Full-automatic optical fiber winding system based on machine vision and winding method thereof

Abstract

The invention provides a full-automatic optical fiber winding system based on machine vision, which comprises a working table, a movable bottom plate is arranged on the working table, an optical fiber releasing rotating module is arranged on the movable bottom plate, a vertical frame is arranged on one side of the optical fiber releasing rotating module, and an optical fiber tension control module and an optical fiber arranging and pressing mechanism are arranged on the vertical frame. A fiber collecting rotating module is further arranged on one side of the movable bottom plate on the workbench, the fiber collecting rotating module is provided with a rotatable annular framework cylinder tool, an optical fiber inlet angle collecting camera and a fiber arranging and pressing mechanism which face the annular framework cylinder tool are further arranged on the vertical frame, and the problems that monitoring is not timely and consistency is poor when manual participation is conducted on optical fiber arrangement are solved.

Description

Technical Field

[0001] This invention relates to the field of optical fiber winding, and in particular to a fully automated optical fiber winding system and method based on machine vision. Background Technology

[0002] Fiber optic loop technology is a core technology in fields such as fiber optic gyroscopes, optical sensors, and hydrophones. Its core principle is to wind optical fibers into a loop according to specific rules to achieve phase-sensitive transmission and detection of optical signals. The core principle utilizes the low-loss and high birefringence characteristics of optical fibers to allow two beams of light to propagate in opposite directions within the loop. By detecting the phase difference between the two beams, changes in rotational angular velocity (in the case of a fiber optic gyroscope) or changes in external physical quantities (such as temperature and stress) can be sensed.

[0003] The uniformity of the winding directly determines the accuracy of phase difference detection. Traditional fiber arrangement mainly relies on manual assistance in fiber arrangement and monitoring of fiber winding arrangement. The judgment of the winding condition is relatively subjective, and the same worker may need to supervise multiple fiber arrangement devices. If there is a deviation in the fiber insertion angle or a fiber arrangement defect, it is difficult to correct it in time, resulting in low quality and consistency of fiber ring or hydrophone ring winding. Summary of the Invention

[0004] This invention provides a fully automated optical fiber winding system and method based on machine vision, which solves the problems of untimely monitoring and poor consistency when manually involved in optical fiber winding.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a fully automatic optical fiber winding system based on machine vision, including a worktable, a movable base plate on the worktable, a fiber feeding rotation module on the movable base plate, a support frame on one side of the fiber feeding rotation module, an optical fiber tension control module and a fiber feeding and pressing mechanism on the support frame, a fiber take-up rotation module on one side of the movable base plate on the worktable, the fiber take-up rotation module is equipped with a rotatable ring skeleton cylindrical fixture, and an optical fiber entry angle acquisition camera and a fiber feeding and pressing mechanism facing the ring skeleton cylindrical fixture are also provided on the support frame.

[0006] In a preferred embodiment, the fiber feeding rotation module includes a base frame with a rotatable connecting shaft on the base frame. One end of the connecting shaft is fitted with a fiber feeding disc, and the other end of the connecting shaft is fitted with a second synchronous pulley. The base frame also includes a fiber feeding motor with a first synchronous pulley fitted on the shaft end of the fiber feeding motor. The first and second synchronous pulleys are driven by a first synchronous belt.

[0007] In a preferred embodiment, the fiber tension control module includes multiple rotatable encoders mounted on a stand. The stand is also equipped with a hinged dance wheel, with a swing arm at the hinge point between the dance wheel and the stand. One end of the dance wheel is equipped with a rotatable guide wheel, and the fiber optic cable passes under the guide wheel.

[0008] In a preferred embodiment, the fiber take-up rotation module includes a base, on which a drive shaft and a driven shaft are arranged coaxially at intervals. Each of the drive shaft and the driven shaft is provided with a shaft clamp at one end that is close to each other. The shaft clamp is used to fit the end of the clamping ring skeleton cylindrical tooling.

[0009] In a preferred embodiment, the base is provided with a guide rail, a slidable slider is fitted on the guide rail, and a translation plate is also provided. The lower end of the translation plate is connected to the slider, and the upper end of the translation plate is provided with a mounting seat for connecting the driven shaft. The base is also provided with a rotatable lead screw, and a threaded nut is fitted on the lead screw. The upper end of the nut is connected to the translation plate.

