Braid defect detection method and system based on artificial intelligence

Abstract

The invention relates to the technical field of new-generation information, in particular to a braid defect detection method and system based on artificial intelligence. The method comprises the following steps: acquiring dynamic images of the front and back surfaces of a braid through an industrial camera, adjusting exposure time and light source brightness, and generating an acquisition image; performing segmentation and enhancement processing on the acquired image, extracting an image feature sequence of a defect area of the braid, and generating standardized image data; inputting the standardized image data into a deep learning model, identifying a defect region, and outputting a defect position, confidence and area; and generating an NG signal based on the defect parameters, transmitting the NG signal to a PLC (Programmable Logic Controller) to control a labeling machine to mark a defect area, storing defect data and generating a visual analysis report. Through combination of high-precision image acquisition and a deep learning algorithm, accurate identification and real-time feedback of complex defects on the surface of the braid are realized, the false alarm rate and the omission ratio are remarkably reduced, and the detection efficiency and the braid production quality are improved.

Description

Technical Field

[0001] The present invention relates to the field of new generation information technology and industrial visual inspection, and in particular to a method and system for detecting defects in a ribbon based on artificial intelligence. Background Art

[0002] In the webbing manufacturing process, surface quality is an important factor affecting product performance and market competitiveness. However, traditional webbing defect detection mainly relies on manual visual inspection or simple image processing methods, which often have the following problems when facing complex and changeable defect characteristics:

[0003] (1) Incomplete defect detection: Due to the various types of surface defects on the ribbon, including hair, breakage, stains, etc., traditional detection methods are difficult to effectively distinguish these subtle features.

[0004] (2) High false alarm and missed detection rates: Due to the limitations of the feature extraction capabilities of traditional algorithms, good products and defective areas are easily confused, resulting in unstable detection results.

[0005] (3) Lack of real-time and adaptability: When faced with different types of webbing or complex production environments, traditional systems find it difficult to quickly adjust parameters or provide real-time feedback on test results, which affects production efficiency and product quality.

[0006] In recent years, the application of artificial intelligence, especially deep learning technology, in the field of visual inspection has demonstrated powerful capabilities, which can automatically extract image features and achieve accurate classification through training models. However, combining deep learning with industrial production environments still requires solving technical problems in multiple links such as data collection, real-time processing, and feedback control. Therefore, there is an urgent need for an intelligent system that can efficiently collect dynamic image data of webbing, use deep learning models for accurate detection and real-time feedback, so as to improve the quality and efficiency of webbing production. Summary of the invention

[0007] The present invention provides a method and system for detecting defects in a ribbon based on artificial intelligence, so as to solve the problem of how to accurately identify a variety of complex defects based on dynamic image data collected from the ribbon surface by using a deep learning model and provide real-time feedback, thereby reducing the false alarm rate and missed detection rate, and thus improving the quality and efficiency of ribbon production.

[0008] In order to solve the above technical problems, the present invention provides a method for detecting webbing defects based on artificial intelligence, comprising:

[0009] Use an industrial camera to obtain dynamic images of the front and back of the ribbon, adjust the exposure time and light source brightness, and generate captured images;

[0010] Segmenting and enhancing the collected image, extracting an image feature sequence of a ribbon defect area, and generating standardized image data;

[0011] Inputting the standardized image data into a deep learning model, identifying defect areas, and outputting defect locations, confidence levels, and areas;

[0012] An NG signal is generated based on the defect parameters and transmitted to the PLC to control the labeling machine to mark the defect area. At the same time, the defect data is stored and a visual analysis report is generated.

[0013] Furthermore, the step of obtaining dynamic images of the front and back sides of the ribbon by an industrial camera includes:

[0014] Get the ribbon type and its attribute parameters to generate the initial shooting configuration;

[0015] Adjusting the exposure time, gain and light source brightness of the industrial camera based on real-time feedback to capture dynamic images of the front and back sides of the ribbon;

[0016] The dynamic image is transmitted to the acquisition server through the global shutter mode.

