A machine vision-based automatic assembly method and system for optical module components

Abstract

The application discloses a kind of based on machine vision's optical module assembly automatic assembly method and system, it is related to optical module assembly technical field, including based on multimodal vision fusion perception and closed-loop feedback control, in turn complete front calibration and template training, component feeding and visual positioning, PCBA and optical chip die bonding assembly, optical fiber array / lens and chip active alignment coupling, shell assembly and whole machine detection, sorting and data traceability, through multimodal vision positioning, multidimensional gradient optimization active alignment, error compensation, realize optical module assembly full-process automation assembly.The application uses the above optical module assembly automatic assembly method to solve the problems that traditional optical module assembly relies on manual, low precision, poor efficiency, insufficient consistency, and single vision cannot cope with transparent / reflection component identification, lack of real-time error compensation.

Description

Technical Field

[0001] This invention relates to the field of optical module assembly technology, and in particular to an automatic assembly method and system for optical module components based on machine vision. Background Technology

[0002] As a core component of optical communication systems, optical modules are responsible for converting optical signals into electrical signals and are widely used in data centers, 5G communications, and fiber optic broadband. The assembly process of optical modules involves the assembly of multiple precision components such as optical chips, fiber arrays, lenses, housings, and PCBAs. It places extremely high demands on the spatial positioning accuracy of these components, the coaxiality of the optical path, and coupling efficiency. Typically, micron-level positioning accuracy and a consistently high coupling efficiency are required to ensure the normal operation and performance stability of the optical module.

[0003] Currently, the assembly of optical modules mainly adopts manual or semi-automatic assembly methods, which have many technical bottlenecks: First, the assembly process is highly dependent on human experience. The skill level and operational stability of the operators directly affect the assembly accuracy and consistency, resulting in low product yield. Moreover, manual assembly is inefficient and cannot meet the requirements of high-speed mass production. Second, the optical chips, fiber arrays, lenses, and other components of optical modules are mostly made of transparent or reflective materials. Single 2D or 3D vision guidance methods are difficult to reliably identify the minute features of the components and are easily affected by surface reflection interference, resulting in large positioning deviations and affecting assembly quality. Third, existing assembly methods lack real-time feedback mechanisms and cannot dynamically compensate for interference factors such as mechanical deviations, environmental vibrations, and temperature changes during the assembly process. Furthermore, the coupling process often adopts a sequential trial-and-error approach, which is inefficient and makes it difficult to guarantee coupling accuracy. Fourth, existing assembly processes are mostly partially automated and have not achieved a closed loop from component loading to finished product sorting and data traceability. There are many manual intervention links, which further affect assembly efficiency and product consistency.

[0004] In the existing technology, some patents and papers disclose vision-based optical module assembly technologies. For example, CN108080952A discloses an automatic assembly process and mechanism for SFP optical modules, which only automates basic steps such as SFP optical module feeding, pressing, and sheet metal assembly. It uses a single CCD vision guide and does not involve core technologies such as 3D vision fusion, active alignment optimization, closed-loop feedback, and thermal deformation compensation, and cannot solve the problems of transparent / reflective component recognition and high-precision coupling. CN116810366A discloses a robot-based optical module precision assembly system and method, which emphasizes the collaborative positioning of robots and vision, but does not disclose the mathematical model of multimodal vision fusion, active alignment, and full-process error compensation system, which does not match the core requirement of high-precision optical coupling of optical modules. Related papers mostly focus on the positioning or coarse alignment of single components, without involving multi-component collaborative assembly, full-process automation, and optical performance-driven assembly optimization. The algorithm complexity and accuracy are far from meeting the assembly requirements of high-speed optical modules.

[0005] To address these issues, there is an urgent need for an automated assembly method and system for optical module components based on machine vision. Summary of the Invention

[0006] To address the aforementioned issues, this application proposes an automated assembly method and system for optical module components based on machine vision. The aim is to achieve fully automated assembly of optical module components, solve the problem of identifying transparent or reflective components, improve assembly accuracy and efficiency, and provide an assembly method with real-time error compensation and full-process traceability capabilities, thereby overcoming the shortcomings of existing technologies.

[0007] On the one hand, this application proposes an automatic assembly method for optical module components based on machine vision, including the following steps: S1. Based on camera calibration technology and improved YOLOv5 model, standard images and system parameters of each component of the optical module are calibrated, features are extracted and model training is performed to obtain a unified pixel-physical world coordinate system, error compensation parameters and adaptive component recognition model. The improved YOLOv5 model integrates the CBAM attention mechanism and the Focal Loss loss function; S2. Based on a vibratory feeder, conveyor line, multimodal vision camera, and improved sub-pixel corner detection algorithm, various optical module components and acquired multi-view images are processed for feeding, imaging, preprocessing, and pose calculation to obtain the component's three-dimensional pose data. and deviation data ; S3. Based on a robotic arm, a vision guidance device, and a UV curing device, PCBA and optical chips are positioned, picked up, mounted, electrically connected, and cured to obtain die-bonded components. S4. Based on passive pre-alignment and active fine alignment modes, improved multi-dimensional gradient optimization algorithm and temperature-controlled curing device, the die bonding component, fiber array or lens, optical power and spot data are aligned, coupled, cured and error-compensated to obtain the target assembly component. The improved multidimensional gradient optimization algorithm integrates particle swarm optimization and gradient descent algorithms; S5. Based on robotic arms, 3D vision scanning equipment and multi-performance testing equipment, perform whole-machine testing, and perform assembly, gap detection, fastening and full-performance testing on target assembly components and optical module shells to obtain optical, mechanical and appearance inspection data of the whole machine and qualified / unqualified optical module whole machines. S6. Based on the sorting device and data management system, the whole machine inspection results and assembly data of each workstation are sorted, recorded and analyzed to obtain the sorted optical modules and the full life cycle assembly traceability log.