[0010] In a preferred embodiment, the fiber feeding and pressing mechanism includes a translation platform, a rotating platform on the translation platform, a lifting cylinder on the rotating platform, and a pressing rod at the end of the lifting cylinder.

[0011] The preferred solution includes a fiber winding method: The fiber feeding rotating module feeds the fiber, and the optical fiber is wound onto the ring skeleton cylindrical tooling after passing around the fiber tension control module. The fiber take-up rotation module drives the ring skeleton cylindrical tooling to rotate, and the fiber arrangement and pressing mechanism performs fiber pressing and fiber arrangement on the optical fiber. Fiber optic insertion angle acquisition camera and fiber optic alignment image acquisition camera monitor fiber compression points in real time; During fiber arrangement, the fiber is made to have a certain fiber entry angle. The fiber entry angle acquisition camera and the fiber arrangement image acquisition camera sample and calculate the degree of the fiber entry angle in real time and transmit it to the control system. The control system compares the real-time fiber inlet angle with the set theoretical angle, adjusts the position of the fiber feeding and pressing mechanism, and compensates for the fiber inlet angle.

[0012] In the preferred scheme, Edge detection operators using image algorithms are used to identify fiber optic edge components in images; The engineered processing ensures the stability of the detection edges; The fiber insertion angle is measured based on the boundary line of the fitted fiber and the outer contour of the ring skeleton cylindrical tool. Determine whether there is a fiber arrangement defect based on the fiber input and the fiber boundary line already wound on the ring skeleton cylindrical tooling. The fiber insertion angle and the theoretical fiber insertion angle are compared and the fiber insertion defects are combined to adjust the fiber insertion and pressing position of the fiber insertion and pressing mechanism.

[0013] The preferred solution includes an edge line fitting algorithm: First, preprocessing is performed using Canny edge detection to calculate gradients:

[0014]

[0015]

[0016]

[0017] in The gradient value in the x-direction. For the Sobel operator in the x-direction, The gradient value in the y-direction. Here, src is the Sobel operator in the y-direction, src is the original input image, G is the gradient magnitude of the pixel, and θ is the gradient direction angle of the pixel. The algorithm performs pixel chain extraction, searching along the quantized direction and connecting consecutive edge pixels with similar directions to form pixel chains with a consistent direction; In the line segment fitting stage, the least squares linear regression technique is used to approximate each curved pixel chain as a straight line segment, and the starting and ending coordinates are used to represent it. Perform length filtering, remove excessively short line segments based on a preset threshold, retain long line segments with significant meaning, and remove noise interference; Merging two close line segments with similar directions into a longer line segment yields a complete and coherent straight line. The formula for calculating the merged point set is:

[0018] The direction angle of the merged line segment. , The coordinates of each pixel on the line segment before merging are given, where x and y are the average coordinates of each pixel on the line segment before merging. The output results in a clear and accurate set of long straight line segments in the image. By combining the image conditions during the looping process, the target optical fiber straight line is extracted from the image after fine-tuning.

[0019] The preferred solution includes a fiber arrangement defect detection algorithm: Sample collection, collecting images of fiber routing scenes; Perform data augmentation operations on the training set images to increase sample diversity; Data partitioning and preprocessing: all images are normalized in size and pixel values ​​to eliminate the effects of scale and illumination differences; The model is constructed using a convolutional neural network (CNN) architecture, consisting of convolutional layers, pooling layers, and fully connected layers. Through layer-by-layer feature extraction, it achieves the mapping from image pixels to defect categories. Core computation, convolutional layer execution In the operation, z is the neuron output, and σ is the activation function. Let the weight be the weight of the i-th input. Let b be the i-th input value, b be the bias term, and n be the number of input dimensions. The fully connected layer maps the extracted high-dimensional features to the probability distribution of 9 types of defects. Loss function, for example, classification loss The core is to penalize the prediction error of the defect category probability, which is calculated only in the image region containing the defect, where... For classifying losses, This is an indicator function; it is 1 if the i-th grid contains an object, and 0 otherwise. Let be the true probability that an object in the i-th grid belongs to category c. To correspond to the predicted probabilities, classes is the set of all classes; Perform iterative training of the model; Set a defect confidence threshold to eliminate low-confidence false positives, and add rule validation for similar defect types to improve classification accuracy; Key region extraction involves extracting the key ROI (Region of Interest) of the fiber routing area from the acquired images using traditional image algorithms such as threshold segmentation and contour filtering, while eliminating background interference. Defect classification involves inputting the preprocessed ROI into the optimized classification network, which outputs the defect category and its corresponding confidence score. Output the defect type, confidence level, and detection time, generate fiber routing defect detection results, and trigger an alarm or winding machine adjustment command when a defect is detected.