[0017] Furthermore, the step of segmenting and enhancing the collected image includes:

[0018] Cropping the collected image to generate independent sub-images of the front and back sides of the ribbon;

[0019] Performing edge enhancement and noise removal on the sub-image based on a deep learning model to extract texture features of the ribbon surface;

[0020] The extracted features are standardized to generate standardized image data that meets the input requirements of the deep learning model.

[0021] Furthermore, the cutting step includes:

[0022] Extracting images of the front and back sides of the ribbon from the captured image according to preset cropping coordinates;

[0023] The front image and the back image are combined into an image sequence as an input for subsequent enhancement processing.

[0024] Furthermore, the edge enhancement and noise removal steps include:

[0025] Performing a convolution operation on the image sequence to generate an edge enhancement map;

[0026] The enhanced image is denoised by Gaussian modeling, and a denoised image is output as a texture feature map.

[0027] Furthermore, the step of inputting the standardized image data into the deep learning model includes:

[0028] Inputting the standardized image data into a pre-trained deep learning model;

[0029] Use convolutional neural network to extract the texture and defect features of the ribbon surface;

[0030] The output class probability matrix is ​​used as the input for defect area extraction.

[0031] Furthermore, the step of extracting the defective area includes:

[0032] Setting a classification confidence threshold based on the class probability matrix;

[0033] Extract the location information and boundary coordinates of the defect area and generate a defect mask;

[0034] The average confidence of the defect area is calculated.

[0035] Furthermore, the step of generating an NG signal includes:

[0036] Based on the defect mask and the detection parameters, determining whether the defect meets a confidence threshold and an area threshold;

[0037] The judgment result is converted into an NG signal and transmitted to the PLC control module.

[0038] Furthermore, the method for generating an NG signal further includes:

[0039] The PLC controls the labeling machine to mark the defective area according to the NG signal;

[0040] Synchronously record the marking position and timestamp to generate a marking log;

[0041] The defect data and marking log are stored in a server, and a visual analysis report is generated.

[0042] Furthermore, a ribbon defect detection system based on artificial intelligence comprises:

[0043] The data acquisition module is used to collect dynamic images of the front and back sides of the ribbon in real time through an industrial camera, adjust the exposure time and light source brightness, and generate high-quality collected images;

[0044] An image processing module, used for performing cropping, enhancement processing and standardization processing on the collected image to generate standardized image data;

[0045] A defect recognition module, used to input the standardized image data into a deep learning model to extract the location information, confidence level and area of ​​the defect area;

[0046] The real-time control and feedback module is used to generate NG signals to control the labeling machine to mark the defective area, store the defect data and generate a visual analysis report.

[0047] The key innovative features of the present invention include:

[0048] (1) Dynamic acquisition optimization: By adjusting the shooting parameters and light source brightness of the industrial camera in real time to adapt to different ribbon types and surface textures, the quality and adaptability of dynamic image acquisition are improved.

[0049] (2) Precise recognition driven by deep learning: By inputting standardized image data into a pre-trained deep learning model, high-precision recognition of complex defects can be achieved, and multi-dimensional parameters such as position, confidence, and area can be output.

[0050] (3) Closed-loop control and feedback mechanism: Combining the generation of NG signals with real-time PLC control, the marking and data storage of defective areas are automatically completed, achieving efficient linkage between inspection and production.

[0051] (4) Data analysis and visualization support: By storing and analyzing defect parameters, intuitive visualization reports are generated, which improves the digitalization and intelligence level of production line management.

[0052] The following are its main beneficial effects:

[0053] (1) Accurate detection: Dynamically adjusted industrial camera exposure time and light source brightness are used to generate high-quality captured images, solving the problem of unstable image quality caused by the complex gloss and texture of the ribbon surface, and significantly improving the reliability of detection data.

