SMT alignment method and device for uncovering fid of rigid-flex board

Abstract

The application discloses a kind of soft and hard combination board uncovering and revealing Fid SMT alignment method and device.Establish positioning coordinate system after plate, collect multiple exposure images and polarization switching images and HDR fusion, remove specular reflection, combined with height data registration to form joint feature map;It is input into improved HG-BCNet segmentation neural network, and selection uncovering strategy;Recapture calculates the occlusion rate and quality score to obtain a set of usable Fid, and trigger secondary micro-uncovering and adaptive recapture when abnormal;Robustly solve compensation parameters for usable Fid, and fit flexible zone deformation with partition affine and thin plate spline;Issue to SMT equipment recapture verification, trigger recalculation / reshaping according to residual value and spatial distribution and output evidence chain.The application improves HG-BCNet height guided segmentation and uncovering closed-loop control, Fid availability quantitative determination and flexible zone deformation compensation by combination, significantly improves SMT alignment accuracy and production line beat efficiency, reduces quality risks such as mounting offset.

Description

Technical Field

[0001] This invention relates to the field of machine vision alignment technology, and more specifically to an SMT alignment method and apparatus for exposing the FID by removing the cover of a rigid-flex board. Background Technology

[0002] Rigid-flex PCBs combine rigid and flexible areas, enabling three-dimensional bending assembly and high-density interconnection within limited space. They are widely used in consumer electronics, automotive electronics, medical, and aerospace industries. Compared to traditional rigid PCBs, rigid-flex PCBs are more susceptible to stress release, temperature and humidity changes, and differences in clamping methods during manufacturing, transfer, and SMT assembly. This manifests as warping, springback, localized expansion and contraction, and shear deformation in the flexible areas, making it more difficult to maintain stable alignment accuracy in printing, mounting, and dispensing processes. To ensure assembly yield and consistency, SMT equipment typically needs to establish a board coordinate system by identifying a Fiducial (Fid) reference mark and performing translation, rotation, and necessary magnification / nonlinearity compensation to offset errors between the board and equipment coordinates.

[0003] However, in the actual production of rigid-flex PCBs, protective films, cover sheets, tapes, or temporary covers (collectively referred to as "cover sheets") are often placed on the board surface to protect the flexible area traces, coverlay opening areas, gold fingers, or specific sensitive pads. Cover sheets can prevent dust, scratches, and reduce the risk of mechanical abrasion during transportation, turnover, stacking, and loading, but they may also obscure critical fiber areas (Fids), especially when the fiber areas are located near the flexible area or close to the opening areas. If the FID is obscured by the cover plate, the SMT equipment will have difficulty accurately identifying the board reference. Existing production lines usually solve this problem in the following ways: First, reserve an unobstructed area for the FID during the design phase or move the FID to a rigid area. However, this method increases wiring constraints, occupies limited board space, and may interfere with structural components. Second, manually remove or lift the cover plate before mounting, and then perform alignment and placement. However, manual operation has a significant impact on cycle time and there are risks of inconsistent lifting force, improper lifting angle causing tearing of the cover film, strain on the flexible area, or adhesive residue contamination. Third, abandon the FID and use positioning holes, board shape, or carrier reference for alignment. However, this method is difficult to compensate for the local non-rigid deformation of the flexible area, which can easily lead to defects such as solder paste printing misalignment, placement misalignment, uneven copper exposure of pads, bridging, and cold solder joints, especially in micro-pitch devices and high-density BGA / connector scenarios.

[0004] Furthermore, even with automated visual recognition, the surface of rigid-flex PVC panels is often subject to interference from ink reflections, specular reflections from the cover film, translucent textures of the tape, contaminant particles, and scratches. This leads to unstable image features at the cover sheet boundaries and in the FID area, resulting in issues such as boundary recognition jitter, misjudgment of graspable points, and misidentification due to partial occlusion or reflection of the FID after the cover is removed. Traditional algorithms based on thresholds, edge detection, or simple template matching lack robustness under complex lighting and material variations; while image algorithms lacking linkage with 3D height information often struggle to distinguish between "true boundary steps" and "reflective pseudo-edges." Simultaneously, the removal of the cover is itself a mechanical process. Without force / displacement / angle closed-loop control, sudden changes in peeling force may cause excessive tension in the flexible area or tearing of the cover sheet, further leading to contamination and scrap. Even after removal, if subsequent alignment still uses a single rigid body transformation model, it is difficult to effectively compensate for local deformation in the flexible area, resulting in non-uniform distribution of re-image residuals in the flexible area that cannot be automatically corrected.

[0005] Therefore, existing technologies urgently need an SMT alignment solution suitable for rigid-flex PCBs that can reliably identify cover glass boundaries and grippable points within the production line cycle time, achieve closed-loop safety control of the cover removal process, quantitatively determine the exposure and availability of FID after cover removal, and combine the partitioning or nonlinear compensation model of flexible area deformation to achieve closed-loop verification and adaptive recalculation of alignment parameters, thereby improving alignment accuracy and production consistency, and reducing reliance on manual labor and quality risks. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses an SMT alignment method and apparatus for exposing cover plates (Fids) on rigid-flex PCBs. After mounting the board, a positioning coordinate system is established. Multi-exposure images and polarization-switched images are acquired and HDR-fused, specular reflections are removed, and height data is used for registration to form a joint feature map. This map is then input into an improved HG-BCNet segmentation neural network, which outputs a cover plate mask, boundary probability map, and uncertainty map. Cover plate boundary polygons, boundary confidence, and grabbable candidate points are extracted, and a cover plate unveiling strategy is selected. Adhesion, lifting, and peeling / lifting are controlled in a closed-loop manner based on peeling force / displacement / angle. Repeated testing calculates the occlusion rate and quality score to obtain a set of usable FIDs. Anomalies trigger secondary micro-uncovering and adaptive repeated testing. Weighted robust compensation parameters are calculated for the usable FIDs, and the deformation of the flexible area is fitted using affine partitioning and thin-plate splines. The results are sent to the SMT equipment for repeated testing verification, triggering recalculation / refitting based on residual values ​​and spatial distribution, and outputting an evidence chain. This application combines anti-reflective multimodal imaging, improved HG-BCNet height-guided segmentation and capping closed-loop control, FID availability quantification and flexible area deformation compensation to achieve automatic and reliable capping and high-precision alignment closed-loop correction. This significantly improves SMT alignment accuracy and production line cycle efficiency, and reduces quality risks such as manual intervention, flexible area damage, contamination and placement misalignment.

[0007] A method for SMT alignment of rigid-flex PCBs to expose the fiber (Fid) by peeling back the cover, comprising the following steps:

[0008] S1: Perform the mounting and initial positioning of the rigid-flex PCB, and establish a coarse positioning coordinate system based on the positioning holes and the PCB outline; S2: Acquire board surface images and 3D height data to identify the cover sheet boundary and its relative position with the target FID. Determine the prior region of the target FID based on the coarse positioning coordinate system. Input the board surface images and 3D height data into the improved HG-BCNet height-guided boundary consistency segmentation neural network to output a cover sheet segmentation mask, boundary probability map, and pixel-level uncertainty map. Extract the cover sheet boundary polygon, boundary confidence, and set of grabbable candidate points from the cover sheet segmentation mask and boundary probability map. Select a cover removal strategy based on the spatial relationship between the cover sheet boundary polygon and the prior region of the target FID, the adhesion level, and the boundary confidence. S3: Controls the cover-opening actuator to perform attaching, pulling, peeling, and lifting actions on the cover sheet, and performs closed-loop control based on peeling force, peeling displacement, and peeling angle during the action; S4: After uncovering, the target Fid is exposed and its availability is determined, and a set of available Fids is generated; S5: Based on the available Fid set, calculate the alignment compensation parameters and generate partition compensation or nonlinear compensation for the deformation of the flexible region; S6: Send the alignment compensation parameters to the SMT equipment and perform a re-scan verification, recalculate and perform a second micro-uncovering based on the re-scan residual trigger parameters.

[0009] Preferably, in step S2, the acquired panel images include a multi-exposure image sequence and a polarization state switching image sequence under the same field of view; the multi-exposure image sequence is HDR fused and the polarization state switching image sequence is despectralized to obtain a reference image with reduced reflection; the reference image is spatiotemporally registered with the three-dimensional height data to construct a joint feature map; on the joint feature map, an improved HG-BCNet height-guided boundary consistency segmentation neural network is used to perform height-guided cover semantic segmentation and boundary refinement to obtain the cover boundary polygon, boundary confidence, and a set of grabbable candidate points; and the uncovering strategy is selected based on the spatial relationship between the cover boundary polygon and the target FID prior region, the adhesion level, and the boundary confidence.

[0010] Preferably, highly guided cover semantic segmentation and boundary refinement are performed on the joint feature map by the improved HG-BCNet highly guided boundary consistent segmentation neural network, which is implemented by a highly guided segmentation model. The improved HG-BCNet highly guided boundary consistent segmentation neural network includes: An image coding branch is used to extract multi-scale image features from the reference image for suppressing reflection; The height encoding branch is used to extract multi-scale height features from the three-dimensional height data and height gradient features; The height boundary gating fusion module is used to generate gating weights based on the height gradient and perform weighted fusion of the multi-scale image features and multi-scale height features. The multi-head output structure outputs at least the cover plate segmentation mask, the boundary probability map, and the pixel-level uncertainty map; The boundary probability map is used to refine the edges of the cover plate segmentation mask, and the uncertainty map is used to form the boundary confidence and participate in the screening of the set of grabbable candidate points.

[0011] Preferably, the boundary refinement head of the improved HG-BCNet height-guided boundary consistency segmentation neural network includes a height gradient-adaptive dilated convolution structure, whose dilation rate adaptively varies with the height gradient magnitude at the pixel location: the larger the height gradient magnitude, the larger the dilation rate is used to enhance boundary continuity; the smaller the height gradient magnitude, the smaller the dilation rate is used to avoid excessive boundary smoothing and drift; wherein, the dilation rate is an integer between 1 and 8, and the adaptive selection from height gradient to dilation rate is achieved through a preset piecewise mapping or linear mapping.

[0012] Preferably, the improved HG-BCNet height-guided boundary consistency segmentation neural network introduces a boundary height step consistency constraint during training or online self-calibration, so that the boundary probability map output by the model is consistent with the height gradient feature in spatial location; and during the inference stage, it adaptively adjusts the simplification tolerance of the boundary polygon fitting and the threshold for selecting grabbillable candidate points based on the pixel-level uncertainty map, so as to reduce missegmentation and misgrabbing under conditions of reflection, pollution, and partial occlusion.

[0013] Preferably, step S3 includes adaptively setting the peel angle window and speed curve based on the confidence level of the cover sheet boundary and the adhesion level, and performing stabilization control of speed suppression, angle correction and micro-retreat displacement when a sudden change in peel force is detected.

