SMT (Surface Mount Technology)-based visual correction method

Abstract

The invention relates to the technical field of surface mounting, and discloses an SMT (Surface Mount Technology)-based patch vision correction method, which comprises the following steps: S1, a calibration stage: constructing a double-coordinate system mapping model based on a PCB (Printed Circuit Board) plane, and establishing a nonlinear distortion compensation matrix according to the patch thickness; s2, a pickup stage: a CCD shooting camera is arranged at the top of the suction nozzle, a top CCD is triggered to shoot when the suction nozzle picks up an element, and pin coplanarity data is extracted; s3, a positioning stage: setting a reference point at the top of the PCB, scanning the PCB reference point through a CCD, and calculating a deviation between an actual position and a theoretical coordinate by using an algorithm processor; and S4, compensation execution: the motion shaft system dynamically adjusts the mounting path according to the deviation vector. The visual correction method based on SMT pasting has the advantages of accurate pasting and the like, and solves the problems of pasting position offset, low efficiency and unstable precision caused by factors such as vibration of an element feeder, mechanical transmission errors and thermal deformation of a PCB (Printed Circuit Board) due to dependence on manual calibration in the existing SMT pasting process.

Description

Technical Field

[0001] This invention relates to the field of surface mount technology, specifically to a visual correction method for SMT (Surface Mount Technology) placement. Background Technology

[0002] SMT (Surface Mount Technology) is an abbreviation for a series of processes performed on a PCB (Printed Circuit Board). SMT is a surface mount technology, one of the most popular technologies and processes in the electronics assembly industry.

[0003] In existing SMT placement processes, manual calibration is still the primary method. Due to factors such as component feeder vibration, mechanical transmission errors, and PCB thermal deformation, placement positions may shift, resulting in low efficiency and unstable accuracy. Therefore, a visual correction method based on SMT placement is proposed to address the aforementioned problems. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a visual correction method for SMT placement, which has advantages such as precise placement. It solves the problem that in existing SMT placement processes, most still rely on manual calibration, which leads to placement position deviations due to factors such as component feeder vibration, mechanical transmission errors, and PCB thermal deformation, resulting in low efficiency and unstable accuracy.

[0006] (II) Technical Solution

[0007] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A visual correction method based on SMT patch, comprising the following steps:

[0008] S1, Calibration Stage: Construct a dual-coordinate system mapping model based on the PCB board plane, and establish a nonlinear distortion compensation matrix according to the patch thickness;

[0009] S2, Pick-up stage: A CCD camera is set on the top of the nozzle. When picking up a component through the nozzle, the top CCD is triggered to capture the image and extract the pin coplanarity data.

[0010] S3, Positioning Stage: Set a reference point on the top of the PCB board, scan the PCB reference point with a CCD, and use an algorithm processor to calculate the deviation between the actual position and the theoretical coordinates.

[0011] S4, Compensation Execution: The motion axis system dynamically adjusts the mounting path according to the deviation vector, establishes a position deviation prediction model (Kalman filtering + time series analysis), generates XYZθ four-dimensional compensation commands, and corrects the nozzle pose in real time through the motion controller;

[0012] S5, Quality Traceability: Stores compensation logs for each PCB for process optimization.

[0013] The beneficial effects of this invention are: simultaneous acquisition by dual CCDs, feature fusion matching algorithm, dynamic prediction compensation mechanism, and closed-loop verification system.

[0014] This SMT placement-based visual correction method has the advantage of precise placement.

[0015] Based on the above technical solution, the present invention can be further improved as follows.

[0016] Furthermore, in step S1, the nonlinear distortion compensation matrix is ​​established sequentially from the PCB board to the top of the patch using a nine-point calibration method. The Levenberg-Marquardt algorithm is used to optimize the distortion compensation matrix to compensate for the radial distortion of the lens (maximum distortion < 0.05 pixels). The output parameters are as follows:

[0017]

[0018] Where K is the affine transformation matrix and φ is the Gaussian radial basis function.