[0054] (2) Efficient identification: The deep learning model is used to extract and classify the features of standardized image data, accurately identifying the complex defect areas on the surface of the ribbon, including various defect types such as lint, breakage, and stains, effectively reducing the false alarm rate and missed detection rate.

[0055] (3) Intelligent feedback: The PLC receives NG signals in real time, drives the labeling machine to accurately mark the defective area, and simultaneously records the defect parameters and marking log, realizing automatic closed-loop control after the mark is detected, significantly improving production efficiency.

[0056] (4) Data-driven optimization: The test results and marked data are stored in the server and combined with the generated visual reports to support quality tracking and production improvement, providing comprehensive data support for the intelligent upgrade of the ribbon manufacturing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic flow chart of a method for detecting webbing defects based on artificial intelligence provided in an embodiment of the present application.

[0058] Figure 2 A structural block diagram of a webbing defect detection system based on artificial intelligence provided in an embodiment of the present application.

[0059] Figure 3 This is an overall framework diagram of a detection system provided in this embodiment.

[0060] Figure 4 A framework diagram of the device layer in the overall framework diagram of a detection system provided in this embodiment.

[0061] Figure 5 A schematic diagram of the camera image acquisition effect of the control layer in the overall framework diagram of the detection system provided in this embodiment.

[0062] Figure 6 This is a parameter configuration diagram of the detection system at the workshop level in the overall framework diagram of the detection system provided in this embodiment. DETAILED DESCRIPTION

[0063] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0064] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0065] Embodiment 1:

[0066] Reference Figure 1 , is a flow chart of a method for detecting defects in a ribbon based on artificial intelligence provided by an embodiment of the present invention, and the flow chart may at least include steps S100-S400:

[0067] S100, obtain dynamic images of the front and back of the ribbon through industrial cameras, adjust exposure time and light source brightness, and generate high-quality acquisition images;

[0068] S200, segmenting and enhancing the collected image, extracting an image feature sequence of a ribbon defect area, and generating standardized image data;

[0069] S300, inputting the standardized image data into a deep learning model, identifying defect areas, and outputting parameters such as defect location, confidence level, and area;

[0070] S400, generating an NG signal based on the defect parameters, transmitting the signal to a PLC to control a labeling machine to mark the defect area, and storing the defect data and generating a visual analysis report.

[0071] Step S100 at least includes steps S110-S130:

[0072] S110: Acquire initial shooting parameters of the industrial camera and generate initial parameter configuration based on the ribbon type.

[0073] Specifically, S110 reads the type of webbing (parameter T) and the specific properties of the webbing (width W t , Color C t , Glossiness G t ), initialize the shooting parameters P0 of the industrial camera. These parameters include but are not limited to exposure time E0, frame rate F0, gain G0 and resolution R0.

[0074] The formula is described as follows:

[0075] P0={E0,F0,G0,R0}=f(T,W t ,C t ,G t )

[0076] Where T is the webbing type, which is input by the user or automatically identified; W t is the width of the webbing, provided by the front-end detection module; C t is the ribbon color, using the standard representation of the RGB space; G t is the glossiness of the ribbon, measured by a gloss sensor and quantified into the range of [0,1].

[0077] Furthermore, based on these initial parameters and in combination with the performance model of the optical device, an initial configuration file Conf ig0 for subsequent shooting is generated and passed to S120 for calling.

[0078] S120: Adjust the exposure time and gain of the industrial camera, dynamically optimize the light source brightness according to the glossiness of the ribbon, and obtain dynamic images of the front and back of the ribbon.

[0079] Further, in S120, based on the initial configuration P0 of S110, the exposure time and gain value are adjusted through a real-time feedback mechanism to adapt to the dynamic webbing environment. Specifically, combined with the collected light source intensity L i , dynamically calculate the optimized exposure time E i and gain G i .

[0080] The formula is as follows:

[0081]

[0082] Among them, L t is the target light source brightness, which is configured by the user or given by the light source equipment standard; L i It is the current light source brightness, measured in real time by the light source sensor.

