A double-sided dispensing visual closed-loop control method for crystal oscillator production

Abstract

The application provides a double-sided dispensing visual closed-loop control method for crystal oscillator production, and belongs to the field of special equipment manufacturing for semiconductor devices. The application solves the problems of existing open-loop control, such as the inability to compensate for dynamic deviation, the disconnection between detection and execution, and the insufficient precision of double-sided space coordination. The method comprises the following steps: during the dispensing of the first side, capturing dynamic images of the glue liquid at a sampling frequency of no less than 500 frames / s, and calculating the shape factor to identify the glue hanging abnormality in real time; after the dispensing is completed, periodically collecting the definition evaluation value of the reference point on the wafer surface, calculating the wafer tilt angle vector, and generating the Z-axis compensation amount. During the dispensing of the second side, the PID algorithm is used to calculate the driving parameter adjustment amount according to the actual relative deviation of the glue point volume, and the control parameter is corrected in real time according to the environmental temperature. The application realizes real-time closed-loop regulation and control of the dispensing process and double-sided space coordination, and improves the consistency of the glue points and the alignment precision.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device manufacturing equipment, and specifically discloses a visual closed-loop control method for double-sided dispensing in crystal oscillator production. Background Technology

[0002] As a core frequency control component in electronic devices, the manufacturing precision of quartz crystal oscillators directly affects the stability of the entire system. In the crystal oscillator packaging process, the dispensing step applies conductive or fixing adhesive to the connection points between the chip and the substrate, achieving mechanical fixation and electrical interconnection of the chip. Dispensing quality, including the accuracy of the adhesive dot placement, volume consistency, and stress balance of the double-sided adhesive layers, directly determines the equivalent resistance, frequency temperature drift characteristics, and long-term reliability of the crystal oscillator.

[0003] As crystal oscillators evolve towards miniaturization (e.g., 2016, 1612 sizes) and higher frequencies, dispensing processes face precision challenges. Currently, the most common automated dispensing solutions in the industry are typically based on vision-guided robot positioning systems. For example, the technology disclosed in patent document CN103824754A uses an industrial camera to capture the position of a base, driving a robotic arm to complete dispensing along a preset trajectory. While this type of solution performs coordinate calibration before dispensing, it lacks real-time monitoring of the actual glue spraying results during execution, essentially constituting open-loop control. In actual production, glue viscosity changes non-linearly with ambient temperature, humidity, and solvent evaporation; the internal pressure of the piezoelectric jet dispensing actuator also fluctuates slightly; and the needle is more prone to glue residue or wear after prolonged operation. These factors all cause deviations in the volume and shape of the glue dots from the preset values, and open-loop systems cannot respond to these dynamic deviations.

[0004] To address these shortcomings, post-dispensing visual inspection equipment has emerged in the industry, such as the patent with publication number CN221260780U, which uses multi-angle vision to screen finished crystal oscillators for defects. However, the inspection and execution processes are completely decoupled in terms of space and control command chain, meaning that quality deviation information cannot be fed back to the dispensing controller in real time. This "production first, inspection later" model results in a large number of defective products accumulating on the production line when the inspection system detects batch defects, leading to material waste and efficiency losses.

[0005] High-performance crystal oscillators often employ double-sided dispensing to achieve structural symmetry and vibration resistance. However, double-sided dispensing is not simply a superposition of single-sided processes. The shrinkage stress generated during the curing process of the first dispensing layer alters the spatial orientation of the crystal, potentially causing micrometer-level tilting. This directly leads to a deviation between the actual spray plane and the pre-defined plane during the second dispensing process. Existing vision systems mostly focus on two-dimensional features of a single plane, making it impossible to establish a physical relationship model between the front and back sides in the same coordinate system, let alone dynamically compensate for spatial orientation changes caused by curing stress.

[0006] Existing technologies have shortcomings in three dimensions: dynamic sensing during the dispensing process, adaptive parameter adjustment, and collaborative control of two-sided space. Summary of the Invention

[0007] The purpose of this invention is to provide a visual closed-loop control method for double-sided dispensing in crystal oscillator production, which solves the technical problems of open-loop control in existing double-sided dispensing processes for crystal oscillators being unable to compensate for dynamic deviations, the disconnect between detection and execution, and insufficient spatial coordination accuracy of double-sided dispensing.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A visual closed-loop control method for double-sided dispensing in crystal oscillator production includes the following steps:

[0010] Step 1: Before dispensing, establish the coordinate transformation relationship between the coordinate system of the top vision acquisition component, the coordinate system of the bottom vision acquisition component, and the physical coordinate system of the multi-axis linear motion platform. Map the pixel coordinates captured by the heterogeneous vision sensors to a unified three-dimensional physical space coordinate system. During the dispensing process on the first side, continuously capture dynamic images of the adhesive detaching from the nozzle and contacting the pad surface at a sampling frequency of no less than 500 frames / second. Calculate the pixel area and pixel perimeter of the sprayed adhesive dots on the two-dimensional plane in real time, and calculate the shape factor based on the pixel area and pixel perimeter. The shape factor is equal to 4π multiplied by the pixel area divided by the square of the pixel perimeter. Identify dispensing anomalies by monitoring the dynamic evolution curve of the shape factor.

[0011] Step 2: After the first side dispensing is completed, periodically collect image sharpness evaluation values ​​at at least three reference points on the wafer surface. Calculate the height change of each reference point based on the change in the image sharpness evaluation values. Solve the tilt vector of the wafer plane based on the height change of each reference point and its coordinates on the wafer plane. Calculate the Z-axis compensation based on the tilt vector and the coordinates of the target dispensing position on the second side, and superimpose the Z-axis compensation into the Z-axis motion command for the second side dispensing. During the second side dispensing process, calculate the driving parameter adjustment amount using a proportional-integral-derivative (PID) control algorithm based on the relative deviation between the actual glue dot volume and the target volume of the first side dispensing. Correct the proportional coefficient, integral coefficient, and derivative coefficient of the PID control algorithm in real time based on the ambient temperature.

[0012] Furthermore, establishing the coordinate transformation relationship in step 1 specifically includes:

[0013] A high-precision calibration plate is used to perform multi-point sampling within the overlapping field of view of the top and bottom visual acquisition components. The Zhang calibration method is used to establish the coordinate transformation relationship between the top visual coordinate system, the bottom visual coordinate system, and the physical coordinate system of the multi-axis linear motion platform. The coordinate transformation relationship includes the camera intrinsic parameter matrix, distortion coefficients, and rotation and translation matrices obtained from hand-eye calibration.

[0014] Furthermore, step 1, identifying dispensing anomalies, specifically includes:

[0015] When the shape factor shows a downward trend over multiple consecutive spraying cycles, it is determined that the needle of the piezoelectric jet dispensing actuator is stuck with glue, and the cleaning program is automatically triggered.

[0016] The cleaning procedure includes a high-frequency, low-amplitude pulse vibration needle step, a high-pressure pulse jetting residual adhesive step, and a vacuum adsorption cleaning needle tip step, performed sequentially.

[0017] Furthermore, before calculating the shape factor in step 1, the method further includes extracting the pad contour using a sub-pixel edge detection algorithm:

[0018] The improved Canny edge detection operator is invoked to achieve coarse localization of pad edges through adaptive double threshold segmentation. In the neighborhood of the coarsely localized edge, the pixel gray-level gradient is reconstructed using a bicubic interpolation algorithm. Subpixel-level edge refinement is achieved by finding the zero point of the second derivative of the gray-level gradient. Based on the refined edge point set, the geometric center of the pad is fitted using the least squares method.