[0019] The beneficial effects of adopting the above-mentioned further scheme are that the nine-point calibration method optimized by the LM algorithm can establish a submicron-level distortion model (>99% confidence level), improve the nonlinear error convergence speed by 40%, and ensure that the full field-of-view distortion compensation residual is <1.5μm at a magnification of 0.3X, thus meeting the positioning requirements of high-density BGA chips.

[0020] Furthermore, in step S2, a coaxial ring light source and a telecentric lens are configured to simultaneously acquire images during the component pickup and PCB positioning stages, and a narrow-band filter is applied to eliminate ambient light interference (wavelength matching LED light source).

[0021] Among them, the coaxial ring LED light source (8 wavelengths adjustable, including 850nm infrared / 470nm blue light / 630nm red light), the telecentric lens is a double telecentric lens (magnification 0.3X, depth of field ±5mm), and an electric filter wheel (bandwidth ±10nm, switching time <50ms).

[0022] The beneficial effects of adopting the above-mentioned further solution are that, by combining the wavelength adaptive light source with the dual telecentric optical path, the contrast of pad recognition is improved to 25dB, the infrared band can penetrate 0.8mm thick IC package to detect coplanarity, and the exposure synchronization accuracy is maintained at 3ms level when the light intensity uniformity is >95%.

[0023] Furthermore, in step S3, the geometric features of the component pins and pads are extracted (SIFT feature points + edge gradient histogram). Feature point matching and outlier removal are achieved based on the RANSAC algorithm. The feature extraction layer consists of an improved SIFT feature (DoG keypoint detection + 128-dimensional descriptor) and a multi-scale edge gradient histogram (8-direction bin, window size 16×16).

[0024] Hierarchical RANSAC matching:

[0025] Primary matching: Bidirectional nearest neighbor search (accelerated by kd-tree);

[0026] Mismatch Removal: Geometric Consistency Verification (Affine Transformation Constraints);

[0027] Precise registration: Nonlinear optimization of the LM algorithm;

[0028] Fusion positioning accuracy: ±0.5 pixels (corresponding to actual accuracy ±3μm).

[0029] The beneficial effects of adopting the above-mentioned further scheme are that the improved SIFT features achieve a 128-dimensional descriptor matching accuracy of >98% in the pad image, the combination of multi-scale gradient histograms enables the orientation resolution to reach ±2°, and the RANSAC hierarchical optimization reduces the number of registration iterations by 60% while maintaining sub-pixel localization capability.

[0030] Furthermore, in step S4, the Kalman filter state equation is:

[0031]

[0032] State variable x=[ΔX, ΔY, Δθ, Δv_x, Δv_y]^T;

[0033] Four-dimensional compensation instruction generation:

[0034] Position compensation: PID feedforward control (response time < 2ms);

[0035] Angle compensation: Incremental rotation matrix correction;

[0036] Vibration suppression: Acceleration feedback compensation (bandwidth 500Hz).

[0037] The beneficial effects of adopting the above-mentioned further scheme are that the Kalman-PID composite control architecture achieves axial / tangential vibration suppression ratios of -35dB / -28dB respectively, phase delay of <0.1ms at a bandwidth of 500Hz, and θ-axis following error of <0.003° when combined with incremental rotation compensation.

[0038] Furthermore, in step S5, a second photo is taken after mounting for verification, and the compensation coefficient is fed back to the learning database to establish a component type-compensation parameter mapping table to achieve adaptive optimization.

[0039] It also includes an online verification module, namely post-mount photography verification (5 megapixels, dual-sided telecentric optical path, pad coverage analysis (sub-pixel edge detection algorithm)).

[0040] The learning database includes:

[0041] Compensation parameter self-learning framework:

[0042] Feature importance analysis based on random forest;

[0043] Compensation amount regression model (MAPE<1.5%).

[0044] The beneficial effects of adopting the above-mentioned further scheme are that feature selection based on random forest reduces the dimensionality of the compensation model by 70%, the online learning system can achieve parameter convergence with MAPE < 1% within 20 mounting cycles, and adaptively covers 0402 to 50mm² QFN element types.

[0045] Furthermore, in step S2, a photoelectric encoder is used to trigger acquisition (resolution 4096ppr), and the multispectral image fusion strategy is as follows: the visible light band (470-630nm) is used to extract the pad contour, and the infrared band (850nm) is used to penetrate the warp of the detection element.