[0083] In addition, the gloss G t Dynamically adjust the light source brightness L opt , to ensure the uniformity of the image on both sides of the ribbon:

[0084] L opt =L t ×(1+G t )

[0085] The adjusted parameters {E i ,G i ,L opt}Apply to industrial cameras and light source equipment to capture dynamic images of the front and back of the ribbon Img t .

[0086] S130: Transmitting the acquired dynamic image to the acquisition server through the global shutter mode to generate a high-quality acquisition image without ghosting.

[0087] Specifically, S130 converts the dynamic image Img obtained in S120 into t The image is transmitted to the acquisition server and the global shutter mode is used to avoid the ghosting phenomenon. Furthermore, the image quality is tested to determine the image resolution R t , Contrast C t , signal-to-noise ratio SNR t Whether the preset standards are met.

[0088] The quality inspection formula is as follows:

[0089] Q t =w1·R t +w2·C t +w3·SNR t

[0090] Among them, Q t is the image quality score; Rt ,C t ,SNR t They are resolution, contrast and signal-to-noise ratio respectively; w1, w2, w3 are weight coefficients, which are adjusted based on experience or experimental data.

[0091] When Q t ≥Q threshold (preset quality threshold), the high-quality image Img opt Stored in the acquisition server and passed to S200 as input.

[0092] S110→S120: The initial configuration Config0 output by S110 provides basic parameters such as exposure time and gain, which are used as input of S120 to dynamically adjust the shooting conditions.

[0093] S120→S130: dynamic image Img output from S120 t and the adjusted shooting parameters {E i ,G i ,L opt}Directly used by S130 for image transmission and quality detection.

[0094] S130→S200: High-quality image Img generated by S130 opt It is the core input of S200 image segmentation and enhancement, ensuring the accuracy and reliability of subsequent processing.

[0095] Step S200 at least includes steps S210-S230:

[0096] S210: Preliminarily cropping the collected image to divide it into independent sub-images of the front side and the back side to generate a left-right image sequence.

[0097] Specifically, S210 obtains the high-quality captured image Img transmitted from S130. opt , for each frame of image, according to the segmentation parameter {x start ,x end ,y start ,y end}Crop and extract independent sub-images of the front and back sides.

[0098] The formula is described as follows:

[0099] Img left =Img opt [x l1 :x l2 ,y l1 :y l2 ]

[0100] Img right =Imgopt [x r1 :x r2 ,y r1 :y r2 ]

[0101] Among them, Img left ,Img right The images are the front and back of the ribbon respectively; x l1 ,x l2 ,y l1 ,y l2 is the cropping coordinate of the front image; x r1 ,x r2 ,y r1 ,y r2 The cropping coordinates of the reverse image.

[0102] Further, the trimmed Img left and Img right Combined into an image sequence S Img =Img left ,Img right} for subsequent processing.

[0103] S220: Perform edge enhancement and noise removal on the image sequence based on the deep learning preprocessing model to extract the surface texture and defect features of the ribbon.

[0104] Further, S220 is based on the image sequence S generated in S210 Img , use the preprocessing model to perform edge enhancement and noise removal operations on each frame of the image. The computational model of the image preprocessing process is as follows:

[0105] Edge enhancement: E(x,y)=G(x,y)*K e

[0106] Where, E(x,y): enhanced edge map; G(x,y): input image; K e Filter kernel matrix for edge enhancement.

[0107] Noise removal: D(x,y)=E(x,y)-N(x,y)

[0108] Where D(x,y) is the image after noise removal; N(x,y) is the noise image obtained by Gaussian distribution modeling.

[0109] Specifically, the results of edge enhancement and noise removal are output as the image sequence S after feature extraction. Feature ={D left ,D right} for subsequent standardization processing.

[0110] S230: Standardize the preprocessed image sequence to generate standardized image data that meets the input requirements of the deep learning model.