[0014] Preferably, step S4 includes acquiring a repeat image of the target FID region and performing FID detection in conjunction with three-dimensional height data. The FID detection includes: locating candidate FID regions on a reference image with reduced reflectivity; detecting the FID center point within the candidate regions and outputting the detection uncertainty; calculating the FID occlusion rate based on the cover film semantic segmentation results; and generating a FID quality score based on contrast, sharpness, specular reflection index, and contamination features. When the occlusion rate or quality score does not meet the threshold, a secondary micro-uncovering is triggered and adaptive repeating is triggered. The adaptive repeating includes adaptive adjustment of exposure, polarization state, and incident light angle.

[0015] Preferably, step S5 includes: assigning weights to each Fid based on its detection uncertainty and quality score; obtaining global rigid body transformation parameters from the plate coordinate system to the device coordinate system using weighted robust estimation; and fitting the local deformation of the flexible region using a partitioned affine model and a thin plate spline nonlinear model under the constraint of a three-dimensional height field, and outputting a partitioned compensation table and a nonlinear compensation mapping.

[0016] Preferably, step S6 includes: when the reshoot residual is non-uniformly distributed in the flexible area and exceeds a threshold, the refitting of the partition compensation table or nonlinear compensation mapping is triggered, and an evidence chain corresponding to the compensation parameters is output. The evidence chain includes at least the available FID set, the quality score of each FID, the occlusion rate, the reshoot residual and its corresponding image position.

[0017] This application also provides an SMT alignment device for exposing the FID (Fid) of a rigid-flex PCB, comprising: The coordinate system establishment module performs the mounting and initial positioning of the rigid-soft bonding board, and establishes a coarse positioning coordinate system based on the positioning holes and the shape of the board. The uncovering strategy selection module acquires board surface images and 3D height data to identify the cover edge and its relative position with the target Fid. Based on the coarse positioning coordinate system, it determines the prior region of the target Fid. The board surface images and 3D height data are input into an improved HG-BCNet height-guided boundary consistency segmentation neural network to output a cover segmentation mask, a boundary probability map, and a pixel-level uncertainty map. The cover segmentation mask and the boundary probability map are used to extract the cover boundary polygon, boundary confidence, and a set of grabbable candidate points. The uncovering strategy is selected based on the spatial relationship between the cover boundary polygon and the prior region of the target Fid, the adhesion level, and the boundary confidence. The cap-opening execution module controls the cap-opening execution mechanism to perform actions such as attaching, pulling, peeling, and lifting the cap sheet, and performs closed-loop control based on peeling force, peeling displacement, and peeling angle during the action. The Fid generation module confirms the exposure and availability of the target Fid after the cover is removed, and generates a set of usable Fids. The compensation module calculates the alignment compensation parameters based on the available FID set and generates partition compensation or nonlinear compensation for the deformation of the flexible region. The secondary fine-tuning module sends the alignment compensation parameters to the SMT equipment and performs a re-examination verification, recalculating and performing a secondary micro-uncovering based on the re-examination residual trigger parameters.

[0018] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) This application introduces multi-exposure and polarization switching imaging before peeling off the cover, performs HDR fusion and specular destigma decomposition, and then registers it with three-dimensional height data to form a joint feature map. The improved HG-BCNet is used to realize semantic segmentation and boundary refinement of the cover sheet. Compared with methods that rely solely on single-frame visible light or traditional edge detection, the joint feature map utilizes both texture information and height step information, which can effectively suppress specular reflection and semi-transparent texture interference caused by cover film, tape, ink, etc., and reduce the probability of "false edges", "broken edges" and missegmentation. The uncertainty and boundary confidence of the network output further quantify the recognition reliability, enabling the system to automatically filter out low-confidence boundary segments and output more reliable boundary polygons and a set of grabbable candidate points, thereby reducing the selection of adsorption points, peeling failure and mis-peeling from the source, and improving the robust adaptation capability to different cover sheet materials, adhesion differences and lighting changes.

[0019] (2) This application incorporates the confidence level of the cover sheet boundary, adhesion level, and candidate point set into the capping control loop. It implements closed-loop control based on peel force, displacement, and / or angle throughout the entire process of application, lifting, and peeling / uncapping. When a sudden change in peel force is detected, stabilization strategies such as speed suppression, angle correction, and micro-retraction displacement are executed. Compared with manual capping or open-loop mechanical capping, this solution can limit peak peel stress, reduce the risk of stretching in the flexible area, tearing of the cover film, adhesive residue dragging, and contamination, and improve the process consistency of rigid-flex boards under high-cycle conditions on the production line. Simultaneously, the system supports adaptive reshooting of exposure, polarization, and incident angle, as well as automatic triggering of secondary micro-capping, enabling rapid recovery of FID exposure and usability even under conditions of partial obstruction, reflection, contamination, or re-adhesion, reducing line downtime and rework.

[0020] (3) This application determines the occlusion rate and quality score of the exposed FID after the cover is removed, and assigns weights to the available FIDs in combination with the detection uncertainty. It adopts weighted robust estimation to solve the global rigid body transformation, which improves the anti-interference ability against abnormal FIDs, local reflections and contamination. At the same time, under the constraint of three-dimensional height field, it introduces partitioned affine and thin plate spline nonlinear compensation, which can model and correct the springback, local expansion and contraction and shear deformation of the flexible area of ​​the rigid-soft bonded plate, which significantly reduces the local mounting offset caused by using only positioning holes / shapes or a single rigid body model. The re-image residual is not only used for threshold determination, but also triggers recalculation and refitting based on the spatial distribution of residuals to form an alignment closed loop. In addition, the output includes the FID set, score, occlusion rate, residual and image position evidence chain, which facilitates quality traceability and process optimization, and achieves simultaneous improvement of accuracy, yield and maintainability. Attached Figure Description

[0021] Figure 1 This is a flowchart of an SMT alignment method for exposing the FID by removing the cover of a rigid-flex board according to the present invention. Figure 2 This is a structural diagram of an SMT alignment device module for exposing the FID (Fid) by removing the cover of a rigid-flex board according to the present invention. Detailed Implementation

[0022] Those skilled in the art will understand that, in order to make the above-mentioned objects, features, and beneficial effects of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This application illustrates an SMT alignment method for exposing the fiber (Fid) on a rigid-flex PCB, comprising the following steps: S1: The rigid-flex PCB is loaded and initially positioned, establishing a coarse positioning coordinate system based on the positioning holes and the PCB's shape. In some embodiments, S1 involves loading and initial positioning (establishing a coarse positioning coordinate system based on the positioning holes and the PCB's shape), after which the rigid-flex PCB, either along with a carrier or directly as a bare board, enters the alignment station of the SMT production line. The alignment station includes a conveying mechanism, a positioning platform, a clamping and positioning mechanism, a top industrial camera, backlight and ring light source, and an industrial control computer controller. The positioning platform can be a vacuum adsorption platform or a mechanical support platform, used to stably support the board during alignment. The clamping and positioning mechanism may include left and right side grippers and front and rear stops, or include positioning pins and locking structures for the carrier. Loading and mechanical pre-positioning: When the board reaches the designated position on the positioning platform, a photoelectric sensor outputs a positioning signal, and the conveying mechanism stops. The clamping and positioning mechanism operates as follows: the left and right grippers clamp inward to the edge of the plate or the carrier, and the front stop cooperates with the side guide rail to restrict the plate's freedom of movement in the front-back and left-right directions. If a carrier is used, the positioning pin is inserted into the carrier's reference hole and locked to bring the plate into a stable mechanical pre-positioning state. At this point, the plate is fixed, but residual errors such as upper plate offset, slight rotation, and carrier clamping errors may still exist, requiring further visual establishment of a coarse positioning coordinate system.

[0023] The system captures images of the positioning holes and extracts their centers. The controller drives the top camera to acquire images of the hole areas under backlight conditions. Backlighting creates distinct dark areas around the holes, facilitating hole boundary extraction. The system identifies holes within a pre-defined search window: first, it performs binarization segmentation on the image to obtain candidate hole regions; then, it filters these regions by connecting components, removing dust particles or stray small areas, retaining only those meeting area and roundness requirements; subsequently, it fits the hole edges to obtain the center position of the positioning hole. If there are two positioning holes, two hole centers are obtained separately, and the extraction results undergo a consistency check, such as whether the hole diameter falls within the allowable range, whether the fitting residual is less than a threshold, and whether the direction of the connection line between the hole centers is reasonable. If the check fails, the hole is marked as an abnormal hole and enters a backup strategy.

[0024] The system captures images of the board's outline and extracts its edges. A controller drives a top camera to acquire images of the board's outline area under ring light or low-angle side lighting conditions. The system performs edge detection on the outline images, extracts the board's outline contour lines, and performs straight-line fitting on at least one long side to obtain the outline direction. Simultaneously, it extracts at least one corner point as a reference for outline verification. The outline information is primarily used to suppress false detections of positioning holes, correct the board's rotation direction, and provide alternative positioning information when hole positions are abnormal.

[0025] The system establishes a coarse positioning coordinate system and outputs quality markers. It primarily relies on the center of the positioning holes and the outer shape direction to calculate the initial translation and rotation angle of the workpiece relative to the positioning platform, thus establishing the coarse positioning coordinate system. Specifically, it uses the center of one of the positioning holes as the coordinate origin reference point, the direction of the line connecting the two holes as the rotation direction reference, and then uses the long side direction of the outer shape to verify the consistency of the rotation direction. When the difference between the outer shape direction and the direction of the hole connection exceeds an allowable threshold, correction is performed primarily based on the outer shape direction and secondarily on the hole direction. Simultaneously, the system outputs coarse positioning quality markers, including: whether hole recognition was successful, whether the hole fitting residual exceeds the limit, whether the outer shape edge fitting residual exceeds the limit, and the consistency error between the hole direction and the outer shape direction. These are used in subsequent steps to adaptively set the search range and re-shooting strategy.

[0026] Anomaly and Degradation Strategies: The system performs degraded positioning but still outputs a coarse positioning coordinate system under the following conditions: If only one positioning hole is identified, a coarse positioning coordinate system is established using the "single hole center as position reference and the long side direction of the outline as rotation reference"; if all positioning holes fail to be identified, a coarse positioning coordinate system is established using the "long side direction of the outline and corner point position," and the quality marker is set to low confidence; if the outline is obscured or the edges are unclear, the carrier reference hole or carrier reference edge is used preferentially for coarse positioning. These degradation strategies ensure production line continuity while providing a usable initial estimate for subsequent fine alignment.

[0027] The system generates a priori region for the target FID in conjunction with subsequent steps. It maps the coarse positioning coordinate system to the nominal position of the target FID recorded in the process template, obtaining the predicted position of the target FID on the current board. A priori region for the target FID is then generated centered on this predicted position. If the coarse positioning quality is marked as high, the priori region is set to a smaller window to improve computational efficiency; if the coarse positioning quality is marked as low, the priori region is automatically expanded, and more retakes are allowed to improve the robustness of subsequent cover identification and FID detection. Through the above embodiments, a stable coarse positioning coordinate system can be established even under conditions such as board offset on rigid-flex boards, carrier repetitive positioning errors, and slight springback in the flexible area, providing a reliable initial reference for subsequent cover identification, cover removal strategy selection, and FID exposure confirmation.

