Intelligent production line for industrial defect detection

Abstract

The invention relates to the technical field of image detection, and discloses an intelligent production line for industrial defect detection aiming at the problems of insufficient pertinence, low defect detection precision, lack of whole-process management and control and the like of a traditional industrial steel plate shot blasting process, and comprehensive detection of the corrosion condition of a steel plate before shot blasting is realized by designing RGB cameras arranged on the upper side and the lower side of the steel plate. Based on a corrosion degree feedback adjustment mechanism, corrosion information collected by an RGB camera is combined with a preset corrosion judgment standard, dynamic adjustment of the shot blasting amount and the conveying speed is achieved, the problem that the pertinence of the shot blasting process is insufficient is solved, after shot blasting is completed, a steel plate surface image is collected, pocking mark feature information is extracted after algorithm processing, and the shot blasting quality is improved. And comparing with a preset production standard to judge whether the steel plate is qualified or not. Detection data are transmitted to the central platform through the industrial Ethernet, and intelligent control over the quality of the steel plate is achieved.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, specifically to an intelligent production line that integrates multi-sensor collaboration, deep learning algorithms, and dynamic parameter adjustment, and is particularly suitable for industrial defect detection and process control throughout the entire production process. Background Technology

[0002] For steel that has been rusted for a long time in industry, we usually use shot blasting. However, controlling the shot blasting amount and adjusting the speed of the conveying machine have always been difficult points in the process: if the shot blasting amount is insufficient or the conveying speed is too fast, the rust cannot be completely removed, which will affect the quality of subsequent processing; if the shot blasting amount is too large or the conveying speed is too slow, it may cause excessive deformation of the steel surface, scratches and pits, and even reduce dimensional accuracy, directly affecting the product qualification rate.

[0003] Traditional industrial cameras, limited by hardware performance and design positioning, typically suffer from the following drawbacks: insufficient imaging resolution and detail capture capability, poor light adaptability, low defect contrast, easy image shift due to dust adhering to the lens or vibration, misjudgment or missed detection of pits, and image analysis that relies too much on manual operation, resulting in low efficiency and consistency.

[0004] In existing technologies, although black-and-white cameras and RGB cameras based on grayscale feature analysis algorithms have been used for related detection, the two types of equipment have not yet formed an effective collaborative mechanism. This makes it difficult to achieve high-precision identification of pitting on steel plates after shot blasting, and also makes it impossible to effectively monitor the entire production process.

[0005] Based on this, the present invention proposes an intelligent production line that can dynamically detect the degree of corrosion, detect defects with high precision, and control the entire process, so as to overcome the limitations of traditional processes. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent production line for industrial defect detection. It employs an RGB camera based on the YOLO algorithm and a monochrome camera based on a grayscale feature analysis algorithm to work together to detect pitting on the surface of steel plates. The monochrome camera captures the outline of tiny pits through high-contrast imaging, while the YOLO algorithm of the RGB camera identifies the texture and oxidation features around the pits. This allows for accurate determination of whether the size and density of the pits meet production specifications (e.g., diameter ≤ 0.3mm and ≤ 5 pits per 100cm² is acceptable). This solves the problem of missed detection and misjudgment in the detection of rough surfaces after shot blasting using a single camera.

[0007] The objective of this invention can be achieved through the following technical solutions: A smart production line for industrial defect detection includes a rust detection camera unit, a shot blasting parameter adjustment unit, a pitting detection unit, a quality judgment unit, and a steel plate production control platform. A corrosion detection unit, deployed at the corrosion detection station, is used to collect surface images of the steel plate before shot blasting and identify the corrosion level. The shot blasting parameter dynamic adjustment unit is connected to the equipment signal of the rust detection unit and the shot blasting station, and is used to dynamically adjust the shot blasting amount and steel plate conveying speed according to the identified rust level. The pitting detection unit is deployed at the pitting detection station to collect surface images of the steel plate after shot blasting and to identify and analyze pitting defects based on a dual-camera collaborative detection mechanism. The quality assessment unit is signal-connected to the pitting detection unit and is used to determine the quality grade of the steel plate based on the pitting detection results. The steel plate production control platform is used to generate corresponding control commands and feed them back to the production line; The corrosion detection unit includes at least two sets of industrial RGB cameras, a dedicated light homogenizing device, and an edge computing terminal; The RGB cameras are symmetrically deployed on the upper and lower sides of the steel plate and are mounted using adjustable brackets. The dedicated light-diffusing device uses high color rendering LED light strips and a diffuser to provide uniform illumination to the steel plate surface at an inclined angle; The edge computing terminal has a built-in GPU module for running a corrosion-level identification model based on the YOLO algorithm.

[0008] The judgment logic of the corrosion level identification model based on the YOLO algorithm is as follows: the level is divided according to the proportion of the corrosion area to the total surface area of ​​the steel plate. The area proportion is less than 5% and there is no deep corrosion, which is light corrosion; the area proportion is between 5% and 15%, which is moderate corrosion; and the area proportion is greater than 15% or there is deep corrosion texture, which is heavy corrosion.

[0009] The adjustment rules of the shot blasting parameter dynamic adjustment unit are as follows: When the corrosion level is light, reduce the shot blasting amount by 20%-30% and increase the conveyor speed by 15%-20%; When the corrosion level is moderate, maintain the current shot blasting volume and conveying speed; When the corrosion level is severe, increase the shot blasting amount by 30%-50% and reduce the conveyor speed by 20%-25%.

[0010] The pitting detection unit includes multiple detection groups deployed on the upper and lower sides of the steel plate. Each detection group includes an RGB camera based on the YOLO algorithm, a black and white camera based on the grayscale feature analysis algorithm, and a set of special spotlight detection device for pitting. The RGB camera and the monochrome camera are rigidly connected and aligned in the field of view through a dual-camera connector with adjustable spacing and angle.

[0011] The dual-camera connector includes: a dual-camera fixed seat provided with two independent mounting positions with adjustable clamping structures; a pitch adjustment component connected to the two mounting positions through a sliding guide rail, capable of achieving stepless adjustment and locking of the horizontal pitch; a reference mounting frame provided with an interface for fixing to an external bracket and supporting overall angle fine-tuning.