[0111] In S230, the feature image sequence S generated in S220 is Feature Standardization is performed to ensure that the image size, brightness, and contrast are uniform and meet the input specifications of the deep learning model.

[0112] Size standardization: I norm (x,y)=I resize (x,y)*R s

[0113] Among them, I norm (x, y) is the image after normalization; I resize (x, y) is the resized image; R s is the target resolution ratio.

[0114] Brightness and contrast normalization:

[0115] I adj (x,y)=α·(I norm (x,y)-β)

[0116] Among them, I adj (x, y) is the adjusted standardized image; α is the contrast adjustment factor; β is the brightness adjustment reference value.

[0117] The final generated standardized image sequence S Norm = {I adj,left ,I adj,right} is passed to S300 for defect detection.

[0118] S210→S220: S210 generates the front and back image sequences S by cropping Img It is the input for S220 to perform edge enhancement and noise removal.

[0119] S220→S230: Feature extraction image sequence S output by S220 Feature Used by S230 as input for normalization processing.

[0120] S230→S300: Standardized image sequence S generated by S230 Norm It is the core input data for S300 defect identification.

[0121] Step S300 at least includes steps S310-S330:

[0122] S310: Input the standardized image data into the pre-trained deep learning model to classify and identify the surface texture and defect features of the ribbon.

[0123] Specifically, S310 receives the standardized image data S generated from S230. Norm =I adj,left ,I adj,right} and pass it as input to the pre-trained deep learning model M pre .

[0124] Model M pre Convolutional Neural Network (CNN) is used to extract the texture and defect features of the ribbon surface. The feature map is expressed as follows:

[0125] F(x,y)=CNN(I adj (x,y)

[0126] Among them, F(x,y) is the feature mapping matrix; CNN is the convolution operation; I adj (x,y) is the input normalized image.

[0127] Furthermore, based on the extracted feature map F(x,y), the model performs classification recognition through a fully connected layer and outputs the category probability matrix P(x,y) for each pixel:

[0128] P(x,y)=Softmax(W·F(x,y)+b)

[0129] Among them, W and b are classification weights and biases; Softmax is the classification activation function.

[0130] The output category probability matrix P(x,y) contains the probability value of each pixel belonging to the defect category, which serves as the input for the next step of defect area extraction.

[0131] S320: extracting the position information of the defect area and the corresponding confidence from the classification result, and generating a defect position parameter and a confidence sequence.

[0132] Further, S320 sets a classification threshold T based on the classification result P(x, y) of S310. c Extract the location information of the defect area and calculate the corresponding confidence.

[0133] Perform threshold segmentation on the classification result matrix P(x,y) to generate the defect mask M defect (x,y):

[0134]

[0135] Among them, T c Classification confidence threshold.

[0136] Using defect mask M defect (x,y), extract the boundary coordinates of the defect area {(x min ,y min ),(x max ,y max )}. Calculate the average confidence C of the defect area avg :

[0137]

[0138] The defect location information generated {(x min ,y min ),(x max ,y max )} and confidence C avg It will be used as input for the next step of area calculation.

[0139] S330: further calculate the area of ​​the defect area and output the detection parameters of the defect position, confidence and area.

[0140] S330 defect area mask M based on S320 defect And position coordinates, further calculate the area A of the defect defect , and output the detection parameters.

[0141] The actual area of ​​the defect region is calculated by counting the total number of pixels with a value of 1 in the defect mask:

[0142]

[0143] Among them, p is the actual physical size corresponding to the pixel.

[0144] Furthermore, the position, confidence and area are integrated to generate the detection parameter D result :

[0145] D result ={(x min ,y min ),(x max ,y max ),C avg ,A defect}

[0146] The final output detection parameter D result It contains information such as defect location, confidence level and area, which will be used as input for subsequent modules.

[0147] S310→S320: The category probability matrix P(x,y) output by S310 is the core input of S320 for extracting defect areas and calculating confidence.

