynamic threshold T2 are automatically adjusted according to the current operating speed V of the loom and the current product specifications. The adjustment rules are as follows: the higher the speed V, the smaller the values ​​of T1 and T2; the higher the mesh number of the product, the smaller the values ​​of T1 and T2.

[0031] Furthermore, this application also proposes that, in step eight, when multiple types of defects are detected to overlap in the same local area, the intelligent control host uses the most severe type as the main classification result for that area and adds the identification code of all detected types to the log.

[0032] Furthermore, this application also proposes that, in step nine, the specific rules for grade classification are as follows: if the defect area S is less than the first area threshold S1 and the number of defects N per unit length is less than the first quantity threshold N1, and the defect type is not a broken warp thread, hole, or unclassified structural defect, then it is classified as a minor defect; if S is between S1 and S2 or N is between N1 and N2, and the defect type is not an unclassified structural defect, then it is classified as a general defect; if S is greater than the second area threshold S2 or N is greater than the second quantity threshold N2 or the defect distribution density ρ is greater than the density threshold ρ0, or the defect type is an unclassified structural defect, then it is classified as a severe defect, where S1 <S2,N1<N2。

[0033] Furthermore, this application also proposes that, in step ten, the specific method of the three-level audible and visual alarm is as follows: the first-level alarm only displays prompt information on the human-machine interface and records the log, without triggering audible and visual signals; the second-level alarm triggers the yellow indicator light to stay on and emits an intermittent buzzing sound; the third-level alarm triggers the red indicator light to flash and emits a continuous high-decibel buzzing sound.

[0034] Furthermore, this application also proposes that in step eleven, the inkjet marking device uses fast-drying industrial marking ink, and the inkjet marking is a long strip mark along the fabric edge direction, with the marking length being positively correlated with the lateral dimension of the defect in the width direction of the fabric surface.

[0035] Furthermore, this application also proposes that, in step thirteen, the self-learning optimization further includes: the operator confirms the authenticity of each alarm through the human-machine interface, the system automatically constructs a false alarm sample set and a missed alarm sample set based on the feedback results, and performs feature clustering analysis on the false alarm samples and missed alarm samples. When the cumulative number of samples of a certain type of feature exceeds the preset number threshold, the system automatically adjusts the weight coefficient of that type of feature in the comprehensive difference degree D.

[0036] Furthermore, this application also proposes a template consistency verification step before step five: the intelligent control host performs statistics on the average background grayscale and grid periodicity features of the continuously acquired N0 frames of images. When the deviation of the statistical value from the previous calibration value exceeds the preset template update threshold, the standard template data is automatically triggered to adapt to image drift caused by light source aging or slow environmental changes.

[0037] As described above, the online detection and alarm method for fabric defects in air conditioning filters provided in this application includes steps such as image acquisition and speed synchronization, image preprocessing, dynamic region of interest extraction, grid feature parameter extraction, multi-template standard comparison, comprehensive difference calculation and initial screening, local secondary verification, intelligent identification of defect types, defect level classification, graded audible and visual alarm and shutdown control, physical marking of defects, data storage and uploading of detection data, self-learning optimization and emergency response. Through the above technical solution, real-time online detection, intelligent identification and graded processing of fabric defects are achieved, effectively solving the problems of low efficiency of manual inspection, single function of traditional sensors, lack of dedicated adaptation of machine vision systems and insufficient dedicated algorithms in the prior art. It can detect multiple fabric defects in the fabrication process of air conditioning filters in real time and accurately, realize automatic graded alarm and physical marking, reduce the labor intensity and missed detection rate of manual inspection, and improve product quality and production efficiency.

[0038] Furthermore, this application effectively solves the technical problem that composite defects or novel unknown defects cannot be classified in the existing classification system by setting unclassified structural defects as a catch-all type in addition to the six known defect categories and enforcing a severe defect level classification for this type. This avoids omissions or downgrades due to the inability to match known types, and ensures that the detection system can still alarm and shut down at the most stringent level for any structural anomalies that exceed the experience range, thereby improving the system's safety redundancy and quality assurance capabilities. Attached Figure Description

[0039] Figure 1 This application provides a schematic diagram of an online detection and alarm process for defects in the fabrication of an air conditioning filter. Detailed Implementation

[0040] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0041] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0042] Traditional methods for detecting defects in air conditioning filter mesh fabric rely primarily on manual inspection, which suffers from high labor intensity, slow detection speed, and a high rate of missed detections. Traditional mechanical or photoelectric broken yarn sensors have limited functionality, cannot identify mesh appearance defects, and are complex to install and debug. Existing machine vision inspection systems lack specific adaptations for the Titan loom installation structure, making installation difficult, and lack dedicated detection algorithms for the periodic mesh texture characteristics of air conditioning filters, making it difficult to integrate broken yarn detection with mesh appearance defect detection.

[0043] like Figure 1 As shown, the method provided in this application relies on an online inspection system comprising a vision sensor, a backlight, an intelligent control host, an audible and visual alarm, an inkjet marking device, and an industrial Ethernet communication module. The intelligent control host, as the core processing unit, communicates with the vision sensor, backlight, audible and visual alarm, inkjet marking device, loom control system, and workshop-level quality management system. This method uses a vision sensor to acquire fabric images in real time and dynamically adjusts the acquisition frame rate according to the loom's operating speed to ensure constant physical resolution. The images undergo preprocessing, dynamic region of interest extraction, and mesh feature parameter extraction. Multiple templates are compared with a standard database to calculate the comprehensive difference for initial screening. For suspected defects, a local secondary verification is performed to improve accuracy. Once confirmed as a genuine defect, the system intelligently identifies the defect type and classifies it into levels, ultimately triggering a graded audible and visual alarm and machine shutdown control, and physically marking the defect using the inkjet marking device. The detection data is stored and uploaded to the quality management system, which also possesses self-learning optimization and emergency response functions. Specifically, the method includes the following steps:

[0044] Step 1: Image Acquisition and Speed ​​Synchronization: The loom drives the air conditioner filter fabric forward at a running speed V. A backlight located below the filter fabric illuminates the fabric surface evenly. A vision sensor located above the filter fabric acquires a full-width grayscale image of the fabric surface in real time, while simultaneously acquiring the current running speed signal of the loom. The acquisition frame rate of the vision sensor is dynamically adjusted according to the running speed signal to maintain a constant physical resolution of the acquired image in the direction of fabric travel.

[0045] Step 2, Image Preprocessing: The intelligent control host sequentially performs median filtering for noise reduction and contrast-limited adaptive histogram equalization on the acquired raw grayscale image to enhance the grayscale contrast between the fabric texture and the background, resulting in the preprocessed image.

[0046] Step 3: Dynamic Region of Interest Extraction: Based on the edge position of the fabric and the periodic texture features of the grid, the intelligent control host automatically locates and extracts the effective detection area in the preprocessed image, eliminating interference from non-detection areas on both sides of the fabric.

[0047] Step 4: Extraction of grid feature parameters: The intelligent control host performs latitude and longitude edge detection on the image within the effective detection area and extracts the following feature parameters of the filter grid: longitude spacing, latitude spacing, single grid area, longitude gray-level gradient change rate, latitude gray-level gradient change rate, and number of grid intersections;

[0048] Step 5: Multi-template standard comparison: The intelligent control host calls the pre-stored standard mesh texture database, which stores the standard feature parameter range and standard mesh topology structure corresponding to the current production variety; the real-time feature parameters extracted in Step 4 are compared with the standard feature parameters item by item, and the deviation rate of each item is calculated.

[0049] Step 6: Calculation and initial screening of comprehensive difference: The intelligent control host calculates the weighted sum of various deviation rates to obtain the comprehensive difference D; when D is greater than the preset first dynamic threshold T1, it is determined that there is a suspected defect in the area, and the location coordinates, frame number and local image block of the suspected defect are recorded.

[0050] Step 7, Local Secondary Verification: The intelligent control host controls the vision sensor to perform optical zoom or digital cropping magnification on the area where the suspected defect is located, and acquires a local high-resolution verification image; Steps 2 to 6 are repeated on the verification image. If the comprehensive verification difference D' is less than the preset second dynamic threshold T2, it is determined to be a false alarm, the suspected signal is ignored and the detection continues; if D' ≥ T2, it is confirmed as a real defect.

[0051] Step 8: Intelligent Defect Type Identification: For areas confirmed as genuine defects, the intelligent control host classifies the type based on local image features: if a grayscale abrupt change occurs along the meridian direction and the latitude continuity is interrupted, it is determined to be a broken meridian thread; if a grayscale abrupt change occurs along the latitude direction and the meridian continuity is interrupted, it is determined to be a broken latitude thread; if the area deviation rate of adjacent grids exceeds the third threshold T3 and is continuously distributed, it is determined to be uneven mesh size; if the grid area abruptly changes to several times the normal value and the boundary is irregularly torn, it is determined to be a hole; if the spacing between meridians or latitudes shows periodic width changes, it is determined to be a sparse-dense path; if a free thread shadow appears locally that is not in the meridian or latitude direction, it is determined to be a skipped thread; if none of the above criteria can be matched, it is determined to be an unclassified structural defect.

[0052] Step Nine, Defect Grading: The intelligent control host classifies defects into three levels—minor defects, general defects, and severe defects—according to preset grading rules based on defect type, defect area S, number of defects N per unit length, and defect distribution density ρ. When the defect type is a broken warp thread or a hole, the grade is at least general defect. When the defect type is an unclassified structural defect, it is directly classified as a severe defect. When the hole area exceeds the second area threshold S2, it is directly classified as a severe defect.

[0053] Step 10, Graded Audible and Visual Alarm and Shutdown Control: The intelligent control host sends corresponding control signals to the audible and visual alarms according to the defect level. Minor defects trigger a Level 1 alarm and are recorded in the log, general defects trigger a Level 2 audible and visual alert, and serious defects trigger a Level 3 audible and visual alarm. When a Level 3 audible and visual alarm is triggered, the intelligent control host first sends a marking execution signal. After the physical marking of the defect in Step 11 is completed, a shutdown signal is sent to the loom control system.