[0020] The beneficial effects of this invention are as follows: it uses a dual-vision device to monitor fiber optic cable placement from different perspectives, and uses an algorithm to identify whether the fiber insertion angle is reasonable, limiting the fiber insertion angle within a safe range; it adjusts the position and angle of the final fiber clamping rod through the movement feedback compensation of the fiber placement mechanism, and corrects the fiber insertion angle in a timely manner; it improves the success rate and accuracy of visual recognition by collecting a large number of fiber placement images and performing machine learning, monitors fiber placement defects such as splitting and overlapping, provides timely alarms, prompts staff to handle the issues, and improves the quality of the winding. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This is a side view of a fully automated fiber optic winding machine.

[0023] Figure 2 This is a front view of a fully automated fiber optic winding equipment.

[0024] Figure 3 This is a partial structural diagram of the fiber tension control module.

[0025] Figure 4 This is a schematic diagram of the fiber feeding and pressing mechanism.

[0026] Figure 5 This is a schematic diagram of the glue application mechanism.

[0027] Figure 6 This is a schematic diagram of a fiber take-up rotary mold.

[0028] Figure 7 This is a schematic diagram of fiber optic arrangement and edge detection.

[0029] Figure 8 This is the logic diagram of the fiber optic compensation algorithm.

[0030] Figure 9 This is the logic diagram of the fiber optic identification algorithm.

[0031] Figure 10 This is a schematic diagram of the fiber entry angle and fiber exit direction.

[0032] Figure 11 This is a schematic diagram of dual-camera fiber routing identification.

[0033] Figure 12 This is a schematic diagram of angle detection.

[0034] Figure 13 This is a diagram of the classification network training process.

[0035] Figure 14 This is a schematic diagram of boundary detection.

[0036] Figure 15 This is a schematic diagram for detecting fiber routing defects.

[0037] In the diagram: Fiber feeding rotation module 1; translation module 101; moving base plate 102; fiber feeding tray 103; fiber feeding motor 104; first synchronous pulley 105; second synchronous pulley 106; first synchronous belt 107; connecting shaft 108; base frame 109; fiber tension control module 2; upright frame 201; guide wheel 202; dance wheel 203; swing arm 204; encoder 205; fiber take-up rotation module 3; base 301; drive shaft 302; driven shaft 303; shaft clamp 304; Shaft seat 305; Fiber take-up motor 306; Third synchronous pulley 307; Fourth synchronous pulley 308; Second synchronous belt 309; Mounting seat 310; Translation plate 311; Slider 312; Guide rail 313; Wire nut seat 314; Lead screw 315; Ring frame cylindrical fixture 4; Fiber optic entry angle acquisition camera 5; Fiber optic arrangement image acquisition camera 6; Fiber optic arrangement and pressing mechanism 7; Translation stage 701; Rotating stage 702; Lifting cylinder 703; Fiber pressing rod 704; Worktable 8; Glue application mechanism 9. Detailed Implementation

[0038] Example 1: like Figure 1-10In this paper, a fully automated fiber optic winding system based on machine vision is provided, including a worktable 8, a movable base plate 102 on the worktable 8, a fiber feeding rotation module 1 on the movable base plate 102, a support frame 201 on one side of the fiber feeding rotation module 1, a fiber tension control module 2 and a fiber placement and pressing mechanism 7 on the support frame 201, a fiber take-up rotation module 3 on one side of the movable base plate 102 on the worktable 8, a rotatable ring skeleton cylindrical fixture 4, and a fiber entry angle acquisition camera 5 facing the ring skeleton cylindrical fixture 4 and a fiber placement and pressing mechanism 7 on the support frame 201.