[0012] The working process of the pitting detection unit includes: the black-and-white camera captures a grayscale image, extracts the pitting contour through grayscale threshold segmentation and morphological processing, and calculates the preliminary quantity and size of the pitting; The RGB camera captures a color image, verifies the features of the suspected pitting areas detected by the black-and-white camera through the YOLO algorithm, identifies the pitting and determines whether there are accompanying oxidation features.

[0013] The decision logic of the quality determination unit is as follows: Qualified: The pitting size does not exceed the standard and the quantity does not exceed the standard, and there are no oxidation features; Size exceeding the standard: The diameter of the minimum circumscribed circle of the pitting is greater than the preset threshold; Density exceeding the standard: The number of effective pitting per unit area is greater than the preset threshold; Seriously exceeding the standard: The pitting is accompanied by oxidation features.

[0014] Preferably: The quality determination unit performs the following feedback control according to the determination result: If it is determined to be qualified, control the conveying system to convey the steel plate to the next process; If it is determined to be slightly unqualified, control the shot blasting equipment to perform parameter fine-tuning and then process the subsequent steel plates; If it is determined to be seriously exceeding the standard, immediately trigger an audible and visual alarm and pause the conveying system, waiting for manual handling; The steel plate production control platform receives and integrates all data from the rust detection unit, the shot blasting parameter dynamic adjustment unit, the pitting detection unit, and the quality determination and feedback unit through the industrial Ethernet, realizes real-time monitoring of the production process, historical data query, over-standard warning, and generation and push of process parameter optimization reports.

[0015] Advantages of the present invention: (1) The present invention improves the pitting detection accuracy: The black-and-white camera solves the problem of missed detection of tiny pitting by RGB under low contrast, and the RGB camera solves the problem of misjudging "normal texture as pitting" by the black-and-white camera. The combination of the two greatly improves the overall detection accuracy; (2) This invention is adapted to complex surfaces after shot blasting: the surface of the steel plate after shot blasting is rough and easily reflects light. The high contrast imaging of the black and white camera suppresses the reflection interference, and the YOLO algorithm of RGB distinguishes texture features, covering the entire scene of "strong light-weak light" and "smooth-rough", avoiding the failure of a single camera under extreme conditions. (3) This invention achieves accurate compliance determination: through three-dimensional determination of size, density and oxidation, it meets the detailed requirements of production specifications (such as the different treatment methods for only size exceeding the standard and the accompanying oxidation), and solves the ambiguity of traditional manual inspection based on "judgment based on experience"; (4) The present invention improves detection efficiency and stability: the dual-camera collaborative detection speed is adapted to the speed of the steel plate production line, and long-term stability is maintained through dynamic parameter adjustment, reducing the cost of manual re-inspection; (5) The present invention reduces the dependence on hardware: the grayscale feature parsing algorithm of the black and white camera reduces the computation of the YOLO model of the RGB camera (only focusing on the suspected area), and real-time detection can be achieved without high-end computing equipment. The hardware cost is significantly reduced compared with the pure RGB deep learning solution. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of an intelligent production line for industrial defect detection according to the present invention; Figure 2 This is a schematic diagram of the structure of a rust detection device in an intelligent production line for industrial defect detection according to the present invention. Figure 3 This is a flowchart illustrating the algorithm for assessing the severity of rust in an intelligent production line for industrial defect detection, as described in this invention. Figure 4 This is a schematic diagram of the dual-machine collaborative calibration structure in an intelligent production line for industrial defect detection according to the present invention; Figure 5 This is a schematic diagram of the status of pitting information in an intelligent production line for industrial defect detection according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1-5 As shown, the present invention is an intelligent production line for industrial defect detection, comprising: Step 1: Structural and hardware design for rust removal: This step focuses on rust detection before shot blasting of steel plates, and adopts a hardware architecture that combines an RGB camera with the YOLO algorithm. The specific configuration involves deploying two sets of industrial-grade RGB cameras on each side of the preprocessing section of the steel plate conveying path. Each set of cameras is evenly distributed along the width of the steel plate, and the spacing between adjacent cameras is set according to the width of the steel plate (with the standard of covering the entire surface of the steel plate without blind spots).

[0020] Meanwhile, to improve the imaging clarity of rusted areas, a "rust detection-specific uniform light device" is deployed next to each set of RGB cameras. This device uses high color rendering LED light strips to form a uniform surface light source through a diffuser. The light shines on the steel plate surface at a 30° angle to avoid the interference of reflection caused by direct sunlight. The light intensity can be adjusted by a dimming module to adapt to the imaging needs of steel plates with different degrees of rust.

[0021] The hardware connection adopts a "camera-industrial switch-edge computing terminal" architecture: all four RGB cameras are connected to the industrial-grade switch via gigabit Ethernet interfaces, and the switch is then connected to the edge computing terminal via optical fiber. The terminal has a built-in GPU module to support the real-time operation of the YOLO algorithm.

[0022] Meanwhile, to adapt to industrial environments, the camera is equipped with a dustproof and impact-resistant metal protective shell (made of aluminum alloy), and the front of the lens is equipped with a quick-removable cleaning lens.

[0023] The edge computing terminal has reserved communication interfaces with shot blasting equipment and roller conveyor system to prepare the hardware for subsequent parameter adjustment.

[0024] For example: The hardware design in step one includes a detection component, an auxiliary imaging component, and a data transmission and processing component; Inspection components: Employs an industrial-grade RGB camera, which has stable imaging capabilities adapted to industrial environments and can capture color images of the steel plate surface to identify rusted areas; Auxiliary imaging components: Equipped with a "corrosion detection-specific light homogenizing device", which includes a high color rendering LED light strip and a diffuser. The LED light strip can provide a stable light source, and the diffuser can convert the light into a uniform surface light source. Meanwhile, the camera is equipped with a metal protective shell and a quick-removable cleaning lens at the front of the lens. Data transmission and processing components: including industrial-grade switches and edge computing terminals. The industrial-grade switches are used for data relay, and the edge computing terminals have built-in GPU modules that can support the YOLO algorithm. Furthermore, such as Figure 2As shown, the RGB camera arrangement is as follows: Two sets of cameras are deployed on each of the upper and lower sides of the steel plate, with each set evenly distributed along the width of the steel plate to ensure that the combined shooting range of all cameras can completely cover the surface of the steel plate without any blind spots. The dedicated light-diffusing device for rust detection is deployed next to each set of RGB cameras, corresponding to one of them. Its light shines on the surface of the steel plate at an angle to avoid direct reflection and ensure clear imaging of the rusted area. Camera and switch connection: All four RGB cameras are connected to an industrial-grade switch via network interfaces to achieve initial transmission of image data; Switch and terminal connection: The industrial-grade switch is connected to the edge computing terminal via optical fiber, transmitting the image data captured by the camera to the terminal for processing; Terminal connection with external devices: The edge computing terminal has a reserved communication interface for subsequent connection with shot blasting equipment and roller conveying system, providing a hardware connection basis for parameter adjustment based on corrosion detection results.