[0054] Step 11, Physical Marking of Defects: The intelligent control host controls the inkjet marking device located behind the vision sensor to spray an identifiable mark at the edge of the fabric corresponding to the defect, while simultaneously recording the mark position and the longitudinal distance of the defect on the fabric surface; after receiving the marking execution signal, the inkjet marking device immediately performs the marking operation, and after marking is completed, it sends a marking completion signal back to the intelligent control host;

[0055] Step 12, Detection Data Storage and Upload: Before triggering the stop signal, the intelligent control host first writes the image, type, grade, occurrence time, loom number, and fabric speed information of the current defect into the non-volatile memory, and uploads it to the workshop-level quality management system via industrial Ethernet.

[0056] Step Thirteen: Self-learning Optimization and Emergency Response: The intelligent control host periodically calculates the false alarm rate and the missed alarm rate. When the false alarm rate exceeds the preset threshold, it automatically raises the first dynamic threshold T1 or the second dynamic threshold T2. When the missed alarm rate exceeds the preset threshold, it automatically lowers T1 or T2 to achieve adaptive optimization of the judgment threshold. When the defect detection frequency per unit time exceeds a preset multiple of the historical average, the system temporarily lowers T1 and T2 and triggers a process anomaly warning.

[0057] For ease of understanding, the following explains some key terms in this embodiment:

[0058] A vision sensor is a device that can convert optical images into electrical signals. In this embodiment, it is used to acquire grayscale images of air conditioner filter fabric in real time for subsequent image processing and defect detection.

[0059] Backlights are typically positioned below the object being inspected to provide uniform transmitted light, allowing the surface texture of the object to exhibit good contrast in the image, thus facilitating image acquisition by the vision sensor.

[0060] The intelligent control host is the core processing unit of the entire detection system. It is responsible for receiving and processing image data and loom operation signals, and executing complex image analysis algorithms, defect judgment logic, alarm control, and data management functions.

[0061] The standard mesh texture database is used to store the standard feature parameter range and standard mesh topology of air conditioner filters of different products, serving as a benchmark for defect comparison and judgment.

[0062] The first dynamic threshold T1 and the second dynamic threshold T2 are dynamically adjustable parameters used for initial defect screening and secondary verification. Their values ​​affect the system's detection sensitivity and false alarm rate.

[0063] The third threshold T3 is a specific parameter used to judge mesh unevenness defects. When the deviation rate of adjacent mesh areas exceeds this threshold, it is judged as mesh unevenness.

[0064] An audible and visual alarm is a device that alerts operators by emitting sound and / or light signals to indicate that the system has detected an abnormality, triggering different levels of alarms based on the severity of the defect.

[0065] An inkjet marking device is a device that can spray identifiable marks at designated locations on a fabric surface to physically mark the locations of confirmed defects, facilitating subsequent processing.

[0066] Non-volatile memory is a storage device that retains data even after power is lost. It is used to store key data such as images, types, and grades of detected defects.

[0067] Industrial Ethernet is an Ethernet technology used in the field of industrial control to enable high-speed data transmission between intelligent control hosts and workshop-level quality management systems.

[0068] A workshop-level quality management system is a platform for centralized management and analysis of quality data during the production process. It receives and processes defect information from the inspection system, and performs quality traceability and statistical analysis.

[0069] The first area threshold S1 and the second area threshold S2 are area parameters used for defect grade classification. The severity of the defect is determined based on the relationship between the defect area and these thresholds.

[0070] The first quantity threshold N1 and the second quantity threshold N2 are quantity parameters used for defect grade classification. The severity of defects is determined based on the relationship between the number of defects per unit length and these thresholds.

[0071] The density threshold ρ0 is a density parameter used to classify defects. When the density of defect distribution exceeds this threshold, it may be classified as a severe defect.

[0072] This embodiment provides an online detection and alarm method for defects in the fabric of air conditioning filters.

[0073] In terms of image acquisition and speed synchronization, the loom drives the air conditioner filter fabric forward at a running speed V. A backlight positioned below the filter fabric provides uniform illumination to the fabric surface, ensuring the quality of image acquisition. A vision sensor positioned above the filter fabric acquires a full-width grayscale image of the fabric surface in real time. Simultaneously, it acquires the current running speed signal of the loom and dynamically adjusts the acquisition frame rate of the vision sensor based on this signal, maintaining a constant physical resolution of the acquired image along the fabric's direction of travel.

[0074] In terms of image preprocessing, the intelligent control host sequentially performs median filtering for noise reduction and contrast-limited adaptive histogram equalization on the acquired raw grayscale image. This aims to enhance the grayscale contrast between the fabric texture and the background, thereby obtaining the preprocessed image.

[0075] In terms of dynamic region of interest extraction, the intelligent control host automatically locates and extracts the effective detection area in the preprocessed image based on the edge position of the fabric and the periodic texture features of the grid, so as to eliminate the interference of non-detection areas on both sides of the fabric.

[0076] In terms of mesh feature parameter extraction, the intelligent control host performs latitude and longitude edge detection on the image within the effective detection area and extracts the following feature parameters of the filter mesh: longitude spacing, latitude spacing, single mesh area, longitude gray-level gradient change rate, latitude gray-level gradient change rate, and number of mesh intersections.

[0077] For multi-template standard comparison, the intelligent control host calls upon a pre-stored standard mesh texture database. This database stores the standard feature parameter range and standard mesh topology corresponding to the current product variety. Subsequently, the real-time feature parameters extracted in step four are compared item by item with the standard feature parameters, and the deviation rate of each item is calculated.

[0078] In terms of comprehensive difference calculation and initial screening, the intelligent control host performs a weighted summation of various deviation rates to obtain the comprehensive difference degree D. When D is greater than the preset first dynamic threshold T1, it is determined that there is a suspected defect in the area, and the location coordinates, frame number, and local image block of the suspected defect are recorded.

[0079] In terms of local secondary verification, the intelligent control host controls the vision sensor to perform optical zoom or digital cropping magnification on the area where the suspected defect is located to acquire a local high-resolution verification image. Steps two to six are repeated on the verification image. If the overall verification difference D' is less than the preset second dynamic threshold T2, it is determined to be a false alarm, the suspected signal is ignored and the detection continues; if D' ≥ T2, it is confirmed as a real defect.

[0080] In terms of intelligent defect type identification, for areas confirmed as genuine defects, the intelligent control host classifies the type based on their local image features. Specifically, if a gray-scale abrupt change occurs along the meridian direction and the continuity of the latitude line is interrupted, it is determined to be a broken meridian wire; if a gray-scale abrupt change occurs along the latitude direction and the continuity of the meridian is interrupted, it is determined to be a broken latitude wire; if the deviation rate of adjacent grid areas exceeds the third threshold T3 and is continuously distributed, it is determined to be uneven mesh size; if the grid area abruptly changes to several times the normal value and the boundary is irregularly torn, it is determined to be a hole; if the spacing between meridians or latitude lines shows periodic width changes, it is determined to be a sparse-dense path; if a free thread shadow that does not follow the meridian or latitude direction appears locally, it is determined to be a skipped thread; if none of the above criteria can be matched, it is determined to be an unclassified structural defect.

[0081] Regarding defect grading, the intelligent control host classifies defects into three levels—minor defects, general defects, and severe defects—according to preset grading rules, based on defect type, defect area S, number of defects N per unit length, and defect distribution density ρ. When the defect type is a broken warp thread or a hole, the grade is at least general defect; when the defect type is an unclassified structural defect, it is directly classified as severe defect; when the hole area exceeds the second area threshold S2, it is directly classified as severe defect.

[0082] Regarding graded audible and visual alarms and machine shutdown control, the intelligent control host sends corresponding control signals to the audible and visual alarms based on the defect level. Minor defects trigger a level one alarm and are logged; general defects trigger a level two audible and visual alert; and severe defects trigger a level three audible and visual alarm. When a level three audible and visual alarm is triggered, the intelligent control host first issues a marking execution signal. After the physical marking of the defect in step eleven is completed, it then sends a shutdown signal to the loom control system.

[0083] Regarding the physical marking of defects, the intelligent control host controls an inkjet marking device located behind the vision sensor to spray a recognizable mark at the edge of the fabric corresponding to the defect, while simultaneously recording the mark position and the longitudinal distance of the defect on the fabric surface. Upon receiving a marking execution signal, the inkjet marking device immediately performs the marking operation and sends a marking completion signal back to the intelligent control host after marking is complete.

[0084] In terms of data storage and uploading, before triggering a stop signal, the intelligent control host prioritizes writing the image, type, grade, occurrence time, loom number, and fabric speed information of the current defect into a non-volatile memory, and then uploads it to the workshop-level quality management system via industrial Ethernet.

[0085] Regarding self-learning optimization and emergency response, the intelligent control host periodically calculates the false alarm rate and the missed alarm rate. When the false alarm rate exceeds a preset threshold, it automatically raises the first dynamic threshold T1 or the second dynamic threshold T2; when the missed alarm rate exceeds a preset threshold, it automatically lowers T1 or T2, achieving adaptive optimization of the judgment threshold. When the defect detection frequency per unit time exceeds a preset multiple of the historical average, the system temporarily lowers T1 and T2 and triggers a process anomaly warning.

[0086] This method effectively solves the problems of low efficiency, high missed detection rate, limited functionality of traditional sensors, and lack of dedicated algorithms and structural adaptation in the air conditioner filter weaving process by employing online real-time image acquisition, intelligent image processing, multi-level defect judgment, graded alarms and physical marking, combined with a self-learning optimization mechanism. It accurately identifies various mesh weaving defects, achieves intelligent defect type identification and grading, and significantly improves the quality control level and production automation of air conditioner filters.

[0087] To address this, this application further proposes that, in the image acquisition and speed synchronization step, the brightness of the backlight is automatically adjusted based on real-time feedback from an ambient light sensor. The backlight typically employs a high-brightness LED array or a high-frequency fluorescent lamp, and its brightness can be precisely adjusted via control signals from the intelligent control host, for example, through pulse width modulation (PWM) or constant current drive. The ambient light sensor, which can be a photoresistor, photodiode, or integrated ambient light sensor chip, is used to monitor the ambient light intensity in the production area in real time. It converts the ambient light intensity into an electrical signal and feeds it back to the intelligent control host. Based on the received ambient light signal and a preset control algorithm, the intelligent control host dynamically adjusts the output power of the backlight to counteract the impact of ambient light changes on the fabric lighting effect.