[0028] In some embodiments, the upper board and initial positioning station are located before the board feeding end of the solder paste printer or pick and place machine, and include: 1) Conveying mechanism: belt / chain conveyor + side guide rail; 2) Positioning reference platform: vacuum adsorption platform or mechanical support platform; 3) Clamping mechanism: left and right side grippers + front and rear stops (or carrier locking mechanism); 4) Coarse positioning vision unit: top industrial camera (12MP or above, telecentric lens preferred) + ring light / side light; 5) Positioning hole detection light source: coaxial + backlight (backlight is used for hole height comparison); 6) Controller and coordinate mapping module: used to calculate the coarse positioning coordinate system and output it to S2.

[0029] In some embodiments, pre-calibration (performed once / during line change), (1) Camera-platform coordinate calibration: Place a calibration plate on the platform, acquire calibration point images, and calculate the extrinsic parameter transformation Tcam→table from camera coordinates to platform coordinates. (2) Process template establishment: Record the nominal position of the positioning hole, the plate outline boundary, and the nominal position of the target Fid (used for the subsequent "Fid prior area") in the CAD / GERBER or MES process data. The template coordinate system is denoted as CStpl. (3) Carrier deviation compensation: If a carrier is used, record the nominal relationship between the carrier reference hole and the platform positioning pin, and establish a carrier number and deviation table.

[0030] S1-1 The upper plate is mechanically pre-positioned. The plate enters the platform with the conveyor and triggers the photoelectric sensor to signal the position. The platform opens the vacuum adsorption or clamping mechanism to close, completing the mechanical pre-positioning: the left and right side grippers limit the plate edges; the front stop positions the front edge of the plate; or rigid positioning is achieved by inserting a positioning pin into the carrier reference hole.

[0031] At this point, a coarse mechanical positioning state is obtained, but several errors still exist: upper plate offset, rotation, edge burrs, and springback in the flexible area. S1-2: Coarse identification and hole center fitting of positioning holes. The top camera captures the positioning hole area (generally 2 holes; 3 holes can improve robustness) under backlight conditions: Threshold segmentation is performed within the ROI to obtain the hole connected domain; ellipse fitting / least square circle fitting is performed on the edge points of the connected domain to obtain the hole center coordinates (xh1, yh1), (xh2, yh2); hole diameter consistency and fitting residuals are calculated. If the residual is greater than the threshold (e.g., >0.15mm), it is judged as an abnormal hole (burr obstruction / contamination), and the backup strategy is entered. Engineering parameter example: Hole center positioning accuracy can reach ±0.03mm (telecentric lens + backlight). S1-3 Plate Shape Edge Detection and Corner Extraction (for Rotation / Scale Assistance): The plate shape area is photographed under ring light or side light. Canny edge or line segment detection is used to extract the shape contour. A straight line fit is performed on the long side to obtain the edge direction angle θedge. At least one corner point (such as the upper left corner) is extracted to suppress false detection of hole positions. The shape information is mainly used for: 1) verifying whether the hole position direction is consistent with the shape direction; 2) serving as a substitute alignment basis when the hole position is missing / occluded. S1-4 Coarse Positioning Coordinate System Solution (Translation + Rotation + Optional Scale): The rigid body transformation is obtained by comparing the nominal hole positions (Xh1,Yh1), (Xh2,Yh2) in the template coordinate system CStpl with the measured hole centers (xh1,yh1), (xh2,yh2): the rotation θ0 is obtained from the angle between the two hole vectors; the translation (tx,ty) is obtained from the difference in hole centers; if slight scale differences are allowed (camera calibration or temperature drift), a scale ss (generally close to 1) can be added. The coarse positioning transformation Ttpl→table(0) is obtained, and a coarse positioning coordinate system CS0 (defined as the plate coordinate system after the template coordinate system is transformed) is established accordingly. The output content includes at least: translation tx,ty; rotation θ0; coarse positioning quality indicators: hole fitting residual, shape fitting residual, and hole-shape consistency error. The prior region of the target Fid is generated, and the nominal position (Xfid,Yfid) of the target Fid in the template is mapped to the platform coordinate using the coarse positioning transformation Ttpl→table(0) to obtain the coarse predicted position (xfid(0),yfid(0)). The search window is adaptively set according to the coarse positioning quality indicators: normal: ROI radius 2–4mm; single hole / shape degradation: ROI radius 5–10mm.

[0032] This ROI is the "target Fid prior region" in the claim. Running HG-BCNet and Fid detection in this region in S2 / S4 can significantly reduce false detections and computational load.

[0033] S2: Acquire the board surface image and 3D height data to identify the cover plate boundary and its relative position with the target Fid. Determine the prior region of the target Fid based on the coarse positioning coordinate system. Input the board surface image and 3D height data into the improved HG-BCNet height-guided boundary consistency segmentation neural network to output the cover plate segmentation mask, boundary probability map and pixel-level uncertainty map. Extract the cover plate boundary polygon, boundary confidence and set of grabbable candidate points from the cover plate segmentation mask and boundary probability map. Select the uncovering strategy based on the spatial relationship between the cover plate boundary polygon and the prior region of the target Fid, the adhesion level and the boundary confidence.

[0034] In some embodiments, 1) Hardware and acquisition configuration (same field of view, same coordinate system): The workstation in step S2 is installed above the cover-opening actuator and includes: a top-mounted area array industrial camera (preferably with a telecentric lens), and a switchable light source: coaxial light, low-angle ring light; polarization components: a linear polarizer at the front end of the camera and a polarizer at the end of the light source (or an electrically controlled polarizing wheel / liquid crystal polarizer) for acquiring images in two polarization states; a three-dimensional height sensor: a structured light or laser triangulation profilometer (shared field of view with the camera or calibrated and registerable). The camera and height sensor are both fixed on the same rigid bracket. A mapping relationship of "image coordinates - platform coordinates - height coordinates" is established through one calibration to ensure strict alignment of subsequent joint feature maps. 2) The target Fid prior region (ROI) is determined by the coarse positioning coordinate system. The coarse positioning coordinate system has been output in S1. In S2, the controller reads the nominal position and size of the target Fid from the process template (CAD / GERBER / MES), maps it to the current board surface based on the coarse positioning coordinate system, and generates the prior region ROI of the target Fid: When the quality mark of S1 is high: the ROI takes a smaller window (e.g., only covering a few millimeters around the Fid) to improve speed; when S1 is degraded positioning: the ROI is automatically expanded (e.g., expanded to 2 to 3 times the original size), and allows for a more conservative subsequent reshoot strategy. 3) Board surface image acquisition (multi-exposure and polarization switching) and de-reflection reference image generation are acquired in the same field of view: multi-exposure sequence: at least 3 frames (short / medium / long exposure) to avoid overexposure in highly reflective areas and loss of details in dark areas; polarization switching sequence: at least 2 frames (polarization directions differ by about 90°) to reduce specular reflection components; synchronous acquisition of three-dimensional height data: to obtain a height map (including the "height step" of the cover edge). The following steps are then performed: Multi-exposure fusion: generating a high dynamic range image to provide usable contrast for the cover plate edge, tape texture, and ink area simultaneously; Polarization de-reflection: suppressing specular reflection based on the difference between the two polarization images to obtain a "suppressed reflection reference image"; Spatiotemporal registration: aligning the reference image and height map to the same pixel grid according to the calibration relationship. 4) Joint feature map construction (for input to HG-BCNet) In this embodiment, the joint feature map contains at least the following channels: suppressed reflection reference image (can be 3 channels); height map (1 channel); height gradient map (1 channel, calculated from the height map, highlighting the step at the cover plate edge); local flatness map (1 channel, used to determine the suction cup attachment stability and gripping feasibility). The above channels are aligned pixel by pixel and then stitched together to obtain the "joint feature map", which is then cropped into: a low-resolution input for the entire board (used to quickly lock the large area of ​​the cover plate); and a high-resolution input for the Fid prior ROI (used for fine boundary and gripping point calculation).

[0035] The improved HG-BCNet adopts a connection method of "dual-branch encoding - multi-scale gating fusion - shared decoding - multi-head output". The data flow is as follows: 1) Input layer: The joint feature map enters the network input layer and is divided into two paths according to the channel: Image path: input is the anti-reflective reference image; Height path: input is the height map, height gradient map, and flatness map. 2) Dual-branch encoder (parallel): Image encoding branch: consists of 4 scale cascaded encoding blocks (each encoding block contains convolution / normalization / activation and downsampling), outputting 4 levels of multi-scale image features. Height encoding branch: the structure is symmetrical with the image branch, but the input is the height-related channel, and it also outputs 4 levels of multi-scale height features. This allows the network to learn "reflective texture information" and "height step geometric information" separately and thoroughly, avoiding mutual contamination. 3) Height Boundary Gated Fusion Module (One layer interpolated at each scale, key connection points): At each scale, the "height gradient features" are first processed by a lightweight convolution to generate a gated weight map, which is then used to fuse two feature paths: For locations with obvious height steps: the gate weights are biased towards the height branch, making the boundary more closely resemble the real step; For locations with reliable texture: the gate weights are biased towards the image branch, making the segmentation of the inner / outer regions of the cover more complete. The fused multi-scale features enter the shared decoder. 4) Shared Decoder (U-Net-style upsampling plus skip connections): The decoder upsamples step by step according to the scale and performs skip connections and splices with the fused features of the same scale to restore spatial details and output a high-resolution fused feature map. 5) Multi-head Output (Parallel Heads): The decoded fused features simultaneously enter three output heads: Cover Segmentation Head: Outputs a cover segmentation mask (cover / non-cover pixel-level classification); Boundary Refinement Head: Outputs a boundary probability map (fine boundary probability, for boundary polygon extraction); Uncertainty Head: Outputs a pixel-level uncertainty map (higher uncertainty in reflections, contamination, and partial occlusion, used for confidence and threshold adaptation). The advantages of this connection are: the segmentation is responsible for the "region", the boundary head is responsible for the "outline", and the uncertainty is responsible for the "credibility". The three complement each other, making it particularly suitable for your "uncovering strategy selection + grasping point filtering" business loop.

[0036] Compared to the conventional U-Net / DeepLab algorithm, the height gradient adaptive dilated convolution boundary head does not use dilated convolution with a fixed dilation rate. Instead, during inference, it adaptively selects the dilation rate based on the strength of the height gradient at each pixel location: a larger dilation rate is used for strong height gradients (significant step changes at the edges of real covers), improving boundary connectivity and reducing breaks; a smaller dilation rate is used for weak height gradients (transparent films, weak boundaries, or noisy regions), avoiding excessive smoothing and drift of the boundary. The dilation rate is an integer from 1 to 8, and the adaptation of "height gradient → dilation rate" is achieved through a preset piecewise mapping or linear mapping. The effect is better in scenarios with strong reflections, complex tape textures, and discontinuous boundaries, resulting in a more continuous boundary probability map and significantly reduced boundary drift, thus making the polygon boundaries more stable and the grasping points more reliable.