[0148] S320→S330: Defect mask M generated in S320 defect and location information {(x min ,y min ),(x max ,y max )} is the direct data source for S330 to calculate area.

[0149] S330→S400: Detection parameter D generated in S330 result It will be used as input data for the S400 real-time control and feedback module for subsequent marking and data storage.

[0150] Step S400 at least includes steps S410-S430:

[0151] S410: Input the defect parameters into a signal processing module, generate an NG signal according to a confidence threshold, and transmit it to a PLC device.

[0152] Specifically, S410 receives the detection parameter D generated from S330 result ={(x min ,y min ),(x max ,y max ),C avg ,A defect}, and the confidence C avg and defect area A defect A threshold comparison is performed to determine whether an NG signal is generated.

[0153] According to the set confidence threshold T c and area threshold T a , generating NG signal NG signal :

[0154]

[0155] Among them, T c is the confidence threshold, which is preset by the user or dynamically adjusted by the system; T a is the area threshold, which is related to the webbing type.

[0156] Furthermore, the generated NG signal is converted into a standard level signal through a relay module and transmitted to a PLC device to control the labeling machine.

[0157] S420: Control the labeling machine to accurately mark the defective area according to the NG signal, and record the marking position and timestamp at the same time.

[0158] Further, S420 uses the NG signal NG generated in S410 signal, the defective area is marked by the PLC-controlled labeling machine.

[0159] According to the position information of the defect area in the detection parameters {(x min ,y min ),(x max ,y max )}, calculate the motion control parameters M of the marking device control :

[0160]

[0161] Among them, X center ,Y center To mark the center position.

[0162] Furthermore, the labeling machine is based on M control Instructions, paste marks at the defective area.

[0163] Synchronous record marks the timestamp T of completion mark , forming a marked log Log mark :

[0164] Log mark =NG signal ,M control ,T mark

[0165] S430: The defect data is stored in the server, and a visual analysis report is generated according to the marking results for subsequent quality tracking.

[0166] S430 collects the output data of S410 and S420, stores the defect information in the server, and generates a visual analysis report.

[0167] The detection parameter D result and tag log mark Stored in server database DB defect :

[0168] DB defect =DB defect ∪{D result ,Log mark}

[0169] Generate visual reports based on stored data, including defect statistics, marker location distribution graphs, and time series analysis:

[0170] Report defect =f(DB defect )

[0171] Where f is the visualization report generation function; Report defectContains statistical tables, distribution graphs, and time series graphs.

[0172] S410→S420: NG signal generated by S410 NG signal It is the core input for S420 to control the labeling machine to perform labeling.

[0173] S420→S430: the mark log recorded in S420 mark Used by S430 for data storage and report generation.

[0174] S430 → Subsequent modules: Visual analysis report generated by S430 defect It can provide a basis for subsequent production optimization or quality improvement.

[0175] Figure 3 This is an overall framework diagram of a detection system provided in this embodiment. Figure 3 As shown, the framework is divided into three levels:

[0176] Equipment layer: mainly involves various types of detection execution equipment

[0177] Control layer: mainly involves the entire parameter configuration module (equipment control system), input module (visual imaging system), processing module (algorithm reasoning system), output module and storage module (communication system), etc.;

[0178] Workshop level: mainly involves connecting with EMS and remotely controlling various processing modules.

[0179] Figure 4 A framework diagram of the device layer in the overall framework diagram of a detection system provided in this embodiment.

[0180] Figure 5 A schematic diagram of the camera image acquisition effect of the control layer in the overall framework diagram of the detection system provided in this embodiment.

[0181] Figure 6 This is a parameter configuration diagram of the detection system at the workshop level in the overall framework diagram of the detection system provided in this embodiment.

[0182] The key innovative features of the present invention include:

[0183] (1) Dynamic acquisition optimization: By adjusting the shooting parameters and light source brightness of the industrial camera in real time to adapt to different ribbon types and surface textures, the quality and adaptability of dynamic image acquisition are improved.