[0088] The specific adjustment method is as follows: the backlight brightness is automatically increased when the ambient light level increases, and automatically decreased when the ambient light level decreases. The intelligent control host operates a closed-loop control logic. When the ambient light sensor detects an increase in ambient light level, the system correspondingly increases the output power of the backlight to compensate for the increased external light and ensure that the total transmitted light intensity remains stable. Conversely, when the ambient light level decreases, the system reduces the backlight brightness. This adjustment strategy aims to maintain the constancy of the overall illumination of the fabric surface, ensuring that the intensity of the light signal received by the visual sensor is not affected by environmental interference.

[0089] Ultimately, through the above adjustment methods, the average grayscale value of the fabric image acquired by the visual sensor is consistently maintained within the preset grayscale target range. After each image acquisition, the intelligent control host calculates the average grayscale value of the current frame in real time. This average grayscale value serves as a feedback signal and is compared with the preset grayscale target range. For example, for an 8-bit grayscale image, the preset grayscale target range could be 100-120. If the average grayscale value of the current image is lower than this range, it indicates that the image is too dark, and the system will instruct an increase in backlight brightness; if it is higher than this range, it indicates that the image is too bright, and the system will instruct a decrease in backlight brightness. Through a PID controller or other adaptive control algorithms, the intelligent control host can precisely and dynamically adjust the backlight brightness, thereby stabilizing the average grayscale value of the acquired fabric image within the target range and ensuring consistent image quality.

[0090] Through the above technical solution, this application can effectively cope with fluctuations in ambient light in the production environment. When the ambient light changes, the brightness of the backlight will automatically adjust accordingly, thereby ensuring that the average gray value of the fabric image acquired by the vision sensor remains stable within the preset gray target range. This stable image quality provides high-quality and consistent input data for subsequent image preprocessing steps (such as median filtering for noise reduction, contrast-limited adaptive histogram equalization), dynamic region of interest extraction, grid feature parameter extraction, and multi-template standard comparison. It avoids images that are too bright or too dark due to changes in ambient light, thereby significantly improving the accuracy, stability, and robustness of defect detection, reducing false alarms and false negatives, and ensuring the reliable operation of the online detection system.

[0091] This application further proposes a method for extracting dynamic regions of interest, specifically: In step three, the method for extracting dynamic regions of interest is as follows: performing row grayscale projection and column grayscale projection on the preprocessed image, determining the left and right boundaries of the fabric based on the abrupt change points at the edges of the projection curves, and further determining the upper and lower boundaries based on the periodic characteristics of the latitude and longitude lines, thus forming a rectangular effective detection window that dynamically changes with the position of the fabric.

[0092] Specifically, the intelligent control host performs row and column grayscale projection on the preprocessed image. Row grayscale projection involves accumulating or averaging the grayscale values ​​of all pixels in each row of the image to obtain a one-dimensional curve reflecting the grayscale distribution of the image in the vertical direction. Similarly, column grayscale projection involves accumulating or averaging the grayscale values ​​of pixels in each column of the image to obtain a one-dimensional curve reflecting the grayscale distribution of the image in the horizontal direction. These projection curves can intuitively show the changing trend of image content in specific directions, especially at the boundary between the fabric and non-fabric areas, where significant statistical abrupt changes in grayscale values ​​occur. The intelligent control host determines the left and right boundaries of the fabric by analyzing the edge abrupt change points of these projection curves. For example, in the column grayscale projection curve, the fabric area usually exhibits a relatively high grayscale value, while the grayscale values ​​of the non-fabric areas on both sides will be significantly different. By calculating the first or second derivative of the projection curve, or by setting an appropriate grayscale threshold, points of abrupt change in grayscale values ​​from the background to the fabric and from the fabric to the background can be accurately detected; these abrupt change points correspond to the precise horizontal edges of the fabric. Building upon this foundation, to further refine the longitudinal range of the effective detection area, the intelligent control unit determines the upper and lower boundaries based on the periodic characteristics of the warp and weft threads. Within the defined transverse region of the fabric, local image analysis is performed, such as Fourier transform or autocorrelation analysis of the grayscale projection curve, to identify the inherent periodic texture of the fabric. The periodicity of the warp and weft threads manifests as stable peaks and troughs on the projection curve. By identifying stable regions of these periodic features, the effective start and end positions of the fabric in the longitudinal direction can be determined, thereby eliminating non-standard weaving areas or deformation areas caused by uneven tension at the beginning and end of the fabric. Finally, based on the precisely determined left, right, upper, and lower boundaries, the intelligent control unit forms a rectangular effective detection window that dynamically changes with the position of the fabric. Since the fabric may experience slight transverse swaying or longitudinal stretching during transport, these boundaries are dynamically adjusted in real time. The intelligent control unit continuously performs projection analysis and feature recognition, updating the coordinates and size of this rectangular window based on the latest boundary information to ensure that subsequent mesh feature parameter extraction is always performed within the most accurate and effective fabric region.

[0093] Through the above technical solution, the intelligent control host can utilize the statistical characteristics of the overall grayscale distribution of the image, and accurately identify the edge abrupt change points between the fabric and the background through row grayscale projection and column grayscale projection, thereby stably determining the left and right boundaries of the fabric. Based on this, the inherent periodic texture characteristics of the warp and weft threads are further combined to determine the upper and lower boundaries, effectively avoiding misjudgments caused by irregularities or local deformations at the beginning and end of the fabric. This dynamic rectangular effective detection window can adapt in real time to the positional shift or slight deformation of the fabric during transport, ensuring that subsequent mesh feature parameter extraction is always performed within the most accurate and effective area. This significantly improves the accuracy of defect detection and the robustness of the system, effectively eliminates interference from non-detection areas on both sides of the fabric, and enhances overall detection efficiency and reliability.

[0094] In a further embodiment, this application proposes that in step four above, the Canny edge detection operator is used for latitude and longitude edge detection. After edge extraction, the main straight line clusters in the longitude and latitude directions are identified through Hough line transform. The pixel distance between adjacent parallel lines is calculated and converted into physical dimensions to obtain the longitude and latitude spacing. The single grid area is calculated by using the closed quadrilateral formed by the intersections of the detected longitude and latitude lines and converted into physical area. Specifically, when the intelligent control host performs latitude and longitude edge detection on the image within the effective detection area, the Canny edge detection operator is used. The Canny edge detection operator is a multi-level edge detection algorithm that smooths the image through Gaussian filtering to remove noise, calculates the gradient intensity and direction of the image, then performs non-maximum suppression to refine the edges, and determines the final edges through double threshold hysteresis processing. This method can effectively identify real edges in the image while suppressing false edges caused by noise, thus clearly outlining the longitude and latitude structure of the filter fabric and providing high-quality edge information for subsequent feature extraction.

[0095] After edge extraction, the intelligent control host identifies the main clusters of straight lines in the longitude and latitudinal directions using the Hough transform. The Hough transform is an effective technique for identifying straight lines in an image; it maps points in image space to curves in parameter space, causing the points constituting a straight line in image space to converge at a single intersection point in parameter space. Since the warp and weft threads of the filter fabric are typically parallel and periodically distributed, the Hough transform can effectively identify these main clusters of straight lines. By analyzing the direction of these lines, the sets of straight lines in the longitude and latitudinal directions can be accurately distinguished.

[0096] Subsequently, the intelligent control host calculates the pixel distance between adjacent parallel lines and converts it into physical dimensions to obtain the meridian spacing and parallel spacing. Once the clusters of lines in the meridian and parallel directions are identified through the Hough linear transform, the lines in each direction can be sorted. For the meridian clusters, the pixel distance between two adjacent parallel lines is calculated; similarly, for the parallel clusters, the pixel distance between two adjacent parallel lines is calculated. These pixel distances are then converted into actual physical dimensions based on the calibration parameters of the visual sensor acquisition system (e.g., the actual physical size corresponding to each pixel), thus obtaining the accurate meridian and parallel spacing.

[0097] Furthermore, the area of ​​a single grid cell is calculated by examining the closed quadrilaterals formed by the intersections of detected meridians and parallels, and then converted into physical area. After identifying the meridians and parallels, the vertices of the grid are determined by calculating the intersections of these lines. Typically, a grid cell is an approximate quadrilateral enclosed by four lines (two meridians and two parallels). By determining the pixel coordinates of these intersections, the pixel area of ​​each closed quadrilateral (i.e., a single grid cell) can be calculated. For example, the area of ​​a quadrilateral enclosed by four intersections can be calculated using the polygon area calculation formula. Finally, the calculated pixel area is also converted into the actual physical area according to the system's calibration parameters.

[0098] Through the above technical solution, the Canny edge detection operator effectively suppresses image noise and accurately extracts the true edge information of the filter fabric, avoiding the false edges or edge breaks that may occur in traditional edge detection methods. Based on this, by combining Hough line transform to identify the main straight line clusters in the warp and weft directions, the periodic structure of the filter can be robustly located, accurately identifying the main warp and weft directions even in cases of slight fabric deformation or local uneven texture. Furthermore, by calculating the pixel distance between adjacent parallel lines and converting it to physical dimensions, and by calculating the physical area of ​​a single grid through the closed quadrilateral formed by the detected warp and weft intersections, this application can obtain high-precision key feature parameters such as warp spacing, weft spacing, and single grid area. These precise parameters provide a reliable quantitative basis for subsequent intelligent identification and classification of defect types, significantly improving the accuracy and stability of defect detection, effectively reducing false alarm and false negative rates, thereby ensuring the weaving quality of the air conditioner filter.