[0037] This application employs a boundary-height step consistency constraint in its network. During training or online self-calibration, the spatial consistency between the boundary head output and the height gradient is used as a constraint: the boundary should fall more frequently at locations with significant height steps. This significantly suppresses false boundaries caused by "reflective false edges" and "printed texture edges," making the boundaries more closely match the actual cover edges and improving generalization ability across board types and lighting conditions. This application uses uncertainty head-driven adaptive post-processing to adaptively adjust the uncertainty map during the inference stage: contour simplification tolerance; candidate point screening threshold. In cases of contamination, glue residue, partial occlusion, and local overexposure, the number of false grab points is significantly reduced, avoiding cover failure or risks to the flexible area. ROI dual-resolution inference (network inference is divided into two stages: low-resolution positioning of the large cover area across the entire board, followed by high-resolution fine segmentation and boundary refinement of the FID prior ROI). This reduces the overall computational load without sacrificing boundary accuracy, meeting SMT cycle time requirements.

[0038] The implementation details the process from network output to "boundary polygons, confidence scores, and grab point sets." For cover plate boundary polygon extraction, the edges of the segmentation mask are first refined using a boundary probability map; then, the contour of the largest connected component in the cover plate region is extracted; the contour is simplified to obtain the polygon boundary (used for subsequent path planning and spatial relationship calculations). Boundary confidence is calculated (used for strategy and loop closure). Boundary confidence is derived from three types of information: the average intensity of the boundary probability map at the boundary (the brighter the boundary, the more reliable); the consistency between the boundary and the height step (the more consistent, the more reliable); and the uncertainty level at the boundary (the lower the uncertainty, the more reliable). A confidence score of 0-1 is output for cover removal strategy selection and S3 parameter adaptation. Candidate point sets can be generated by sampling on the boundary polygon at fixed steps and selecting points according to the following rules: flatness meets the requirements (more stable suction cup contact); far from critical wiring in the flexible area, weak areas of the cover film, and prohibited areas for components (safety distance meets the threshold); priority is given to corner points or locations with greater curvature (easier to peel off); boundary confidence meets the standard and uncertainty is low. A Top-K candidate point set is obtained through scoring, which serves as the attachment / gripping target points for the capping actuator.

[0039] This embodiment uses a "priority + identification" fusion method to obtain the adhesion level: the prior is to read the material and adhesive parameters of the cover sheet for the plate type / batch from the process database to obtain the initial adhesion level; the identification correction combines the characteristics of the cover sheet material (e.g., transparent film / tape / rigid cover sheet), boundary confidence, flatness distribution, etc., to correct the adhesion level (e.g., predicting it as low / medium / high). The peeling strategy selection rules are as follows: high boundary confidence and low adhesion level: corner peeling is used, faster speed, larger angle window; medium boundary confidence and medium adhesion level: "partial lifting and peeling along the boundary path" is used, medium speed; low boundary confidence or high adhesion level: a conservative strategy is used (first pre-lift at a small angle, change candidate points if necessary, increase re-exposure confirmation), and a stricter force / angle window is set for S3. This invention utilizes a high-step gradient to suppress reflective false edges, resulting in a more stable polygonal cover plate boundary. Uncertainty-driven candidate point screening significantly reduces false adsorption / pinching. Boundary confidence and adhesion level directly drive the capping strategy and S3 closed-loop parameters, improving the first-time success rate. ROI dual-resolution inference reduces computational load and meets SMT cycle time. Grab points avoid weak flexible areas / dense wiring areas, reducing the risk of tearing and adhesive residue contamination.

[0040] S3: Controls the cover-opening actuator to perform attaching, pulling, peeling, and lifting actions on the cover sheet, and performs closed-loop control based on peeling force, peeling displacement, and peeling angle during the action; In some embodiments, in this example, the peeling actuator is mounted above the alignment station and consists of a three-axis linear module, a rotary module, an end effector, a force sensor, and a controller. The three-axis linear module is used to move within the plane of the plate and achieve vertical approach; the rotary module is used to adjust the peeling angle; the end effector has a switchable structure, including at least one of a vacuum suction head and a micro-gripper. The force sensor is located between the end effector and the rotary module to collect the peeling force in real time; the displacement is obtained by an encoder; the peeling angle is obtained by a rotary encoder or an angle sensor. The controller is connected to the S2 module to receive the set of grabbable candidate points, boundary confidence, and adhesion level, and sets the initial control parameters accordingly.

[0041] (1) Initial parameter settings (adaptively given by S2): The controller generates the process parameters for this peeling based on the boundary confidence and adhesion level, including: upper limit of adhesion pressure, vacuum build-up time, pre-pull-up displacement, peel angle window, peel speed curve, peel force threshold and abrupt change criterion, micro-retraction displacement, etc. Example strategy: High boundary confidence and low adhesion level: allow higher peel speed and wider angle window; Low boundary confidence or high adhesion level: adopt conservative speed, narrower angle window, and reduce adhesion pressure and increase the number of re-application confirmations.

[0042] (2) Adhesion stage (positioning and safe attachment of candidate points) The capping actuator moves to the highest-scoring candidate point in the output of S2 and performs the following actions: 1) Descent approach: The end approaches the cover surface at low speed and stops descending when the displacement reaches the preset height or the contact force reaches the minimum threshold; 2) Establish attachment: If it is an adsorption head, the vacuum is turned on and maintained for a preset time to ensure adsorption stability; if it is a gripper, the edge of the cover or the reserved pull-up area is clamped with a micro-clamping force, and the maximum clamping force is limited to avoid damaging the flexible area; 3) Attachment verification: The vacuum negative pressure or gripper displacement feedback is read. If the attachment verification fails, the next point in the candidate point set is switched and the attachment stage is repeated until success or the maximum number of attempts is reached.

[0043] (3) Pre-pull-up stage (small displacement test pull and adhesion confirmation): After successful adhesion, the actuator pulls the cover plate upward with a small displacement to form the initial lifting angle. The purpose of this stage is to: verify whether the cover plate is truly gripped; judge whether the adhesion strength is consistent with the S2 estimate by the force and displacement response; and form a controllable peeling edge to avoid tearing caused by direct forceful peeling. The controller monitors the peeling force curve in real time: if the peeling force is always very small and the vacuum is stable within the pre-pull-up displacement, the peeling is determined to be successful; if the peeling force increases rapidly or fluctuates abnormally, it is determined that the adhesion is strong or there is a risk of re-adhesion, and the subsequent peeling speed is automatically reduced and the angle window is narrowed.

[0044] (4) Peeling / Lifting Stage (Force-Displacement-Angle Three-Loop Collaborative Closed Loop): After successful pre-lifting, the main peeling / lifting stage begins. The controller performs collaborative control of "target angle window constraint + target force range constraint + speed curve constraint": 1) Angle control: The rotating module controls the peeling angle within the preset window to avoid excessive peeling force due to a small angle, and also to avoid instantaneous popping or tearing of the cover sheet due to an excessive angle; 2) Speed ​​control: Advance along the peeling path according to the speed curve. The path can be a local segment of the polygonal boundary of the cover sheet or a preset straight path; 3) Force control: Read the peeling force in real time. When the peeling force approaches the upper limit threshold, the advancing speed is automatically reduced and / or the peeling angle is reduced. When the peeling force is lower than the lower limit threshold and the peeling progress is slow, the speed can be appropriately increased to improve the cycle time; 4) Displacement verification: Based on displacement and angle feedback, it is judged whether the peeling is progressing as expected. If the phenomenon of "displacement increases but peeling length does not increase" occurs, it is determined that back-adhesion or slippage may occur, triggering stabilization control.

[0045] (5) Peel force mutation and stabilization control: During the peeling process, when a sudden increase in peel force is detected within a short period of time (exceeding the mutation threshold or the rate of increase threshold), the controller immediately executes stabilization control, which includes at least one or a combination of the following actions: speed suppression: reducing the advance speed to a conservative value; angle reversal: reversing the peel angle to an angle more conducive to reducing peel force; micro-retraction displacement: retracting a small displacement in the opposite direction of peeling to release stress, and then continuing peeling at a lower speed; micro-vibration debonding: applying a small periodic displacement disturbance to the end effector to reduce the peak adhesion force. The above stabilization control can suppress stress peaks without stopping the production line, reducing the risks of flexible area stretching, cover film tearing, and adhesive residue dragging.

[0046] (6) Peeling in place and maintaining the lifted state: After peeling / lifting reaches the preset length or preset area (at least ensuring that there is no obstruction above the target Fid prior area), the actuator maintains the lifted state and locks the angle and displacement to avoid rebound and re-adhesion. The controller outputs "exposed area range, current lifted angle, current displacement and peeling force stability" as status variables to S4 for Fid exposure confirmation and availability determination.

[0047] (7) Abnormal handling and degradation strategy: If multiple attachment failures occur: switch the end effector type (switch the suction head to a gripper or change the suction cup diameter), and reselect the point from the candidate point set; if the peeling force continues to exceed the limit: stop peeling, maintain the current lifted state and request S4 to re-examine whether the Fid is sufficiently exposed; if it is still not exposed, trigger a second micro-lifting or change the peeling starting point; if a risk of cover tearing or adhesive residue is detected (e.g., abnormal force fluctuation and abnormal displacement feedback): immediately terminate peeling and alarm record to protect the board and equipment. Through the above embodiments, S3 can achieve controllable execution of the lifting action under different cover materials, different adhesion strengths and reflective interference conditions, and significantly improve the success rate and safety of the first lifting by using closed-loop collaborative control of peeling force, displacement and angle, providing stable preconditions for the subsequent Fid exposure confirmation in S4 and the precise alignment calculation in S5.

[0048] S4: After uncovering, the target Fid is exposed and its availability is determined, and a set of usable Fids is generated; in some embodiments, S4 is the exposure and availability determination of the target Fid (generating a set of usable Fids).

[0049] In some embodiments, after the cover is opened / lifted in step S3, the cover-opening actuator maintains the lifting angle and displacement unchanged to prevent the cover from springing back and sticking. Based on the coarse positioning coordinate system established in step S1 and the target Fid prior area determined in step S2, the controller drives the camera to re-photograph the target Fid area, and determines the exposure status and availability of the target Fid, generating a set of available Fids for subsequent step S5 to calculate the alignment compensation parameters.

[0050] (1) Repeat image acquisition and adaptive imaging: The system acquires repeat images within the target Fid prior area. Repeat acquisition includes at least: 1) a standard exposure image; 2) adaptive repeat acquisition triggered when needed. The adaptive repeat acquisition includes at least one of the following: adjusting exposure time, switching polarization state, changing the incident angle of illumination (switching between coaxial light and low-angle ring light), and adding a short exposure frame to suppress overexposure and reflection. In this embodiment, whether adaptive repeat acquisition is triggered is determined by two types of indicators: firstly, whether the overexposure ratio, contrast, and sharpness of the Fid area in the repeat image meet the standards; secondly, whether the average value of the uncertainty map output in step S2 within the Fid prior area is higher than a threshold. When the indicators do not meet the standards, the imaging parameters are automatically switched and re-acquired to improve the stability of subsequent detection.