[0025] Step Two: Development of the Corrosion Severity Algorithm: Based on the RGB images acquired in step one, a corrosion severity determination model was developed using the YOLO algorithm. First, the acquired images are preprocessed, including image denoising (using median filtering) and size normalization, uniformly adjusted to 640×640 pixels; Then, the YOLO algorithm is used to locate the rusted areas in the image and output the bounding boxes of the rusted areas; Next, feature parameters of the rusted area are extracted, such as the area ratio of the rusted area, average color saturation, and texture complexity. Finally, based on these characteristic parameters, a standard for classifying rust levels was established, dividing the degree of rust into light, moderate, and severe.

[0026] Among them, mild corrosion is defined as corrosion area accounting for less than 5% and with shallow texture; Moderate corrosion is defined as corrosion affecting 5%-15% of the total area. Severe corrosion is defined as corrosion covering more than 15% of the area, or the presence of obvious deep corrosion textures. The model was trained using a large number of labeled samples (no fewer than 2000 images), achieving an accuracy rate of over 90% in corrosion level determination.

[0027] For example: like Figure 3 As shown, the process of developing a corrosion severity determination model based on the RGB image acquired in step one and relying on the YOLO algorithm is as follows: Image preprocessing and median filtering denoising principle: Replace the pixel value with the median of the pixel's neighborhood gray values ​​to suppress salt-and-pepper noise. Formula: Let the original image be f(x,y), the filtered image be g(x,y), and the neighborhood window size be... (Usually odd numbers are used, such as ( ),but: ; In this formula, the parameters are defined as follows: g(x,y): The pixel gray value at coordinates (x,y) in the filtered image.

[0028] f(x,y): The pixel grayscale value of the original image at coordinates (x,y).

[0029] median: The median function is used to calculate the median gray value of a pixel within a neighborhood.

[0030] m×n: The size of the neighborhood window, usually an odd number (e.g., 3x3), where m is the number of rows in the window and n is the number of columns in the window.

[0031] i: The offset of the neighborhood window in the row direction, with a value range of... ,.., .

[0032] j: The offset of the neighborhood window in the column direction, with a value range of... ,.., .

[0033] For each pixel, take the median gray value of all pixels in its neighborhood as the new value of that pixel. Size normalization principle: Images are scaled to a uniform size (e.g., 640×640) to fit the input requirements of the YOLO model. Formula: Let the original image width and height be ((W,H)), and the target size be (). , The mapping relationship between the pixel (x,y) of the scaled image and the coordinates (x,y) of the original image (taking bilinear interpolation as an example): (Grayscale values ​​are calculated by weighting neighboring pixels; the specific logic is encapsulated by an image processing library.) In this size normalization formula, the parameters are defined as follows: x', y': The coordinates of pixels in the scaled image.

[0034] x, y: The coordinates of the pixels (x', y') in the original image and the corresponding pixels in the scaled image.

[0035] W, H: Width and height of the original image.

[0036] Wtar, Htar: target size (i.e., the width and height of the scaled image, such as 640 in 640×640).

[0037] The YOLO corrosion region localization algorithm's core logic is based on deep learning. It divides the image into a grid, and for each grid, it predicts the bounding box (x, y, w, h) and class confidence of the target (corrosion region). The bounding box loss function (taking CIoU loss as an example) is as follows: Let the prediction box be: ; In these formulas, the parameters are defined as follows: The predicted bounding box is the bounding box of the rusted area predicted by the YOLO algorithm.

[0038] , : Center coordinates of the prediction box (horizontal and vertical coordinates in the image grid).

[0039] , : The width and height of the prediction box.

[0040] The actual bounding box is: ; Ground Truth Box: The actual bounding box of the rusted area in the image.

[0041] , : The center coordinates (horizontal and vertical coordinates) of the true bounding box.

[0042] , The width and height of the actual frame.

[0043] but: Where: IoU (Intersection over Union): The ratio of the overlapping area of ​​the predicted bounding box and the ground truth bounding box to the total area of ​​the two, which measures the localization accuracy; : The squared Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; c: The length of the diagonal of the smallest rectangle enclosing the two frames; Weighting coefficients; v: Measures the consistency of the aspect ratio between the two frames; The principle behind the proportion of rusted area: It involves calculating the percentage of pixels in the rusted area relative to the total number of pixels in the image. The formula is: ; In this formula, the parameters are defined as follows: : Area percentage of rusted region, which is the proportion of the number of pixels in the rusted region to the total number of pixels in the entire image.

[0044] : Number of pixels in the rusted area.

[0045] The total number of pixels in the entire image.

[0046] Average color saturation and texture complexity: Principle: The image is converted to HSV space to obtain saturation information and the mean is calculated; texture features are extracted through the gray-level co-occurrence matrix to reflect the complexity of the texture; Corrosion level determination model: Principle: The corrosion level is determined by combining the area ratio of the rusted area and the texture and other characteristic parameters. Judgment rules: Light corrosion: AreaRatio < 5% and shallow texture; Moderate corrosion: 5% ≤ AreaRatio ≤ 15%; Heavy corrosion: AreaRatio > 15% or obvious deep corrosion texture.