[0019] Furthermore, the calculation of the height change of each reference point in step 2 is performed in the following way:

[0020] Multiply the lens depth-of-field calibration factor by a logarithmic function base e, where the independent variable of the logarithmic function is 1 plus the ratio of the change in the sharpness evaluation value to the sharpness evaluation value of the reference point at the initial moment.

[0021] The change in the sharpness evaluation value is equal to the sharpness evaluation value at the current moment minus the sharpness evaluation value at the initial moment.

[0022] Furthermore, in step 2, the tilt vector of the wafer plane is calculated using the least squares plane fitting method, and the plane model is set as follows:

[0023] The height change of any reference point is equal to the tilt angle of the wafer around the X-axis multiplied by the x-coordinate of the reference point, plus the tilt angle of the wafer around the Y-axis multiplied by the y-coordinate of the reference point, plus the overall height change of the wafer center; where both the x-coordinate and y-coordinate are based on the wafer center as the origin.

[0024] Furthermore, the method for calculating the adjustment amount of the driving parameters using the incremental proportional-integral-derivative control algorithm in step 2 is as follows:

[0025] The adjustment amount of the driving parameter is equal to the proportional coefficient multiplied by the difference between the two previous volume relative deviations, plus the integral coefficient multiplied by the current volume relative deviation, plus the derivative coefficient multiplied by the current volume relative deviation minus twice the volume relative deviation of the previous cycle plus the difference between the volume relative deviations of the two previous cycles.

[0026] Wherein, the relative volume deviation is the ratio of the deviation between the actual glue dot volume and the target volume in the current spraying cycle to the target volume.

[0027] Furthermore, it also includes an active inhibition step for fiber drawing:

[0028] The process of the adhesive droplet detaching from the needle is captured by a lateral vision acquisition component at a frame rate of no less than 1000 frames per second. Multiple frames are continuously acquired in each spraying cycle. The connected region between the adhesive droplet and the needle is extracted and the skeleton length of the connected region is calculated as the string length. When the string length exceeds a preset threshold, a reverse pulse is applied immediately after the piezoelectric jet dispensing actuator finishes its drive pulse. The duration of the reverse pulse is set to 0.3 to 0.8 times the width of the forward pulse, and the voltage amplitude of the reverse pulse is set to 30% to 50% of the forward pulse voltage.

[0029] Furthermore, it also includes a fluorescence tracing imaging step:

[0030] A fluorescent tracer is added to the adhesive, and a narrow-band filter is disposed in front of the top visual acquisition component. The center wavelength of the narrow-band filter matches the excitation spectrum of the fluorescent tracer, and an ultraviolet laser light source is used as auxiliary illumination to acquire high-brightness fluorescent images of the adhesive dots under ultraviolet light excitation.

[0031] Furthermore, it also includes a step for measuring the three-dimensional volume of adhesive dots based on multi-view photometric stereo method:

[0032] At least four independently controllable ring light sources are evenly arranged around the top visual acquisition component. Each ring light source is lit in sequence and an image of the glue dot is acquired. For each pixel in the image, the surface normal vector is solved according to the Lambert reflection model. The relative height map is reconstructed from the surface normal field through a global integration algorithm. The height values ​​of all pixels in the glue dot area are integrated to obtain the true three-dimensional volume of the glue dot. The true three-dimensional volume is used to replace the two-dimensional area to convert the volume for calculating the volume deviation.

[0033] Furthermore, the reconstruction of the relative height map from the surface normal field employs the Frankot-Chellappa global integral algorithm to solve the height field in the Fourier domain;

[0034] The volume integration method is as follows: multiply the square of the physical size corresponding to the pixel by the sum of the height values ​​of all pixels in the glue dot area to obtain the true three-dimensional volume of the glue dot;

[0035] The sum of the height values ​​of all pixels within the glue dot region is obtained by accumulating the height values ​​of each pixel in the pixel set of the glue dot region.

[0036] Furthermore, it also includes adaptive feedforward-feedback composite control steps based on online rheological parameter identification:

[0037] The droplet ejection dynamics are observed using a lateral vision acquisition component. The initial velocity and diameter of the droplet upon exiting the nozzle are detected. The current dynamic viscosity is inferred from the driving voltage and the threshold voltage. The dynamic viscosity is corrected using the deformation ellipticity during droplet flight. The driving voltage for the next ejection cycle is pre-adjusted based on the ratio of the corrected dynamic viscosity to the reference viscosity. Simultaneously, the proportional, integral, and derivative coefficients of the proportional-integral-derivative controller are dynamically adjusted based on the real-time viscosity.

[0038] Furthermore, the method for inferring the current dynamic viscosity is as follows:

[0039] Multiply the proportional coefficient by the difference between the current peak driving voltage and the turn-on threshold voltage, and then divide by the initial droplet velocity to obtain the current dynamic viscosity;

[0040] The method of correcting dynamic viscosity using deformation ellipticity is as follows: multiply the current dynamic viscosity by 1 and add the sum of empirical coefficients multiplied by deformation ellipticity to obtain the corrected dynamic viscosity.

[0041] The deformation ellipticity is equal to the difference between the major axis and the minor axis of the droplet divided by the sum of the major and minor axes.

[0042] Furthermore, it also includes self-optimizing control steps based on deep reinforcement learning:

[0043] A deep Q-network is constructed as a reinforcement learning agent, defining a state space and a discrete action space. The state space includes the current driving voltage, current pulse width, statistical values ​​of volume deviation of the most recent injections, statistical values ​​of shape factor, ambient temperature, dynamic viscosity, and cumulative number of injections. The action space corresponds to different adjustment amounts of the driving parameters.

[0044] The reward function is designed to encourage minimization of volume deviation and jet stability. An ε-greedy strategy is used to select actions. After each jet cycle, experience tuples are stored and mini-batch gradient descent training is performed. The network parameters are updated online.

[0045] Furthermore, it also includes real-time monitoring and compensation steps for the degree of cure:

[0046] After the first side is coated with adhesive, a first fluorescent tracer and a second fluorescent tracer are added to the adhesive. Images of the two fluorescent bands are acquired alternately by multispectral imaging, the fluorescence ratio is calculated, and the real-time curing degree of the adhesive dots is obtained according to the pre-calibrated mapping model between the curing degree and the fluorescence ratio.

[0047] Based on the real-time degree of curing, the final volume shrinkage rate when the adhesive is fully cured is predicted using a curing kinetics model, and then the remaining amount of Z-axis displacement of the wafer center caused by curing shrinkage is calculated.

[0048] The Z-axis compensation amount of the second surface dispensing is dynamically updated based on the remaining amount, and the second surface dispensing process is triggered when the degree of curing reaches a preset threshold or the curing rate drops below a preset value.

[0049] Furthermore, the mapping model between the degree of curing and the fluorescence ratio adopts a logarithmic function form, wherein the independent variable of the logarithmic function is the fluorescence ratio, and the fluorescence ratio is equal to the fluorescence intensity of the first fluorescence band divided by the fluorescence intensity of the second fluorescence band.

[0050] In the curing kinetics model, the relationship between volume shrinkage rate and degree of curing is an exponential function, and the exponential term of the exponential function includes the product of the curing rate constant and the degree of curing.

[0051] The curing rate constant follows the Arrhenius equation and varies with ambient temperature; when the curing rate is detected to be too low and it is necessary to speed up the production cycle, the UV-assisted curing light source is triggered to irradiate the adhesive dot area to accelerate the curing process.

[0052] Furthermore, the first fluorescent tracer is Rhodamine B, whose fluorescence intensity is sensitive to changes in the polarity of the microenvironment during the adhesive curing process; the second fluorescent tracer is fluorescein, whose fluorescence intensity remains relatively stable during the curing process.