[0046] The beneficial effects of adopting the above-mentioned further scheme are that the photoelectric encoder triggering results in a multispectral image timing deviation of <1μs, a visible light / infrared image registration error of <0.3 pixels, and a sensitivity of over 95% for solder paste coverage detection in the 850nm band.

[0047] Furthermore, in step S3, an FPGA+GPU heterogeneous computing architecture is adopted, with image processing latency <5ms, supporting positioning accuracy of 0.01mm and angle correction of ±0.1°, and compatible with positioning of irregularly shaped components such as BGA / QFN.

[0048] The beneficial effects of adopting the above-mentioned further solutions are that FPGA enables 5Gbps real-time image stream processing, GPU acceleration enables 128-dimensional feature matching throughput to reach 8000fps, and supports 60000CPH placement cycle time with a positioning accuracy of 0.01mm. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0050] Figure 2 This is the element type-parameter mapping table for the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] In the embodiments, by Figure 1-2 This invention presents a visual correction method based on SMT (Surface Mount Technology) patching, comprising the following steps:

[0053] S1, Calibration Stage: Construct a dual-coordinate system mapping model based on the PCB board plane, and establish a nonlinear distortion compensation matrix according to the patch thickness;

[0054] S2, Pick-up stage: A CCD camera is set on the top of the nozzle. When picking up a component through the nozzle, the top CCD is triggered to capture the image and extract the pin coplanarity data.

[0055] S3, Positioning Stage: Set a reference point on the top of the PCB board, scan the PCB reference point with a CCD, and use an algorithm processor to calculate the deviation between the actual position and the theoretical coordinates.

[0056] S4, Compensation Execution: The motion axis system dynamically adjusts the mounting path according to the deviation vector, establishes a position deviation prediction model (Kalman filtering + time series analysis), generates XYZθ four-dimensional compensation commands, and corrects the nozzle pose in real time through the motion controller;

[0057] S5, Quality Traceability: Stores compensation logs for each PCB for process optimization;

[0058] In step S1, the nonlinear distortion compensation matrix is ​​established sequentially from the PCB board to the top of the patch using a nine-point calibration method. The Levenberg-Marquardt algorithm is used to optimize the distortion compensation matrix to compensate for lens radial distortion (maximum distortion < 0.05 pixels). The output parameters are:

[0059]

[0060] Where K is the affine transformation matrix and φ is the Gaussian radial basis function;

[0061] The nine-point calibration method optimized by the LM algorithm can establish a submicron-level distortion model (>99% confidence level), improve the nonlinear error convergence speed by 40%, and ensure that the full field-of-view distortion compensation residual is <1.5μm at a magnification of 0.3X, meeting the positioning requirements of high-density BGA chips;

[0062] In step S2, a coaxial ring light source and a telecentric lens are configured to acquire images synchronously during component pickup and PCB positioning. Narrow-band filters are used to eliminate ambient light interference (wavelength matching LED light source).

[0063] Among them, the coaxial ring LED light source (8 wavelengths adjustable, including 850nm infrared / 470nm blue light / 630nm red light), the telecentric lens is a double telecentric lens (magnification 0.3X, depth of field ±5mm), and an electric filter wheel is applied (bandwidth ±10nm, switching time <50ms).

[0064] By combining a wavelength-adaptive light source with a dual telecentric optical path, the contrast of pad recognition is improved to 25dB. The infrared band can penetrate 0.8mm thick IC packages to detect coplanarity, and the light intensity uniformity is >95% while maintaining 3ms-level exposure synchronization accuracy.

[0065] In step S3, the geometric features of component pins and pads are extracted (SIFT feature points + edge gradient histogram). Feature point matching and outlier removal are achieved based on the RANSAC algorithm. The feature extraction layer consists of an improved SIFT feature (DoG keypoint detection + 128-dimensional descriptor) and a multi-scale edge gradient histogram (8-direction bin, window size 16×16).