[0039] The worktable 8 is equipped with a translation module 101, which consists of a sliding table, a guide rail slider mechanism, a lead screw and nut mechanism, and a servo drive motor. The sliding table is connected to a movable base plate 102, and the servo drive motor drives the movable base plate 102 to move linearly.

[0040] It also includes an adhesive application mechanism 9, which is used to apply adhesive to the optical fiber passing through the optical fiber tension control module 2. The adhesive application mechanism 9 has a linear movement mechanism, which can adjust the adhesive application position.

[0041] In a preferred embodiment, the fiber feeding rotation module 1 includes a base frame 109, on which a rotatable connecting shaft 108 is provided. One end of the connecting shaft 108 is fitted with a fiber feeding disc 103, and the other end of the connecting shaft 108 is provided with a second synchronous pulley 106. The base frame 109 is also provided with a fiber feeding motor 104, and the shaft end of the fiber feeding motor 104 is fitted with a first synchronous pulley 105. The first synchronous pulley 105 and the second synchronous pulley 106 are driven by a first synchronous belt 107.

[0042] The fiber feeding motor 104 drives the fiber feeding disc 103 to rotate. The folding mechanism makes the design more compact and the layout space smaller.

[0043] In a preferred embodiment, the fiber tension control module 2 includes multiple rotatable encoders 205 mounted on the stand 201. The stand 201 is also equipped with a hinged dance wheel 203. A swing arm 204 is provided at the hinge point between the dance wheel 203 and the stand 201. One end of the dance wheel 203 is equipped with a rotatable guide wheel 202, and the fiber passes under the guide wheel 202.

[0044] In a preferred embodiment, the fiber take-up rotating module 3 includes a base 301, on which a drive shaft 302 and a driven shaft 303 are arranged coaxially at intervals. A shaft clamp 304 is provided at the end of the drive shaft 302 and the driven shaft 303 that are close to each other. The shaft clamp 304 is used to fit the end of the clamping ring skeleton cylindrical tool 4.

[0045] Driven shaft 303, ring skeleton cylindrical tooling 4 and shaft clamp 304 rotate synchronously.

[0046] One end of the base 301 is provided with a bearing seat 305, which is rotatably connected to the drive shaft 302. One end of the drive shaft 302 is fitted with a fourth synchronous pulley 308. The base 301 is also provided with a fiber take-up motor 306. The shaft end of the fiber take-up motor 306 is provided with a third synchronous pulley 307. The third synchronous pulley 307 and the fourth synchronous pulley 308 are driven by a second synchronous belt 309.

[0047] In a preferred embodiment, the base 301 is provided with a guide rail 313, a slidable slider 312 is fitted on the guide rail 313, and a translation plate 311 is also provided. The lower end of the translation plate 311 is connected to the slider 312, and the upper end of the translation plate 311 is provided with a mounting seat 310 for connecting the driven shaft 303. The base 301 is also provided with a rotatable lead screw 315, and a threaded nut seat 314 is fitted on the lead screw 315. The upper end of the nut seat 314 is connected to the translation plate 311.

[0048] One end of the lead screw 315 is connected to an operating handle, which can be manually rotated to adjust the relative position of the driven shaft 303 and the driving shaft 302. This facilitates the assembly and disassembly of the ring skeleton cylindrical tooling 4.

[0049] In a preferred embodiment, the fiber feeding and pressing mechanism 7 includes a translation stage 701, a rotating stage 702 on the translation stage 701, a lifting cylinder 703 on the rotating stage 702, and a pressing rod 704 at the end of the lifting cylinder 703.