[0047] Step 3: Analysis of the corrosion degree results and roller linkage Construct a closed-loop control logic of "detection-analysis-decision-execution"; After receiving the corrosion degree results output from step two, the edge computing terminal performs analysis and processing: When the corrosion is determined to be minor, a prompt is generated to reduce the shot blasting amount; When the corrosion is determined to be moderate, a command is generated to maintain the constant speed and shot blasting amount. When the corrosion is determined to be severe, a command to reduce speed and increase shot blasting amount is generated; Control commands are sent to the controllers of the roller conveyor system and the shot blasting equipment via the Modbus RTU protocol. The specific linkage rules are as follows: For minor corrosion, send a speed-up command to the roller controller (increase the conveying speed by 15%-20% compared to the reference speed) and a shot blasting equipment command to reduce the shot blasting amount (reduce the shot blasting amount by 20%-30% compared to the reference amount). When there is moderate corrosion, a maintenance command is sent to the roller controller to maintain the roller conveying speed and shot blasting amount at a constant level. In cases of severe corrosion, a deceleration command is sent to the roller controller (the conveying speed is reduced by 20%-25% compared to the reference speed), and an increase in shot blasting quantity is sent to the shot blasting equipment (the shot blasting quantity is increased by 30%-50% compared to the reference quantity). After receiving the command, the controller changes the roller speed by adjusting the motor frequency and adjusts the shot blasting amount by controlling the opening and closing of the feeding valve, thus achieving real-time linkage control. For example: Foundation for closed-loop control logic construction: With "detection-analysis-decision-execution" as the core closed-loop control logic, and relying on edge computing terminals, roller conveyor system controllers, shot blasting equipment controllers and related communication protocols, real-time linkage between rust detection results and equipment operation is achieved; The edge computing terminal, as the core processing node, receives the corrosion degree results output in step two and completes subsequent analysis, decision-making, and instruction sending. Corrosion degree result analysis and instruction generation: Result reception and parsing: The edge computing terminal receives the corrosion degree results (light, moderate, and severe) output from step two through its internal data interface, and verifies the validity of the results to ensure that the data is accurate.

[0048] Command transmission and reception: Communication protocol adaptation: The edge computing terminal communicates with the roller conveyor system controller and the shot blasting equipment controller using the Modbus RTU protocol. The terminal needs to be pre-configured with protocol-related parameters, such as baud rate, data bits, stop bits, and parity check, to ensure that they match the controller parameters. Command transmission: The edge computing terminal encapsulates the generated control commands according to the Modbus RTU protocol format and sends them to the roller conveyor system controller and shot blasting equipment controller respectively through dedicated communication lines; Command reception and feedback: After receiving a command, the controller parses and verifies it. If the command format is correct and the content is valid, it returns a successful reception feedback to the edge computing terminal. If there is an error, it returns an error message, and the terminal needs to resend the command.

[0049] Equipment execution and adjustment: Roller conveyor system adjustment: The roller controller adjusts the motor speed by changing the power supply frequency of the drive motor according to the received speed adjustment command; when it receives an acceleration command, it increases the motor frequency; when it receives a deceleration command, it decreases the motor frequency; when it receives a maintenance command, it keeps the current motor frequency unchanged, thereby realizing the corresponding change in roller conveyor speed. Shot blasting equipment adjustment: The shot blasting equipment controller controls the opening and closing degree of the feeding valve according to the shot blasting quantity adjustment command; when receiving a command to reduce the shot blasting quantity, it reduces the valve opening and closing degree; when receiving a command to increase the shot blasting quantity, it increases the valve opening and closing degree; when receiving a maintenance command, it keeps the valve opening and closing degree unchanged, thereby changing the supply of shot blasting material to achieve the purpose of adjusting the shot blasting quantity.

[0050] Linkage control monitoring and optimization: Operational status monitoring: Install sensors (such as speed sensors and flow sensors) on the roller conveyor system and shot blasting equipment to collect equipment operating parameters (conveyor speed, shot blasting volume) in real time, and feed them back to the edge computing terminal through communication lines; Closed-loop optimization: The edge computing terminal compares and analyzes the actual operating parameters fed back with the target parameters required by the instructions. If the deviation is within the allowable range, the current control state is maintained; if the deviation exceeds the range, the adjustment instructions are regenerated and sent to the controller until the device operating parameters meet the requirements, ensuring the accuracy and stability of the linkage control.

[0051] Step 4: Structural and hardware design for pit detection For the detection of pitting on steel plates after shot blasting, two sets of detection units are deployed on each of the upper and lower sides of the conveyor section after shot blasting. Each detection unit includes one RGB camera based on the YOLO algorithm, one black and white camera based on the grayscale feature analysis algorithm, and one set of "dedicated spotlight device for pitting detection". This device uses an array of LED point light sources and focuses on the camera detection area at a 45° oblique angle. The strong light highlights the concave and convex contours of the pitting (the pitting depressions form shadow contrasts). The brightness of the light source supports graded adjustment and can be dynamically adapted according to the reflective characteristics of the steel plate surface after shot blasting. The two sets of detection units are arranged sequentially along the steel plate conveying direction, with the spacing between them preferably to avoid interference with image acquisition. In terms of hardware connectivity, a total of 8 cameras (4 RGB cameras and 4 monochrome cameras) and 4 sets of focusing devices on the top and bottom sides are connected to the image processing server through industrial Ethernet switches (the focusing devices adjust brightness through the PWM signal output by the server). The switches and the server are connected by gigabit Ethernet cables to ensure stable transmission of image data. To achieve synchronous image acquisition and processing, a multi-channel image acquisition card is installed in the server. Each camera and focusing device is connected to the acquisition card via a trigger line to ensure that the image of the steel plate at the same location is acquired at the same time and the light source is turned on synchronously. Meanwhile, the black and white camera adopts a high-resolution model (resolution of no less than 2 million pixels), while the RGB camera adopts a low-light model to adapt to different lighting environments.