[0053] The concentrations of the first fluorescent tracer and the second fluorescent tracer are both 0.05% of the total mass of the adhesive; the multispectral imaging adopts a time-division multiplexing method, alternately using ultraviolet laser light sources with center wavelengths of 365nm and 450nm for excitation, and using narrowband filters with center wavelengths of 595nm and 520nm for reception;

[0054] The curing degree threshold is 0.85, and when the curing rate drops to less than 0.01 / second, it is determined that the shrinkage tends to stabilize;

[0055] The UV-assisted curing light source has a center wavelength of 365nm, a power density of 50mW / cm², and an irradiation time of 30 seconds.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] This invention extends visual perception from pre-dispensing positioning to in-dispensing monitoring, constructing a real-time closed-loop system of spraying-sensing-comparison-correction. It transforms an open-loop material deposition process into a closed-loop servo control process, so that the control error of the glue dot volume no longer depends on the accuracy of the open-loop model, but is continuously corrected through feedback. This transformation makes the system inherently robust to disturbances such as glue viscosity fluctuations, needle wear, and ambient temperature changes, effectively solving the problem of uncompensated dynamic deviations.

[0058] This invention directly embeds the detection sensor (high-speed camera) into the time series of the execution process, enabling the detection results to drive the adjustment of execution parameters within the same control cycle. This eliminates the time gap in the detection-feedback link, shifting the intervention point of quality control from reactive to real-time. From a control theory perspective, this represents a leap from open-loop to closed-loop, generating a new system capability of real-time correction.

[0059] This invention indirectly measures the wafer tilt angle caused by backside curing shrinkage by monitoring changes in wafer surface sharpness from the top, thus achieving indirect perception of variables that cannot be directly observed (backside stress field). A complete causal chain is established: "sharpness change → height change → tilt angle vector → Z-axis compensation," transforming the originally uncontrollable coupled stress into a measurable geometric deviation, which is then eliminated through motion compensation. This method breaks through the traditional mindset of treating the front and back sides as independent tasks in double-sided dispensing, achieving collaborative control across processes and interfaces. Attached Figure Description

[0060] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0061] Figure 1 This is the main flow of the method described in this invention.

[0062] Figure 2 This is a flowchart of the multi-view photometric stereo method for measuring the three-dimensional volume of adhesive dots according to the present invention.

[0063] Figure 3 This is a simplified flowchart of the real-time monitoring and compensation process for curing degree of the present invention. Detailed Implementation

[0064] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0065] The following is in conjunction with the appendix Figures 1-3 The embodiments of the present invention will be described in detail below.

[0066] The method described in this invention is based on an integrated double-sided high-precision dispensing control system, which includes a top vision acquisition component, a bottom vision acquisition component, a side vision acquisition component, a piezoelectric jet dispensing actuator, a multi-axis linear motion platform, a closed-loop control unit, an ambient temperature sensor, and an ultraviolet laser light source.

[0067] Both the top and bottom vision acquisition units utilize industrial-grade complementary metal-oxide-semiconductor (CMOS) image sensors, equipped with telecentric lenses to eliminate geometric measurement errors caused by changes in object focal length. The piezoelectric jet dispensing actuator integrates a multilayer piezoelectric ceramic actuator. The multi-axis linear motion platform is driven by linear motors and equipped with a metal grating ruler for position feedback. The lateral vision acquisition unit employs a high-speed camera with a frame rate of at least 1000 frames per second.

[0068] During the coordinate system establishment phase of system initialization, this invention performs a precise calibration process for the global spatial coordinate system. Multi-point sampling is performed within the overlapping field of view of the top and bottom visual acquisition components using a high-precision calibration board. The Zhang calibration method is then used to establish the coordinate transformation relationship between the top visual coordinate system, the bottom visual coordinate system, and the physical coordinate system of the multi-axis linear motion platform. This transformation relationship includes the camera intrinsic parameter matrix, distortion coefficients, and the rotation and translation matrix obtained from hand-eye calibration. It maps the two-dimensional pixel coordinates captured by heterogeneous visual sensors to a unified Euclidean three-dimensional physical spatial coordinate system, thereby eliminating spatial geometric errors introduced by non-perpendicularity of axes, optical axis offset, and residual lens distortion due to limitations in mechanical installation precision. To achieve logical association between the dispensing areas on both sides of the crystal oscillator base, a three-dimensional spatial mapping model based on the homogeneous coordinate transformation principle is established to ensure that the calculation accuracy of the spatial relative position deviation during double-sided dispensing remains at the micrometer level.

[0069] In the initial calibration stage of the pre-dispensing phase, the top and bottom vision acquisition components acquire images of the front and back sides of the crystal oscillator base to be processed, respectively. The closed-loop control unit uses a composite algorithm based on sub-pixel edge detection to extract the physical contour of the pads to be dispensed.

[0070] First, the improved Canny edge detection operator is invoked to achieve coarse localization of pad edges through adaptive double threshold segmentation. Then, in the neighborhood of the coarsely localized edge, the pixel grayscale gradient is reconstructed using a bicubic interpolation algorithm. By finding the zero point of the second derivative of the grayscale gradient, subpixel-level edge refinement is achieved.

[0071] Based on the refined edge point set, the geometric center of the pad is fitted using the least squares method, thereby extracting the centroid coordinates and azimuth angle relative to the system reference axis. The closed-loop control unit compares these real-time extracted position parameters with the gold standard template stored in memory item by item to calculate the initial position deviation vector of the current workpiece. If any component of the deviation vector exceeds the preset initial process threshold, the system will automatically trigger the motion compensation program, driving the multi-axis linear motion platform to perform dynamic compensation until the deviation vector converges to the preset process allowable range in the visual closed-loop feedback.

[0072] Real-time fluid sensing and closed-loop control during the first dispensing process:

[0073] In the real-time dispensing execution and online monitoring stages, this invention employs a synchronous parallel control strategy, seamlessly embedding the sensing process into the execution cycle. When the piezoelectric jet dispensing actuator receives a jetting pulse, the top vision acquisition component captures the jetting process at a sampling frequency of no less than 500 frames per second, focusing on monitoring the dynamic evolution of the adhesive fluid as it detaches from the nozzle and contacts the solder pad surface. The high-speed field-programmable gate array (FPGA) image processor built into the closed-loop control unit separates the moving adhesive dot target from the complex background using background subtraction and optical flow estimation algorithms.

[0074] By establishing a real-time quality assessment model for adhesive dots based on geometric morphological feature parameters, the pixel area of ​​the adhesive dots on a two-dimensional plane is calculated. ,perimeter and shape factor Shape factor The calculation formula is:

[0075]

[0076] in, This represents the pixel area (in pixels²) of the glue dot projection onto the image. This represents the pixel perimeter of the glue dot projection (unit: pixels). This refers to the shape factor (dimensionless, ranging from 0 to 1). In the logic of this invention, the shape factor... It is defined as a key dimension for measuring dispensing quality. When When the value approaches 1, it indicates that the glue dots exhibit good roundness under the action of surface tension, indicating that the glue solution spreads evenly and wets well.

[0077] The closed-loop control unit monitors in real time. The dynamic evolution curve of the value can indirectly identify whether there are defects such as glue residue, stringing, or abnormal wettability of the pad surface in the needle of the piezoelectric jet dispensing actuator.

[0078] Specifically, if A decreasing trend in the value over multiple consecutive periods usually indicates that the needle tip is experiencing glue residue due to adhesive curing, causing a shift in the jetting kinetic energy vector; if If the value suddenly increases and then decreases rapidly, it may indicate that there are contaminants on the pad surface causing abnormal wettability.