[0099] In practical applications, simply weighting and summing the deviation rates may not fully reflect the differences in importance of different characteristic parameters for defect identification, nor can it accurately reflect the severity of the impact of different types of defects on product quality. This may lead to insufficient sensitivity to certain critical defects or overreaction to non-critical changes, affecting the accuracy and reliability of detection. In this regard, this application further proposes that in step six, the formula for calculating the comprehensive difference degree D is: D = α·Δd_warp + β·Δd_weft + γ·ΔA + δ·ΔG_warp + ε·ΔG_weft + ζ·ΔP, where Δd_warp is the deviation rate of meridian spacing, Δd_weft is the deviation rate of parallel spacing, ΔA is the deviation rate of single grid area, ΔG_warp is the deviation rate of meridian gray-level gradient change rate, ΔG_weft is the deviation rate of parallel gray-level gradient change rate, and ΔP is the deviation rate of the number of grid intersections; α, β, γ, δ, ε, and ζ are the weight coefficients of the corresponding terms, and α+β+γ+δ+ε+ζ=1.

[0100] This calculation formula defines the quantification method of the overall difference degree D. Its core lies in comprehensively and precisely quantifying the overall deviation between the real-time detected fabric grid features and the preset standard by weighted summation of the deviation rates of multiple key grid feature parameters. This weighted summation mechanism allows the system to assign different weights to different feature parameters based on their actual importance in defect judgment, thereby reflecting the quality status of the fabric surface more accurately and flexibly.

[0101] Specifically, Δd_warp represents the relative difference between the real-time detected warp spacing and the standard warp spacing. It is typically calculated as: (Real-time warp spacing - Standard warp spacing) / Standard warp spacing. This parameter directly reflects the weaving uniformity of the fabric in the warp direction; abnormal changes may indicate problems such as uneven warp density, broken yarns, or yarn bunching. Δd_weft represents the relative difference between the real-time detected weft spacing and the standard weft spacing. Its calculation method is similar to the warp spacing deviation rate, reflecting the weaving uniformity of the fabric in the weft direction. Abnormal changes may indicate problems such as uneven weft density, weft breaks, or weft skipping. ΔA represents the relative difference between the real-time detected area of ​​a single grid and the area of ​​a standard single grid. It is typically calculated as: (Real-time single grid area - Standard single grid area) / Standard single grid area. This parameter comprehensively reflects the coordinated changes in warp and weft spacing and is highly sensitive to defects involving changes in the overall grid structure, such as uneven mesh size and holes. ΔG_warp represents the relative difference between the real-time detected grayscale gradient change rate along the warp direction and the standard value. The grayscale gradient change rate reflects the smoothness, uniformity of thickness, and presence of foreign matter attachment in the yarn. Anomalies in this value may indicate warp surface defects, contamination, or localized abnormal weaving density. ΔG_weft represents the relative difference between the real-time detected grayscale gradient change rate along the weft direction and the standard value. Its calculation method is similar to the warp grayscale gradient change rate deviation rate. Anomalies in this value may indicate weft surface defects, contamination, or localized abnormal weaving density. ΔP represents the relative difference between the real-time detected number of grid intersections per unit area and the standard value. The number of grid intersections is an important indicator of the integrity and density of the grid structure. Anomalies in this value may indicate missing grid structures, holes, skipped threads, or severe weaving irregularities.

[0102] Weighting coefficients α, β, γ, δ, ε, and ζ are used to adjust the relative importance of each deviation rate in the calculation of the overall difference degree D. By assigning different weights to different characteristic parameters, the system can flexibly adjust the focus of defect judgment based on actual production experience, the impact of different defect types on product quality, and the sensitivity of specific varieties to certain characteristics. For example, for products with high requirements for mesh uniformity, the weight of ΔA can be appropriately increased; for products with high requirements for yarn continuity, the weights of Δd_warp and Δd_weft can be increased. The sum of the weighting coefficients is 1, ensuring the dimensional consistency and comparability of the overall difference degree D.

[0103] Through the above technical solution, this application can more precisely quantify the deviation between the fabric grid features and the standard. Compared with simple deviation rate summation, the introduction of weighting coefficients α, β, γ, δ, ε, and ζ allows the intelligent control host to flexibly adjust the importance of each parameter in the comprehensive evaluation based on the differences in the impact of different defect types on product quality and the sensitivity of different feature parameters to defect performance. For example, for serious defects such as broken warp threads or holes, the weights of parameters such as the corresponding warp spacing deviation rate Δd_warp and single grid area deviation rate ΔA can be set higher, thereby ensuring that these key defects can be identified earlier and more accurately. This weighted summation method allows the comprehensive difference degree D to more accurately reflect the actual quality status of the fabric, effectively avoiding misjudgments caused by fluctuations in a single parameter or the accumulation of slight fluctuations in multiple secondary parameters. It also improves the detection capability for complex, multi-feature defects, significantly enhancing the accuracy and reliability of defect detection.

[0104] In actual production, the loom's operating speed V and the specifications (such as mesh count) of the produced filter screens constantly change. If the first dynamic threshold T1 and the second dynamic threshold T2 are fixed values, they may not be able to adapt to these dynamic changes, leading to excessive false alarms or missed alarms during high-speed production or high-mesh count product testing, affecting the accuracy and stability of the detection system. To address this, this application further proposes that in step six, the comprehensive difference calculation and initial screening judgment step, the first dynamic threshold T1 and the second dynamic threshold T2 are automatically adjusted according to the current loom operating speed V and the current product specifications. The adjustment rules are: the higher the speed V, the smaller the values ​​of T1 and T2; the higher the mesh count of the product, the smaller the values ​​of T1 and T2.

[0105] Specifically, the first dynamic threshold T1 is used in the initial screening stage of defect detection. When the overall difference D of the fabric area exceeds T1, the area is initially judged as a suspected defect. The second dynamic threshold T2 is used for secondary verification of suspected defects. Only when the overall difference D' after verification exceeds T2 is it finally confirmed as a real defect. These two thresholds are key parameters for the system to determine whether there are defects on the fabric surface and the severity of the defects. Their values ​​directly affect the sensitivity and specificity of the detection. The current operating speed V of the loom refers to the real-time speed at which the weaving equipment drives the air conditioner filter fabric forward. In high-speed production mode, the time window for the vision sensor to acquire images is relatively short, and the images may have a slight blurring or increased noise trend. At the same time, the real-time processing capability of the detection system is more demanding. Therefore, the detection threshold needs to be adjusted accordingly to adapt to the impact of this high-speed movement. The current production product specifications mainly refer to the weaving density or mesh count of the air conditioner filter. Different types of filters have significant differences in their mesh structure, yarn thickness, mesh size, and other characteristic parameters. For example, high-mesh-count (i.e., high-density) filters have a finer mesh structure and may have a lower tolerance for defects; in other words, even very small deviations may be considered defects. Therefore, different inspection standards and thresholds are needed for different product varieties and specifications. Automatic adjustment refers to the intelligent control host autonomously calculating and updating the values ​​of the first dynamic threshold T1 and the second dynamic threshold T2 based on real-time loom operating speed V and current product variety and specification information, using preset algorithms or rules. This adjustment requires no manual intervention and ensures that the detection system is always in optimal working condition, adapting to changes in the production environment. This adjustment rule reflects the adaptability of the detection system to different production conditions. When the loom operating speed V is higher, in order to ensure accurate capture of defects in a shorter detection time, the system needs to increase detection sensitivity, that is, respond more quickly to small deviations. Therefore, the values ​​of T1 and T2 are reduced, so that even a smaller overall difference D or D' can trigger defect judgment. Similarly, when the mesh count of the produced filter is higher, it means the mesh structure is finer and the tolerance for defects is lower. The system needs to more strictly identify subtle structural anomalies, so the values ​​of T1 and T2 need to be reduced to improve the ability to detect fine defects. Conversely, when producing products at low speed or with low mesh count, the threshold can be appropriately relaxed to reduce unnecessary false alarms.

[0106] Through the above technical solution, the first dynamic threshold T1 and the second dynamic threshold T2 can be automatically adjusted according to the current operating speed V of the loom and the current product specifications. When the loom operating speed V increases, the system will correspondingly decrease the values ​​of T1 and T2, thereby improving the detection sensitivity and ensuring that even under high-speed production conditions, minute defects can still be effectively captured, avoiding missed detections. Simultaneously, when the mesh count of the produced filter mesh is higher, the system will also decrease the values ​​of T1 and T2, enabling more stringent defect judgment for high-precision, high-density filter mesh products, effectively identifying minute structural anomalies, and ensuring product quality. This dynamic adaptive threshold adjustment mechanism allows the defect detection system to better adapt to complex changes in the production environment, significantly reducing false alarm and missed detection rates, improving detection accuracy and reliability, and thus enhancing the overall production quality and efficiency of air conditioning filter meshes.

[0107] This application further proposes that when multiple types of defects are detected overlapping in the same local area, the intelligent control host uses the most severe type as the primary classification result for that area and appends the identification codes of all detected types to the log. Specifically, the intelligent control host is the core processing unit of the entire detection system, responsible for receiving image data collected by the vision sensor and performing a series of complex tasks such as image preprocessing, feature extraction, defect judgment, type identification, grade classification, alarm control, and data storage and uploading. It is typically composed of a high-performance industrial computer, embedded processor, or dedicated FPGA / DSP chip, equipped with corresponding image processing algorithms and control software, capable of processing large amounts of data and making decisions in real time and efficiently. In this embodiment, the intelligent control host further undertakes the function of intelligently judging and managing information on complex defect situations. The simultaneous detection of multiple types of defects overlapping in the same local area refers to the system identifying, through image analysis and feature extraction, that a specific small area of ​​the air conditioner filter fabric simultaneously meets the judgment criteria for two or more different defect types. For example, an area might simultaneously exhibit characteristics of "holes" (abrupt changes in grid area, irregular tearing of boundaries) and "broken warp threads" (abrupt grayscale bands along the warp direction, interruption of latitude continuity). Such overlapping defects usually indicate a more complex and severe quality problem in that area. When the intelligent control host identifies multiple overlapping defects in the same local area, it selects the most severe type as the final, primary classification result for that area based on preset defect severity levels or priority rules. For instance, if an area contains both "sparse-dense lines" and "holes," and "holes" are defined as a more severe defect than "sparse-dense lines" in quality standards, the intelligent control host will determine the primary classification result for that area as "holes." This approach ensures that, with limited alarm or tagging resources, priority is given to addressing and resolving issues with the greatest impact on product quality. Furthermore, in addition to determining a primary classification result, the intelligent control host also records the unique identifiers or descriptions of all identified defect types (including primary and secondary classification types) within that local area into the system's detection log or database. For example, if an area is classified as "hole" by the main category, but "broken warp thread" is also detected at the same time, the log will record "Main category: hole, additional type: broken warp thread". The purpose of this approach is to retain complete defect information, which is convenient for subsequent quality traceability, process analysis, defect pattern research, and system self-learning optimization. Even if the main category only reflects the most serious problem, other potential problems will not be missed.