[0052] For example: Detection unit deployment plan: Deployment location: Set up an inspection area in the steel plate conveying section after shot blasting. Deploy 2 sets of inspection units on the upper and lower sides of the steel plate. During installation, ensure that it does not affect the steel plate conveying and covers the complete surface. Group setup: Each group includes an RGB camera (based on the YOLO algorithm), a monochrome camera (based on the grayscale feature analysis algorithm), and a spotlight device. The positions need to be adjusted so that the illumination range coincides with the shooting range. Inter-group spacing: Two sets of detection units are arranged along the steel plate conveying direction, with the spacing based on the standard of no image acquisition interference; Camera selection: The RGB camera is equipped with a YOLO model, pre-trained with three types of features: "pockmarks", "normal textures", and "oxidation spots"; Light source selection: An array of LED point light sources is used, with a preset initial spot diameter that can be adjusted according to the actual detection situation; Irradiation angle and brightness: Focus on the detection area at a 45° oblique angle; brightness supports graded adjustment. During the debugging phase, test the optimal brightness under different reflective characteristics and establish an adjustment reference table. Core architecture: An industrial Ethernet switch serves as the data hub, connecting 8 cameras and 4 sets of focusing devices; the switch is connected to the image processing server via gigabit Ethernet cable, and the cable and interface are checked before connection; Concentrator control: Connects to the server via a line, receives PWM signals to adjust brightness, and tests signal transmission stability after connection; Synchronous triggering: The server is equipped with a multi-channel image acquisition card, and the camera and focusing device are connected to it through a trigger line. The line is straightened to avoid tangling and to ensure signal synchronization. Synchronous debugging: The acquisition card sends a synchronization signal to control the camera and light source to work synchronously, monitors the response time, and ensures that the synchronization error is within the allowable range; Dual camera connector: Core structural design: The connector adopts a frame structure and mainly includes three functional modules: Dual camera mount: The design includes two independent camera mounting positions, one for a monochrome camera and the other for an RGB camera. Each mounting position features an adjustable clamping structure that secures the camera via elastic clips or a lockable movable baffle, adapting to different camera housing shapes. The bottom of each mounting position provides space to allow for lens clearance, ensuring the lens remains unobstructed and its optical axis remains parallel (deviations can be corrected with subsequent fine-tuning). Spacing adjustment component: The two camera mounting positions are connected by a sliding guide rail, allowing for horizontal spacing adjustment. During adjustment, scale markings assist in positioning, and the position is secured by a locking knob after adjustment. This structure allows for flexible spacing adjustment based on inspection needs (such as the detection range of pitting on steel plates or the field-of-view coverage requirements of the two devices), ensuring that the overlapping area of ​​their fields of view meets inspection requirements. Reference mounting bracket: The connector has a connection interface at the bottom for use with testing equipment (such as a testing line bracket), and uses a universal snap-fit ​​or bolt fixing structure. A fine-tuning angle is provided at the interface, allowing adjustment of the connector's level using shims or a rotatable connection structure to ensure that the optical axes of the two camera lenses are perpendicular to the steel plate surface. Connector functional design: Position fine-tuning function: Each camera mounting position supports minute angle adjustments (such as tilt and horizontal rotation) via adjustment knobs on the edge of the mounting position. After adjustment, it can be locked in place to ensure the camera position does not shift during testing, guaranteeing precise spatial correspondence between the images acquired by the two devices and providing a foundation for subsequent data fusion. Cable management structure The connector has a cable slot on the side, which can be inserted and fixed to the camera's power cable and data cable to prevent the cable from shaking and interfering with the detection or getting tangled in the equipment, while keeping the overall structure neat. Adaptability to the testing process After the connector is installed, it needs to be debugged in conjunction with the camera calibration process: by adjusting the spacing and angle, the positional deviation of the same mark point in the calibration board images captured by the two cameras is within the set range. During subsequent testing, the stability of the connector can ensure that the fields of view of the two cameras always overlap, avoiding detection deviations caused by relative displacement of the equipment, and ensuring the accuracy of pitting identification, size calculation and data fusion.

[0053] like Figure 4 As shown, camera calibration (establishing the correspondence between pixels and physical dimensions) The first step is to perform intrinsic parameter calibration. Place the checkerboard calibration plate flat on the steel plate to be tested, and take images of the calibration plate from different angles using two cameras (at least eight images). Calculate the intrinsic parameter matrix using the Zhang Zhengyou calibration method. The formula is: In this formula, K is the camera intrinsic parameter matrix, which describes the camera's internal optical characteristics and imaging geometry, and is used to establish the transformation relationship between the camera coordinate system and the image pixel coordinate system. and These are the focal lengths (in pixels) in the x and y directions, respectively. and Principal point coordinates (pixels); The second step is distortion correction. To address lens imaging distortion, a radial distortion correction formula is used to correct the image. In the formula, (Distance from pixel to principal point).

[0054] , The distortion coefficient describes the degree of radial distortion of the lens and is obtained through camera calibration.

[0055] , : The coordinates of pixels in the image after distortion correction.

[0056] x, y: The coordinates of pixels in the image before distortion correction.

[0057] The third step is to establish the correspondence between pixels and physical dimensions. Calibration is used to determine the actual physical size corresponding to each unit pixel, which is then used for subsequent calculations of the actual size of the pitting. The fourth step is dual-camera collaborative calibration. Using the same calibration board as a reference, the positional transformation relationship between the two cameras is calculated to ensure that the detection areas of the two cameras correspond accurately.

[0058] Step 5: Grayscale Feature Analysis Algorithm for Black and White Cameras Develop traditional machine vision detection algorithms based on grayscale images captured by black-and-white cameras. First, the image is preprocessed by using Gaussian filtering to remove industrial noise from the industrial environment. Then, contrast enhancement (using histogram equalization) is used to highlight the grayscale difference between the pits and the steel plate surface. Finally, a threshold segmentation method (using dynamic thresholding to determine the segmentation threshold) is used to separate the pitted area from the steel plate surface image. Next, morphological operations (erosion followed by dilation) are used to remove small noise regions after segmentation; finally, the number of pits is counted using a connected component labeling algorithm, and parameters such as the maximum diameter and area of ​​each pit are calculated. To improve the algorithm's anti-interference capability, texture suppression processing was added to address potential texture interference on the steel plate surface after shot blasting. Texture regions were excluded by calculating the local variance of the image.

[0059] For example: Image acquisition and preprocessing: The first step is to use a black and white camera to capture images of the steel plate surface and obtain grayscale images of the detection area. The second step is Gaussian filtering for noise reduction. A Gaussian filter is used to process the image; the formula is: : Represents the pixel grayscale value at coordinates (x, y) in the image after Gaussian filtering.

[0060] : is the normalization coefficient, used to ensure that the overall brightness of the filtered image does not change significantly due to the weighted summation (determined by the weights of the Gaussian filter template, where the template weights are 16, so it is divided by 16).