[0079] Double-sided curing coupling compensation and wafer tilt angle calculation:

[0080] To address the impact of the curing shrinkage of the first adhesive dot on the dispensing position of the second side during double-sided dispensing, this invention introduces a vision-based wafer tilt angle detection and Z-axis compensation method. After the first side dispensing is completed, the wafer is repositioned via a flipping mechanism to enter the preparation stage for the second side dispensing.

[0081] During this stage, the first adhesive dot is in the early stage of curing. The stress generated by the shrinkage of the adhesive will gradually act on the wafer, which may cause the wafer to tilt at the micrometer level relative to the base plane.

[0082] To quantitatively detect this tilt, the top vision acquisition component periodically acquires images of the wafer surface at a preset frequency (e.g., 2 Hz). Four corner points on the wafer surface are pre-selected as reference points, and the image sharpness evaluation value of each reference point is recorded at the initial moment (immediately after flipping) and at subsequent moments. The image sharpness evaluation value is calculated using the Tenengrad operator, which characterizes the focus of the image by calculating the sum of squares of the image grayscale gradient: higher sharpness indicates that the point is closer to the focal plane of the lens; decreased sharpness indicates that the point has shifted along the optical axis.

[0083] Through offline calibration experiments, a mapping relationship between changes in sharpness and changes in reference point height was established. This mapping relationship is logarithmic, and its specific expression is as follows:

[0084]

[0085] in, This represents the change in height at the reference point (unit: μm). The change in the sharpness evaluation value (dimensionless). The sharpness rating (dimensionless) at the initial moment is used as a reference point. The current sharpness rating (dimensionless). The lens depth-of-field calibration coefficient (unit: μm) was determined offline using a precision displacement stage and a standard sample. Based on the height changes of at least three reference points and their coordinates on the wafer plane, the tilt vector of the wafer plane was calculated using a least-squares plane fitting method. The plane model was set as follows:

[0086]

[0087] in, For the first The coordinates of each reference point in a coordinate system with the wafer center as the origin (unit: mm). This represents the change in height at that point (unit: μm). The tilt angle about the X-axis (unit: rad). The tilt angle about the Y-axis (unit: rad). The overall height variation at the wafer center (unit: μm) is obtained by solving the normal equation using the least squares method. and .

[0088] After obtaining the tilt angle vector, the coordinates of the target glue dot position on the second surface are used. (Unit: mm), calculate the Z-axis height deviation at this point caused by wafer tilt:

[0089]

[0090] in, This is the Z-axis compensation amount (unit: mm). This compensation amount is superimposed on the Z-axis motion command of the second-side dispensing, ensuring that the second-side dispensing needle accurately lands on the correct plane after the height change caused by the tilt angle during spraying. This cross-interface real-time closed-loop intervention allows the front adhesive dots to compensate for the height fluctuations caused by the curing stress on the back side at the moment of spraying, ensuring the spatial overlap of the double-sided adhesive dots and the consistency of the dispensing thickness.

[0091] The driving parameters for dispensing the second side are adaptively adjusted:

[0092] During the second dispensing process, this invention adaptively corrects the drive parameters of the piezoelectric jet dispensing mechanism based on the deviation between the actual glue dot volume and the target volume from the first dispensing volume. Specifically, after each dispensing operation, the top vision acquisition component acquires the actual area of ​​the glue dot and converts the area into volume using an offline calibrated conversion coefficient. The relative volume deviation is calculated, and an incremental proportional-integral-derivative (PID) control algorithm is used to calculate the adjustment amount of the drive parameters.

[0093] The expression for the incremental PID control algorithm is:

[0094]

[0095] in, For the first The adjustment amount of the driving parameters per injection cycle (dimensionless). For the first The volumetric relative deviation per injection cycle (dimensionless, e.g., 0.04 represents 4%). This represents the relative volume deviation from the previous cycle. This represents the relative volume deviation between the first two cycles. This is the proportionality constant (dimensionless). The integral coefficient is dimensionless. These are the differential coefficients (dimensionless). The proportional coefficient is responsible for quickly responding to the current deviation trend, the integral coefficient is used to eliminate the steady-state error of the system, and the differential coefficient is used to predict the changing trend of the deviation and provide damping to prevent overshoot.

[0096] These three key parameters are not fixed values, but are determined by the closed-loop control unit based on the real-time temperature values ​​fed back by the ambient temperature sensor. (Unit: °C), dynamically corrected by querying a preset viscosity-temperature weighted matrix. The correction formula is:

[0097]

[0098] in, , , The reference parameter is dimensionless. For temperature The corresponding viscosity correction factor (dimensionless) was determined through offline experiments. This weighted matrix records the driving voltage compensation value required to achieve the target adhesive volume under different ambient temperatures. A closed-loop mechanism considering the thermal sensitivity of fluid viscosity ensures that the dispensing system maintains high spray volume linearity even under fluctuating workshop ambient temperatures.

[0099] Adjustment amount These values ​​are mapped to corrections for the drive voltage and pulse width of the piezoelectric jet dispensing actuator. When the actual volume of multiple consecutive dispensing dots is lower than the target volume, the system automatically increases the peak value of the drive voltage or extends the pulse width; conversely, when the actual volume is higher than the target volume, the system decreases the drive voltage or shortens the pulse width.

[0100] To further suppress stringing during the dispensing process, a lateral vision acquisition component and a string breakage compensation logic are integrated into the system. The lateral vision acquisition component is positioned perpendicular to the jetting direction, capturing the droplet ejection process from the needle at a frame rate of at least 1000 frames per second. Multiple frames are continuously acquired within each jetting cycle. The connected region between the droplet and the needle is extracted through image segmentation, and the skeleton length of the connected region is calculated as the stringing length. When the stringing length exceeds a preset threshold, the closed-loop control unit immediately applies a reverse pulse after the piezoelectric jet dispensing actuator's drive pulse ends. The duration of the reverse pulse is set to 0.3 to 0.8 times the width of the forward pulse, and the voltage amplitude is set to 30% to 50% of the forward pulse voltage. This reverse pulse generates a momentary negative pressure within the piezoelectric jet dispensing actuator's cavity, drawing back the unbroken string and preventing it from falling into non-target areas.

[0101] In a preferred embodiment of the present invention, a fluorescent tracer is added to the adhesive, and a narrow-band filter is positioned in front of the top visual acquisition component. The center wavelength of the narrow-band filter matches the excitation spectrum of the fluorescent tracer, and an ultraviolet laser light source is used as auxiliary illumination. Under ultraviolet light excitation, the adhesive dots exhibit high-brightness fluorescence, forming a strong contrast with the background of the metal pads, making it possible to clearly identify trace amounts of adhesive overflow and extremely fine stringing, further improving detection sensitivity.

[0102] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further described below with reference to specific embodiments.

[0103] Example 1: This example describes the double-sided dispensing process for a high-frequency crystal oscillator, model SMD 2016. The crystal oscillator base is made of ceramic, and the pads are gold-plated. A high-viscosity silver conductive adhesive with an initial viscosity of 5500 mPa·s is used.

[0104] The system hardware configuration is as follows: The top and bottom vision acquisition components have a resolution of 20 megapixels (5472×3648), a pixel size of 2.4μm×2.4μm, and are equipped with telecentric lenses with a distortion rate ≤0.01%. The piezoelectric jet dispensing actuator has a maximum jetting frequency of 1000 Hz, a minimum jetting volume of 0.5 nanoliters, and a response time <1 millisecond. The multi-axis linear motion platform is driven by a linear motor and equipped with a 0.1μm resolution metal grating ruler, with a positioning accuracy of ±0.5μm and a repeatability of ±0.2μm. The lateral vision acquisition component has a frame rate of 1000 frames / second. The ambient temperature sensor has an accuracy of ±0.1℃.