[0008] Preferably, the specific content in S1 includes: Based on camera calibration technology, the system parameters are calibrated to obtain a unified pixel-physical world coordinate system and error compensation parameters; The camera calibration techniques include camera intrinsic parameter calibration, binocular extrinsic parameter calibration, and hand-eye calibration. The system parameters include camera parameters, robotic arm parameters, and motion platform parameters; Based on the improved YOLOv5 model, standard image features of each component of the optical module are extracted and the model is trained to obtain an adaptive component recognition model. The optical module components include an optical chip, fiber array, lens, housing, and PCBA; Specifically, standard images of various components are collected, and key features of component Mark points, edges, and keyways are extracted through image preprocessing techniques. The extraction of minute features of fiber core and chip photosensitive area is particularly enhanced to obtain feature data. The standard image and the extracted feature data are input into the improved YOLOv5 model for training to obtain an adaptive component recognition model; The improved YOLOv5 model integrates the CBAM attention mechanism and the Focal Loss loss function; The CBAM attention mechanism module is embedded in the feature fusion link between the neck and head of the model, and is divided into a channel attention unit (CA) and a spatial attention unit (SA), with CA and SA executed in series. The Focal Loss function replaces the original YOLOv5 cross-entropy loss to address the feature imbalance of optical module components and improve the accuracy of small target recognition.

[0009] Preferably, the specific content of the channel attention unit (CA) is as follows: ; ; The specific content of the Spatial Attention Unit (SA) is as follows: ; ; in, F The input is the original feature map of CBAM, with dimension . C is the number of channels, H is the feature map height, and W is the feature map width. GAP ( F ) for feature maps F Perform global average pooling. GMP ( F ) for feature maps F Perform global max pooling. To merge the results of average pooling and max pooling, FC () represents a fully connected layer operation. This refers to the adjustment coefficient for the weights of small feature channels. For small feature channel weight vectors, dimension , To Perform L2 normalization. It is the Sigmoid activation function. To improve the post-channel attention weight map, the dimensions are... , The output feature map after channel attention processing. This is the original response value vector of the small feature channels. To Perform Softmax normalization. This is element-wise multiplication. Here is the feature map after channel attention processing, and here is the input for spatial attention. For feature splicing operations, This is a 1×1 convolution operation. As a spatial location sensitive factor, The coordinate weight matrix of the small feature space has dimensions. , To Perform bilinear interpolation mapping. The standard deviation of the Gaussian kernel. To improve the spatial attention weight map, the dimension is... , Based on Generate Gaussian space weights. This is the final output feature map of the CBAM module.

[0010] Preferably, the Focal Loss function introduces an adaptive weighting term for the difficulty in distinguishing small feature samples and a dynamic adjustment factor for class weights to solve the problem of unbalanced loss distribution between the difficulty in distinguishing small feature samples and regular samples and background samples. The expression is as follows: ; in, To improve the total Focal Loss value, As a positive and negative sample balance factor, A balance factor for samples with minor and conventional features. For samples with small features that are difficult to distinguish, the weight adaptive term is used. The target class probability predicted by the model. Based on the focusing parameters, To optimize the coefficients of the focus term, To improve the focus item, For the logarithmic loss term, To locate the loss weight coefficients, for CIoU Locating the loss, For small feature real regions, For the small feature model to predict the region, For small feature real region and predicted region IoU value; in, .

[0011] Preferably, in step S2, based on a vibratory feeder, conveyor line, multimodal vision camera, and improved sub-pixel corner detection algorithm, various optical module components and acquired multi-view images are processed for feeding, imaging, preprocessing, and pose calculation to obtain the component's three-dimensional pose data. and deviation data The specific content also includes: The optical module housing, PCBA, optical chip, fiber array, and lens assembly are sequentially fed into the assembly station via a vibratory feeder and conveyor line. A multimodal vision camera is used to acquire multi-view images of the components. After Gaussian filtering for noise reduction, Canny edge detection, and histogram equalization preprocessing, a standard image is obtained. The multimodal vision camera includes a top-view high-resolution CCD camera, a side-view 3D structured light camera, and a coaxial illumination module. The target corner coordinates of the standard image are obtained by improving the accuracy of the sub-pixel corner detection algorithm. An adaptive component recognition model is used to perform preliminary identification of component types and poses on standard images, thereby obtaining the three-dimensional pose data of each component in the world coordinate system. and deviation data .

[0012] The preferred improved sub-pixel corner detection algorithm is as follows: For a standard image, calculate the Harris response value of each pixel, set the response threshold to 0.01-0.03, filter out pixels with Harris response values ​​higher than the threshold as candidate corner points, initially locate the positions of component Mark points and edge corner points, and remove non-corner point pixels; For the initially located candidate corner points, the Zernike matrix subpixel localization algorithm is used to achieve subpixel-level correction of the corner point coordinates. The steps are as follows: Centered on the candidate corner point, select a 3×3 or 5×5 local image window and calculate the Zernike moments of the image within the window; By using the phase and amplitude information of the Zernike moments, the true sub-pixel coordinates (x, y) of the corner points are fitted. sub ,y sub ), correcting the initial positioning corner point deviation; For tiny corner points of optical modules, optimize the Zernike moment calculation weights to enhance the Zernike moment response of tiny corner points; To suppress localization noise using Kalman filtering and obtain stable corner coordinates, i.e., target corner coordinates, the following steps are taken: Establish a Kalman filter model and convert the sub-pixel corner coordinates (x... sub ,y sub Using the observed values ​​as the basis, set up the state equation and the observation equation; A Kalman filter prediction-update process is performed on the sub-pixel corner coordinates of each frame of the image to remove outliers caused by positioning noise, and output stable and accurate final corner coordinates (x, y, z). final ,y final ).

[0013] Preferably, the specific content of S4 is as follows: The fiber optic array or lens is actively aligned and coupled with the chip, employing passive pre-alignment, active precision alignment, and closed-loop feedback. First, visual guidance moves the fiber optic array or lens near the chip's optical window, controlling the deviation within a certain range. Within 30μm; The optical signal source is turned on, and the output optical power P and the light spot image are collected in real time through the optical power meter and the spectrum analyzer. The vision system simultaneously monitors the relative displacement and angular deviation between the fiber end face and the photosensitive area of ​​the chip, and executes the improved multidimensional gradient optimization algorithm to perform nanometer-level step search in 6 degrees of freedom until the optical power reaches the preset threshold and the light spot roundness reaches the target roundness. After locking the current position, apply low-shrinkage UV adhesive and cure in stages with temperature control. During the curing process, continuously monitor the light power to compensate for thermal deformation errors.