[0051] (2) Exposure Confirmation (Occlusion Rate Calculation): The system uses the improved HG-BCNet output results from step S2 to determine the occlusion of the Fid region: First, the mask portion overlapping with the prior Fid region is extracted from the cover plate segmentation mask to obtain the "coverage range of the cover plate in the Fid region"; then, the boundary probability map is combined to determine whether the edge of the cover plate crosses the critical Fid region; if the system is equipped with three-dimensional height data, it further checks whether there is still a "height protrusion / covering layer" in the Fid region to eliminate misjudgments caused by transparent film residue or local re-adhesion. The Fid occlusion rate is calculated based on the above information and compared with a preset threshold: if the occlusion rate is lower than the threshold, it is determined as "exposed"; if the occlusion rate is higher than the threshold, it is determined as "not sufficiently exposed".

[0052] (3) FID Detection (Center Point / Structure Localization and Detection Uncertainty Output): When the image is determined to be "exposed" or "basically exposed," the system performs FID detection in the re-shot image: Locates the candidate FID region within the prior area of ​​the target FID (preferably combining template position and local contrast peaks for rapid localization); performs FID center point detection within the candidate region and extracts key geometric features according to the FID type (circular, cross, ring, or square marker); outputs the FID center point coordinates and detection uncertainty. Detection uncertainty can be obtained by combining the following factors: candidate region matching consistency, key point localization stability, center point deviation under multiple imaging, and the statistical value of the pixel-level uncertainty map in the FID region. The smaller the detection uncertainty, the higher the reliability of the FID for alignment.

[0053] (4) Usability Determination: To avoid the situation of "displayed but unusable", this embodiment performs a quality score on the FID. The score consists of at least the following dimensions: 1) Contrast index: grayscale difference between FID and background, edge gradient intensity; 2) Sharpness index: edge sharpness, blur degree (e.g., evaluated by high-frequency energy or edge width); 3) Reflectivity index: overexposed area ratio, specular reflection index (brightness instability at the same location under different polarization / exposure); 4) Contamination / Defect index: stains, scratches, glue residue obstruction, structural incompleteness caused by local gaps; 5) Consistency index: similarity between the geometry of the FID and the expected template. The quality score and detection uncertainty are used together for usability determination: when the quality score reaches the threshold and the detection uncertainty is lower than the threshold, the FID is determined to be "usable"; otherwise, it is determined to be "unusable".

[0054] (5) Generation and weighting of available Fid sets: The system adds Fids that pass the availability determination to the available Fid set. If there are multiple Fids on a board (e.g., multiple reference points located in the rigid and flexible areas), the system outputs the following for each available Fid: Fid center point coordinates; occlusion rate; quality score; detection uncertainty; and corresponding image evidence location (image cropping index or coordinate range of the Fid region). It also provides a basis for weighting in subsequent S5: Fids with high quality scores and low uncertainty have higher weights; Fids with low quality scores or high uncertainty have lower weights and are directly removed if necessary.

[0055] (6) Secondary micro-lifting trigger and re-examination closed loop: The system triggers secondary micro-lifting and enters re-examination closed loop when any of the following situations occur: the FID occlusion rate exceeds the threshold and is judged as not exposed; the occlusion rate meets the standard but the quality score is insufficient or the reflectivity index is abnormal; the detection uncertainty is too large and it still cannot be stabilized after multiple adaptive re-examinations. The execution method of secondary micro-lifting is as follows: under the premise of keeping the clamping positioning unchanged, the lifting execution mechanism is controlled to make a small adjustment to the lifting angle or peeling displacement, or switch to the next point in the candidate point set to start peeling again. After the secondary micro-lifting is completed, the re-examination and judgment of this step are repeated until a set of usable FIDs that meet the conditions is obtained or the maximum number of attempts is reached, and then the system enters the degradation mode (e.g., using positioning holes / shape alignment and prompting manual processing).

[0056] This embodiment simultaneously records evidence chain data when generating the set of usable Fids, including: re-image index, Fid prior region range, occlusion rate, quality score, detection uncertainty, usability determination result, and reason label for triggering adaptive re-image or secondary micro-uncovering, for subsequent quality traceability and process optimization. Through the above embodiment, S4 can reliably confirm whether the Fid is truly exposed and usable for alignment under complex conditions such as reflection, contamination, partial occlusion, or transparent film residue. At the same time, it provides robust input to S5 in the form of a set of usable Fids and their scores / uncertainties, improving the stability and accuracy of subsequent alignment compensation calculations.

[0057] S5: Based on the available Fid set, calculate the alignment compensation parameters and generate partition compensation or nonlinear compensation for the deformation of the flexible region; In some embodiments, step S5 calculates alignment compensation parameters based on the available Fid set and generates flexible area deformation compensation. After completing step S4, the system obtains the available Fid set. Each available Fid includes at least: Fid center point coordinates, occlusion rate, quality score, detection uncertainty, and corresponding image evidence location. The controller performs alignment compensation calculation on the industrial computer, outputs alignment compensation parameters for use by SMT equipment, and generates a partition compensation table or nonlinear compensation mapping for local deformation of the flexible area of ​​the rigid-flex board.

[0058] (1) Preprocessing and weighting of available FID sets: The system first performs a consistency check on available FIDs: checking whether the relative geometric relationship of each FID is consistent with the process template, and removing abnormal FIDs that deviate significantly; checking whether there is high uncertainty caused by strong reflection or contamination in the area where the FID is located, and reducing its weight if necessary. Then, the weight is calculated for each FID. The weight is determined by a combination of quality score, detection uncertainty, and occlusion rate: the higher the quality score, the smaller the detection uncertainty, and the lower the occlusion rate, the higher the weight; conversely, the lower the weight. The system also allows setting "forced rejection" rules, such as directly excluding the FID from the calculation when the occlusion rate exceeds the upper limit or the uncertainty exceeds the upper limit.

[0059] (2) Robust calculation of global alignment compensation parameters (rigid body / similar transformation): The system reads the nominal coordinates of each Fid from the process template and establishes a correspondence with the measured center point coordinates of the Fid. Weighted robust estimation is used for global calculation: First, all available Fids are used for initial fitting to obtain preliminary translation and rotation compensation; then the fitting residual of each Fid is calculated. If the residual of some Fids is very large, their weight is reduced or they are removed according to the robust rule; a second fitting is performed to obtain stable global compensation parameters. The output global compensation parameters include at least: translation compensation in the X direction, translation compensation in the Y direction, and rotation compensation; when needed, magnification compensation or shear compensation can be output to compensate for camera calibration error, thermal expansion of the carrier, or overall expansion and contraction of the plate.

[0060] (3) Deformation identification of flexible area (determining whether to use partitioning or nonlinearity): In order to determine whether deformation compensation of flexible area is needed, the system performs deformation discrimination: compare the residual distribution of each Fid. If the residual shows a significant difference between the rigid area and the flexible area or shows a consistent offset trend in the flexible area, it is determined that there is local deformation. Combine the three-dimensional height data (warping / springback characteristics) to evaluate the deformation risk level of the flexible area. If it only shows a linear offset trend in a local area, partitioning compensation is preferred. If it shows bending, torsion or multiple inconsistent trends, nonlinear compensation is selected.

[0061] (4) Partition Compensation Implementation Example (Partition Affine / Local Rigid Body): When partition compensation is used, the system divides the board surface into several sub-regions, including at least a rigid region and a flexible region. Within the flexible region, it can be further divided into multiple sub-regions according to the key mounting area, connector area, and bending area. The partitioning criteria can be: the predefined region boundaries in the process template; the warping abrupt change zone in the three-dimensional height data; or automatic partitioning based on the spatial proximity relationship between the Fid and the key mounting area. For each sub-region, the system selects the available Fid within or near the region as constraint points, and calculates the local translation and rotation compensation of the sub-region respectively. When the Fid of a certain sub-region is insufficient, the Fid of the neighboring region is used for extrapolation and the confidence level is reduced. The final output is a "Partition Compensation Table", which includes at least: sub-region ID, sub-region boundary range, local compensation parameters of the region, and region confidence level. When the pick-and-place machine / printer is running, it calls the corresponding compensation parameters according to the region where the current processing coordinates fall, and realizes the local correction of the flexible region.

[0062] (5) Nonlinear Compensation Implementation Example (Thin Plate Spline Mapping): When nonlinear compensation is used, the system constructs a nonlinear mapping from template coordinates to measured coordinates. The sources of control points for the mapping include: available FID center points; key texture points or structural feature points in the flexible area (optional, obtained by image feature tracking); and virtual control points supplemented at the boundary of the flexible area when necessary (generated by height field and shape constraints). The system fits the thin plate spline mapping with control points as constraints, so that the mapping can be accurately matched at the control points, while maintaining smoothness and continuity throughout the entire area, avoiding abrupt changes introduced by compensation. The output is "nonlinear compensation mapping", and its data format can be: a displacement field discretized in a grid manner (e.g., outputting the compensation amount of each grid node according to a fixed grid spacing); or output in the form of a set of control points + interpolation parameters (interpolated in real time by the device or controller). When the SMT equipment mounts / prints at a certain coordinate point, the controller first obtains the initial coordinates according to global compensation, and then performs local correction on the point through nonlinear mapping, thereby offsetting the spatial nonlinear deviation caused by the bending and springback of the flexible area.

[0063] (6) Output alignment compensation parameters and evidence chain (for S6 re-shooting and loop closure). The final output data in this embodiment includes: global compensation parameters (translation, rotation, and optional magnification / shearing); partition compensation table or nonlinear compensation mapping; weight, residual, quality score, and uncertainty of each Fid; deformation risk level and selected compensation mode (partition / nonlinear / no compensation). At the same time, the above outputs are bound to the Fid image evidence position to form an evidence chain, which is used to quickly locate "whether the source of error is Fid instability, improper uncovering, or deformation of the flexible area" when recalculating or refitting after step S6 re-shooting verification.

[0064] By using "available FID weighting + robust solution", the impact of abnormal FID caused by reflection, contamination or weak exposure on global compensation can be reduced; by using "zonal compensation or thin plate spline nonlinear compensation", spatial correction can be made for local springback, expansion and torsion in the flexible area of ​​the rigid-flex board, so as to significantly reduce the residual in the key mounting area.

[0065] S6: Send the alignment compensation parameters to the SMT equipment and perform a re-scan verification, recalculate and perform a second micro-uncovering based on the re-scan residual trigger parameters.