[0105] Step 1: System Calibration and Initial Trial;

[0106] A high-precision calibration board is placed at the common center of the field of view of the top and bottom vision acquisition components. A multi-axis linear motion platform drives the calibration board to perform 9-point translation and 4-angle rotation sampling. Zhang's calibration method is used to solve for the camera intrinsic parameters and the hand-eye calibration matrix. After calibration, the closed-loop control unit stores the calibration results in non-volatile memory.

[0107] During the pre-dispensing stage, the top vision acquisition component captures images of the front pads on the base. The closed-loop control unit uses a sub-pixel edge detection algorithm to extract the pad contours, calculate the centroid coordinates and azimuth angle, and compares them with the gold standard template. The detected deviations for the front pads are +3.5μm in the X direction, -1.2μm in the Y direction, and 0.02° in the angle. Because the X-direction deviation exceeds the preset threshold of ±2μm, the system triggers a motion compensation program, driving a multi-axis linear motion platform for physical correction. After correction, the vision image is acquired again, and the deviation vector converges to +0.12μm in the X direction, -0.05μm in the Y direction, and 0.003° in the angle, meeting the process requirements.

[0108] Step 2: Real-time monitoring of adhesive application on the first surface;

[0109] During the first dispensing process, the top vision acquisition component captures dynamic images of the adhesive liquid at a sampling frequency of 500 frames per second. The closed-loop control unit calculates the shape factor of the adhesive dot for each dispensing cycle in real time. In 20 consecutive sprays, A value that is stable between 0.96 and 0.98 indicates that the glue dot has good roundness and no abnormal glue adhesion.

[0110] When production reached the 1200th spray, The value gradually decreased from 0.97 to 0.89. The closed-loop control unit detected this abnormal trend, determined it to be adhesive residue on the needle, and automatically triggered a cleaning program: first, the needle was vibrated with a high-frequency, low-amplitude pulse (frequency 2000 Hz, voltage 30 volts, duration 0.5 seconds); then, residual adhesive was sprayed with a high-voltage pulse (100 volts, single pulse); finally, the needle tip was cleaned by vacuum suction. The entire cleaning process took 3 seconds. The value has been restored to 0.97.

[0111] Step 3: Two-sided coupling compensation;

[0112] After the first layer of adhesive is applied, the wafer is rotated 180° using a flipping mechanism. At this point, the first layer of adhesive is in the early stages of curing. The top vision acquisition component periodically acquires images of the wafer surface at a frequency of 2 Hz, selecting four corner points of the wafer as reference points, with coordinates as follows: Point 1 , point 2 , point 3 , point 4 Unit: mm. Sharpness evaluation values ​​for each reference point at the initial moment (acquired immediately after flipping). The values ​​were 1256, 1243, 1271, and 1268, respectively. After a 30-second freeze period, the sharpness evaluation values ​​were collected again. Through the sharpness-height mapping relationship (formula) ,in The height change at each reference point was calculated after calibration to 2.8 μm. as follows:

[0113] , , , .

[0114] Substitute the coordinates of the four points and the change in height into the planar model. Least squares fitting (including constant terms) is used. Solving for the inclination vector yields the inclination angle vector. , , The coordinates of the target adhesive dot location on the second side are: (Using the center of the wafer as the origin). Calculate the Z-axis compensation:

[0115]

[0116] The system increases the Z-axis motion command for dispensing adhesive on the second side by 1.05 μm.

[0117] Step 4: Adaptive adjustment of dispensing parameters on the second side;

[0118] During the first dispensing process, the average actual volume of the first 100 sprays was 0.48 nanoliters, while the target volume was 0.50 nanoliters, resulting in a relative volume deviation of 4%. The ambient temperature sensor reported a current temperature of 24.5℃. The viscosity correction coefficient was obtained by consulting the viscosity-temperature weighted matrix (see Table 1, which was determined through offline experiments). The baseline PID parameters are: , , After temperature correction , , The adjustment is calculated using an incremental PID formula, where the current deviation is... Previous cycle deviation Deviation in the first two cycles The adjustment amount was calculated. (Calculation process:) ).

[0119] Map the adjustment amount to the drive voltage correction: initial drive voltage 70 volts, new voltage Volts (where 0.3 is an empirical coefficient set according to the characteristic curve of the piezoelectric jet dispensing actuator); Pulse width correction: initial pulse width 0.4 ms, new pulse width Milliseconds (0.7 is an empirical coefficient). During the second dispensing process, the actual spray volume remained stable between 0.49 and 0.51 nanoliters.

[0120] Table 1. Viscosity-temperature weighted matrix;

[0121] 23.0-23.5 1.08 23.5-24.0 1.04 24.0-24.5 1.02 24.5-25.0 1.00 25.0-25.5 0.97

[0122] Step 5: Active suppression of wire drawing;

[0123] Throughout the production process, the lateral vision acquisition component continuously monitored the jet morphology at 1000 frames per second. At the 850th jet, image analysis showed that the filament length reached 0.12 mm, exceeding the preset threshold of 0.10 mm. The closed-loop control unit immediately applied a reverse pulse after the drive pulse ended: the reverse pulse duration was 0.2 milliseconds (50% of the forward pulse width of 0.4 milliseconds), and the voltage amplitude was -28 volts (40% of the forward voltage of 70 volts). The filament was successfully reabsorbed, and subsequent high-magnification microscopic examination revealed no residual filaments.

[0124] After 24 hours of continuous operation, the production data was compiled: a total of 18,000 crystal oscillators were coated with adhesive on both sides. All quality indicators are summarized in Table 2.

[0125] Table 2 Production quality statistics for Example 1;

[0126] Deviation in adhesive dot diameter consistency (3σ) ±0.82μm Based on 18,000 samples Inconsistency in glue dot weight ±0.91% Based on 18,000 samples Double-sided center overlap error 1.05μm Average value, based on 18,000 samples Stringing and glue overflow rate 0.018% There were 23 instances of stringing, with only 2 leaving residue. scrap rate 0.13% Includes all defect types

[0127] Example 2: This example is basically the same as Example 1, except that the adhesive batch was changed to a medium-viscosity silver conductive adhesive from another supplier, with an initial viscosity of 3200 mPa·s. The ambient temperature was also set at 26.5℃, higher than in Example 1.

[0128] During system initialization, the closed-loop control unit automatically loads the corresponding viscosity-temperature weighted matrix based on the batch information of the new adhesive. After pre-dispensing calibration, the shape factor during the first dispensing process... It remained stable between 0.95 and 0.97 (due to the decrease in glue viscosity, the spreading speed was slightly faster, and the roundness decreased slightly but was still within the threshold range).

[0129] During the double-sided coupling compensation stage, the wafer tilt vector caused by back-side curing shrinkage was measured. , (Due to the low viscosity of the glue, the shrinkage stress is small), the calculated Z-axis compensation is 0.84μm.

[0130] During the second dispensing process, due to the high ambient temperature (26.5℃), the adhesive viscosity decreased, resulting in an actual volume deviation of +3% on the first side (the actual volume was slightly higher than the target value). After temperature compensation, the driving voltage decreased from 70 volts to 68.5 volts, and the pulse width decreased from 0.4 milliseconds to 0.385 milliseconds. Statistics from 24 hours of production showed: adhesive dot diameter consistency deviation ±0.91 μm, adhesive dot weight consistency error ±1.05%, double-sided center overlap error 1.12 μm, and a scrap rate of 0.18%. Although these indicators were slightly inferior to those of Example 1, they were still superior to the traditional method.