[0066] Hierarchical RANSAC matching:

[0067] Primary matching: Bidirectional nearest neighbor search (accelerated by kd-tree);

[0068] Mismatch Removal: Geometric Consistency Verification (Affine Transformation Constraints);

[0069] Precise registration: Nonlinear optimization of the LM algorithm;

[0070] Fusion positioning accuracy: ±0.5 pixels (corresponding to actual accuracy ±3μm);

[0071] Improved SIFT features achieve a 128-dimensional descriptor matching accuracy of >98% in pad images. Combined with multi-scale gradient histograms, the orientation resolution reaches ±2°. RANSAC hierarchical optimization reduces the number of registration iterations by 60% while maintaining sub-pixel localization capability.

[0072] In step S4, the Kalman filter state equation is:

[0073]

[0074] State variable x=[ΔX, ΔY, Δθ, Δv_x, Δv_y]^T;

[0075] Four-dimensional compensation instruction generation:

[0076] Position compensation: PID feedforward control (response time < 2ms);

[0077] Angle compensation: Incremental rotation matrix correction;

[0078] Vibration suppression: Acceleration feedback compensation (bandwidth 500Hz);

[0079] The Kalman-PID composite control architecture achieves axial / tangential vibration suppression ratios of -35dB / -28dB respectively, phase delay of <0.1ms at a 500Hz bandwidth, and incremental rotational compensation to make the θ-axis following error <0.003°.

[0080] In step S5, a second photo is taken after mounting for verification, and the compensation coefficient is fed back to the learning database to establish a component type-compensation parameter mapping table to achieve adaptive optimization.

[0081] It also includes an online verification module, namely post-mount photography verification (5 megapixels, dual-sided telecentric optical path, pad coverage analysis (sub-pixel edge detection algorithm)).

[0082] The learning database includes:

[0083] Compensation parameter self-learning framework:

[0084] Feature importance analysis based on random forest;

[0085] Compensation amount regression model (MAPE<1.5%);

[0086] Feature selection based on random forest reduces the dimensionality of the compensation model by 70%, and the online learning system can achieve parameter convergence of MAPE<1% within 20 mounting cycles, adaptively covering 0402 to 50mm² QFN element types;

[0087] In step S2, a photoelectric encoder is used to trigger the acquisition (resolution 4096ppr). The multispectral image fusion strategy is as follows: the visible light band (470-630nm) is used to extract the pad contour, and the infrared band (850nm) is used to penetrate the warp of the detection element.

[0088] The photoelectric encoder triggering ensures a multispectral image timing deviation of <1μs, a visible / infrared image registration error of <0.3 pixels, and a solder paste coverage detection sensitivity of over 95% in the 850nm band.

[0089] In step S3, an FPGA+GPU heterogeneous computing architecture is adopted, with image processing latency <5ms, supporting positioning accuracy of 0.01mm and angle correction of ±0.1°, and compatible with positioning of irregularly shaped components such as BGA / QFN.

[0090] The FPGA enables 5Gbps real-time image stream processing, while the GPU acceleration achieves a throughput of 8000fps for 128-dimensional feature matching, supporting a placement cycle of 60000CPH with a positioning accuracy of 0.01mm.

[0091] Working principle:

[0092] Step 1: Construct a dual-coordinate system mapping model, and use the nine-point calibration method to establish a nonlinear distortion compensation matrix to calibrate the positions of the PCB board and SMT components;

[0093] Step 2: CCD cameras are set at the top and bottom of the robotic arm. When the component is picked up by the nozzle, the top CCD is triggered to capture the image and extract the pin coplanarity data. A reference point is set on the top of the PCB board. The reference point of the PCB is scanned by the CCD and the algorithm processor is used to calculate the deviation between the actual position and the theoretical coordinates.