[0108] Through the above technical solution, when multiple types of defects overlap in the same local area on the air conditioner filter fabric, the intelligent control host can intelligently use the most severe type as the primary classification result for that area. This avoids classification ambiguity or decision-making difficulties caused by the coexistence of multiple defects, ensuring that the quality control system can prioritize identifying and responding to defects that have the greatest impact on product quality, thereby improving the effectiveness and targeting of alarm and shutdown control. Simultaneously, the addition of identification codes for all detected types to the log ensures that all detected defect information is completely preserved even after the primary classification is determined. This provides comprehensive and detailed data support for subsequent quality analysis, process improvement, defect pattern research, and system self-learning optimization, greatly enhancing the intelligence level and data value of the defect detection system, and contributing to a deeper understanding and solution of complex quality problems in the weaving process.

[0109] In its implementation, simply identifying the type of defect may not be sufficient to fully assess its impact on product quality. For example, a tiny hole and a large, minor stain, although different in type, may have significantly different degrees of impact on product performance and subsequent processing. Without detailed grading, all defects may be treated the same, resulting in a lack of specificity in quality control strategies, an inability to effectively differentiate between priorities, and consequently, an impact on production efficiency and the rationality of resource allocation. In response, this application further proposes the following specific rules for grade classification in step nine: if the defect area S is less than the first area threshold S1 and the number of defects N per unit length is less than the first quantity threshold N1, and the defect type is not a broken warp thread, hole, or unclassified structural defect, then it is classified as a minor defect; if S is between S1 and S2 or N is between N1 and N2, and the defect type is not an unclassified structural defect, then it is classified as a general defect; if S is greater than the second area threshold S2 or N is greater than the second quantity threshold N2 or the defect distribution density ρ is greater than the density threshold ρ0, or the defect type is an unclassified structural defect, then it is classified as a severe defect, where S1 is less than S2 and N1 is less than N2.

[0110] Specifically, the defect area S refers to the physical area occupied by the region identified as a real defect by the intelligent control host in step eight. This area can be calculated by multiplying the number of pixels of the defect in the image by the actual physical area corresponding to a single pixel. For example, in an image acquired by a vision sensor, the pixels of the defect area are identified and counted, and then converted into the actual physical area according to the calibration parameters of the vision system (such as the number of pixels per millimeter), usually in square millimeters (mm²). The number of defects per unit length N refers to the total number of defects detected within a certain preset length range of the fabric (e.g., per meter or per ten meters). When recording the location coordinates of the defects, the intelligent control host can simultaneously record the longitudinal position of the defects in the direction of travel on the fabric surface. By statistically analyzing the number of defects appearing within a specific longitudinal length range, the number of defects per unit length N can be obtained, which helps to assess the degree of concentration of defects on the fabric surface. The defect distribution density ρ refers to the ratio of the number of defects in a specific area to the area of ​​that area, reflecting the density of defects on the fabric surface. The intelligent control host can define a sliding window or a fixed detection area, count the number of defects within that area, and divide the count by the area of ​​that area to calculate the defect distribution density ρ. For example, the number of defects can be counted within a 10cm x 10cm area to obtain the number of defects per square centimeter.

[0111] The first area threshold S1 and the second area threshold S2 are area boundary values ​​used to distinguish defects of different severity, where S1 is less than S2. These thresholds are preset based on experience, product quality standards, or historical data. For example, S1 can be set to 5 mm², and S2 can be set to 20 mm². When the defect area is less than S1, it may be considered a minor defect; when the area is between S1 and S2, it may be a general defect; when the area is greater than S2, it may be a serious defect. These thresholds can be adjusted according to different product specifications and quality requirements. The first quantity threshold N1 and the second quantity threshold N2 are quantity boundary values ​​used to distinguish defects of different severity, where N1 is less than N2. These thresholds are also preset based on experience, product quality standards, or historical data. For example, N1 can be set to 2 defects per meter, and N2 can be set to 5 defects per meter. When the number of defects per unit length is less than N1, it may be considered a minor defect; when the number is between N1 and N2, it may be a general defect; when the number is greater than N2, it may be a serious defect. The density threshold ρ0 is a boundary value used to distinguish the distribution density of defects of different severity. This threshold is preset based on experience or quality standards. For example, ρ0 can be set to 0.1 defects per square centimeter. When the defect distribution density exceeds ρ0, even if the area of ​​a single defect is not large, it may indicate a serious quality problem in that area, and thus be classified as a severe defect.

[0112] Minor defects, general defects, and severe defects are classifications of defects, representing different degrees of impact on product quality. These classifications are determined based on a comprehensive assessment of the area, quantity, density thresholds, and defect type. Minor defects typically refer to defects with minimal impact on product function and appearance; general defects refer to defects that have some impact on product function or appearance but are still acceptable or require rework; severe defects refer to defects that have a significant impact on product function or appearance and may lead to product scrapping. When determining minor defects, certain inherently more serious defect types are excluded, i.e., defects that do not fall under the categories of broken warp threads, holes, or unclassified structural defects. The intelligent control host has already identified the defect type in step eight. When classifying the defects, it first checks whether the defect type is a broken warp thread, a hole, or an unclassified structural defect. If it belongs to these types, even if the area and quantity meet the criteria for a minor defect, it will not be classified as a minor defect, but rather at least as a general defect or directly as a severe defect, to ensure strict handling of these critical defects. Similarly, in general defect assessment, unclassified structural defects are excluded, meaning defects whose type does not belong to the unclassified structural defect category. If the defect type is an unclassified structural defect, it is directly classified as a serious defect, because such defects usually indicate structural flaws that the system cannot identify and that may have a potentially serious impact.

[0113] Through the above technical solution, this application can perform refined classification of detected defects, overcoming the limitations of making a single judgment based solely on defect type. This classification mechanism comprehensively considers multiple dimensions such as defect area S, number of defects N per unit length, defect distribution density ρ, and defect type, so that even defects of the same type can be accurately classified into minor defects, general defects, or serious defects based on their specific size, quantity, density, and inherent severity (such as broken warp threads, holes, and unclassified structural defects). For example, a tiny hole may be more damaging than a large area of ​​minor stain. This solution ensures priority attention to critical defects by directly elevating holes to at least the general defect level.

[0114] In step eight, "unclassified structural defects" is set as a catch-all category to address two special situations that may occur during actual weaving: First, defects present a complex form, such as a broken yarn that wraps around and tears multiple surrounding meshes. This area simultaneously exhibits some characteristics of warp yarn breakage and holes, but neither fully meets their respective typical criteria. Second, equipment exhibits novel or sporadic anomalies, such as a broken reed causing penetrating abrasions on the fabric surface, or foreign object compression causing overall mesh distortion. Such defects have no corresponding matching item in the six typical criteria in step eight. In the above situations, this area has been confirmed as a genuine defect through the comprehensive difference degree D'≥T2 in step seven. That is, the system has confirmed that this area has a significant structural anomaly exceeding the allowable range of the standard template, but cannot classify it into any known specific type. If it is forcibly classified into one of the six categories at this time, it will lead to the distortion of type information and mislead subsequent level determination and manual repair decisions; if it is directly ignored or simply recorded as "unknown" without level classification, the actual structural defect may be downgraded or even missed in the classification process.

[0115] For the reasons stated above, this application categorizes genuine defects that cannot match the six criteria separately as "unclassified structural defects." It is important to clarify that "unclassified" does not mean the system is uncertain about the nature of the defect, but rather that the system has clearly identified a structural anomaly in the area that does not conform to the known six morphological characteristics. In step nine, the mandatory rule of directly classifying this type as a severe defect is based on the following engineering considerations: known, categorizable defects often have mature contingency plans and clear repair or handling methods, while "unclassified" means that the anomaly exceeds the system's experiential knowledge, and its potential quality risks are unpredictable. From the perspective of industrial quality control safety principles, when the system state cannot accurately match known patterns, it should lean towards the most conservative approach, i.e., handle it according to the strictest level, to ensure that any unknown type of structural defect is not underestimated or overlooked due to "unclassification." This "strict handling" strategy conforms to the fail-safe principle commonly used in the field of industrial automation testing, i.e., when the system cannot make a deterministic judgment, it must choose the safest response method for product quality.

[0116] This refined classification of levels enables subsequent alarm and shutdown controls (step ten) to respond more accurately to actual quality requirements, avoids excessive intervention in minor defects, and ensures timely handling of serious defects, thereby significantly improving the level of intelligent management of the production process and the effectiveness of product quality control.

[0117] In a further preferred embodiment, this application further proposes that the specific method of the three-level audible and visual alarm in step ten is as follows: the first-level alarm only displays prompt information on the human-machine interface and records the log, without triggering audible and visual signals; the second-level alarm triggers the yellow indicator light to stay on and emits an intermittent buzzing sound; the third-level alarm triggers the red indicator light to flash and emits a continuous high-decibel buzzing sound.

[0118] Specifically, when the intelligent control host determines a minor defect and triggers a Level 1 alarm, it only displays relevant prompts on the human-machine interface (HMI), such as the defect type, location, and time of occurrence, while simultaneously recording this information in the system log. The HMI, serving as the platform for operator interaction with the intelligent control host, is typically presented via a touchscreen or monitor, clearly conveying non-urgent information through text, icons, or charts. Log recording ensures the traceability of all minor defect events, facilitating subsequent data analysis, quality management, and process optimization. In this situation, the system does not trigger any audible or visual signals to avoid unnecessary interference with operators and prevent decreased operator alertness due to frequent non-urgent alarms, thereby maintaining a relatively quiet production environment and operator focus.