[0050] In the preferred scheme, The fiber feeding and rotating module 1 feeds the fiber, and the fiber is wound onto the ring skeleton cylindrical fixture 4 after passing around the fiber tension control module 2. The fiber take-up rotating module 3 drives the ring skeleton cylindrical tooling 4 to rotate, and the fiber arrangement and pressing mechanism 7 performs fiber pressing and fiber arrangement on the optical fiber. Fiber optic insertion angle acquisition camera 5 and fiber optic alignment image acquisition camera 6 monitor fiber compression points in real time; During fiber arrangement, the fiber is made to have a certain entry angle. The fiber entry angle acquisition camera 5 and the fiber arrangement image acquisition camera 6 sample and calculate the degree of the entry angle in real time and transmit it to the control system. The control system compares the real-time fiber inlet angle with the set theoretical angle, adjusts the position of the fiber feeding and pressing mechanism 7, and compensates for the fiber inlet angle.

[0051] In the preferred scheme, Edge detection operators using image algorithms are used to identify fiber optic edge components in images; The engineered processing ensures the stability of the detected edges; The fiber insertion angle is measured based on the boundary line of the fitted fiber and the outer contour of the ring skeleton cylindrical tool 4. Based on the fiber boundary lines of the fiber input and the fiber already wound on the ring skeleton cylindrical tool 4, determine whether there is a fiber arrangement defect. The fiber insertion angle and the theoretical fiber insertion angle are compared and the fiber insertion defect is considered to adjust the fiber insertion and pressing position of the fiber insertion and pressing mechanism 7.

[0052] Example 2: The machine vision intelligent recognition module has two CCD cameras arranged at an angle to each other, which respectively capture images of the fiber entry angle and fiber arrangement. The fiber starts from the wire feeding rotation module, passes through the tension control module, the automatic fiber arrangement module and finally reaches the fiber take-up module, and is precisely arranged and wound on the ring skeleton cylindrical tooling.

[0053] Based on image acquisition from two sets of CCD cameras, the machine vision intelligent recognition module acquires image data such as fiber optic entry angle, remaining number of turns to the edge, and fiber arrangement. It then analyzes and processes the relevant data using fiber optic angle detection algorithms, fiber optic boundary detection algorithms, and fiber arrangement defect detection algorithms. After analysis and judgment, the information is fed back to the PLC for logical operations to control the motion module to achieve fully automatic winding of fiber optic gyroscopes or hydrophones.

[0054] The automatic fiber optic alignment fixture, supported by translation, lifting, and rotation modules, can perform translational and rotational movements in the X and Z directions. The fixture utilizes ceramic materials for superior wear resistance and lubrication, precisely controlling fiber alignment and positioning. An air-blowing device removes adhesive and debris from the fiber surface, facilitating clear image data acquisition by the CCD camera. The X-axis translation module maintains a fixed proportional relationship with the frame rotation speed, creating the necessary conditions for the orderly alignment of the fibers.

[0055] When laying fibers from left to right, the translation stepper moves the fiber to a certain entry angle with the normal plane, which is more conducive to fiber laying. Similarly, when laying fibers from right to left, the rotation stepper rotates 180° to change the fiber pressing posture and entry angle of the L-shaped fixture.

[0056] The camera samples and calculates the fiber inlet angle in real time and transmits it to the PLC. The PLC compensates for the lateral movement of the transverse servo by comparing the real-time angle with the angle set by the HMI, ensuring that the optical fiber is arranged in an orderly manner.

[0057] The fiber tension control module senses the fiber tension through a tension wheel fixed to the sensor, and works with a swing wheel to achieve dynamic balance of fiber winding under the set tension, thus achieving precise tension control. The fiber feeding rotation module uses a synchronous wheel drive to drive the fiber feeding disc to rotate and feed the fiber.

[0058] The fiber take-up rotation module uses a top-down arrangement of the drive shaft and driven shaft to clamp the ring skeleton cylindrical tooling. The drive shaft provides the rotational driving force, and the driven shaft rotates coaxially with it.

[0059] Example 3: To further improve the level of automated winding of the winding machine, an image detection module has been added to the key stages of winding to assist in achieving automated winding. The main functions include: real-time fiber entry angle detection throughout the winding process, edge boundary detection of the skeleton, and fiber placement defect detection during the intermediate flat winding process.

[0060] The overall hardware structure diagram is as follows: Figure 11 As shown, two industrial cameras, working in conjunction with a telecentric lens, clearly capture the outline of the fiber optic cable wound around the skeleton. Simultaneously, a high-quality, stable, and interference-free winding image is required. The selected industrial cameras have a resolution of 1920x1200. One industrial camera is responsible for real-time angle detection, while the other is responsible for boundary detection and fiber routing defect detection. The main defect types detected are slots and fiber overlaps.