[0094] Step 3: The motion axis system dynamically adjusts the mounting path based on the deviation vector, establishes a position deviation prediction model, generates XYZθ four-dimensional compensation instructions, corrects the nozzle pose in real time through the motion controller, and stores the compensation log for each PCB for process optimization.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A vision-based SMT patch rectification method, characterized in that, Comprising the following steps: S1, calibration stage: based on the PCB board plane to build a double coordinate system mapping model, and according to the patch thickness to establish a non-linear distortion compensation matrix; S2, pick-up stage: the top of the suction nozzle is provided with a CCD shooting camera, the CCD shooting at the top is triggered when the suction nozzle picks up the element, and the pin coplanarity data is extracted; S3, positioning stage: a reference point is arranged on the top of the PCB board, the PCB reference point is scanned by the CCD, and an algorithm processor is used to calculate the deviation between the actual position and the theoretical coordinates; S4, compensation execution: the motion shaft system dynamically adjusts the mounting path according to the deviation vector, establishes a position deviation prediction model (Kalman filter + time series analysis), generates a four-dimensional compensation instruction of XYZθ, and corrects the suction nozzle pose in real time through the motion controller; S5, quality traceability: store the compensation log of each PCB for process optimization.

2. The SMT patch vision correction method according to claim 1, characterized in that: In step S1, the non-linear distortion compensation matrix is established from the PCB board to the top of the patch in turn by nine-point calibration method, the Levenberg-Marquardt algorithm is used to optimize the distortion compensation matrix, the lens radial distortion (maximum distortion <0.05 pixels) is compensated, and the output parameters are: Where K is the affine transformation matrix, and φ is the Gaussian radial basis function.

3. The SMT patch vision rectification method according to claim 1, characterized in that: In the S2 step, a coaxial ring light source and a telecentric lens are configured, images are collected synchronously during the element picking and PCB positioning stages, and a narrow-band filter is applied to eliminate environmental light interference (wavelength matching LED light source); Wherein, the coaxial ring LED light source (8 wavelength adjustable, containing 850nm infrared / 470nm blue light / 630nm red light), the telecentric lens is a double telecentric lens (magnification 0.3X, depth of field ±5mm), and an electric filter wheel (bandwidth ±10nm, switching time <50ms) is applied.

4. The SMT patch vision rectification method according to claim 1, characterized in that: In the S3 step, the geometric features (SIFT feature points + edge gradient histogram) of the element pin and pad are extracted, the feature point matching and outlier rejection are realized based on RANSAC algorithm, wherein the feature extraction layer is an improved SIFT feature (DoG key point detection + 128-dimensional descriptor), and a multi-scale edge gradient histogram (8 direction bin, window size 16x16); Hierarchical RANSAC matching: Primary matching: bidirectional nearest neighbor search (k-d tree acceleration); Mis-matching rejection: geometric consistency verification (affine transformation constraint); Accurate registration: LM algorithm nonlinear optimization; Fusion positioning accuracy: ±0.5 pixels (corresponding to actual accuracy ±3μm).

5. The SMT patch vision rectification method according to claim 1, characterized in that: In the step S4, the Kalman filter state equation is: State variable x=[ΔX, ΔY, Δθ, Δv_x, Δv_y]^T; Four-dimensional compensation instruction generation: Position compensation: PID feedforward control (response time <2ms); Angle compensation: incremental rotation matrix correction; Vibration suppression: acceleration feedback compensation (bandwidth 500Hz).

6. The SMT patch vision rectification method according to claim 1, characterized in that: In the step S5, after mounting, the compensation coefficient is fed back to the learning database to establish an element type-compensation parameter mapping table for adaptive optimization; There is also provided with online verification module, that is, after mounting shooting verification (5 million pixels, double telecentric light path, pad coverage analysis (sub-pixel edge detection algorithm); The learning database comprises: The compensation parameter self-learning framework comprises: Random forest-based feature importance analysis; The compensation amount regression model (MAPE<1.5%).

7. The SMT patch vision rectification method according to claim 1, characterized in that: In the step S2, photoelectric encoder triggering acquisition (resolution 4096ppr) is adopted, and a multispectral image fusion strategy is adopted, that is, a pad contour is extracted in a visible light band (470-630nm), and an infrared band (850nm) is used for penetrating detection element warping.

8. The SMT patch vision rectification method according to claim 1, characterized in that: In the step S3, an FPGA+GPU heterogeneous computing architecture is adopted, image processing delay is less than 5ms, 0.01mm positioning accuracy and ±0.1° angle correction are supported, and BGA / QFN and other special-shaped component positioning are compatible.

Images