[0119] When the intelligent control host determines a minor defect and triggers a level-two alarm, the system will activate a solid yellow indicator light and emit an intermittent buzzer sound. A solid yellow indicator light is a visual alarm; yellow commonly represents warning or caution in industrial safety signage, and its continuous illumination alerts operators to a problem that requires attention but is not an immediate emergency. The intermittent buzzer sound is an auditory alarm; the sound is emitted intermittently, such as once or several times at regular intervals. The volume and frequency are set within a range that will attract the operator's attention but is not overly harsh, aiming to moderately alert the operator to the presence of a minor defect, prompting them to inspect or take preventative measures.

[0120] When the intelligent control host determines a serious defect and triggers a level three alarm, the system will activate a flashing red indicator light and emit a continuous high-decibel buzzer. The flashing red indicator light is a strong visual alarm; red in industrial safety signage typically represents danger or urgency, and its flashing quickly attracts the operator's attention, indicating a serious problem requiring immediate attention. The continuous high-decibel buzzer is an emergency auditory alarm; its continuous and loud sound aims to inform the operator of the serious defect in the most direct and forceful way, requiring immediate intervention, such as inspecting the fabric or stopping the machine.

[0121] By combining different levels of defect alarms with specific and intuitive audio-visual signals and human-machine interface prompts, the above technical solution effectively addresses the problems of insufficient differentiation and untimely response that may exist in traditional alarm methods. Level 1 alarms are displayed on the human-machine interface and logged, maintaining traceability of information without interfering with operation. Level 2 alarms use a solid yellow indicator light and intermittent buzzer sound to moderately attract the operator's attention, prompting them to pay attention to general defects and avoiding excessive alarms. Level 3 alarms use a flashing red indicator light and a continuous high-decibel buzzer sound to forcefully remind the operator of serious defects in the most conspicuous and urgent way, prompting them to take immediate intervention measures. This significantly improves the operator's efficiency in identifying defects of different severity levels and their response speed, ensuring quality control and safe production during the production process and avoiding greater losses caused by untimely defect handling. This graded and differentiated alarm mechanism makes alarm information clearer and more intuitive, improving the system's practicality and reliability.

[0122] In a further preferred embodiment, this application further proposes that in step eleven, the inkjet marking device uses fast-drying industrial marking ink, and the inkjet marking is a long strip mark along the fabric edge direction, with the marking length being positively correlated with the lateral dimension of the defect in the width direction of the fabric surface.

[0123] Specifically, fast-drying industrial marking inks refer to specialized inks with rapid drying properties suitable for industrial production environments. Their main characteristic is that they cure or penetrate and dry in a very short time after being sprayed onto the fabric surface, thus preventing ink spread, blurring, or erasure by subsequent processes, ensuring the clarity and durability of the markings. Furthermore, industrial marking inks typically possess excellent abrasion resistance, water resistance, and chemical corrosion resistance to adapt to various environmental conditions during weaving and post-processing. Sprayed markings are elongated strips along the fabric edge, meaning the marking has a certain length and a relatively small width, with its longitudinal axis parallel to the fabric edge. This design aims to make the markings clearly visible in the fabric edge area without occupying the effective detection area of ​​the fabric, avoiding unnecessary impact on subsequent processing or product appearance. Compared to dot markings, elongated strip markings have higher visibility and recognizability, facilitating quick location of defect areas by operators. The marking length is positively correlated with the lateral dimension of the defect in the fabric width direction, meaning the length of the sprayed strip marking is adjusted according to the actual size of the detected defect in the fabric width direction. When the defect is large in width, the length of the inkjet marking increases accordingly; conversely, when the defect is small, the marking length is shorter. This positive correlation can be achieved by the intelligent control unit dynamically adjusting the inkjet marking device's spraying time or spraying stroke based on the defect size information identified in step eight. For example, the intelligent control unit can look up the appropriate marking length from a preset table of correspondence between marking length and defect width or calculate it using a formula, and then instruct the inkjet marking device to execute the marking.

[0124] Through the above technical solutions, the inkjet marking device uses fast-drying industrial marking ink, ensuring the clarity, durability, and stability of the markings in industrial environments, effectively avoiding problems such as blurring or detachment. Simultaneously, the inkjet markings are designed as elongated strips along the fabric edge, providing higher visibility and easy identification without affecting the main fabric area, facilitating quick defect location by operators. Furthermore, by making the marking length positively correlated with the lateral dimension of the defect in the fabric width direction, the marking not only indicates the location of the defect but also intuitively reflects its size, providing richer and more accurate reference data for subsequent manual inspection, repair, or automated cutting. This significantly improves the efficiency and accuracy of defect handling, thereby enhancing the intelligence level of the entire production process and product quality management capabilities.

[0125] This application further proposes a self-learning optimization method, which includes: operators confirming the authenticity of each alarm through a human-machine interface; the system automatically constructs a false alarm sample set and a missed alarm sample set based on the feedback results; and performs feature clustering analysis on the false alarm and missed alarm samples. When the cumulative number of samples of a certain feature exceeds a preset threshold, the system automatically adjusts the weight coefficient of that feature in the comprehensive difference degree D. Specifically, operators confirm the authenticity of each alarm through a human-machine interface. This human-machine interface can be a touch screen or a monitor, displaying detailed information such as the image, type, and location of the alarm defect, and providing options such as "confirm real defect," "confirm false alarm," or "confirm missed alarm" for operators to select. By introducing professional human judgment, the system can obtain high-quality supervisory data, providing reliable real labels for subsequent refined learning. Based on the operator's feedback results, the intelligent control host automatically constructs the false alarm sample set and the missed alarm sample set. When an operator marks an alarm as a false alarm, the system stores the corresponding local image patch, extracted feature parameters, and calculated comprehensive difference degree D as false alarm samples in a dedicated false alarm sample library. Similarly, when an operator confirms a missed alarm, the system stores the feature data of the missed fault information manually entered into the missed alarm sample library. These sample sets form the basis for deeper feature analysis. Subsequently, the intelligent control host performs feature clustering analysis on the constructed false alarm and missed alarm sample sets. This analysis aims to identify common feature patterns that lead to false alarms or missed alarms. For example, the system can employ clustering algorithms such as K-means and DBSCAN, using the feature vectors of each sample (e.g., meridian spacing deviation rate Δd_warp, parallel spacing deviation rate Δd_weft, single grid area deviation rate ΔA, meridian gray-level gradient change rate deviation rate ΔG_warp, parallel gray-level gradient change rate deviation rate ΔG_weft, grid intersection number deviation rate ΔP, etc.) as input to identify which feature combinations or feature value ranges frequently lead to erroneous judgments. Through clustering, the system can group false positives or false negatives with similar features into one category, thereby identifying the underlying causes. When the cumulative number of samples of a certain feature exceeds a preset threshold, the intelligent control host automatically adjusts the weight coefficient of that feature in the comprehensive difference degree D. Specifically, if cluster analysis reveals that a certain type of false positive sample performs exceptionally well on a specific feature (e.g., the deviation rate of the grayscale gradient of meridians ΔG_warp), and the number of samples in this type reaches a preset threshold, the system will appropriately reduce the weight coefficient δ corresponding to that feature to decrease its contribution to the overall difference D, thereby reducing the occurrence of false positives of that type. Conversely, if a certain type of missed positive sample does not perform well on a certain feature, the system will appropriately increase the weight of that feature to enhance its sensitivity. This feature-specific weight adjustment allows the system to more accurately optimize the detection model, avoiding the potential side effects of global adjustments.

[0126] By incorporating an operator feedback mechanism through the aforementioned technical solution, the system can acquire high-quality supervisory data and construct accurate false alarm and false negative sample sets. Furthermore, feature clustering analysis is performed on these samples, enabling the intelligent control host to identify specific feature patterns leading to false alarms or false negatives, rather than relying solely on global statistical data. When a specific error pattern accumulates to a certain level, the system can selectively adjust the weight coefficients of the corresponding features in the comprehensive difference degree D. This refined self-learning optimization mechanism allows the system to more accurately identify different types of defects, effectively reducing false alarm and false negative rates, significantly improving the accuracy and robustness of defect detection, and avoiding over-adjustment or under-adjustment problems that may arise from global threshold adjustments, thereby ensuring the stability and reliability of the air conditioning filter weaving quality.

[0127] In some embodiments described above, the detection of defects in the fabric of air conditioner filters is achieved by comparing the real-time extracted feature parameters with pre-stored standard feature parameters. However, during long-term operation, factors such as light source aging, changes in ambient light, or loom vibration may cause overall grayscale shifts or subtle changes in grid features in the images acquired by the vision sensor. This can lead to discrepancies between the pre-stored standard template data and the actual fabric condition, thereby affecting the accuracy and stability of defect detection and increasing the risk of false alarms or missed alarms.

[0128] To address this, this application further proposes a template consistency verification step prior to the multi-template standard comparison in step five. This step aims to monitor the overall characteristics of the fabric image in real time and dynamically adjust or update the standard template data based on the monitoring results, thereby effectively addressing image drift issues caused by environmental changes or equipment aging and ensuring the reliability of subsequent defect determination.

[0129] Specifically, the intelligent control host performs statistical analysis on the mean background grayscale value and grid periodicity features of the continuously acquired N0 frames of images. The mean background grayscale value reflects the overall brightness level of the fabric image and can be obtained by averaging the pixel grayscale values ​​in the non-grid areas of the N0 frames. The grid periodicity features characterize the structural regularity of the filter fabric, such as the average value and fluctuation of the warp and weft spacing, and can be extracted by performing Fourier transform, autocorrelation analysis, or statistical analysis based on edge detection results on the N0 frames. By statistically analyzing multiple frames, the impact of noise from a single frame on feature extraction can be effectively reduced, improving the robustness of the statistical results.