[0061] For the fiber entry angle detection algorithm module, edge detection operators from image algorithms are used to identify the fiber edge in the image. After engineering processing, the detected edges are stabilized. The main edge detection operator used is OpenCV's Fast Line Detector (FLD) algorithm, which combines multiple computer vision techniques. First, preprocessing is performed using Canny edge detection to calculate gradients.

[0062]

[0063]

[0064]

[0065] Next, gradient direction quantization is performed, grouping the gradient direction (0°-180°) of each edge pixel into a few predefined directions (such as horizontal, vertical, and two diagonal directions), thus discretizing the continuous gradient directions and preparing for subsequent grouping. Based on this, the algorithm performs pixel chain extraction, searching along the quantized directions and connecting consecutive edge pixels with similar directions to form pixel chains with consistent orientations. After obtaining the pixel chains, the algorithm enters the line segment fitting stage, typically using linear regression techniques such as least squares to approximate each curved pixel chain as a straight line segment, represented by the coordinates of the start and end points. Then, length filtering is performed, removing excessively short line segments based on a preset threshold, retaining long line segments with significant meaning. This step effectively removes a large amount of noise interference. Finally, distance merging is performed. If two line segments are spatially close enough and have similar directions, the algorithm merges them into a longer line segment to obtain complete and coherent straight line features. The formula for calculating the merged point set is:

[0066] Finally, the results of this series of processing steps are output, yielding a clear and accurate set of long straight line segments from the image. Combined with the visual conditions during the looping process, fine-tuning allows for efficient and precise extraction of the target fiber optic line from the image, with real-time angle calculation accuracy reaching 0.01°. The angle detection effect is as follows: Figure 12 As shown: For the boundary detection module and the fiber arrangement defect detection module, deep learning object detection network and classification network were used respectively. The algorithm network used in this study mainly consists of convolutional neural network modules, and its core convolution calculation formula is:

[0067] Where z represents the neuron output. The weight represents the weight of the i-th input. σ represents the i-th input value, b is the bias term, and σ represents the activation function.

[0068] Regarding the calculation of network loss, this network achieves this by optimizing a multi-part loss function. This loss function comprehensively considers the localization error, confidence error, and classification error of the bounding box. For bounding box prediction in object detection, it directly predicts the offset of the bounding box relative to the grid cell. The loss calculation formula is as follows:

[0069] This represents the indicator function, which is 1 when the j-th bounding box predictor in the i-th grid cell is responsible for detecting the object. The weighting coefficients representing coordinate errors. The bounding box localization loss is given by S, where S is the side length of the grid in the image, and B is the number of bounding boxes per grid. , ) represents the center coordinates of the true bounding box. , ) represents the true bounding box width and height, , ) represents the predicted bounding box center coordinates, , () is used to predict the width and height of the bounding box.

[0070] The confidence loss is used to penalize the error in the bounding box confidence C. It consists of two parts: mesh cells containing objects and mesh cells not containing objects, calculated as follows:

[0071]

[0072] The confidence loss when objects are included. For true confidence level, To predict confidence levels, The confidence loss when there are no objects is represented. The weighting coefficient is the one without objects. This is an indicator function; it is 1 if the i-th grid and j-th bounding box are not responsible for detecting objects, and 0 otherwise. The classification loss, used to penalize errors in the object class probability, is calculated only on the grid cells containing the object image:

[0073] For classifying losses, This is an indicator function; it is 1 if the i-th grid contains an object, and 0 otherwise. Let be the true probability that an object in the i-th grid belongs to category c. To correspond to the predicted probabilities, classes is the set of all classes. To train the model using actual production data on fiber routing defects, a total of 20,858 images were collected. The data volumes for each category are as follows:

[0074] Training process loss and accuracy changes, such as Figure 13 After approximately 160 training iterations, the average loss on both the training and test sets of the fiber optic defect classification network stabilized, with accuracies of 97.99% and 97.74% respectively, indicating that the network can fit the image data well after the first stage of preprocessing. After further engineering processing, the inference speed is 30ms / f, which meets the performance requirements of actual detection. As for the object detection network, after collecting 400 sample data points and training for approximately 60 epochs, the model stabilized.