[0066] In some embodiments, step S5 outputs global alignment compensation parameters, a partitioned compensation table, and / or a nonlinear compensation mapping. The controller sends the compensation parameters to the SMT equipment (including solder paste printers, pick-and-place machines, or dispensing machines) via a communication interface and triggers the equipment to perform a repeat check. The controller performs closed-loop decision-making based on the amplitude and spatial distribution of the repeat residual, triggering parameter recalculation, deformation compensation refitting, and / or secondary micro-uncovering, thereby ensuring alignment accuracy and cycle stability.

[0067] (1) Compensation parameter distribution and equipment-side activation method: The controller and SMT equipment controller are connected via Ethernet communication, and parameter distribution is achieved using the equipment open interface or standard industrial protocol. The distributed content includes at least: global compensation parameters: used to correct the translation and rotation from board coordinates to equipment coordinates; partition compensation table: containing sub-region boundaries and region compensation parameters (if partition compensation is used); nonlinear compensation mapping: containing grid displacement field or control point interpolation parameters (if nonlinear compensation is used); parameter version number and timestamp: used for traceability and to prevent accidental overwriting. After receiving the parameters, the equipment side executes according to the preset activation strategy: first apply global compensation to the entire board, and then apply partition or nonlinear local correction to the coordinates falling into the flexible area or critical mounting area. The controller also sends a re-scan verification command to the equipment, specifying the re-scan points, re-scan order and imaging parameter suggestions.

[0068] (2) Setting up verification points for repeat testing (for residual calculation): This embodiment adopts a combination of "baseline Fid repeat testing + critical area repeat testing": Baseline Fid repeat testing: Repeat testing is performed on at least two available Fids to verify whether the overall translation and rotation are correct; Critical area repeat testing: Several verification points are set in the critical mounting area or bending sensitive area of ​​the flexible area to verify the zoning or nonlinear compensation effect. The repeat testing points can be preset by the process template, or the controller can dynamically increase the number of points according to the deformation risk level of S5: the higher the risk, the more repeat testing points and the larger the coverage area.

[0069] (3) Residual acquisition and quality verification: After applying compensation parameters, the SMT equipment performs re-sampling and outputs the detection results of each re-sampling point, including: the coordinates of the detected mark center, the detection confidence level, and the deviation from the expected position. The controller receives the above results and performs secondary verification: when the equipment reports low confidence or poor image quality, the controller requires the equipment to switch exposure / light source / polarization or repeat sampling; when the confidence level is stable, the deviation is written as the re-sampling residual into the residual list and associated with the area to which the point belongs (rigid area / flexible area / specific sub-area).

[0070] (4) Residual judgment rules: The controller makes two-level judgments on the residuals: 1) Amplitude judgment: If the residual of any repeat point exceeds the first threshold, it is judged that there is an abnormality in the alignment; if it exceeds the second threshold, it is judged as a serious abnormality and triggers the protection strategy (stopping production or entering the manual confirmation process). 2) Spatial distribution judgment: The residuals are aggregated and analyzed by region: If the residuals of both the rigid area and the flexible area are large and the direction is consistent, it is given priority to judge that the global compensation is inaccurate; if the residual of the rigid area is small and the residual of the flexible area is non-uniformly distributed, or a local residual cluster is formed in the flexible area, it is given priority to judge that the deformation compensation of the flexible area is insufficient or the local deformation state changes; if the residuals of some points are suddenly large and the corresponding point position confidence is low, it is given priority to judge that the FID recognition is unstable or the detection drift is caused by local reflection / contamination.

[0071] (5) Trigger parameter recalculation and deformation compensation refit (corresponding to S5 closed loop). When the global compensation is determined to be abnormal, the controller triggers parameter recalculation: re-call the robust solution process of S5, re-filter and re-weight the available Fid set and recalculate the global compensation parameters; if it is found that the residual of individual Fid is abnormally large, the weight of the Fid is reduced or removed, and the solution is re-solved. When the residual of the flexible area is determined to be clustered or non-uniformly distributed, the controller triggers deformation compensation refit: if partition compensation is used: automatically adjust the partition boundary or increase the number of sub-regions of the flexible area, and recalculate the local compensation parameters of each region; if nonlinear compensation is used: increase the control point density of the flexible area or reduce the grid spacing, and refit the nonlinear compensation mapping; after refitting, the updated parameters are reissued and the re-sampling verification is performed again until the residual meets the threshold or reaches the maximum number of closed loops.

[0072] (6) Conditions and execution method for triggering secondary micro-uncovering (linked with S3 / S4): When the residual abnormally exhibits one of the following characteristics, the controller will prioritize triggering secondary micro-uncovering instead of simply recalculating: the re-examination points are concentrated near the target Fid and the residual fluctuates greatly over time, indicating that the cover may re-adhere or the edge of the cover may be shaking; the confidence of the equipment re-examination is consistently low, and the occlusion rate or reflectivity recorded in S4 is close to the threshold; there are signs of the cover edge intruding into the Fid area in the re-examination image. The execution method for secondary micro-uncovering is as follows: without changing the plate clamping position, the uncovering execution mechanism is controlled to slightly increase the lifting angle or slightly advance the peeling displacement, or switch to the next point in the candidate point set to start peeling again; then immediately return to S4 for exposure confirmation and availability determination, and if necessary, update the available Fid set and execute S5 and this step S6 again.

[0073] (7) Degradation strategy and evidence chain output: When the number of closed loops reaches the upper limit and the residual cannot be reduced to within the threshold, the system enters the degradation strategy: use positioning holes / shape alignment and limit the mounting range, or prompt manual handling of cover plates and contamination; mark abnormal boards and record batch information to avoid mixing in good products.

[0074] Meanwhile, this embodiment outputs a complete chain of evidence, including at least: compensation parameter version number, list of retake points, residuals and confidence levels at each point, reason labels for triggering recalculation / refitting / secondary micro-uncovering, and corresponding image location indexes, for quality traceability and process optimization.

[0075] Preferably, in step S2, the acquired panel images include a multi-exposure image sequence and a polarization state switching image sequence under the same field of view; the multi-exposure image sequence is HDR fused and the polarization state switching image sequence is despectralized to obtain a reference image with reduced reflection; the reference image is spatiotemporally registered with the three-dimensional height data to construct a joint feature map; on the joint feature map, an improved HG-BCNet height-guided boundary consistency segmentation neural network is used to perform height-guided cover semantic segmentation and boundary refinement to obtain the cover boundary polygon, boundary confidence, and a set of grabbable candidate points; and the uncovering strategy is selected based on the spatial relationship between the cover boundary polygon and the target FID prior region, the adhesion level, and the boundary confidence.

[0076] Preferably, highly guided cover semantic segmentation and boundary refinement are performed on the joint feature map by the improved HG-BCNet highly guided boundary consistent segmentation neural network, which is implemented by a highly guided segmentation model. The improved HG-BCNet highly guided boundary consistent segmentation neural network includes: An image coding branch is used to extract multi-scale image features from the reference image for suppressing reflection; The height encoding branch is used to extract multi-scale height features from the three-dimensional height data and height gradient features; The height boundary gating fusion module is used to generate gating weights based on the height gradient and perform weighted fusion of the multi-scale image features and multi-scale height features. The multi-head output structure outputs at least the cover plate segmentation mask, the boundary probability map, and the pixel-level uncertainty map; The boundary probability map is used to refine the edges of the cover plate segmentation mask, and the uncertainty map is used to form the boundary confidence and participate in the screening of the set of grabbable candidate points.

[0077] Preferably, the boundary refinement head of the improved HG-BCNet height-guided boundary consistency segmentation neural network includes a height gradient-adaptive dilated convolution structure, whose dilation rate adaptively varies with the height gradient magnitude at the pixel location: the larger the height gradient magnitude, the larger the dilation rate is used to enhance boundary continuity; the smaller the height gradient magnitude, the smaller the dilation rate is used to avoid excessive boundary smoothing and drift; wherein, the dilation rate is an integer between 1 and 8, and the adaptive selection from height gradient to dilation rate is achieved through a preset piecewise mapping or linear mapping.

[0078] In some embodiments, a further structural improvement embodiment of the improved HG-BCNet height-guided boundary consistent segmentation neural network, based on the above-mentioned improved HG-BCNet height-guided boundary consistent segmentation neural network, in order to further improve the positioning accuracy of cover plate boundaries, the ability to suppress reflective false edges, and the robustness to small registration errors of image-height data in the rigid-flex PCB scenario, this embodiment further inserts a Height-Conditioned Deformable Alignment Layer (HC-DCAL) before the height boundary gated fusion module, so as to realize dynamic alignment of image features and height features before fusion. The improved HG-BCNet height-guided boundary consistency segmentation neural network sets a height-conditional deformable alignment layer before the height boundary gated fusion module. The height-conditional deformable alignment layer generates an offset field and modulation weights with height gradient features as conditional input, and performs deformable convolution alignment on the multi-scale image features output by the image coding branch based on the offset field and the modulation weights to obtain aligned image features. The height boundary gated fusion module performs weighted fusion on the aligned image features and the multi-scale height features output by the height coding branch to improve the consistency between the cover plate boundary and the height step and reduce reflective false edge interference.

[0079] (1) The part of the network structure that remains unchanged still includes the following: image coding branch, height coding branch, height boundary gating fusion module, shared decoder and multi-head output structure; wherein, the multi-head output structure outputs at least a cover segmentation mask, a boundary probability map and a pixel-level uncertainty map; the boundary probability map is used to refine the edge of the cover segmentation mask, and the uncertainty map is used to form the boundary confidence and participate in the screening of the set of grabbable candidate points; and the boundary refinement head can still adopt a height gradient adaptive dilated convolution structure, with the dilation rate between 1 and 8 adaptively selected according to the height gradient magnitude to enhance boundary continuity and suppress boundary drift.

[0080] (2) The insertion position and connection method of the newly added height-conditionally deformable alignment layer, wherein the height-conditionally deformable alignment layer is set between the "multi-scale image features output by the image coding branch" and the "height boundary gating fusion module", and preferably a set of height-conditionally deformable alignment layers is set in each scale feature layer. The connection method is as follows: 1) The image coding branch outputs the first multi-scale image features at each scale; 2) The height coding branch outputs the second multi-scale height features at each scale and outputs the height gradient features; 3) The first multi-scale image features and the height gradient features are input together into the height-conditionally deformable alignment layer to obtain the aligned image features; 4) The aligned image features and the second multi-scale height features are input into the height boundary gating fusion module for weighted fusion to obtain the fused features; 5) The fused features are input into the shared decoder, and the multi-head output structure outputs the cover segmentation mask, the boundary probability map and the pixel-level uncertainty map.

[0081] (3) The specific structural composition of the height-condition deformable alignment layer, which includes at least an offset prediction sub-layer, a modulation weight sub-layer and a deformable alignment sub-layer: (a) Offset prediction sub-layer: with height gradient features as input, it passes through at least one convolutional layer to obtain an offset field, which is used to characterize the alignment displacement direction and displacement magnitude of image features at different spatial locations; (b) Modulation weight sub-layer: with the splicing features of height gradient features and first multi-scale image features as input, it passes through at least one convolutional layer to obtain a modulation weight map, which is used to characterize the reliability of offset alignment at different spatial locations; (c) Deformable alignment sub-layer: according to the offset field and modulation weight map, deformable convolution sampling is performed on the first multi-scale image features to obtain aligned image features, and output to the subsequent height boundary gating fusion module.