[0184] (2) Precise recognition driven by deep learning: By inputting standardized image data into a pre-trained deep learning model, high-precision recognition of complex defects can be achieved, and multi-dimensional parameters such as position, confidence, and area can be output.

[0185] (3) Closed-loop control and feedback mechanism: Combining the generation of NG signals with real-time PLC control, the marking and data storage of defective areas are automatically completed, achieving efficient linkage between inspection and production.

[0186] (4) Data analysis and visualization support: By storing and analyzing defect parameters, intuitive visualization reports are generated, which improves the digitalization and intelligence level of production line management.

[0187] The following are its main beneficial effects:

[0188] (1) Accurate detection: Dynamically adjusted industrial camera exposure time and light source brightness are used to generate high-quality captured images, solving the problem of unstable image quality caused by the complex gloss and texture of the ribbon surface, and significantly improving the reliability of detection data.

[0189] (2) Efficient identification: The deep learning model is used to extract and classify the features of standardized image data, accurately identifying the complex defect areas on the surface of the ribbon, including various defect types such as hair, breakage, and stains, effectively reducing the false alarm rate and missed detection rate.

[0190] (3) Intelligent feedback: The PLC receives NG signals in real time, drives the labeling machine to accurately mark the defective area, and simultaneously records the defect parameters and marking log, realizing automatic closed-loop control after the mark is detected, significantly improving production efficiency.

[0191] (4) Data-driven optimization: The test results and marked data are stored in the server and combined with the generated visual reports to support quality tracking and production improvement, providing comprehensive data support for the intelligent upgrade of the ribbon manufacturing process.

[0192] Embodiment 2: Figure 2 FIG. 2 shows a structural block diagram of a ribbon defect detection system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system may include:

[0193] The data acquisition module 10 collects dynamic images of the front and back sides of the ribbon in real time through an industrial camera, dynamically adjusts the exposure time, gain and light source brightness, and obtains high-definition images.

[0194] According to the width, color and glossiness of the ribbon, the shooting parameters P0 = {E0, F0, G0, R0} are initialized and the parameters are optimized to adapt to different ribbon types.

[0195] The dynamic images without ghosting are transmitted to the acquisition server to meet the standards of resolution, contrast and signal-to-noise ratio.

[0196] The image processing module 20 segments the image data transmitted by the acquisition module, cuts the front and back sides of the ribbon into independent sub-images, and generates a left and right image sequence S Img ={Img left ,Img right}.

[0197] Based on the deep learning preprocessing model, the image is edge enhanced and noise removed to extract the surface texture and defect features of the ribbon. Feature .

[0198] The preprocessed images are standardized, including resizing, brightness and contrast standardization, to generate a standardized image sequence S that meets the input requirements of the deep learning model. Norm .

[0199] The defect recognition module 30 inputs the standardized image data into a pre-trained deep learning model, extracts the texture features F(x, y) of the ribbon surface, and classifies and recognizes each pixel to generate a category probability matrix P(x, y).

[0200] Extract the location information of the defect area from the classification result {(x min ,y min ),(x max ,y max )} and the corresponding confidence C avg , generate defect mask A defect .

[0201] Count the defect area A based on the defect mask defect , comprehensive generation of detection parameters D result ={(x min ,y min ),(x max ,y max ),C avg ,A defect}.

[0202] The real-time control and feedback module 40 receives the detection parameter D output by the defect recognition module. result , according to the confidence threshold T c and area threshold T a Generate NG signal NG signal , and transmitted to the PLC device.

[0203] The PLC controls the labeling machine to mark the defective area according to the NG signal, and records the timestamp and location information of the mark. mark .

[0204] Store the detection parameters and marking logs to the server and generate a visual analysis report defect, used for defect statistics and production improvement.

[0205] Beneficial effects of the embodiments:

[0206] (1) Efficient detection: Through the deep learning model, the system can accurately identify various defect types of complex webbings and significantly reduce the false alarm rate and missed alarm rate.