[0075] Specifically, in the application process, key areas in the camera image are first extracted using specific methods. Based on this, traditional image algorithms combined with deep learning algorithms are used to detect side positions or identify specific types of defects. The detection results are illustrated below. Figure 14 , Figure 15 As shown.

[0076] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A fully automated optical fiber winding system based on machine vision, characterized in that: The system includes a workbench (8), a movable base plate (102) on the workbench (8), a fiber feeding rotation module (1) on the movable base plate (102), a stand (201) on one side of the fiber feeding rotation module (1), a fiber tension control module (2) and a fiber feeding and pressing mechanism (7) on the stand (201), a fiber take-up rotation module (3) located on one side of the movable base plate (102) on the workbench (8), a rotatable ring skeleton cylindrical fixture (4) on the take-up rotation module (3), and a fiber insertion angle acquisition camera (5) and a fiber feeding and pressing mechanism (7) facing the ring skeleton cylindrical fixture (4) on the stand (201).

2. The fully automated fiber optic winding system based on machine vision according to claim 1, characterized in that: The fiber feeding rotation module (1) includes a base frame (109), on which a rotatable connecting shaft (108) is provided. One end of the connecting shaft (108) is fitted with a fiber feeding disc (103), and the other end of the connecting shaft (108) is provided with a second synchronous pulley (106). The base frame (109) is also provided with a fiber feeding motor (104), and the shaft end of the fiber feeding motor (104) is fitted with a first synchronous pulley (105). The first synchronous pulley (105) and the second synchronous pulley (106) are driven by a first synchronous belt (107).

3. The fully automated fiber optic winding system based on machine vision according to claim 1, characterized in that: The fiber tension control module (2) includes multiple rotatable encoders (205) mounted on the stand (201). The stand (201) is also equipped with a hinged dance wheel (203). A swing arm (204) is provided at the hinge point between the dance wheel (203) and the stand (201). A rotatable guide wheel (202) is provided at one end of the dance wheel (203). The fiber passes under the guide wheel (202).

4. The fully automated optical fiber winding system based on machine vision according to claim 1, characterized in that: The fiber take-up rotating module (3) includes a base (301), on which a drive shaft (302) and a driven shaft (303) are arranged coaxially at intervals. A shaft clamp (304) is provided at the end of the drive shaft (302) and the driven shaft (303) that are close to each other. The shaft clamp (304) is used to fit the end of the clamping ring skeleton cylindrical tool (4).

5. The fully automated fiber optic winding system based on machine vision according to claim 4, characterized in that: The base (301) is provided with a guide rail (313), a sliding slider (312) is fitted on the guide rail (313), and a translation plate (311) is also provided. The lower end of the translation plate (311) is connected to the slider (312), and the upper end of the translation plate (311) is provided with a mounting seat (310) for connecting the driven shaft (303). The base (301) is also provided with a rotatable lead screw (315), and a threaded nut seat (314) is fitted on the lead screw (315). The upper end of the nut seat (314) is connected to the translation plate (311).

6. The fully automated fiber optic winding system based on machine vision according to claim 1, characterized in that: The fiber pressing mechanism (7) includes a translation stage (701), a rotating stage (702) on the translation stage (701), a lifting cylinder (703) on the rotating stage (702), and a pressing rod (704) at the end of the lifting cylinder (703).

7. The winding method of the fully automated optical fiber winding system based on machine vision according to claim 1, characterized in that: The fiber feeding and rotating module (1) feeds the fiber, and the fiber is wound around the fiber tension control module (2) and then wound onto the ring skeleton cylindrical fixture (4); The fiber take-up rotating module (3) drives the ring skeleton cylindrical tool (4) to rotate, and the fiber pressing and arranging mechanism (7) presses and arranges the optical fiber; The fiber optic insertion angle acquisition camera (5) and the fiber optic arrangement image acquisition camera (6) monitor the fiber compression and fiber arrangement points in real time; During fiber arrangement, the fiber is made to have a certain fiber entry angle. The fiber entry angle acquisition camera (5) and the fiber arrangement image acquisition camera (6) sample and calculate the degree of the fiber entry angle in real time and transmit it to the control system. The control system compares the real-time fiber inlet angle with the set theoretical angle, adjusts the position of the fiber feeding and pressing mechanism (7), and compensates for the fiber inlet angle.