[0131] Example 3:

[0132] This embodiment is basically the same as Embodiment 1, except that the active wire drawing suppression function is turned off. That is, the lateral vision acquisition component still monitors but does not apply a reverse pulse, in order to verify the independent effect of the wire drawing suppression module.

[0133] During production, the lateral vision acquisition component detected 26 wire-drawing events (wire length ≥ 0.10 mm). Due to the absence of a reverse pulse, 21 of these events resulted in residual wire drawing on the product surface (lengths ranging from 0.05 to 0.15 mm), representing a wire drawing residue rate of 80.8%. In subsequent electrical performance testing, 17 of these products exhibited abnormal equivalent resistance or short circuits, accounting for 81% of the products with residual wire drawing. The overall scrap rate for Example 3 was 0.87%, higher than the 0.13% for Example 1. This indicates that the active wire drawing suppression module contributes to reducing the scrap rate.

[0134] Comparative Example 1:

[0135] Comparative Example 1 uses a traditional open-loop dispensing method to produce SMD 2016 crystal oscillators of the same model. The hardware configuration is the same as in Example 1, but all visual closed-loop feedback functions are disabled (including top / bottom vision used only for initial positioning and not for process monitoring; double-sided coupling compensation disabled; temperature compensation disabled; and active stringing suppression disabled). The dispensing coordinates are based entirely on the preset taught position, the piezoelectric jet dispensing actuator drive voltage is fixed at 70 volts, the pulse width is fixed at 0.4 milliseconds, and there are no real-time parameter corrections.

[0136] Operating procedure: The operator manually moves the multi-axis linear motion platform to the center of the solder pad using the teach pendant and records the coordinates; the system sequentially performs dispensing on the first side, flipping, and dispensing on the second side according to a fixed trajectory; every 2 hours, 10 products are randomly selected for inspection by the operator, and if any deviation is found, the teach coordinates or drive parameters are manually adjusted.

[0137] After running continuously for 24 hours, production data were collected, and the results are shown in Table 3.

[0138] Table 3. Production quality statistics for Comparative Example 1;

[0139] Deviation in adhesive dot diameter consistency (3σ) ±4.3μm Based on 18,000 samples Inconsistency in glue dot weight ±5.2% Based on 18,000 samples Double-sided center overlap error 6.7μm Average value, based on 18,000 samples Stringing and glue overflow rate 1.38% There were 47 instances of wire pulling, with 38 instances of residue. scrap rate 3.9% Includes all defect types

[0140] Comparing Example 1 with Comparative Example 1, it can be seen that the method of the present invention improves the consistency deviation of glue dot diameter from ±4.3μm to ±0.82μm (an improvement of 80.9%), the consistency error of glue dot weight from ±5.2% to ±0.91% (an improvement of 82.5%), the double-sided center overlap error from 6.7μm to 1.05μm (an improvement of 84.3%), reduces the occurrence rate of stringing and glue overflow from 1.38% to 0.018% (a reduction of 98.7%), and reduces the scrap rate from 3.9% to 0.13% (a reduction of 96.7%).

[0141] Comparative Example 2: Comparative Example 2 is basically the same as Example 1, except that the two-sided coupling compensation function is turned off (i.e., tilt detection and Z-axis compensation in step three are disabled), but fluid closed-loop regulation and active wire drawing suppression are retained. This comparative example aims to verify the independent contribution of the two-sided coupling compensation module.

[0142] The hardware configuration and operation process are the same as in Example 1. However, after the first side is dispensing, the system does not perform wafer tilt angle detection, and the second side dispensing is performed at the preset height along the Z-axis without compensating for the curing stress on the back side.

[0143] After 24 hours of continuous operation, the production data were statistically analyzed: the consistency deviation of adhesive dot diameter was ±0.95μm, and the consistency error of adhesive dot weight was ±1.02% (these two indicators are close to those of Example 1 because the fluid closed loop is still operating); the double-sided center overlap error was 4.8μm, which is worse than the 1.05μm of Example 1; the scrap rate was 1.2%, of which about 70% of the scrap was due to uneven wafer stress caused by the misalignment of the double-sided adhesive dots. These results demonstrate that the double-sided coupling compensation module plays an important role in improving the double-sided center overlap.

[0144] Comparative Example 3: Comparative Example 3 is basically the same as Example 1, except that the fluid closed-loop regulation function is turned off, that is, the PID temperature compensation and volume deviation regulation in step four are disabled, but the double-sided coupling compensation and active wire drawing suppression are retained.

[0145] The piezoelectric jet dispensing actuator has a fixed drive voltage of 70 volts and a fixed pulse width of 0.4 milliseconds, and is not adjusted according to the actual volume deviation of the glue dots. The system does not compensate for fluctuations in ambient temperature.

[0146] During 24 hours of continuous operation, the workshop temperature fluctuated between 23.5℃ and 26.0℃. Production data showed: glue dot weight consistency error reached ±3.8%, worse than ±0.91% in Example 1; glue dot diameter consistency deviation was ±2.1μm; however, the double-sided center overlap error remained at 1.3μm (due to the double-sided coupling compensation still operating). The scrap rate was 2.1%, mostly due to electrical performance discrepancies caused by inconsistent glue volume. These results demonstrate the crucial role of the fluid closed-loop control module in maintaining glue volume consistency.

[0147] For ease of comparison, the key indicators of each embodiment and comparative example are summarized in Table 4.

[0148] Table 4. Comparison of key indicators between each embodiment and the comparative example;

[0149] Example 1 ±0.82μm ±0.91% 1.05μm 0.018% 0.13% Example 2 ±0.91μm ±1.05% 1.12μm 0.019% 0.18% Example 3 ±0.85μm ±0.95% 1.08μm 0.85% 0.87% Comparative Example 1 ±4.3μm ±5.2% 6.7μm 1.38% 3.9% Comparative Example 2 ±0.95μm ±1.02% 4.8μm 0.020% 1.2% Comparative Example 3 ±2.1μm ±3.8% 1.3μm 0.019% 2.1%

[0150] Example 4:

[0151] This embodiment improves upon Embodiment 1 by addressing the inherent error in the "two-dimensional area to volume conversion" method for measuring adhesive dots. This error stems from the nonlinear influence of the actual three-dimensional morphology of the adhesive dots (such as contact angle and spreading asymmetry) on the area-to-volume conversion coefficient. Especially when there are batch fluctuations on the solder pad surface or changes in adhesive surface tension, relying solely on area conversion leads to amplified volume deviations. This embodiment proposes a method for reconstructing the three-dimensional morphology of adhesive dots and accurately measuring their volume based on a multi-view photometric stereo method.

[0152] Step A: Configuration of the multi-angle light source system;

[0153] Four independently controllable LED ring light sources are evenly arranged around the top vision acquisition unit. The azimuth angles of the four light sources are 0°, 90°, 180°, and 270°, respectively, and the elevation angle is fixed at 30° for each. The center wavelength of the spectrum of each light source is 470nm (blue light), and the half-width at half-maximum (HWHM) is 20nm. A narrow-band filter (center wavelength 470nm, HWHM 10nm) matching the wavelength of the light source is installed in front of the lens of the top vision acquisition unit to eliminate ambient stray light interference. The side vision acquisition unit remains unchanged.