[0061] i = -1,1, j = -1,1: This represents the range of offsets of the neighborhood window in the row and column directions. Combined with the center pixel (x,y), it indicates a neighborhood that extends by 1 pixel in both the horizontal and vertical directions with (x,y) as the center (i.e., a 3x3 neighborhood window).

[0062] w(i,j): represents the weight coefficients of the Gaussian filter template, used to weight pixels at different positions in the neighborhood (the weights of the Gaussian filter are generated by the Gaussian function, and the closer the pixel is to the center, the greater the weight). : Represents the pixel gray value of the original image I at coordinates (x+i, y+j), that is, the original gray value of each pixel in the neighborhood.

[0063] The third step is contrast enhancement. The CLAHE algorithm is used to process the filtered image to improve the grayscale difference between the speckles and the surrounding areas, in preparation for subsequent detection. Pockmark contour extraction and preliminary identification: The first step is edge detection, which uses the Sobel operator to calculate the image gradient, as shown in the following formula: : Represents the gradient response of the image in the horizontal direction (x direction), reflecting the intensity of grayscale changes of pixels in the horizontal direction.

[0064] : Represents the gradient response of the image in the vertical direction (y direction), reflecting the intensity of grayscale changes of pixels in the vertical direction.

[0065] and : is the template for the Sobel operator, used to calculate the gradients in the horizontal and vertical directions, respectively. The values ​​in the template are weight coefficients, and the gradient is calculated by convolving with neighboring pixels of the image.

[0066] I: Represents the original image, which is the input image for which edge detection needs to be performed.

[0067] G: represents the gradient magnitude, used to measure the edge strength at a pixel. The larger the value, the more likely the pixel is to be an edge point.

[0068] : is the formula for calculating the gradient magnitude. It combines gradient information from both directions by taking the square root of the sum of the squares of the horizontal and vertical gradients to characterize the edge intensity.

[0069] Threshold: A criterion used to filter edge points. When the gradient magnitude is greater than the threshold, the pixel is considered an edge point.

[0070] Through the gradient magnitude formula Calculate edge intensity, set threshold to filter edge points, and initially outline the pockmark contours; The second step is binarization. The Otsu algorithm is used to determine the optimal threshold T, as shown in the formula: : indicates the threshold T that maximizes the expression within the parentheses, i.e., finding the optimal binarization threshold T.

[0071] T: is the binarization threshold used to divide image pixels into two categories (such as foreground and background, or mottled areas and non-mottled areas).

[0072] : Indicates the proportion of first-class pixels (such as background or non-spotted areas) in the entire image when the threshold is 0.

[0073] : Represents the proportion of second-class pixels (such as foreground or speckled areas) in the entire image when the threshold is T, and satisfies + = 1.

[0074] : This represents the intra-class variance of the first class of pixels when the threshold is T, reflecting the degree of dispersion of the grayscale values ​​of this class of pixels.

[0075] : This represents the intra-class variance of the second type of pixels when the threshold is T, which also reflects the degree of dispersion of the gray values ​​of this type of pixels.

[0076] This is the weighted sum of the intra-class variances. The core idea of ​​the Otsu algorithm is to find the threshold T that minimizes this weighted sum (or maximizes the inter-class variance), thereby achieving optimal binarization segmentation.

[0077] in , The proportions of two types of pixels, , The image is converted into a binary image based on a threshold to represent the intra-class variance, so that the speckled areas are displayed in white. The third step is connected component identification. The 8-neighborhood labeling method is used to mark connected white regions in the binary graph. The determination criteria are as follows: like Within the image range and (x,y): Represents the coordinates of the current pixel in the binary image.

[0078] dx,dy: are the neighborhood offsets of a pixel, with values ​​ranging from {-1,0,1}, representing the pixel offset within an 8-neighborhood (up, down, left, right, and four diagonal directions) centered at (x,y).

[0079] : Represents the coordinates of a pixel within the neighborhood, which must be within the image area.

[0080] This is the conditional expression for determining connected components, where: I represents the pixel value of the binary image (here it is determined to be 1, that is, a white pixel, which is part of the connected component).

[0081] It can be understood as a metric parameter for pixel distance or connectivity (with a value of 2, used to define the rules for determining connected components).

[0082] , These are the maximum and minimum coordinates of the region in the x-direction, respectively.

[0083] , These are the maximum and minimum coordinates of the region in the y-direction, respectively.

[0084] The expression within the square root is a quantization calculation of the shape or range of a region, used to determine whether pixels belong to the same connected region.

[0085] 8-neighborhood labeling is a neighborhood rule for connecting components, which determines whether a pixel is connected to the center pixel in 8 adjacent directions (including the diagonal).

[0086] Connected white regions: These are regions in a binary image where the pixel value is white (usually 1) and they are interconnected by the 8-neighborhood rule. Each such region is marked as a suspected mole.

[0087] Each connected component is treated as a suspected pockmark, and its location information is recorded. Calculation and selection of pit size: The first step is to calculate the pixel size of the suspected pits, based on the bounding rectangle of the connected components, using the formula... The smallest outer diameter (in pixels) of the suspected pit.

[0088] , : The maximum and minimum coordinates of the connected component along the x-direction.

[0089] , : The maximum and minimum coordinates in the y-direction of the connected component.

[0090] Calculate the minimum circumcircle diameter (in pixels); The second step is to convert to actual size. Based on the correspondence between pixels and physical size, the actual diameter of the pit is obtained using (s is the actual size per unit pixel). The third step is to screen valid targets, eliminate noisy areas that are too small, and retain suspected spots that meet the set range.

[0091] Step Six: Detection Process of RGB Camera Based on YOLO Algorithm First, a dataset of speckled images is constructed, containing speckled images of different sizes and shapes, and then labeled (marking the location and category of the speckles). Then, YOLOv5s was selected as the base model for training. During the training process, the network structure of the model was adjusted to meet the needs of speckle detection (such as adding a small object detection layer). After training, the model is optimized by removing duplicate detection boxes using non-maximum suppression (NMS) and setting a confidence threshold to filter out low-confidence detection results.

[0092] This algorithm can quickly locate the spots in an image and output the bounding box coordinates and confidence scores of the spots. At the same time, it uses the color information of the RGB image to eliminate color interference areas (such as oil stains) that are not spots.