8. The winding method of the fully automated optical fiber winding system based on machine vision according to claim 7, characterized in that: Edge detection operators using image algorithms are used to identify fiber optic edge components in images; The engineered processing ensures the stability of the detection edges; The fiber insertion angle is measured based on the boundary line of the fitted fiber and the outer contour of the ring skeleton cylindrical tool (4). Based on the fiber boundary line of the fiber inserted and the fiber already wound on the ring skeleton cylindrical tool (4), determine whether there is a fiber arrangement defect; The fiber insertion angle and the theoretical fiber insertion angle are compared and the fiber insertion defect is combined to adjust the fiber insertion and pressing position of the fiber insertion and pressing mechanism (7).

9. The winding method of the fully automated optical fiber winding system based on machine vision according to claim 8, characterized in that: Including edge line fitting algorithms: First, preprocessing is performed using Canny edge detection to calculate gradients: Gradient direction quantization is performed by discretizing the continuous gradient directions, where... The gradient value in the x-direction. For the Sobel operator in the x-direction, The gradient value in the y-direction. Here, src is the Sobel operator in the y-direction, src is the original input image, G is the gradient magnitude of the pixel, and θ is the gradient direction angle of the pixel. The algorithm performs pixel chain extraction, searching along the quantized direction and connecting consecutive edge pixels with similar directions to form pixel chains with a consistent direction; In the line segment fitting stage, the least squares linear regression technique is used to approximate each curved pixel chain as a straight line segment, and the starting and ending coordinates are used to represent it. Perform length filtering, remove excessively short line segments based on a preset threshold, retain long line segments with significant meaning, and remove noise interference; Merging two close line segments with similar directions into a longer line segment yields a complete and coherent straight line. The formula for calculating the merged point set is: The direction angle of the merged line segment. , The coordinates of each pixel on the line segment before merging are given, where x and y are the average coordinates of each pixel on the line segment before merging. The output results in a clear and accurate set of long straight line segments in the image. By combining the image conditions during the looping process, the target optical fiber straight line is extracted from the image after fine-tuning.

10. The winding method of the fully automated optical fiber winding system based on machine vision according to claim 9, characterized in that: Including fiber routing defect detection algorithms: Sample collection, collecting images of fiber routing scenes; Perform data augmentation operations on the training set images to increase sample diversity; Data partitioning and preprocessing: all images are normalized in size and pixel values ​​to eliminate the effects of scale and illumination differences; The model is constructed using a convolutional neural network (CNN) architecture, consisting of convolutional layers, pooling layers, and fully connected layers. Through layer-by-layer feature extraction, it achieves the mapping from image pixels to defect categories. Core computation, convolutional layer execution The operation is performed, where z is the neuron output and σ is the activation function. Let the weight be the weight of the i-th input. Let b be the i-th input value, b be the bias term, and n be the number of input dimensions. The fully connected layer maps the extracted high-dimensional features to the probability distribution of multiple types of defects. Loss function, for example, classification loss The core is to penalize the prediction error of the defect category probability, which is calculated only in image regions containing defects. For classifying losses, This is an indicator function; it is 1 if the i-th grid contains an object, and 0 otherwise. Let be the true probability that an object in the i-th grid belongs to category c. To correspond to the predicted probabilities, classes is the set of all classes; Perform iterative training of the model; Set a defect confidence threshold to eliminate low-confidence false positives, and add rule validation for similar defect types to improve classification accuracy; Key region extraction: From the acquired images, key ROIs of the fiber routing area are extracted using traditional image algorithms such as threshold segmentation and contour filtering, while eliminating background interference; Defect classification involves inputting the preprocessed ROI into the optimized classification network, which outputs the defect category and its corresponding confidence score. Output the defect type, confidence level, and detection time, generate fiber routing defect detection results, and trigger an alarm or winding machine adjustment command when a defect is detected.

Images