[0014] Preferably, the improved multidimensional gradient optimization algorithm integrates particle swarm optimization and gradient descent algorithms, introduces adaptive inertia weights, and the objective function is expressed as follows: ; in, Let be the objective function. P This represents the actual coupled optical power output of the fiber array. The maximum optical power under ideal coupling conditions is given by k1, where k1 is the plane translation deviation. The corresponding coupling coefficient, k2 is the vertical translation deviation () The corresponding coupling coefficient, k3 is the angle deviation ( The corresponding coupling coefficient, This represents the translational deviation along the x-axis. This represents the translational deviation along the y-axis. The translational deviation is in the z-axis direction. This represents the rotational deviation about the x-axis. This represents the rotational deviation about the y-axis. This represents the rotational deviation about the z-axis.

[0015] On the other hand, this application proposes an automated assembly system for optical module components based on machine vision, comprising: Basic model building unit: Based on camera calibration technology and improved YOLOv5 model, standard images and system parameters of each component of the optical module are calibrated, features are extracted and model training is performed to obtain a unified pixel-physical world coordinate system, error compensation parameters and adaptive component recognition model. The improved YOLOv5 model integrates the CBAM attention mechanism and the Focal Loss loss function; Adaptive recognition unit: Based on vibratory feeder, conveyor line, multimodal vision camera and improved sub-pixel corner detection algorithm, it performs loading, imaging, preprocessing and pose calculation on various optical module components and acquired multi-view images to obtain the component's three-dimensional pose data. and deviation data ; Assembly Unit: Based on a robotic arm, vision guidance device and UV curing device, PCBA and optical chip are positioned, picked up, mounted, electrically connected and cured to obtain die-bonded components. Based on passive pre-alignment and active fine alignment modes, improved multi-dimensional gradient optimization algorithm and temperature-controlled curing device, the die-bonded components, fiber arrays or lenses, optical power and spot data are aligned, coupled, cured and error-compensated to obtain target assembled components. The improved multidimensional gradient optimization algorithm integrates particle swarm optimization and gradient descent algorithms; Inspection and Adjustment Unit: Based on robotic arms, 3D vision scanning equipment, and multi-performance testing equipment, the unit performs whole-machine inspection, including assembly, gap detection, fastening, and full-performance testing of target assembly components and optical module shells. This process yields optical, mechanical, and appearance inspection data for the whole machine, as well as qualified / unqualified optical modules. Based on a sorting device and data management system, the unit sorts, records, and analyzes the whole-machine inspection results and assembly data from each workstation, resulting in categorized and sorted optical modules and a full lifecycle assembly traceability log.

[0016] In summary, the automatic assembly method and system for optical module components based on machine vision of the present invention has the following advantages compared with traditional technologies: 1. This invention covers the entire process of optical module component loading, positioning, die bonding, coupling, shell assembly, testing, sorting, and data traceability. It requires no manual intervention, solves the problems of traditional assembly relying on human experience and poor operational stability, significantly improves assembly efficiency, and is suitable for high-speed mass production scenarios. 2. The multimodal vision fusion imaging method combined with deep learning feature enhancement technology effectively suppresses the reflection interference of transparent / reflective optical surfaces, realizes pixel-level positioning of tiny features of components, and solves the problem that existing single vision cannot stably identify transparent or reflective components and has large positioning deviations, thus ensuring stable assembly accuracy. 3. By adopting a collaborative approach of passive pre-alignment and active fine alignment, combined with a multi-dimensional gradient optimization algorithm, optical performance-driven assembly optimization is achieved. By acquiring optical power and spot data in real time, the position and attitude of the components are dynamically adjusted to stabilize the coupling efficiency. 4. Based on a deep learning model, the system can automatically identify defects in the appearance and performance of optical modules with high detection accuracy. Combined with the full-process data traceability function, the system enables the lifecycle of each product to be traceable.

[0017] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the steps of an automatic assembly method for optical module components based on machine vision according to the present invention; Figure 2 This is a schematic diagram of an automatic assembly system unit for optical module components based on machine vision according to the present invention. Detailed Implementation

[0019] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0020] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0021] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0022] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0024] Example 1 An automated assembly method for optical module components based on machine vision, such as Figure 1 As shown, it includes the following steps: S1. Preliminary preparation: Complete system calibration and template library construction. Based on camera calibration technology and improved YOLOv5 model, calibrate the standard images and system parameters of each component of the optical module, extract features and train the model to obtain a unified pixel-physical world coordinate system, error compensation parameters and adaptive component recognition model.

[0025] The improved YOLOv5 model integrates the CBAM attention mechanism and the Focal Loss loss function.

[0026] Furthermore, the specific content of S1 includes: Based on camera calibration technology, the system parameters are calibrated to obtain a unified pixel-physical world coordinate system. The error compensation parameters simultaneously correct camera distortion error, robotic arm geometric error, and motion platform positioning error, ensuring that the positioning accuracy reaches the sub-pixel level, and providing a precise coordinate reference for subsequent component positioning and assembly.

[0027] The camera calibration techniques include camera intrinsic parameter calibration, binocular extrinsic parameter calibration, and hand-eye calibration. The system parameters include camera parameters, robotic arm parameters, and motion platform parameters; Based on the improved YOLOv5 model, standard image features of each component of the optical module are extracted and the model is trained to obtain an adaptive component recognition model. The optical module components include an optical chip, fiber array, lens, housing, and PCBA; Specifically, standard images of various components are collected, and key features of component Mark points, edges, and keyways are extracted through image preprocessing techniques. The extraction of minute features of fiber core and chip photosensitive area is particularly enhanced to obtain feature data. The standard image and the extracted feature data are input into the improved YOLOv5 model for training to obtain an adaptive component recognition model; The improved YOLOv5 model integrates the CBAM attention mechanism and the Focal Loss loss function. The CBAM attention mechanism can adaptively strengthen the feature weights of small features and suppress background interference. The Focal Loss loss function can solve the problem of component feature imbalance (low proportion of small features and high proportion of background features), and improve the recognition accuracy and speed of small target components. The CBAM attention mechanism module is embedded in the feature fusion link between the neck and head of the model, and is divided into a channel attention unit (CA) and a spatial attention unit (SA), with CA and SA executed in series. The Focal Loss function replaces the original YOLOv5 cross-entropy loss to address the feature imbalance of optical module components and improve the accuracy of small target recognition.