[0093] For example: Target region extraction: Based on the suspected spot locations output by the black and white camera, the region of interest (ROI) is cropped out from the image acquired by the RGB camera with the suspected spot as the center, thus narrowing the detection range; Feature recognition and verification The first step is YOLO algorithm detection. The ROI image is input into the YOLO model, which outputs the target category (spots, normal texture, etc.) and confidence score. When the confidence score reaches the set value, it is considered a valid recognition. The second step is texture feature verification, calculating the gray-level co-occurrence matrix contrast (Contrast) of the ROI region, using the formula: Where P(i,j) represents the probability of gray levels i and j appearing adjacently, the texture roughness of the region is analyzed. Finally, by combining this contrast result with the recognition result of the YOLO algorithm, it is determined whether the region is a real speckle.

[0094] Step 7: Data Fusion and Result Determination Based on dual data fusion and compliance assessment, a data fusion module was designed to fuse the detection data from black and white cameras and RGB cameras. Based on the location of the speckles detected by the RGB camera using the YOLO algorithm, the speckle parameters (quantity, size, etc.) at the corresponding locations detected by the black and white camera are correlated and matched to form complete speckle detection data; Compliance determination is based on pre-set production standards. The fused speckle data is compared with these standards. The system integrates the outline data of the black and white camera and the feature data of the RGB camera. If all indicators meet the requirements, the steel plate is deemed qualified and judged according to production specifications. Size determination: The smallest circumscribed circle diameter detected by the black and white camera shall be used as the standard; Density determination: Count the number of valid pits within a certain area after RGB verification; Association judgment: If RGB detects pitting accompanied by oxidation characteristics, it is directly judged as "severely exceeding the standard" (rework is required); the final output is one of four categories of results: "qualified", "size exceeding the standard", "density exceeding the standard" and "severely exceeding the standard".

[0095] For example: like Figure 5 As shown, data association matching: based on the location (boundary box coordinates) of the speckles detected by the RGB camera using the YOLO algorithm, the speckle information at the corresponding location is searched in the detection data of the black and white camera; By calculating the coordinate deviation of the pockmark positions detected by the two cameras (if the center of the pockmark detected by RGB is...), The center of the black and white detection corresponding to the location of the pockmark is The deviation value is ; When the deviation is within the preset range, it is determined to be the same pit, and the number, size and other parameters of the pit detected by the black and white camera are associated with the RGB detection results; Data fusion generation: Integrate the location, quantity, size and feature information (such as whether there is oxidation feature) of the associated pits to form a pit detection dataset containing complete information.

[0096] Compliance assessment: Judgment criteria and preset standards: Based on production standards, preset threshold values ​​for pitting size are established. Density threshold Other judgment indicators.

[0097] Specific judgment rules: Size determination: The smallest circumcircle diameter of the pit detected by a black and white camera. Based on this, if D > It was determined to be "oversized"; Density determination: Count the number of valid pits within a set area after RGB verification. If N> It was determined to be "density exceeding the standard"; Association judgment: If the RGB camera detects pitting accompanied by oxidation characteristics, it is directly judged as "severely exceeding the standard" (rework is required); Acceptance criteria: If the size and density of the pits do not exceed the standard and there is no correlation with oxidation characteristics, it is judged as "acceptable". The four categories of results are "seriously exceeding the standard".

[0098] Step 8: Results Feedback and System Adjustment Feedback on test results and dynamic adjustment of parameters: Establish a feedback mechanism for test results to enable dynamic adjustment of the production process; When the steel plate is determined to be qualified, the system sends a continue conveying instruction to the roller conveying system and uploads the test data to the production management system for archiving. If the defect is determined to be minor, the system analyzes the cause of the defect (e.g., a slightly higher number of pits) and sends a fine-tuning instruction to the shot blasting equipment (e.g., appropriately increasing the shot blasting amount by 5%-10%). If the system determines that the problem is serious, it will immediately trigger an alarm signal (audio and visual alarm) and send a pause command to the roller conveyor system. Subsequent operations will only be carried out after manual confirmation and processing. In addition, the system will regularly perform statistical analysis on the test data. If non-conforming products or specific types of defects occur continuously, it will automatically generate optimization suggestions for shot blasting process parameters (such as adjusting the shot blasting angle, changing the shot blasting material, etc.) and push them to the relevant management terminals. For example: Test result feedback: Steel plate inspection result processing plan: Test result confirmation: When the system determines that the steel plate is qualified, the display device clearly shows the test image of the steel plate (with no excessive pitting marks) and the "qualified" judgment result; Processing method: The system automatically records the inspection time, serial number, specifications, and other information of the steel plate. The steel plate is then automatically conveyed to the next production process via a conveyor belt, requiring no manual intervention. Subsequent connection: After the equipment in the next process receives the qualified steel plate, it will automatically start the corresponding production operation to ensure the continuity of the production process; Process for handling steel plates with slight defects: Alarm Response: When the system determines that the steel plate is slightly out of control, it will immediately light up the yellow indicator light and sound an alarm for 10 seconds to ensure that the operator can notice it in time. Handling procedure: After hearing the prompt sound and seeing the yellow indicator light, the operator should proceed to the testing area to check; The steel plate is removed from the conveyor belt and returned to the shot blasting process for reprocessing. During the re-shot blasting, enhanced parameters must be set, increasing the shot blasting intensity by 20% and extending the shot blasting time by 15 seconds compared to the standard, to improve the treatment effect. Re-inspection: The steel plate that has been shot blasted again is conveyed back to the inspection area, where the system performs a second inspection. Subsequent procedures: If the second inspection is passed, proceed to the next production step according to the qualified steel plate processing procedure; if it is still slightly out of standard, repeat the above shot blasting and inspection process, up to a maximum of 3 times. If it is still slightly out of standard after 3 times, it will be treated as a seriously out-of-standard steel plate. Process for handling steel plates with severely substandard specifications: Alarm Response: When the system determines that the steel plate is seriously out of standard, it will immediately light up the red indicator light and emit a continuous alarm sound. The alarm sound will not stop automatically until the operator manually turns it off. At the same time, the system will send information about serious exceedances to the monitoring terminal of the management personnel for a secondary reminder; Handling procedure: Upon receiving the alarm information, operators and managers should quickly arrive at the inspection area. They should review the steel plate inspection image and any parameters exceeding the standard (such as "density exceeds standard: 7 particles / 100cm²"), confirm a serious exceedance, remove the steel plate from the conveyor belt, place it in the scrap area, and mark the reason for scrapping and its serial number on the steel plate. Recording and Analysis: The system automatically records relevant information and reasons for scrapping the steel plate. Management personnel periodically analyze cases of severely substandard steel plates to identify potential problems in the production process. The system also includes methods for handling dynamic optimization. Adjustment prompt: When the system detects that the speckle recognition deviation between the black and white camera and the RGB camera is too large in 5 consecutive frames of images, the system will display a prompt message on the display device: "The system is adjusting parameters, please wait." Inspection pause: During the period when the system automatically adjusts the exposure time of the black and white camera (or the brightness of the LED light source), the inspection operation of the steel plate is paused, and the conveyor belt is also stopped to prevent unqualified or inaccurately inspected steel plates from entering the next process. Resumption of testing: Once the system adjustment is complete and the identification deviation is within the normal testing range, the display device will prompt "System adjustment complete, resume testing," and the conveyor belt will automatically start, resuming the normal testing and processing flow.