[0130] Subsequently, the intelligent control host compares the currently obtained background grayscale mean and grid periodicity features with the corresponding values ​​recorded during the last system calibration or template update. Here, "deviation" can be an absolute difference, a relative difference, or a statistical distance. When the deviation of the statistical value relative to the previous calibration value exceeds a preset template update threshold, it indicates that the overall characteristics of the fabric image have changed significantly and may no longer match the current standard template data. Once a statistical deviation exceeding the preset threshold is detected, the intelligent control host will automatically trigger the standard template data update process. This update process may include re-acquiring a new batch of fabric images and recalculating and storing new standard background grayscale mean, grid periodicity features, and other relevant standard feature parameters based on these new images, thereby forming a new standard grid texture database. Alternatively, updates can be achieved by fine-tuning or correcting existing standard template data, for example, compensating for standard grayscale values ​​based on the detected grayscale drift. This aims to adapt to image drift caused by light source aging or slow environmental changes. Light source aging leads to a gradual decrease in illumination intensity, resulting in a reduction in the overall grayscale value of the acquired image. Slow environmental changes, such as changes in workshop temperature, humidity, or dust accumulation, can also affect the brightness and contrast of the image. These factors can all cause differences between the image acquired by the vision sensor and the image during initial calibration, a phenomenon known as "image drift."

[0131] Through the aforementioned template consistency verification steps, this application effectively solves the image drift problem caused by light source aging or slow environmental changes. The intelligent control host statistically analyzes the average background grayscale and grid periodicity characteristics of continuously acquired images in real time and compares them with historical calibration values. Once a deviation exceeds a preset threshold, it automatically triggers an update of the standard template data. This ensures that the standard grid texture database used for defect judgment remains consistent with the fabric surface condition in the actual production environment, avoiding misjudgments or missed judgments caused by outdated standard templates. Therefore, this solution significantly improves the robustness and accuracy of the online defect detection system for air conditioner filter mesh weaving, reduces the need for manual intervention and frequent calibration, and guarantees long-term stable and reliable detection performance.

[0132] The following example will provide a more detailed explanation of the above technical solution:

[0133] Inside an air conditioner filter production workshop, a Titan series rapier loom is running at high speed, weaving polypropylene air conditioner filter fabric. To solve the problems of high labor intensity, slow detection, and high missed detection rate of traditional manual inspection, as well as the limited functionality of traditional sensors that cannot identify various appearance defects, this loom is equipped with an online detection and alarm system.

[0134] Taking the production of standard 24-mesh (24 mesh holes per inch) air conditioner filter screens on a Titan T8 loom as an example, the loom's operating speed is set to 60cm / min. The vision sensor used is an industrial area array camera with a resolution of 2048×2048 pixels, a lens focal length of 12mm, and a vertical distance of 250mm from the camera's optical center to the fabric surface, resulting in a theoretical pixel resolution of 0.1mm / pixel. The camera's frame rate is dynamically adjusted according to the current loom speed. When the speed is 60cm / min, the frame rate is set to 30fps, ensuring that each frame corresponds to a fabric travel length of approximately 20mm, meeting the requirement for complete acquisition of the periodic texture of the mesh.

[0135] The backlight uses a long strip LED parallel light source with a color temperature of 5000K, and the initial brightness is set to 60% of the maximum brightness. The ambient light sensor monitors the ambient illuminance in real time. When the ambient illuminance fluctuates within the range of 100~500Lux, the brightness automatic adjustment module uses PID closed-loop control to stabilize the average gray value of the acquired image in the range of 150~180 (8-bit grayscale, 0~255 range).

[0136] Median filtering uses a 3×3 pixel window. The cropping limit for contrast-limited adaptive histogram equalization is set to 2.0, dynamically stretching the image grayscale to the full range. During dynamic ROI extraction, row and column grayscale projections are performed on the image. The two extreme points with the largest absolute values ​​of the first derivative of the projection curve are selected as the left and right boundaries of the image, cropping 5% of the non-detection area on each side. The effective detection area width is approximately 90% of the total image width.

[0137] The Canny edge detection low threshold is set to 30 and the high threshold to 80 (based on an 8-bit grayscale image). The Hough linear transform angular resolution is set to 1°, the distance resolution to 1 pixel, and the voting threshold to image height × 0.6. The detected meridian direction angle range is -5° to +5°, and the parallel direction angle range is 85° to 95°. The standard feature parameters of a normal 24-mesh filter are calibrated as follows: meridian spacing 2.15mm (approximately 21.5 pixels), parallel spacing 2.15mm (approximately 21.5 pixels), single grid area 4.62mm² (approximately 462 pixels²), average meridian grayscale gradient change rate 180 (grayscale / pixel), average parallel grayscale gradient change rate 175 (grayscale / pixel), and the theoretical number of grid intersections per unit field of view (100mm × 100mm) is 2304.

[0138] In the formula for calculating the overall difference degree D, the weighting coefficients are set as follows: α=0.25, β=0.25, γ=0.20, δ=0.15, ε=0.10, ζ=0.05. The basis for these settings is as follows: the deviation in warp / weft spacing has the greatest impact on the filter's filtration accuracy, so each is weighted at 0.25; the grid area deviation directly affects the effective filtration area, so its weight is set at 0.20; the grayscale gradient change rate reflects the uniformity of the yarn tension, so its weights are set at 0.15 and 0.10 respectively; the number of intersections is a comprehensive redundancy verification index, so its weight is set at 0.05. The first dynamic threshold T1 is preferably 0.25 for V=60cm / min and 24-mesh, and T2 is preferably 0.15. Initial screening is triggered when D>0.25; a true defect is confirmed when the verification difference degree D'≥0.15. The third threshold T3 is set at 30%, meaning that when the deviation between adjacent grid areas exceeds ±30% of the standard area, it is judged as uneven mesh size.

[0139] Speed ​​Adaptive Rule: When the loom speed V varies within the range of 30~120cm / min, T1 and T2 are automatically adjusted by linear interpolation—T1=0.35, T2=0.25 when V=30cm / min; T1=0.15, T2=0.10 when V=120cm / min. Intermediate speeds are calculated by linear interpolation. Variety Adaptive Rule: When producing 40-mesh high-density filter screens, all line spacing and area thresholds are reduced proportionally (proportioning factor is 24 / 40=0.6), and T1 and T2 are lowered by a fixed level (approximately reduced by 0.05) to accommodate finer texture features.

[0140] Area thresholds: S1 = 5 mm², S2 = 20 mm²; Quantity thresholds: N1 = 2 (pieces / meter), N2 = 5 (pieces / meter); Density threshold ρ0 = 3 (pieces / 100 cm²). Grading rules: When S < 5 mm² and N < 2, and the type is not a broken warp thread, hole, or unclassified structural defect, it is classified as a minor defect; when 5 mm² ≤ S < 20 mm² or 2 ≤ N < 5, and the type is not an unclassified structural defect, it is classified as a general defect; when S ≥ 20 mm² or N ≥ 5 or ρ ≥ 3, or the type is an unclassified structural defect, it is classified as a severe defect. Broken warp threads or holes are at least classified as general defects; holes with an area exceeding S2 are directly classified as severe defects.

[0141] Self-learning trigger conditions: When the system's false alarm rate exceeds 5% within a continuous 8-hour operating cycle, T1 and T2 are automatically increased by 0.02~0.05 each; when the false alarm rate exceeds 2%, T1 and T2 are automatically decreased by 0.02~0.05 each. The adjustment range is proportional to the deviation, and the maximum adjustment range in a single instance does not exceed 0.05, and the cumulative adjustment range per year does not exceed 0.10. Emergency response trigger conditions: When the defect detection frequency exceeds three times the historical average within 5 minutes, the system temporarily decreases T1 and T2 by 0.03 each and triggers a process anomaly warning.

[0142] During operation, the loom runs at a constant speed of 60 cm / min. The system acquires images of the fabric surface in real time. When a frame of the image detects a sudden change in the area of ​​the mesh to 4.2 times the standard area, with an irregular tear at the boundary, the overall difference degree D of this area is calculated to be 0.38, which is greater than T1 (0.25), and it is judged as a suspected defect and its location is recorded. The system then performs local optical zoom magnification on this area to acquire and verify the image. The overall difference degree D' is calculated to be 0.42, which is greater than T2 (0.15), confirming it as a real defect. Subsequently, the system proceeds to the type identification step. This area does not meet the criteria for warp or weft yarn breakage, but it meets the criteria of "the mesh area suddenly changing to several times the normal value and the boundary being irregularly torn," so it is judged as a hole. In step nine, the hole type is at least classified as a general defect, and the area of ​​the hole (4.2 × 4.62 = 19.4 mm²) is close to S2 (20 mm²), so after comprehensive consideration, it is classified as a serious defect. Step ten triggers a level three audible and visual alarm (flashing red indicator light + continuous high-decibel beep). The intelligent control host first sends a marking execution signal to the inkjet marking device. The inkjet marking device sprays a long strip mark approximately 10mm in length at the edge of the fabric corresponding to the defect. After marking is completed, a feedback signal is sent, and the system then sends a stop signal to the loom control system. Before the stop signal is issued, the system prioritizes writing the image, type (hole), grade (severity), occurrence time, loom number, and fabric speed information of the defect into non-volatile memory and uploading it to the workshop-level quality management system via industrial Ethernet.

[0143] For example, the system detects an area with abrupt grayscale changes along the warp direction and a break in the continuity of the weft, with a D value of 0.31. After verification, it is confirmed as a genuine defect, identified as a broken warp thread, with an area S of 3 mm² and a defect number N of 1 per unit length. Although both the area and quantity indicators meet the numerical conditions for a minor defect (S < 5 mm², N < 2), according to the rule that "broken warp threads are at least classified as general defects," this defect is ultimately classified as a general defect, triggering a level two alarm (a yellow indicator light stays on + intermittent buzzer sound), but not triggering a shutdown.

[0144] For example, during continuous operation, if the system detects that the defect detection frequency suddenly rises to 4 times the historical average within a certain hour, the emergency response mechanism is triggered. The system temporarily lowers T1 and T2 by 0.03 each and triggers a process anomaly warning, reminding production managers to check the warp tension or weft yarn supply system.

[0145] Through the aforementioned series of closely coordinated steps, the system can detect various weaving defects in air conditioning filters online in real time with high precision. These defects include appearance defects such as uneven mesh size, sparse or dense patterns, holes, and skipped threads, which are difficult for traditional sensors to identify. The system can also perform graded alarms and shutdown control based on the severity of the defects. Furthermore, it has self-learning and optimization capabilities, which significantly improves the production quality control level of air conditioning filters and solves many drawbacks of manual inspection and the limitations of traditional inspection equipment.