[0028] Furthermore, the specific content of the channel attention unit (CA) is as follows: ; ; The specific content of the Spatial Attention Unit (SA) is as follows: ; ; in, F The input is the original feature map of CBAM, with dimension . (C = number of channels, H = feature map height, W = feature map width) GAP ( F ) for feature maps F Perform global average pooling to preserve the overall feature information of each channel. GMP ( F ) for feature maps F Perform global max pooling to highlight the key local features of the channel. To merge the results of average pooling and max pooling, FC () represents fully connected layer operations (including ReLU activation, dimensionality reduction, or dimensionality increase). This refers to the adjustment coefficient for the weights of small feature channels. For small feature channel weight vectors, dimension , To Perform L2 normalization. It is the Sigmoid activation function. To improve the post-channel attention weight map, the dimensions are... , The output feature map after channel attention processing. This is the original response value vector of the small feature channels. To Perform Softmax normalization. This is element-wise multiplication. Here is the feature map after channel attention processing, and here is the input for spatial attention. For feature splicing operations, This is a 1×1 convolution operation (including BatchNorm batch normalization). As a spatial location sensitive factor, The coordinate weight matrix of the small feature space has dimensions. , To Perform bilinear interpolation mapping. The standard deviation of the Gaussian kernel. To improve the spatial attention weight map, the dimension is... , Based on Generate Gaussian space weights. This is the final output feature map of the CBAM module.

[0029] Furthermore, the Focal Loss function introduces an adaptive weighting term for the difficulty in distinguishing small feature samples and a dynamic adjustment factor for class weights to address the imbalance in loss distribution between the difficulty in distinguishing small feature samples and regular samples / background samples. The expression is as follows: ; in, To improve the total Focal Loss value, As a positive and negative sample balance factor, A balance factor for samples with minor and conventional features. For samples with small features that are difficult to distinguish, the weight adaptive term is used. The target class probability predicted by the model. Based on the focusing parameters, To optimize the coefficients of the focus term, To improve the focus item, For the logarithmic loss term, To locate the loss weight coefficients, for CIoU Locating the loss, For small feature real regions, For the small feature model to predict the region, For small feature real region and predicted region IoU value; in, .

[0030] Input standard images and feature data of various components to complete model training, enabling the model to quickly and accurately identify components of different types and postures, achieve adaptive recognition of component posture and type, and adapt to the assembly requirements of multiple types of optical modules.

[0031] S2. Based on a vibratory feeder, conveyor line, multimodal vision camera, and improved sub-pixel corner detection algorithm, various optical module components and acquired multi-view images are processed for feeding, imaging, preprocessing, and pose calculation to obtain the component's three-dimensional pose data. and deviation data ; Furthermore, S2, based on a vibratory feeder, conveyor line, multimodal vision camera, and improved sub-pixel corner detection algorithm, performs loading, imaging, preprocessing, and pose calculation on various optical module components and acquired multi-view images to obtain the component's three-dimensional pose data. and deviation data The specific content also includes: Automated feeding: The optical module housing, PCBA, optical chip, fiber array, lens and other components are fed into the assembly station in sequence by vibratory feeder and conveyor line, realizing automated and continuous feeding of components and reducing the error and efficiency loss of manual feeding.

[0032] Multi-view imaging: The visual perception unit is activated, and the high-resolution CCD camera looks up to collect the planar features of the top of the component (such as Mark points, pins, and positioning holes). The 3D structured light camera looks to the side to collect the height information and surface contour of the component, which is used to detect parameters such as the thickness and warpage of the transparent component. The coaxial illumination module works with coaxial spectral imaging to suppress the reflection interference of transparent / reflective optical surfaces, clearly capture the true features of the component, and avoid feature extraction deviations caused by reflected light.

[0033] A multimodal vision camera is used to acquire multi-view images of the components. After preprocessing, Gaussian filtering for noise reduction (eliminating image noise and improving image clarity), Canny edge detection (extracting edge features of the components and clarifying the outline boundaries of the components), and histogram equalization (enhancing image contrast and highlighting minute features) are used to obtain a standard image. The multimodal vision camera includes a top-view high-resolution CCD camera, a side-view 3D structured light camera, and a coaxial illumination module. The target corner coordinates of the standard image are obtained by improving the accuracy of the sub-pixel corner detection algorithm. Zernike subpixel localization improves corner point localization accuracy to the 0.1 pixel level. Kalman filtering is used to suppress random noise during the localization process. Combined with template matching and an improved YOLOv5 adaptive component recognition model, components are accurately located and matched with standard feature data in the template library. Based on the pixel-physical world coordinate system established by system calibration, component type and pose are initially identified in standard images to obtain the 3D pose data of each component in the world coordinate system. and deviation data The deviation data is transmitted to the central control system in real time, providing accurate guidance for subsequent assembly operations.

[0034] Furthermore, the details of the improved sub-pixel corner detection algorithm are as follows: For a standard image, calculate the Harris response value of each pixel, set the response threshold to 0.01-0.03, filter out pixels with Harris response values ​​higher than the threshold as candidate corner points, initially locate the positions of component Mark points and edge corner points, and remove non-corner point pixels; For the initially located candidate corner points, the Zernike matrix subpixel localization algorithm is used to achieve subpixel-level correction of the corner point coordinates. The steps are as follows: Centered on the candidate corner point, select a 3×3 or 5×5 local image window (to accommodate small corner points), and calculate the Zernike moments of the image within the window (selecting Zernike moments of order 0-4 to balance positioning accuracy and computational efficiency). By using the phase and amplitude information of the Zernike moments, the true sub-pixel coordinates (x, y) of the corner points are fitted. sub ,y sub This corrects the initial positioning corner deviation and solves the problem that traditional corner detection can only achieve pixel-level positioning and has insufficient accuracy; For tiny corner points of optical modules (such as Mark points of fiber arrays), optimize the Zernike moment calculation weights, enhance the Zernike moment response of tiny corner points, and avoid positioning failure of tiny corner points; Kalman filtering noise reduction (core of stability improvement): Due to vibration and light fluctuations in the assembly environment, the corner coordinates after sub-pixel positioning may have slight fluctuations. Kalman filtering is used to suppress positioning noise and obtain stable corner coordinates, i.e., target corner coordinates. The steps are as follows. Establish a Kalman filter model and convert the sub-pixel corner coordinates (x... sub ,y sub As the observed value, set the state equation (dynamic change of corner coordinates) and the observation equation (positioning error); Initialize the filter parameters: process noise variance Q = 0.001-0.005, observation noise variance R = 0.01-0.02, to adapt to vibration interference in the assembly environment; A Kalman filter prediction-update process is performed on the sub-pixel corner coordinates of each frame of the image to remove outliers caused by positioning noise, and output stable and accurate final corner coordinates (x, y, z). final ,y final This ensures the repeatability and stability of corner point positioning.