[0154] Step B: Multi-source image acquisition and surface normal reconstruction;

[0155] At the end of each spraying cycle (when the adhesive dots are fully spread but not yet cured), the closed-loop control unit sequentially triggers the individual illumination of four ring-shaped light sources. Each time the light is illuminated, the top vision acquisition component captures one image, resulting in a total of four images of the adhesive dots under different lighting directions. , , , For each pixel in the image, its grayscale value... With surface normal vector and the direction vector of the light source Satisfies the Lambert reflection model:

[0156]

[0157] in, For the first The grayscale value of a pixel when a light source is lit (dimensionless, 0~255). The albedo of a pixel (dimensionless, 0~1). It is the unit surface normal vector (dimensionless). For the first The direction vectors of each light source (dimensionless, normalized). Since the direction vectors of the four light sources are known... Furthermore, the four images provide four equations, and for each pixel, an overdetermined system of equations can be solved to obtain the surface normal. and albedo The solution is obtained using the least squares method, and the objective function is:

[0158]

[0159] The constraints are The solution yields the unit surface normal vector for each pixel.

[0160] Step C: Height map integration reconstruction;

[0161] By surface normal field ( Reconstructing the relative height map (using pixel coordinates, unit: pixels) (Unit: μm) The Frankot-Chellappa global integration algorithm is used. This algorithm solves for the globally optimal height field in the Fourier domain that satisfies the following least-squares problem:

[0162]

[0163] in, , The surfaces are respectively direction and The gradient of the direction (dimensionless). Solved using Fourier transform:

[0164]

[0165] In the formula, height map Fourier transform, and They are respectively and Fourier transform, , Spatial frequency (unit: rad / pixel). The imaginary unit. (For) The height map is obtained by performing an inverse Fourier transform. .

[0166] Step D: Calculation of glue dot volume;

[0167] In the reconstructed height map, the glue dot regions (albedo) are first extracted by threshold segmentation. (The set of pixels above the pad background). Integrate the volume over the height values ​​of all pixels within the adhesive dot region:

[0168]

[0169] in, This represents the true three-dimensional volume of the adhesive dot (unit: nL). The physical size corresponding to a pixel (unit: μm / pixel) is obtained through prior camera calibration. This is the set of pixels representing the glue dot region; pixel coordinates The height value at this location (unit: μm). It should be noted that the volume unit conversion is... In actual calculations, it can be Divide by The volume is obtained in nanoliters. It directly reflects the true three-dimensional volume of the adhesive dots and is not affected by changes in the contact angle of the adhesive dots or the asymmetry of the spreading.

[0170] Step E: Closed-loop regulation driven by volume deviation;

[0171] The calculated true volume With target volume Compare and calculate the relative volume deviation. The closed-loop control unit uses the same incremental PID algorithm as in Example 1 to calculate the adjustment amount, which is then mapped to correction values ​​for the drive voltage and pulse width of the piezoelectric jet dispensing actuator. Because the volume measurement accuracy is improved from ±5% of the area conversion to ±1.5% of the 3D reconstruction, the volume consistency of the dispensing on the second surface is further improved.

[0172] Using the method described in Example 4, the system was continuously operated for 24 hours under the same production conditions as in Example 1 (same batch of adhesive, same ambient temperature). Statistical production data showed that the adhesive dot weight consistency error decreased from ±0.91% in Example 1 to ±0.52%, the double-sided center overlap error slightly improved from 1.05μm to 0.98μm (due to the indirect reduction of curing stress fluctuations caused by improved volume consistency), and the scrap rate decreased from 0.13% to 0.09%. The additional computation time (8 milliseconds per adhesive dot) remained within the system's allowable range (1 millisecond spray cycle, parallel processing). This example demonstrates that the photometric stereoscopic three-dimensional morphology reconstruction method can improve the accuracy of adhesive dot volume measurement, thereby optimizing the closed-loop control effect.

[0173] Example 5: Based on Example 1, this example addresses the problem that fixed PID parameters cannot be optimally matched due to the change in the rheological properties of the adhesive with shear history (thixotropy). It proposes an adaptive feedforward-feedback composite control method based on online rheological parameter identification.

[0174] This embodiment observes the droplet ejection dynamics using a lateral vision acquisition component, infers the current viscosity in real time, and dynamically adjusts the drive waveform and PID parameters.

[0175] Step A: Jet dynamics observation system;

[0176] A lateral vision acquisition unit (10,000 frames / second, 512×512 pixels resolution) is positioned perpendicular to the spray direction, covering a field of view from 0 mm to 2 mm below the nozzle. During each spray cycle, the camera acquires a continuous set of images (20 frames per cycle). The FPGA image processor within the closed-loop control unit performs the following processing in real time:

[0177] First frame (0.05 milliseconds after the start of the drive pulse): Detect the instantaneous position of the droplet tip leaving the nozzle and record the initial velocity. (Unit: m / s).

[0178] Fifth frame (0.1 milliseconds after the drive pulse ends): Detects the diameter of the droplet body when it leaves the nozzle. (Unit: μm) and axial length (Unit: μm)

[0179] 10th frame (0.2 ms after detachment): Detect the ellipticity of the droplet during flight. ,in Major axis (unit: μm). The minor axis (unit: μm). Dimensionless.

[0180] Step B: Online viscosity identification model;

[0181] Based on the piezoelectric jetting principle, the droplet ejection speed With driving voltage Nozzle diameter Adhesive viscosity The following relationship exists between them:

[0182]

[0183] in, The observed initial velocity of the droplet (unit: m / s). The proportionality coefficient (unit: Pa·s·mm / (V·ms)) related to nozzle geometry and piezoelectric constant is obtained through offline calibration (typical value is...). ); The peak value of the driving voltage during the current injection cycle (unit: V); The threshold voltage (unit: V) for the piezoelectric jet dispensing actuator was measured to be 38V offline. The current dynamic viscosity (unit: Pa·s). From the above formula, we can deduce:

[0184]

[0185] At the same time, the deformation ellipticity after the droplet detaches is utilized Further verification of the reliability of the viscosity estimate is needed. According to fluid dynamics theory, high-viscosity droplets tend to maintain a spherical shape during flight. Low-viscosity droplets are easier to elongate. (Increase). Establish an empirical correction function:

[0186]

[0187] in, This is an empirical coefficient (dimensionless), determined experimentally (typical value is 2.5). The corrected dynamic viscosity (unit: Pa·s).

[0188] Step C: Adaptive feedforward control;

[0189] The online identification Compared with reference viscosity Compare the viscosity (the glue's nominal value at the manufacturer's factory or the initial viscosity measured before the first application, in Pa·s) to calculate the viscosity drift rate. (Dimensionless). Based on the drift rate, the closed-loop control unit pre-adjusts the drive voltage before the start of the next injection cycle:

[0190]

[0191] in, This is the feedforward drive voltage (unit: V). This is the standard drive voltage (unit: V) without drift. The sensitivity index (dimensionless) was obtained through experimental calibration (typical value is 0.3). This feedforward adjustment can compensate for viscosity variations caused by volumetric deviations without waiting for volumetric feedback.

[0192] Step D: Online adaptive PID parameter control;

[0193] Real-time viscosity The input is fed into the parameter adaptive module, which dynamically adjusts the parameters based on the viscosity-PID parameter mapping function established offline. , , :

[0194]

[0195] in, , , Reference parameters (corresponding) mPa·s). , , The viscosity sensitivity index (dimensionless) is optimized for the proportional, integral, and derivative terms to changes in viscosity. This mapping function is obtained through pre-fitting experiments with step responses of adhesives of different viscosities.