[0082] With the above structure, image features can be dynamically aligned with the true height step boundary indicated by the height gradient before entering gating fusion, avoiding misalignment of image edges and height edges due to local warping, parallax, or calibration errors.

[0083] In the scenario of opening and closing rigid-flex PCBs, the cover plate boundary often corresponds to a step change in the 3D height data. However, the image may also contain false edges caused by ink reflection, specular reflection of the cover film, tape texture, and contaminant particles. If only image features are relied upon for fusion, boundary offset, boundary breakage, or misidentification of reflective false edges as real boundaries can easily occur. On the other hand, there may be slight registration errors between the image and height data during actual assembly and adjustment, and board warping can introduce local parallax, causing the same boundary to appear slightly misaligned in the two modalities.

[0084] This embodiment introduces a height-conditionally deformable alignment layer before fusion, enabling dynamic alignment of image features guided by height gradients: when a region has a significant height gradient, the alignment layer automatically enhances the alignment displacement and increases the alignment weight, making the image features conform to the true height step boundary; when the height gradient is not significant or the region has high noise, the alignment layer reduces the alignment displacement or reduces the alignment weight, avoiding meaningless alignment perturbations. The aligned image features then enter the height boundary-gated fusion module, ensuring that the fusion process simultaneously possesses "boundary realism consistency" and "texture semantic integrity," thereby improving the overall stability of cover segmentation and boundary refinement.

[0085] Compared with structures without a height-condition deformable alignment layer, this embodiment has at least the following advantages: 1) The cover plate boundary positioning is more accurate, the boundary polygon stability is higher, and boundary breakage and drift are significantly reduced; 2) It is more robust to small registration errors of image-height data and can absorb the misalignment effects caused by calibration errors, local warping, and parallax; 3) It is less sensitive to reflective false edges and can effectively reduce missegmentation and false boundaries caused by specular reflection, ink gloss, tape texture, etc.; 4) The boundary confidence generated based on the boundary probability map and uncertainty map is more reliable, which reduces the misselection rate of the candidate point set that can be grasped, thereby improving the success rate of subsequent cover removal and peeling, reducing the number of secondary micro-cover removal triggers, and improving the stability of the production line cycle.

[0086] Preferably, the improved HG-BCNet height-guided boundary consistency segmentation neural network introduces a boundary height step consistency constraint during training or online self-calibration, so that the boundary probability map output by the model is consistent with the height gradient feature in spatial location; and during the inference stage, it adaptively adjusts the simplification tolerance of the boundary polygon fitting and the threshold for selecting grabbillable candidate points based on the pixel-level uncertainty map, so as to reduce missegmentation and misgrabbing under conditions of reflection, pollution, and partial occlusion.

[0087] Preferably, step S3 includes adaptively setting the peel angle window and speed curve based on the confidence level of the cover sheet boundary and the adhesion level, and performing stabilization control of speed suppression, angle correction and micro-retreat displacement when a sudden change in peel force is detected.

[0088] Preferably, step S4 includes acquiring a repeat image of the target FID region and performing FID detection in conjunction with three-dimensional height data. The FID detection includes: locating candidate FID regions on a reference image with reduced reflectivity; detecting the FID center point within the candidate regions and outputting the detection uncertainty; calculating the FID occlusion rate based on the cover film semantic segmentation results; and generating a FID quality score based on contrast, sharpness, specular reflection index, and contamination features. When the occlusion rate or quality score does not meet the threshold, a secondary micro-uncovering is triggered and adaptive repeating is triggered. The adaptive repeating includes adaptive adjustment of exposure, polarization state, and incident light angle.

[0089] Preferably, step S5 includes: assigning weights to each Fid based on its detection uncertainty and quality score; obtaining global rigid body transformation parameters from the plate coordinate system to the device coordinate system using weighted robust estimation; and fitting the local deformation of the flexible region using a partitioned affine model and a thin plate spline nonlinear model under the constraint of a three-dimensional height field, and outputting a partitioned compensation table and a nonlinear compensation mapping.

[0090] Preferably, step S6 includes: when the reshoot residual is non-uniformly distributed in the flexible area and exceeds a threshold, the refitting of the partition compensation table or nonlinear compensation mapping is triggered, and an evidence chain corresponding to the compensation parameters is output. The evidence chain includes at least the available FID set, the quality score of each FID, the occlusion rate, the reshoot residual and its corresponding image position.

[0091] This application also provides an SMT alignment device for exposing the FID (Fid) by peeling up the cover of a rigid-flex PCB, such as... Figure 2As shown, the device includes: an SMT alignment device for revealing the FID (Find Indicator) on the rigid-flex PCB, preferably comprising: a frame and positioning platform assembly; a frame, a positioning platform (vacuum adsorption stage or support stage); a board edge clamping mechanism (side grippers / front stops / positioning pins); and a conveying mechanism (belt / chain / track). A two-dimensional imaging assembly (board surface image acquisition) includes a top industrial camera (which can be an area scan camera) and a lens (preferably a telecentric lens); multiple light source units: coaxial light, ring light, low-angle side light; a polarization switching assembly: a camera-end polarizer and a light source-end polarizer, or an electronically controlled polarization wheel / liquid crystal polarizer, used to acquire images with polarization state switching. A three-dimensional height acquisition assembly includes a structured light three-dimensional camera or a laser triangulation profilometer (which can be a line scan profilometer + motion axis), used to output three-dimensional height data of the board surface; and a calibration fixture between the three-dimensional sensor and the two-dimensional camera (for one-time calibration). The peeling actuator assembly includes a multi-axis motion mechanism: at least an XY translation axis and a Z lifting axis, preferably also including a rotary axis (for peeling angle control); an end effector: a vacuum suction head and / or micro-grippers (switchable / replaceable specifications); a vacuum generator / solenoid valve / vacuum pressure sensor (for suction status detection); a force sensor (installed between the end effector and the motion axis) for acquiring peeling force; displacement feedback: a motor encoder / linear grating ruler; angle feedback: a rotary encoder. The computing and control components (core) include an industrial computer / edge computing unit (including a GPU or AI accelerator card, running HG-BCNet inference and data processing); a motion controller (PLC / motion control card) for controlling axis motion and closed-loop control; a data acquisition module (acquiring force, displacement, angle, vacuum pressure, etc.); and a storage module (for recording evidence chains, parameter versions, and logs). Communication and interface components: Communication interfaces with SMT equipment: Industrial Ethernet (Ethernet / IP, PROFINET, Modbus TCP or device open API); Interfaces with cameras / 3D sensors: GigEVision / USB3 Vision / CameraLink, etc.; I / O interfaces: photoelectric sensors, emergency stop, air supply, valve island, etc.

[0092] The connection methods between the various hardware components are as follows: 1) Mechanical connection: The conveying mechanism is mounted on the frame and connected to the front and rear ends of the positioning platform, responsible for the loading and unloading of the upper plate. The positioning platform is fixed to the frame, and the clamping mechanism is installed around the positioning platform to limit and clamp the plate edge / carrier. The 2D imaging component and the 3D height acquisition component are fixedly mounted on the bracket above the positioning platform, maintaining a fixed relative position with the positioning platform. The cover-opening actuator component is mounted on the crossbeam or independent column above the positioning platform, and its end effector is located in the accessible area above the plate surface, covering the target FID prior area.

[0093] 2) Control and signal connections: The industrial PC connects to the 2D camera via the camera interface to trigger image capture and receive image data; the industrial PC connects to the polarization switching component (polarization wheel / LCD polarizer) via I / O or serial / Ethernet to switch polarization states; the industrial PC connects to the multi-light source controller to switch between coaxial light / ring light / side light and adjust brightness; the industrial PC connects to the 3D height acquisition component to trigger acquisition and receive height map data; the industrial PC connects to the motion controller to send motion trajectory and lifting strategy parameters; the motion controller connects to the servo drives of the X-axis, Y-axis, Z-axis, and rotary axis to control displacement and angle; feedback signals from force sensors, vacuum pressure sensors, encoders / grating rulers, etc., are input to the data acquisition module or motion controller to achieve closed-loop control; the motion controller connects to the vacuum generator / valve island to control the adsorption switch and adsorption pressure; the industrial PC connects to the SMT equipment controller via industrial Ethernet to send alignment compensation parameters and receive re-shooting results / residuals.

[0094] The data stream 2D ​​camera outputs a multi-exposure image sequence and a polarization-switched image sequence → the industrial control computer generates a reflective baseline image; the 3D height acquisition component outputs a height map → the industrial control computer registers the baseline image with the data stream to form a joint feature map; the industrial control computer runs HG-BCNet to output the mask / boundary / uncertainty → extracts the boundary polygon, confidence level, candidate point set and generates a capping strategy; the industrial control computer sends the capping strategy and closed-loop control parameters to the motion controller → the motion controller drives the capping actuator to perform attachment, pull-up, peeling / lifting, and performs closed-loop adjustment based on force / displacement / angle; after capping, the industrial control computer triggers a re-shoot to perform FID exposure confirmation and availability determination → outputs a set of available FIDs; the industrial control computer calculates global compensation and partition / nonlinear compensation → sends the results to the SMT equipment via Ethernet; the SMT equipment returns the residuals after re-shooting → the industrial control computer recalculates / refits and / or performs secondary micro-capping based on the residual trigger parameters.

[0095] The coordinate system establishment module performs the mounting and initial positioning of the rigid-soft bonding board, and establishes a coarse positioning coordinate system based on the positioning holes and the shape of the board. The uncovering strategy selection module acquires board surface images and 3D height data to identify the cover edge and its relative position with the target Fid. Based on the coarse positioning coordinate system, it determines the prior region of the target Fid. The board surface images and 3D height data are input into an improved HG-BCNet height-guided boundary consistency segmentation neural network to output a cover segmentation mask, a boundary probability map, and a pixel-level uncertainty map. The cover segmentation mask and the boundary probability map are used to extract the cover boundary polygon, boundary confidence, and a set of grabbable candidate points. The uncovering strategy is selected based on the spatial relationship between the cover boundary polygon and the prior region of the target Fid, the adhesion level, and the boundary confidence. The cap-opening execution module controls the cap-opening execution mechanism to perform actions such as attaching, pulling, peeling, and lifting the cap sheet, and performs closed-loop control based on peeling force, peeling displacement, and peeling angle during the action. The Fid generation module confirms the exposure and availability of the target Fid after the cover is removed, and generates a set of usable Fids. The compensation module calculates the alignment compensation parameters based on the available FID set and generates partition compensation or nonlinear compensation for the deformation of the flexible region. The secondary fine-tuning module sends the alignment compensation parameters to the SMT equipment and performs a re-examination verification, recalculating and performing a secondary micro-uncovering based on the re-examination residual trigger parameters.