[0035] S3. Based on the robotic arm, vision guidance device and UV curing device, the PCBA and optical chip are positioned, picked up, mounted, electrically connected and cured to obtain the die bonded assembly.

[0036] Substrate positioning: Based on the PCBA pose deviation data output by S2, the central control system controls the robotic arm to grasp the PCBA assembly and transfer it to the precision positioning stage. The vision system then identifies the Mark points on the PCBA again, corrects the PCBA's reference position, ensures accurate PCBA positioning, and avoids reference deviation from affecting the subsequent chip mounting accuracy.

[0037] Chip Pickup and Orientation Adjustment: The robotic arm, equipped with a vacuum nozzle, moves above the optical chip tray under the guidance of the vision system to accurately pick up the optical chip and avoid chip damage. Through the coordinated action of the rotary platform and the Z-axis fine-tuning slide, the angular deviation (θ) and planar offset (x,y) of the optical chip are compensated based on the optical chip pose deviation data output by S2, so that the orientation of the optical chip is consistent with the orientation of the target pad on the PCBA, thus preparing for accurate placement.

[0038] Die bonding alignment and mounting: Based on hand-eye calibration data, the central control system guides the robotic arm to move the optical chip to the target pad position on the PCBA. The vision system monitors the relative position of the chip and the pad in real time and dynamically fine-tunes the position and posture of the robotic arm to ensure accurate alignment of the chip and the pad, with a mounting accuracy of ≤±2μm. After alignment, the electrical connection between the chip and the substrate is completed to ensure stable electrical performance.

[0039] Bonding and Curing: Activate the UV adhesive dispensing valve to apply an appropriate amount of UV adhesive to the connection between the chip and the PCBA to fix the chip. The vision system monitors the adhesive amount and dispensing position in real time to prevent adhesive overflow, insufficient application, or dispensing misalignment, ensuring a secure bond. After dispensing, transfer the assembled PCBA to the UV curing unit and initiate a staged temperature-controlled curing process. Curing is performed using a gradient temperature increase from 25℃ to 80℃ to prevent rapid curing and shrinkage of the UV adhesive, which could lead to chip misalignment or optical path deviation. During curing, the vision system continuously monitors the chip's pose to ensure that the chip position does not shift during curing.

[0040] S4. Based on passive pre-alignment and active fine alignment modes, improved multi-dimensional gradient optimization algorithm and temperature-controlled curing device, the die bonding component, fiber array or lens, optical power and spot data are aligned, coupled, cured and error-compensated to obtain the target assembly component. The improved multidimensional gradient optimization algorithm integrates particle swarm optimization and gradient descent algorithms; Furthermore, the specific content of S4 is as follows: The fiber array or lens is actively aligned and coupled with the chip, using passive pre-alignment, active fine alignment and closed-loop feedback. First, the fiber array or lens is moved to the vicinity of the chip's optical window through visual guidance, and the deviation is controlled within ±30μm. Passive pre-alignment: Based on the pose deviation data of the fiber array / lens and optical chip output by S2, the central control system controls the robotic arm and the precision motion platform to move in coordination. The vision system guides the fiber array / lens to move near the optical window of the optical chip to achieve coarse positioning. The relative deviation between the fiber array / lens and the optical chip is controlled within ±30μm, which lays the foundation for subsequent active fine alignment, shortens the fine alignment time, and improves assembly efficiency.

[0041] Active precision alignment: Turn on the optical signal source, and collect the output optical power P and the light spot image in real time through the optical power meter and the spectrum analyzer. The vision system simultaneously monitors the relative displacement and angular deviation between the fiber end face and the photosensitive area of ​​the chip, and executes the improved multi-dimensional gradient optimization algorithm to perform nanometer-level step search in 6 degrees of freedom until the optical power reaches the preset threshold and the light spot roundness reaches the target roundness. Specifically, the optical signal source is turned on to drive the optical chip to emit light. The optical power meter and spectrum analyzer of the signal detection unit collect the optical power P and the light spot image output by the fiber array in real time, and transmit the collected data to the central control system in real time. The vision system simultaneously monitors the relative displacement and angular deviation between the fiber end face and the photosensitive area of ​​the chip, and feeds it back to the central control system. The central control system executes a multi-dimensional gradient optimization algorithm, using the 6-DOF parameters of the fiber array / lens as optimization variables, and taking the maximization of optical power and the optimal light spot roundness as optimization objectives, to perform a nanometer-level step search, gradually adjusting the position and orientation of the fiber array / lens and correcting the deviation.

[0042] After locking the current position, apply low-shrinkage UV adhesive and cure in stages with temperature control. During the curing process, continuously monitor the light power to compensate for thermal deformation errors.

[0043] Furthermore, the improved multidimensional gradient optimization algorithm integrates particle swarm optimization and gradient descent algorithms, and introduces adaptive inertia weights. The expression for the objective function is as follows: ; in, The objective function is "to maximize the coupled optical power", which means adjusting the translation deviation ( x, y, z) and angle deviation ( This aligns the fiber optic array / lens with the photosensitive area of ​​the chip, maximizing the output optical power and ultimately achieving a coupling efficiency of ≥95%. P The actual coupled optical power output of the fiber array (unit: dBm) is the core evaluation indicator of the alignment effect. The higher the optical power, the higher the alignment accuracy and the better the coupling effect.