[0099] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An intelligent production line for industrial defect detection, characterized in that, Includes a rust detection camera unit, a shot blasting parameter adjustment unit, a pitting detection unit, a quality judgment unit, and a steel plate production control platform: A corrosion detection unit, deployed at the corrosion detection station, is used to collect surface images of the steel plate before shot blasting and identify the corrosion level. The shot blasting parameter dynamic adjustment unit is connected to the equipment signal of the rust detection unit and the shot blasting station, and is used to dynamically adjust the shot blasting amount and steel plate conveying speed according to the identified rust level. The pitting detection unit is deployed at the pitting detection station to collect surface images of the steel plate after shot blasting and to identify and analyze pitting defects based on a dual-camera collaborative detection mechanism. The quality assessment unit is signal-connected to the pitting detection unit and is used to determine the quality grade of the steel plate based on the pitting detection results. The steel plate production control platform is used to generate corresponding control commands and feed them back to the production line.

2. The intelligent production line for industrial defect detection according to claim 1, characterized in that, The corrosion detection unit includes at least two sets of industrial RGB cameras, a dedicated light homogenizing device, and an edge computing terminal; The RGB cameras are symmetrically deployed on the upper and lower sides of the steel plate and are mounted using adjustable brackets. The dedicated light-diffusing device uses high color rendering LED light strips and a diffuser to provide uniform illumination to the steel plate surface at an inclined angle; The edge computing terminal has a built-in GPU module for running a corrosion-level identification model based on the YOLO algorithm.

3. The intelligent production line for industrial defect detection according to claim 2, characterized in that, The judgment logic of the corrosion level identification model based on the YOLO algorithm is as follows: the level is divided according to the proportion of the corrosion area to the total surface area of ​​the steel plate. The area proportion is less than 5% and there is no deep corrosion, which is light corrosion; the area proportion is between 5% and 15%, which is moderate corrosion; and the area proportion is greater than 15% or there is deep corrosion texture, which is heavy corrosion.

4. The intelligent production line for industrial defect detection according to claim 3, characterized in that, The adjustment rules of the shot blasting parameter dynamic adjustment unit are as follows: When the corrosion level is light, reduce the shot blasting amount by 20%-30% and increase the conveyor speed by 15%-20%; When the corrosion level is moderate, maintain the current shot blasting volume and conveying speed; When the corrosion level is severe, increase the shot blasting amount by 30%-50% and reduce the conveyor speed by 20%-25%.

5. The intelligent production line for industrial defect detection according to claim 1, characterized in that, The pitting detection unit includes multiple detection groups deployed on the upper and lower sides of the steel plate. Each detection group includes an RGB camera based on the YOLO algorithm, a black and white camera based on the grayscale feature analysis algorithm, and a set of special spotlight detection device for pitting. The RGB camera and the monochrome camera are rigidly connected and aligned in the field of view through a dual-camera connector that allows for adjustment of spacing and angle.

6. The intelligent production line for industrial defect detection according to claim 1, characterized in that, The dual-camera connector includes: a dual-camera mounting base, which has two independent mounting positions with adjustable clamping structures; The spacing adjustment component connects two mounting positions via a sliding guide rail, enabling stepless adjustment and locking of the horizontal spacing; The reference mounting bracket has an interface for fixing to an external bracket and supports fine-tuning of the overall angle.

7. The intelligent production line for industrial defect detection according to claim 1, characterized in that, The workflow of the pit detection unit includes: the black and white camera acquiring grayscale images, extracting pit outlines through grayscale threshold segmentation and morphological processing, and initially calculating the preliminary number and size of the pits; The RGB camera captures color images, and uses the YOLO algorithm to verify the features of the suspected pitted areas detected by the black-and-white camera, identify the pits and determine whether there are accompanying oxidation features, so as to achieve data fusion of the two and output complete pitted information.

8. The intelligent production line for industrial defect detection according to claim 1, characterized in that, The decision logic of the quality determination unit is as follows: Qualified: The size and quantity of the pits do not exceed the standard, and there is no oxidation feature; Size exceeding the standard: The diameter of the minimum circumscribed circle of the pit is greater than the preset threshold; Density exceeding the standard: The number of effective pits per unit area is greater than the preset threshold; Seriously exceeding the standard: The pits are accompanied by oxidation features.

9. A smart production line for industrial defect detection according to claim 8, characterized in that, The quality determination unit performs the following feedback control according to the determination result: If it is determined to be qualified, control the conveying system to convey the steel plate to the next process; If it is determined to be slightly unqualified, control the shot blasting equipment to fine-tune the parameters and then process the subsequent steel plates; If it is determined to be seriously exceeding the standard, immediately trigger an audible and visual alarm and pause the conveying system, waiting for manual handling.

10. A smart production line for industrial defect detection according to claim 1, characterized in that, The steel plate production control platform receives and integrates all data from the rust detection unit, shot blasting parameter dynamic adjustment unit, pitting detection unit and quality determination and feedback unit through the industrial Ethernet, and realizes real-time monitoring of the production process, query of historical data, early warning of exceeding the standard, and generation and push of process parameter optimization reports.

Images