[0096] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (4) This application introduces multi-exposure and polarization switching imaging before peeling off the cover, performs HDR fusion and specular destigma decomposition, and then registers it with three-dimensional height data to form a joint feature map. The improved HG-BCNet is used to realize semantic segmentation and boundary refinement of the cover sheet. Compared with methods that rely solely on single-frame visible light or traditional edge detection, the joint feature map utilizes both texture information and height step information, which can effectively suppress specular reflection and semi-transparent texture interference caused by cover film, tape, ink, etc., and reduce the probability of "false edges", "broken edges" and missegmentation. The uncertainty and boundary confidence of the network output further quantify the recognition reliability, enabling the system to automatically filter out low-confidence boundary segments and output more reliable boundary polygons and a set of grabbable candidate points, thereby reducing the selection of adsorption points, peeling failure and mis-peeling from the source, and improving the robust adaptation capability to different cover sheet materials, adhesion differences and lighting changes.

[0097] (5) This application incorporates the confidence level of the cover sheet boundary, adhesion level, and candidate point set into the capping control loop. It implements closed-loop control based on peel force, displacement, and / or angle for the entire process of attachment, lifting, and peeling / lifting. When a sudden change in peel force is detected, stabilization strategies such as speed suppression, angle correction, and micro-retraction displacement are executed. Compared with manual capping or open-loop mechanical capping, this solution can limit peak peel stress, reduce the risk of flexible area stretching, cover film tearing, adhesive residue dragging, and contamination, and improve the process consistency of rigid-flex boards under high-cycle conditions on the production line. At the same time, the system supports adaptive reshooting of exposure, polarization, and incident angle, as well as automatic triggering of secondary micro-capping, so that the FID exposure and usability can be quickly restored even in cases of partial obscuration, reflection, contamination, or re-adhesion, reducing downtime and rework.

[0098] (6) This application determines the occlusion rate and quality score of the exposed FID after the cover is removed, and assigns weights to the available FIDs based on the detection uncertainty. It uses weighted robust estimation to solve the global rigid body transformation, which improves the anti-interference ability against abnormal FIDs, local reflections and contamination. At the same time, under the constraint of three-dimensional height field, it introduces partitioned affine and thin plate spline nonlinear compensation, which can model and correct the springback, local expansion and contraction and shear deformation of the flexible area of ​​the rigid-soft bonded plate, which significantly reduces the local mounting offset caused by using only positioning holes / shapes or a single rigid body model. The re-image residual is not only used for threshold determination, but also triggers recalculation and refitting based on the spatial distribution of the residual, forming an alignment closed loop. In addition, the output includes the FID set, score, occlusion rate, residual and image position evidence chain, which facilitates quality traceability and process optimization, and achieves simultaneous improvement of accuracy, yield and maintainability.

[0099] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products, and therefore this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0100] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for SMT alignment of rigid-flex PCBs to expose the fiber (Fid), characterized in that, Including the following steps: S1: Perform the mounting and initial positioning of the rigid-flex PCB, and establish a coarse positioning coordinate system based on the positioning holes and the PCB outline; S2: Acquire board surface images and 3D height data to identify the cover sheet boundary and its relative position with the target FID. Determine the prior region of the target FID based on the coarse positioning coordinate system. Input the board surface images and 3D height data into the improved HG-BCNet height-guided boundary consistency segmentation neural network to output a cover sheet segmentation mask, boundary probability map, and pixel-level uncertainty map. Extract the cover sheet boundary polygon, boundary confidence, and set of grabbable candidate points from the cover sheet segmentation mask and boundary probability map. Select a cover removal strategy based on the spatial relationship between the cover sheet boundary polygon and the prior region of the target FID, the adhesion level, and the boundary confidence. S3: Controls the cover-opening actuator to perform attaching, pulling, peeling, and lifting actions on the cover sheet, and performs closed-loop control based on peeling force, peeling displacement, and peeling angle during the action; S4: After uncovering, the target Fid is exposed and its availability is determined, and a set of available Fids is generated; S5: Based on the available Fid set, calculate the alignment compensation parameters and generate partition compensation or nonlinear compensation for the deformation of the flexible region; S6: Send the alignment compensation parameters to the SMT equipment and perform a re-scan verification, recalculate and perform a second micro-uncovering based on the re-scan residual trigger parameters.

2. The SMT alignment method for exposing the FID (Fid) by peeling back the cover of a rigid-flex PCB according to claim 1, characterized in that, In step S2, the acquired panel images include a multi-exposure image sequence and a polarization state switching image sequence under the same field of view; HDR fusion is performed on the multi-exposure image sequence and specular destigmatization is performed on the polarization state switching image sequence to obtain a reference image with reduced reflection. The reference image and 3D height data are spatiotemporally registered to construct a joint feature map. On the joint feature map, the improved HG-BCNet height-guided boundary consistency segmentation neural network performs height-guided cover semantic segmentation and boundary refinement to obtain cover boundary polygons, boundary confidence, and a set of grabbable candidate points. The uncovering strategy is selected based on the spatial relationship between the cover boundary polygons and the target FID prior region, the adhesion level, and the boundary confidence.

3. The SMT alignment method for exposing the FID (Find) by peeling back the cover of a rigid-flex PCB according to claim 2, characterized in that, Highly guided cover semantic segmentation and boundary refinement are performed on the joint feature map by the improved HG-BCNet highly guided boundary consistent segmentation neural network. The improved HG-BCNet highly guided boundary consistent segmentation neural network includes: An image coding branch is used to extract multi-scale image features from the reference image for suppressing reflection; The height encoding branch is used to extract multi-scale height features from the three-dimensional height data and height gradient features; The height boundary gating fusion module is used to generate gating weights based on the height gradient and perform weighted fusion of the multi-scale image features and multi-scale height features. The multi-head output structure outputs at least the cover plate segmentation mask, the boundary probability map, and the pixel-level uncertainty map; The boundary probability map is used to refine the edges of the cover plate segmentation mask, and the uncertainty map is used to form the boundary confidence and participate in the screening of the set of grabbable candidate points.

4. The SMT alignment method for exposing the FID (Find) by peeling back the cover of a rigid-flex PCB according to claim 3, characterized in that, The improved HG-BCNet height-guided boundary consistency segmentation neural network's boundary refinement head includes a height gradient-adaptive dilated convolution structure. Its dilation rate adaptively varies with the height gradient magnitude at the pixel location: a larger height gradient magnitude results in a larger dilation rate to enhance boundary continuity; a smaller height gradient magnitude results in a smaller dilation rate to avoid excessive boundary smoothing and drift. The dilation rate ranges from 1 to 8, and the adaptive selection from height gradient to dilation rate is achieved through a preset piecewise mapping or linear mapping.

5. The SMT alignment method for exposing the FID (Find) by peeling back the cover of a rigid-flex PCB according to claim 4, characterized in that, The improved HG-BCNet height-guided boundary consistency segmentation neural network introduces a boundary height step consistency constraint during training or online self-calibration, ensuring that the boundary probability map output by the model is consistent with the height gradient feature in spatial location. Furthermore, during the inference phase, it adaptively adjusts the simplification tolerance of the boundary polygon fitting and the threshold for selecting grabbearing candidate points based on the pixel-level uncertainty map, thereby reducing missegmentation and misgrabbing under conditions of reflection, contamination, and partial occlusion.

6. The SMT alignment method for exposing the FID (Fid) by peeling back the cover of a rigid-flex PCB according to claim 1, characterized in that, Step S3 includes adaptively setting the peel angle window and speed curve based on the confidence level and adhesion level of the cover plate boundary, and performing speed suppression, angle correction and micro-retreat displacement stabilization control when a sudden change in peel force is detected.

7. The SMT alignment method for exposing the FID (Fid) by peeling back the cover of a rigid-flex PCB according to claim 3, characterized in that, Step S4 includes acquiring a repeat image of the target FID region and performing FID detection in conjunction with three-dimensional height data. The FID detection includes: locating the candidate FID region on the anti-reflective reference image, detecting the FID center point within the candidate region and outputting the detection uncertainty; calculating the FID occlusion rate based on the cover film semantic segmentation results, and generating a FID quality score based on contrast, sharpness, specular reflection index and contamination features; when the occlusion rate or quality score does not meet the threshold, triggering a secondary micro-uncovering and triggering adaptive repeating, the adaptive repeating including adaptive adjustment of exposure, polarization state and incident light angle.

8. The SMT alignment method for exposing the FID (Fid) by peeling back the cover of a rigid-flex PCB according to claim 8, characterized in that, Step S5 includes assigning weights to each FID based on its detection uncertainty and quality score, and using weighted robust estimation to obtain the global rigid body transformation parameters from the plate coordinate system to the device coordinate system. Under the constraint of a three-dimensional height field, the local deformation of the flexible region is fitted using a partitioned affine model and a thin-plate spline nonlinear model, and the partitioned compensation table and nonlinear compensation mapping are output.

9. The SMT alignment method for exposing the FID (Fid) by peeling back the cover of a rigid-flex PCB according to claim 1, characterized in that, Step S6 includes: when the reshoot residual is non-uniformly distributed in the flexible area and exceeds the threshold, the refit of the partition compensation table or nonlinear compensation mapping is triggered, and the evidence chain corresponding to the compensation parameters is output. The evidence chain includes at least the available FID set, the quality score of each FID, the occlusion rate, the reshoot residual and its corresponding image position.

10. An SMT alignment device for exposing the FID (Find) by removing the cover of a rigid-flex PCB, characterized in that, include: The coordinate system establishment module performs the mounting and initial positioning of the rigid-soft bonding board, and establishes a coarse positioning coordinate system based on the positioning holes and the shape of the board. The uncovering strategy selection module acquires board surface images and 3D height data to identify the cover edge and its relative position with the target Fid. Based on the coarse positioning coordinate system, it determines the prior region of the target Fid. The board surface images and 3D height data are input into an improved HG-BCNet height-guided boundary consistency segmentation neural network to output a cover segmentation mask, a boundary probability map, and a pixel-level uncertainty map. The cover segmentation mask and the boundary probability map are used to extract the cover boundary polygon, boundary confidence, and a set of grabbable candidate points. The uncovering strategy is selected based on the spatial relationship between the cover boundary polygon and the prior region of the target Fid, the adhesion level, and the boundary confidence. The cap-opening execution module controls the cap-opening execution mechanism to perform actions such as attaching, pulling, peeling, and lifting the cap sheet, and performs closed-loop control based on peeling force, peeling displacement, and peeling angle during the action. The Fid generation module confirms the exposure and availability of the target Fid after the cover is removed, and generates a set of usable Fids. The compensation module calculates the alignment compensation parameters based on the available FID set and generates partition compensation or nonlinear compensation for the deformation of the flexible region. The secondary fine-tuning module sends the alignment compensation parameters to the SMT equipment and performs a re-examination verification, recalculating and performing a secondary micro-uncovering based on the re-examination residual trigger parameters.

Images