[0044] The maximum optical power (in dBm) under ideal coupling conditions is the theoretical maximum output optical power when the fiber array / lens is perfectly aligned with the photosensitive area of ​​the chip (without any deviation). It is determined by the optical characteristics of the optical chip and fiber array (such as emission power and transmission efficiency) and serves as the target benchmark for optical power optimization. k1 represents the planar translation deviation. The corresponding coupling coefficient reflects the degree of attenuation of optical power due to translational deviations in the x and y directions (horizontal direction), and k2 is the vertical translational deviation ( The corresponding coupling coefficient reflects the degree of attenuation of optical power due to translational deviation in the z-direction (height direction), and k3 is the angular deviation ( The corresponding coupling coefficient reflects the degree of optical power attenuation caused by component attitude deviation. The coupling coefficient (unitless, a positive number) is directly related to the characteristics of the optical module's optical system and is determined by the optical parameters (such as numerical aperture, focal length, and photosensitive area size) of the fiber array, lens, and chip. Its core function is to "quantify the weight of the impact of different types of deviations on optical power". This represents the three-dimensional translational deviation (unit: μm), corresponding to the spatial positional deviation between the fiber array / lens and the chip's photosensitive area. It is calculated using S2 vision positioning and is a core adjustment parameter for alignment optimization. This refers to the axial translational deviation, specifically the positional deviation between the fiber array / lens and the chip in the horizontal X-direction. This represents the translational deviation along the y-axis, i.e., the positional deviation between the fiber array / lens and the chip in the horizontal Y-direction. This represents the translational deviation along the z-axis, i.e., the positional deviation between the fiber array / lens and the chip in the vertical height direction. The three-dimensional attitude deviation (unit: °, angle) corresponds to the spatial attitude deviation between the fiber array / lens and the chip, calculated by S2 vision positioning, and directly affects the optical path coaxiality. This represents the rotational deviation (pitch angle deviation) around the x-axis, reflecting the attitude deviation of the component in pitch. This represents the rotational deviation (yaw angle deviation) around the y-axis, reflecting the attitude deviation of the component's left and right yaw. This represents the rotational deviation (roll angle deviation) around the z-axis, reflecting the attitude deviation of the component's rotational tilt.

[0045] This improved algorithm integrates particle swarm optimization and gradient descent algorithms, introducing adaptive inertial weights. First, the particle swarm optimization algorithm quickly traverses the 6-DOF parameter space to lock the optimal search region with high coupling efficiency, avoiding the problem of traditional gradient descent algorithms getting trapped in local optima. Then, the gradient descent algorithm performs a small-step local fine search within the optimal search region to improve positioning accuracy. The adaptive inertial weights are dynamically adjusted according to the real-time acquired optical power changes. When the optical power deviation is large, the weights are increased to speed up the search, and when the optical power is close to the preset threshold, the weights are decreased to improve search accuracy, balancing search efficiency and positioning accuracy. This avoids the inefficiency of traditional sequential trial-and-error algorithms, achieving rapid convergence of coupling efficiency and shortening the alignment time to within 2 seconds. At the same time, the resistance to environmental vibration, temperature fluctuations and other disturbances is improved by more than 30%. When the optical power P reaches the preset threshold (≥95% of the theoretical maximum value) and the spot circularity is optimal, the central control system locks the current position and attitude of the fiber array / lens, completing active fine alignment.

[0046] Adhesive curing and thermal deformation compensation: After precise alignment, the UV adhesive dispensing valve is activated to apply low-shrinkage UV adhesive at the connection between the fiber array / lens and the optical chip. The vision system monitors the amount of adhesive and the dispensing position to ensure a firm bond without affecting the optical path. A staged temperature-controlled curing method (gradual temperature increase from 25℃ to 80℃) is then used to cure the UV adhesive. During the curing process, the signal detection unit continuously monitors changes in optical power. The central control system, based on a thermal expansion model, monitors changes in curing temperature and ambient temperature in real time, dynamically corrects the relative position parameters of the fiber array / lens and the optical chip, compensates for optical path offset caused by thermal deformation, and ensures stable coupling accuracy.

[0047] S5. Based on robotic arms, 3D vision scanning equipment and multi-performance testing equipment, perform whole-machine testing, and perform assembly, gap detection, fastening and full-performance testing on target assembly components and optical module shells to obtain optical, mechanical and appearance inspection data of the whole machine and qualified / unqualified optical module whole machines. Housing pre-positioning: The vision system identifies features such as positioning slots and buckles on the optical module housing. The central control system controls the robotic arm to grasp the housing and, under vision guidance, initially fits the housing into the internal components that have completed die bonding and coupling assembly. This ensures that the housing and internal components are initially aligned and prevents the housing from damaging the internal components during the assembly process.

[0048] Gap Measurement and Attitude Fine-tuning: A 3D vision camera scans the mating gap between the shell and internal components to obtain gap distribution data. The central control system calculates the optimal assembly angle and path based on the gap data and controls the robotic arm to dynamically adjust the attitude and position of the shell, achieving stress-free assembly of the shell and internal components. This avoids component deformation or optical path deviation caused by assembly stress and ensures the stability of the assembled optical module structure.

[0049] Locking and fixing: After the shell posture is adjusted, the automatic screw locking device or buckle pressing device is activated to lock and fix the shell to the internal components. The vision system monitors the screw locking depth, buckle pressing force and flatness in real time to ensure that the locking and pressing are firm and to avoid loosening.

[0050] Full performance testing: A comprehensive test is performed on the assembled optical module unit, specifically including: Optical performance testing: The optical power, extinction ratio, insertion loss and other parameters of the optical module are tested by an optical power meter and a spectrum analyzer to ensure that the optical performance meets the preset standards; Mechanical performance testing: Testing parameters such as the sealing performance and insertion / removal force of the optical module housing to ensure the stability of the mechanical structure and meet usage requirements; Appearance defect detection: Based on a deep learning model, the system automatically identifies appearance defects such as surface scratches, abnormal glue content, bubbles, and pin damage on the optical module, distinguishing between qualified and unqualified products. During the detection process, various detection data are recorded simultaneously and transmitted to the central control system.

[0051] S6. Based on the sorting device and data management system, the whole machine inspection results and assembly data of each workstation are sorted, recorded and analyzed to obtain the sorted optical modules and the full life cycle assembly traceability log.

[0052] Automatic sorting: The central control system controls the sorting device to move qualified optical modules to the qualified product storage area and unqualified products to the unqualified product storage area based on the whole machine inspection results. The unqualified products are also classified and marked with the defect type (such as substandard optical performance, appearance defects, loose mechanical structure, etc.).

[0053] Data Traceability: The central control system generates assembly data logs, recording all relevant data such as visual images, component pose parameters, coupling efficiency, dispensing amount, curing parameters, and test results at each assembly station. Each optical module is assigned a unique traceability code, enabling full lifecycle data traceability from component loading to finished product delivery. This facilitates subsequent product quality checks, process optimization, and after-sales maintenance. Simultaneously, it performs statistical analysis on defect and deviation data of non-conforming products, outputting optimization suggestions to adjust assembly parameters, optimize model algorithms, and improve subsequent assembly yield.