[0146] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for online detection and alarm of weaving defects in air conditioning filter mesh, characterized in that, Includes the following steps: Step 1: Image Acquisition and Speed ​​Synchronization: The loom drives the air conditioner filter fabric forward at a running speed V. A backlight located below the filter fabric illuminates the fabric surface evenly. A vision sensor located above the filter fabric acquires a full-width grayscale image of the fabric surface in real time, while simultaneously acquiring the current running speed signal of the loom. The acquisition frame rate of the vision sensor is dynamically adjusted according to the running speed signal to maintain a constant physical resolution of the acquired image in the direction of fabric travel. Step 2, Image Preprocessing: The intelligent control host sequentially performs median filtering for noise reduction and contrast-limited adaptive histogram equalization on the acquired raw grayscale image to enhance the grayscale contrast between the fabric texture and the background, resulting in the preprocessed image. Step 3: Dynamic Region of Interest Extraction: The intelligent control host automatically locates and extracts the effective detection area in the preprocessed image based on the edge position of the fabric and the periodic texture features of the grid, eliminating interference from non-detection areas on both sides of the fabric. Step 4: Extraction of grid feature parameters: The intelligent control host performs latitude and longitude edge detection on the image within the effective detection area and extracts the following feature parameters of the filter grid: longitude spacing, latitude spacing, single grid area, longitude gray-level gradient change rate, latitude gray-level gradient change rate, and number of grid intersections. Step 5, Multi-template Standard Comparison: The intelligent control host calls the pre-stored standard mesh texture database, which stores the standard feature parameter range and standard mesh topology structure corresponding to the current production variety; the real-time feature parameters extracted in Step 4 are compared with the standard feature parameters item by item, and the deviation rate of each item is calculated. Step 6: Calculation and initial screening of comprehensive difference: The intelligent control host performs a weighted summation of various deviation rates to obtain the comprehensive difference D; when D is greater than the preset first dynamic threshold T1, it is determined that there is a suspected defect in the area, and the location coordinates, frame number and local image block of the suspected defect are recorded. Step 7, Local Secondary Verification: The intelligent control host controls the vision sensor to perform optical zoom or digital cropping magnification on the area where the suspected defect is located, and acquires a local high-resolution verification image; Steps 2 to 6 are repeated on the verification image. If the comprehensive verification difference D' is less than the preset second dynamic threshold T2, it is determined to be a false alarm, the suspected signal is ignored and the detection continues; if D' ≥ T2, it is confirmed as a real defect. Step 8, Intelligent Defect Type Identification: For areas confirmed as genuine defects, the intelligent control host determines the type based on their local image features: if a grayscale abrupt change band appears along the warp direction and the latitude continuity is interrupted, it is determined to be a broken warp thread; If a grayscale abrupt change occurs along the latitude direction and the continuity of the longitude is interrupted, it is judged as a broken latitude wire; if the area deviation rate of adjacent grids exceeds the third threshold T3 and is continuously distributed, it is judged as uneven mesh size; if the grid area abruptly changes to several times the normal value and the boundary is irregularly torn, it is judged as a hole; if the spacing between the longitude or latitude lines shows periodic changes in width, it is judged as a sparse-dense path; if a free wire shadow that does not follow the longitude or latitude direction appears locally, it is judged as a skipped wire; if none of the above criteria can be matched, it is judged as an unclassified structural defect. Step Nine, Defect Grading: The intelligent control host classifies defects into three levels—minor defects, general defects, and severe defects—according to preset grading rules based on defect type, defect area S, number of defects N per unit length, and defect distribution density ρ. When the defect type is a broken warp thread or a hole, the grade is at least general defect. When the defect type is an unclassified structural defect, it is directly classified as severe defect. When the hole area exceeds the second area threshold S2, it is directly classified as severe defect. Step 10, Graded Audible and Visual Alarm and Shutdown Control: The intelligent control host sends corresponding control signals to the audible and visual alarms according to the defect level. Minor defects trigger a Level 1 alarm and are recorded in the log, general defects trigger a Level 2 audible and visual alert, and serious defects trigger a Level 3 audible and visual alarm. When a Level 3 audible and visual alarm is triggered, the intelligent control host first sends a marking execution signal. After the physical marking of the defect in Step 11 is completed, a shutdown signal is sent to the loom control system. Step 11, Physical Marking of Defects: The intelligent control host controls the inkjet marking device located behind the vision sensor to spray an identifiable mark at the edge of the fabric corresponding to the defect, and simultaneously records the mark position and the longitudinal distance of the defect on the fabric surface; after receiving the marking execution signal, the inkjet marking device immediately performs the marking operation, and after marking is completed, it sends a marking completion signal back to the intelligent control host; Step 12, Detection Data Storage and Upload: Before triggering the stop signal, the intelligent control host first writes the image, type, grade, occurrence time, loom number, and fabric speed information of the current defect into the non-volatile memory, and uploads it to the workshop-level quality management system via industrial Ethernet. Step Thirteen: Self-learning Optimization and Emergency Response: The intelligent control host periodically calculates the false alarm rate and the missed alarm rate. When the false alarm rate exceeds the preset threshold, it automatically raises the first dynamic threshold T1 or the second dynamic threshold T2. When the missed alarm rate exceeds the preset threshold, it automatically lowers T1 or T2 to achieve adaptive optimization of the judgment threshold. When the defect detection frequency per unit time exceeds a preset multiple of the historical average, the system temporarily lowers T1 and T2 and triggers a process anomaly warning.

2. The method according to 1, characterized in that: In step one, the brightness of the backlight is automatically adjusted based on real-time feedback from the ambient light sensor. The adjustment method is as follows: when the ambient light intensity increases, the backlight brightness is automatically increased; when the ambient light intensity decreases, the backlight brightness is automatically decreased, so that the average gray value of the fabric image collected by the visual sensor is always maintained within the preset gray target range.

3. The method according to 1, characterized in that: In step three, the method for extracting the dynamic region of interest is as follows: perform row grayscale projection and column grayscale projection on the preprocessed image, determine the left and right boundaries of the fabric based on the abrupt change points of the projection curve, and further determine the upper and lower boundaries based on the periodic characteristics of the latitude and longitude lines, forming a rectangular effective detection window that dynamically changes with the position of the fabric.

4. The method according to 1, characterized in that: In step four, the edge detection of latitude and longitude lines uses the Canny edge detection operator. After extracting the edges, the main straight line clusters in the direction of longitude and latitude are identified by Hough line transform. The pixel distance between adjacent parallel lines is calculated and converted into physical size to obtain the spacing between longitude lines and the spacing between latitude lines. The area of ​​a single grid cell is calculated by using the closed quadrilateral formed by the intersections of the detected meridians and parallels to calculate the pixel area and convert it into the physical area.

5. The method according to 1, characterized in that: In step six, the formula for calculating the overall difference degree D is: D = α·Δd_warp + β·Δd_weft + γ·ΔA + δ·ΔG_warp + ε·ΔG_weft + ζ·ΔP, Where Δd_warp is the deviation rate of meridian spacing, Δd_weft is the deviation rate of parallel spacing, ΔA is the deviation rate of single grid area, ΔG_warp is the deviation rate of meridian gray-level gradient change rate, ΔG_weft is the deviation rate of parallel gray-level gradient change rate, and ΔP is the deviation rate of grid intersection number; α, β, γ, δ, ε, and ζ are the weight coefficients of the corresponding terms, and α+β+γ+δ+ε+ζ=1.

6. The method according to 1, characterized in that: In step six, the first dynamic threshold T1 and the second dynamic threshold T2 are automatically adjusted according to the current operating speed V of the loom and the current product specifications. The adjustment rules are as follows: the higher the speed V, the smaller the values ​​of T1 and T2; the higher the mesh number of the product, the smaller the values ​​of T1 and T2.

7. The method according to 1, characterized in that: In step eight, when multiple types of defects are detected to overlap in the same local area, the intelligent control host uses the most severe type as the main classification result for that area and adds the identification code of all detected types to the log.

8. The method according to 1, characterized in that: In step nine, the specific rules for grade classification are as follows: If the defect area S is less than the first area threshold S1 and the number of defects N per unit length is less than the first quantity threshold N1, and the defect type is not a broken warp thread, hole, or unclassified structural defect, it is classified as a minor defect; if S is between S1 and S2 or N is between N1 and N2, and the defect type is not an unclassified structural defect, it is classified as a general defect; if S is greater than the second area threshold S2 or N is greater than the second quantity threshold N2 or the defect distribution density ρ is greater than the density threshold ρ0, or the defect type is an unclassified structural defect, it is classified as a severe defect, where S1 <S2,N1<N2。 9. The method according to 1, characterized in that: In step ten, the specific methods for the three-level audible and visual alarms are as follows: Level 1 alarms only display prompt information on the human-machine interface and record logs, without triggering audible and visual signals; Level 2 alarms trigger a yellow indicator light to stay on and emit an intermittent buzzing sound; Level 3 alarms trigger a red indicator light to flash and emit a continuous high-decibel buzzing sound.

10. The method according to 1, characterized in that: In step eleven, the inkjet marking device uses fast-drying industrial marking ink, and the inkjet marking is a long strip mark along the fabric edge direction. The marking length is positively correlated with the lateral dimension of the defect in the width direction of the fabric surface.

11. The method according to 1, characterized in that: In step thirteen, the self-learning optimization further includes: the operator confirms the authenticity of each alarm through the human-computer interaction interface, the system automatically constructs a false alarm sample set and a missed alarm sample set based on the feedback results, and performs feature clustering analysis on the false alarm samples and missed alarm samples. When the cumulative number of samples of a certain type of feature exceeds the preset number threshold, the system automatically adjusts the weight coefficient of that type of feature in the comprehensive difference degree D.

12. The method according to 1, characterized in that: Before step five, there is a template consistency verification step: the intelligent control host performs statistics on the average background grayscale and grid periodicity features of the continuously acquired N0 frames of images. When the deviation of the statistical value from the previous calibration value exceeds the preset template update threshold, the standard template data is automatically triggered to adapt to image drift caused by light source aging or slow environmental changes.

Images