[0207] (2) Intelligent control: Combined with PLC and real-time feedback mechanism, it can automatically mark defective areas, reduce manual intervention and improve production efficiency.

[0208] (3) Data-driven optimization: The system stores and analyzes production data, providing a comprehensive and visual basis for process optimization and decision support.

[0209] (4) Modular design: Each module operates independently and collaborates with each other, which facilitates system maintenance and expansion and meets the testing needs of various webbing types.

[0210] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A method for detecting ribbon defects based on artificial intelligence, characterized in that: The following steps are involved: Use an industrial camera to obtain dynamic images of the front and back of the ribbon, adjust the exposure time and light source brightness, and generate captured images; Segmenting and enhancing the collected image, extracting an image feature sequence of a ribbon defect area, and generating standardized image data; Inputting the standardized image data into a deep learning model, identifying defect areas, and outputting defect locations, confidence levels, and areas; An NG signal is generated based on the defect parameters and transmitted to the PLC to control the labeling machine to mark the defect area. At the same time, the defect data is stored and a visual analysis report is generated.

2. The method according to claim 1, characterized in that The step of obtaining dynamic images of the front and back sides of the ribbon by an industrial camera comprises: Get the ribbon type and its attribute parameters to generate the initial shooting configuration; Adjusting the exposure time, gain and light source brightness of the industrial camera based on real-time feedback to capture dynamic images of the front and back sides of the ribbon; The dynamic image is transmitted to the acquisition server through the global shutter mode.

3. The method according to claim 1, characterized in that The step of segmenting and enhancing the collected image comprises: Cropping the collected image to generate independent sub-images of the front and back sides of the ribbon; Performing edge enhancement and noise removal on the sub-image based on a deep learning model to extract texture features of the ribbon surface; The extracted features are standardized to generate standardized image data that meets the input requirements of the deep learning model.

4. The method according to claim 3, characterized in that The cutting step comprises: Extracting images of the front and back sides of the ribbon from the captured image according to preset cropping coordinates; The front image and the back image are combined into an image sequence as an input for subsequent enhancement processing.

5. The method according to claim 3, characterized in that: The edge enhancement and noise removal steps include: Performing a convolution operation on the image sequence to generate an edge enhancement map; The enhanced image is denoised by Gaussian modeling, and a denoised image is output as a texture feature map.

6. The method according to claim 1, characterized in that The step of inputting the standardized image data into the deep learning model comprises: Inputting the standardized image data into a pre-trained deep learning model; Use convolutional neural network to extract the texture and defect features of the ribbon surface; The output class probability matrix is ​​used as the input for defect area extraction.

7. The method according to claim 6, characterized in that The step of extracting the defect area comprises: Setting a classification confidence threshold based on the class probability matrix; Extract the location information and boundary coordinates of the defect area and generate a defect mask; The average confidence of the defect area is calculated.

8. The method according to claim 1, characterized in that The step of generating an NG signal comprises: Based on the defect mask and the detection parameters, determining whether the defect meets a confidence threshold and an area threshold; The judgment result is converted into an NG signal and transmitted to the PLC control module.

9. The method according to claim 8, characterized in that Also includes: The PLC controls the labeling machine to mark the defective area according to the NG signal; Synchronously record the marking position and timestamp to generate a marking log; The defect data and marking log are stored in a server, and a visual analysis report is generated.

10. A ribbon defect detection system based on artificial intelligence, characterized in that: include: The data acquisition module is used to collect dynamic images of the front and back sides of the ribbon in real time through an industrial camera, adjust the exposure time and light source brightness, and generate high-quality collected images; An image processing module, used for performing cropping, enhancement processing and standardization processing on the collected image to generate standardized image data; A defect recognition module, used to input the standardized image data into a deep learning model to extract the location information, confidence level and area of ​​the defect area; The real-time control and feedback module is used to generate NG signals to control the labeling machine to mark the defective area, store the defect data and generate a visual analysis report.

Images