[0054] Example 2 An automated assembly system for optical module components based on machine vision, such as Figure 2 As shown, it includes: Basic model building unit: Based on camera calibration technology and improved YOLOv5 model, standard images and system parameters of each component of the optical module are calibrated, features are extracted and model training is performed to obtain a unified pixel-physical world coordinate system, error compensation parameters and adaptive component recognition model. The improved YOLOv5 model integrates the CBAM attention mechanism and the Focal Loss loss function; Adaptive recognition unit: Based on vibratory feeder, conveyor line, multimodal vision camera and improved sub-pixel corner detection algorithm, it performs loading, imaging, preprocessing and pose calculation on various optical module components and acquired multi-view images to obtain the component's three-dimensional pose data. and deviation data ; Assembly Unit: Based on a robotic arm, vision guidance device and UV curing device, PCBA and optical chip are positioned, picked up, mounted, electrically connected and cured to obtain die-bonded components. Based on passive pre-alignment and active fine alignment modes, improved multi-dimensional gradient optimization algorithm and temperature-controlled curing device, the die-bonded components, fiber arrays or lenses, optical power and spot data are aligned, coupled, cured and error-compensated to obtain target assembled components. The improved multidimensional gradient optimization algorithm integrates particle swarm optimization and gradient descent algorithms; Inspection and Adjustment Unit: Based on robotic arms, 3D vision scanning equipment, and multi-performance testing equipment, the unit performs whole-machine inspection, including assembly, gap detection, fastening, and full-performance testing of target assembly components and optical module shells. This process yields optical, mechanical, and appearance inspection data for the whole machine, as well as qualified / unqualified optical modules. Based on a sorting device and data management system, the unit sorts, records, and analyzes the whole-machine inspection results and assembly data from each workstation, resulting in categorized and sorted optical modules and a full lifecycle assembly traceability log.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for automatic assembly of optical module components based on machine vision, characterized in that, Includes the following steps: S1. Based on camera calibration technology and improved YOLOv5 model, standard images and system parameters of each component of the optical module are calibrated, features are extracted and model training is performed to obtain a unified pixel-physical world coordinate system, error compensation parameters and adaptive component recognition model. The improved YOLOv5 model integrates the CBAM attention mechanism and the Focal Loss loss function; S2. Based on a vibratory feeder, conveyor line, multimodal vision camera, and improved sub-pixel corner detection algorithm, various optical module components and acquired multi-view images are processed for feeding, imaging, preprocessing, and pose calculation to obtain the component's three-dimensional pose data. and deviation data ; S3. Based on a robotic arm, a vision guidance device, and a UV curing device, PCBA and optical chips are positioned, picked up, mounted, electrically connected, and cured to obtain die-bonded components. S4. Based on passive pre-alignment and active fine alignment modes, improved multi-dimensional gradient optimization algorithm and temperature-controlled curing device, the die bonding component, fiber array or lens, optical power and spot data are aligned, coupled, cured and error-compensated to obtain the target assembly component. The improved multidimensional gradient optimization algorithm integrates particle swarm optimization and gradient descent algorithms; S5. Based on robotic arms, 3D vision scanning equipment and multi-performance testing equipment, perform whole machine testing, and perform assembly, gap detection, fastening and full performance testing on target assembly components and optical module shells to obtain optical, mechanical and appearance inspection data of the whole machine and qualified or unqualified optical module whole machine. S6. Based on the sorting device and data management system, the whole machine inspection results and assembly data of each workstation are sorted, recorded and analyzed to obtain the sorted optical modules and the full life cycle assembly traceability log.

2. The automatic assembly method for optical module components based on machine vision according to claim 1, characterized in that, The specific contents of S1 include: Based on camera calibration technology, the system parameters are calibrated to obtain a unified pixel-physical world coordinate system and error compensation parameters; The camera calibration techniques include camera intrinsic parameter calibration, binocular extrinsic parameter calibration, and hand-eye calibration. The system parameters include camera parameters, robotic arm parameters, and motion platform parameters; Based on the improved YOLOv5 model, standard image features of each component of the optical module are extracted and the model is trained to obtain an adaptive component recognition model. The optical module components include an optical chip, fiber array, lens, housing, and PCBA; Specifically, standard images of various components are collected, and key features of component Mark points, edges, and keyways are extracted through image preprocessing techniques. The extraction of minute features of fiber core and chip photosensitive area is particularly enhanced to obtain feature data. The standard image and the extracted feature data are input into the improved YOLOv5 model for training to obtain an adaptive component recognition model; The improved YOLOv5 model integrates the CBAM attention mechanism and the Focal Loss loss function; The CBAM attention mechanism module is embedded in the feature fusion link between the neck and head of the model, and is divided into a channel attention unit (CA) and a spatial attention unit (SA), with CA and SA executed in series. The Focal Loss function replaces the original YOLOv5 cross-entropy loss to address the feature imbalance of optical module components and improve the accuracy of small target recognition.

3. The automatic assembly method for optical module components based on machine vision according to claim 2, characterized in that, The specific content of the channel attention unit (CA) is as follows: ； ； The specific content of the Spatial Attention Unit (SA) is as follows: ； ； in, F The input is the original feature map of CBAM, with dimension . C is the number of channels, H is the feature map height, and W is the feature map width. GAP ( F ) for feature maps F Perform global average pooling. GMP ( F ) for feature maps F Perform global max pooling. To merge the results of average pooling and max pooling, FC () represents a fully connected layer operation. This refers to the adjustment coefficient for the weights of small feature channels. For small feature channel weight vectors, the dimension is... , To Perform L2 normalization. It is the Sigmoid activation function. To improve the post-channel attention weight map, the dimensions are... , The output feature map after channel attention processing. This is the original response value vector of the small feature channels. To Perform Softmax normalization. This is element-wise multiplication. Here is the feature map after channel attention processing, and here is the input for spatial attention. For feature splicing operations, This is a 1×1 convolution operation. As a spatial location sensitive factor, The coordinate weight matrix of the small feature space has dimensions. , To Perform bilinear interpolation mapping. The standard deviation of the Gaussian kernel. To improve the spatial attention weight map, the dimension is... , Based on Generate Gaussian space weights. This is the final output feature map of the CBAM module.

4. The automatic assembly method for optical module components based on machine vision according to claim 3, characterized in that, The Focal Loss function introduces an adaptive weighting term for the difficulty in distinguishing small feature samples and a dynamic adjustment factor for class weights to address the imbalance in loss distribution between the difficulty in distinguishing small feature samples and regular samples and background samples. The expression is as follows: ； in, To improve the total Focal Loss value, As a positive and negative sample balance factor, A balance factor for samples with minor and conventional features. For samples with small features that are difficult to distinguish, the weight adaptive term is used. The target class probability predicted by the model. Based on the focusing parameters, To optimize the coefficients of the focus term, To improve the focus item, For the logarithmic loss term, To locate the loss weight coefficients, for CIoU Locating the loss, For small feature real regions, For the small feature model to predict the region, For small feature real region and predicted region IoU value; in, .

5. The automatic assembly method for optical module components based on machine vision according to claim 4, characterized in that, S2. Based on a vibratory feeder, conveyor line, multimodal vision camera, and improved sub-pixel corner detection algorithm, various optical module components and acquired multi-view images are processed for feeding, imaging, preprocessing, and pose calculation to obtain the component's three-dimensional pose data. and deviation data The specific content also includes: The optical module housing, PCBA, optical chip, fiber array, and lens assembly are sequentially fed into the assembly station via a vibratory feeder and conveyor line. A multimodal vision camera is used to acquire multi-view images of the components. After Gaussian filtering for noise reduction, Canny edge detection, and histogram equalization preprocessing, a standard image is obtained. The multimodal vision camera includes a top-view high-resolution CCD camera, a side-view 3D structured light camera, and a coaxial illumination module. The target corner coordinates of the standard image are obtained by improving the accuracy of the sub-pixel corner detection algorithm. An adaptive component recognition model is used to perform preliminary identification of component types and poses on standard images, thereby obtaining the three-dimensional pose data of each component in the world coordinate system. and deviation data .

6. The automatic assembly method for optical module components based on machine vision according to claim 5, characterized in that, The details of the improved sub-pixel corner detection algorithm are as follows: For a standard image, calculate the Harris response value of each pixel, set the response threshold to 0.01-0.03, filter out pixels with Harris response values ​​higher than the threshold as candidate corner points, initially locate the positions of component Mark points and edge corner points, and remove non-corner point pixels; For the initially located candidate corner points, the Zernike matrix subpixel localization algorithm is used to achieve subpixel-level correction of the corner point coordinates. The steps are as follows: Centered on the candidate corner point, select a 3×3 or 5×5 local image window and calculate the Zernike moments of the image within the window; By using the phase and amplitude information of the Zernike moments, the true sub-pixel coordinates (x, y) of the corner points are fitted. sub ,y sub ), correcting the initial positioning corner point deviation; For tiny corner points of optical modules, optimize the Zernike moment calculation weights to enhance the Zernike moment response of tiny corner points; To suppress localization noise using Kalman filtering and obtain stable corner coordinates, i.e., target corner coordinates, the following steps are taken: Establish a Kalman filter model and convert the sub-pixel corner coordinates (x... sub ,y sub As the observed values, set the state equation and the observation equation; A Kalman filter prediction-update process is performed on the sub-pixel corner coordinates of each frame of the image to remove outliers caused by positioning noise, and output stable and accurate final corner coordinates (x, y, y). final ,y final ).

7. The automatic assembly method for optical module components based on machine vision according to claim 6, characterized in that, The specific content of S4 is as follows: The fiber optic array or lens is actively aligned and coupled with the chip, employing passive pre-alignment, active precision alignment, and closed-loop feedback. First, visual guidance moves the fiber optic array or lens near the chip's optical window, controlling the deviation within a certain range. Within 30μm; The optical signal source is turned on, and the output optical power P and the light spot image are collected in real time through the optical power meter and the spectrum analyzer. The vision system simultaneously monitors the relative displacement and angular deviation between the fiber end face and the photosensitive area of ​​the chip, and executes the improved multidimensional gradient optimization algorithm to perform nanometer-level step search in 6 degrees of freedom until the optical power reaches the preset threshold and the light spot roundness reaches the target roundness. After locking the current position, apply low-shrinkage UV adhesive and cure in stages with temperature control. During the curing process, continuously monitor the light power to compensate for thermal deformation errors.

8. The automatic assembly method for optical module components based on machine vision according to claim 6, characterized in that, The improved multidimensional gradient optimization algorithm integrates particle swarm optimization and gradient descent algorithms, introduces adaptive inertia weights, and the objective function is expressed as follows: ； in, Let be the objective function. P This represents the actual coupled optical power output of the fiber array. The maximum optical power under ideal coupling conditions is given by k1, where k1 is the plane translation deviation. The corresponding coupling coefficient, k2 is the vertical translation deviation () The corresponding coupling coefficient, k3 is the angle deviation ( The corresponding coupling coefficient, This represents the translational deviation along the x-axis. This represents the translational deviation along the y-axis. The translational deviation is in the z-axis direction. This represents the rotational deviation about the x-axis. This represents the rotational deviation about the y-axis. This represents the rotational deviation about the z-axis.

9. An automated assembly system for optical module components based on machine vision, characterized in that, include: Basic model building unit: Based on camera calibration technology and improved YOLOv5 model, standard images and system parameters of each component of the optical module are calibrated, features are extracted and model training is performed to obtain a unified pixel-physical world coordinate system, error compensation parameters and adaptive component recognition model. The improved YOLOv5 model integrates the CBAM attention mechanism and the Focal Loss loss function; Adaptive recognition unit: Based on vibratory feeder, conveyor line, multimodal vision camera and improved sub-pixel corner detection algorithm, it performs loading, imaging, preprocessing and pose calculation on various optical module components and acquired multi-view images to obtain the component's three-dimensional pose data. and deviation data ; Assembly Unit: Based on a robotic arm, vision guidance device and UV curing device, PCBA and optical chip are positioned, picked up, mounted, electrically connected and cured to obtain die-bonded components. Based on passive pre-alignment and active fine alignment modes, improved multi-dimensional gradient optimization algorithm and temperature-controlled curing device, the die-bonded components, fiber arrays or lenses, optical power and spot data are aligned, coupled, cured and error-compensated to obtain target assembled components. The improved multidimensional gradient optimization algorithm integrates particle swarm optimization and gradient descent algorithms; Inspection and Adjustment Unit: Based on robotic arms, 3D vision scanning equipment, and multi-performance testing equipment, the unit performs whole-machine inspection, including assembly, gap detection, fastening, and full-performance testing of target assembly components and optical module shells. This process yields optical, mechanical, and appearance inspection data for the whole machine, as well as qualified / unqualified optical modules. Based on a sorting device and data management system, the unit sorts, records, and analyzes the whole-machine inspection results and assembly data from each workstation, resulting in categorized and sorted optical modules and a full lifecycle assembly traceability log.

Images