[0196] Step E: Execution of composite control;

[0197] The complete control process for each injection cycle is as follows:

[0198] Feedforward stage: Based on the identification of the previous cycle calculate This serves as the reference for the driving voltage in this operation.

[0199] Injection and Observation: Perform injection while simultaneously acquiring images and identifying the new viscosity for the current cycle. .

[0200] Feedback Phase: After the adhesive dot falls onto the pad, the top vision acquisition component acquires the area of ​​the adhesive dot (or uses the three-dimensional volume of Example 4) and calculates the volume deviation. .

[0201] PID Update: Adjust volume deviation Substitute the values ​​into the viscosity-adaptive PID controller and calculate the adjustment amount. .

[0202] Final drive voltage: .

[0203] This embodiment operates under the same production conditions as Example 1, but introduces an additional artificial temperature cycle (23℃→28℃→23℃, repeated every 2 hours) to exacerbate viscosity fluctuations. Production data is collected after 24 hours of continuous operation.

[0204] Viscosity values ​​identified online The correlation coefficient with the offline rotational viscometer measurement reached 0.96, and the maximum deviation did not exceed ±4%.

[0205] Adhesive dot weight consistency error: ±0.38% (better than ±0.91% in Example 1 and ±0.52% in Example 4).

[0206] When the temperature changes rapidly (3°C increase within 15 minutes), the system does not exhibit over- or under-adjustment of adhesive volume, and the feedforward control effectively suppresses the response lag caused by thermal inertia.

[0207] Double-sided center overlap error: 0.95μm (slightly improved).

[0208] Scrap rate: 0.07%.

[0209] Example 6: Based on Example 1, this example proposes a method for real-time monitoring of curing degree and prediction and compensation of curing stress based on multispectral fluorescence imaging to address the dynamic evolution of curing shrinkage stress of the first adhesive dot.

[0210] In Examples 1 and 3, curing stress was indirectly compensated by measuring the wafer tilt angle. However, this method only compensates once in the early stage of curing (after flipping), and fails to capture the dynamic changes of shrinkage stress during the curing process.

[0211] In actual production, the curing rate of adhesive is affected by many factors such as temperature, adhesive layer thickness, and oxygen concentration. The curing shrinkage curves of different batches vary, and a single compensation cannot completely eliminate the final stress. This embodiment establishes a real-time mapping model of curing degree and shrinkage rate by monitoring the degree of curing (DoC) of the adhesive dots online, and dynamically predicts and compensates for the final spatial pose changes before the second surface is applied.

[0212] Step A: Configuration of dual-fluorescent tracer doping and multispectral imaging system;

[0213] Two fluorescent tracers were added to the conductive adhesive: Rhodamine B was selected as the first fluorescent tracer, whose fluorescence intensity is sensitive to changes in the microenvironment polarity during the adhesive curing process, and its fluorescence quantum yield gradually increases as the crosslinking reaction proceeds; Fluorescein was selected as the second fluorescent tracer, whose fluorescence intensity remains relatively stable during the curing process and serves as a reference channel. The concentration of both fluorescent tracers was 0.05% of the total adhesive mass, and ultrasonic dispersion was used to ensure uniform mixing. Experiments verified that this concentration did not affect the conductivity of the adhesive or the bond strength after curing.

[0214] A multispectral excitation and imaging module is integrated into the coaxial optical path of the top vision acquisition component. It is equipped with two ultraviolet laser sources: the first source has a center wavelength of 365 nm and is used to excite Rhodamine B (emission peak 595 nm); the second source has a center wavelength of 450 nm and is used to excite fluorescein (emission peak 520 nm). The two sources are alternately illuminated using a time-division multiplexing method at a timing frequency of 100 Hz. A motorized filter wheel is installed in front of the image sensor of the top vision acquisition component, equipped with two narrow-band filters: filter F1 (center wavelength 595 nm, half-width at half-maximum 20 nm) is used to receive the fluorescence signal of Rhodamine B; filter F2 (center wavelength 520 nm, half-width at half-maximum 20 nm) is used to receive the fluorescence signal of fluorescein. A closed-loop control unit synchronously controls the switching of the light sources and filters, ensuring that two fluorescence images of different wavelengths are acquired sequentially in each cycle.

[0215] Step B: Online calibration and real-time calculation of curing degree;

[0216] During the system initialization phase, offline calibration of the cure degree-fluorescence ratio was performed. Using the same adhesive sample as in production, the cure degree of the adhesive at different curing times was measured using differential scanning calorimetry (DSC) under constant temperature (25℃) and constant humidity (50% relative humidity) conditions. (Defined as the ratio of the heat released during the reaction to the total heat released, ranging from 0 to 1). Simultaneously, dual-wavelength fluorescence images were acquired at the same time points, and the fluorescence ratio of each pixel was calculated. ,in The fluorescence intensity (dimensionless gray value) of the Rhodamine B channel. The fluorescence intensity (dimensionless gray value) of the fluorescein channel is used. A mapping model between curing degree and fluorescence ratio is established:

[0217]

[0218] In the formula, and The calibration coefficients (dimensionless) are obtained through least squares fitting (typical values: , Correlation coefficient The logarithmic form of this model reflects the nonlinear growth of fluorescence intensity with curing degree.

[0219] After the first layer of adhesive is applied, the system performs dual-spectral fluorescence imaging of each adhesive dot region at a frequency of 1 Hz. For each adhesive dot, its fluorescence ratio is calculated. Over time The change curve is substituted into the mapping model to obtain the real-time curing degree. Simultaneously, the geometric dimensions of each adhesive dot (obtained through three-dimensional volume reconstruction in Example 4) and the ambient temperature were recorded. .

[0220] Step C: Dynamic prediction model for curing shrinkage rate;

[0221] According to the theory of polymer curing kinetics, the volume shrinkage rate of the adhesive With curing degree The following relationship exists between them:

[0222]

[0223] In the formula, The volume shrinkage rate is dimensionless. The maximum volume shrinkage rate (dimensionless) upon complete curing was determined through offline experiments (for the silver conductive adhesive used in this embodiment). ); The curing rate constant (dimensionless) is temperature-dependent and follows the Arrhenius equation: ,in It is a pre-exponential factor (dimensionless). Activation energy (unit: J / mol) The gas constant is... The values ​​represent absolute temperatures (in K). These parameters were obtained through offline DSC experiments using curing kinetic analysis at different temperatures.

[0224] The system at each sampling time According to the current degree of curing Calculate the volume shrinkage rate that has already occurred. The volume shrinkage of the adhesive dots leads to a reduction in the adhesive layer thickness, which in turn alters the stress state experienced by the wafer. According to a simplified model of elasticity, the Z-axis displacement of the wafer center point... It has a linear relationship with the volume shrinkage rate of the glue dots:

[0225]

[0226] In the formula, Z-axis displacement of the wafer center caused by curing shrinkage (unit: μm). The initial thickness of the adhesive layer (unit: μm) was obtained by reconstructing the height map in Example 4 (typical value 15 μm). The stress transfer coefficient (dimensionless) is calibrated through finite element simulation (typical value 0.35).

[0227] The system monitors the degree of curing in real time. and its rate of change When the degree of curing reaches the preset threshold (At this point, approximately 80% of the shrinkage has been completed) and the curing rate has decreased to At that time, the system predicts the total shrinkage displacement when final curing is complete. Simultaneously, calculate the remaining shrinkage displacement required from the current moment until curing is complete. .

[0228] Step D: Dynamic compensation and optimization of the timing of dispensing adhesive on the second surface;

[0229] Unlike Examples 1 and 3, where compensation was only performed once after flipping, this example continuously monitors the curing process throughout the waiting period before dispensing on the second side, and dynamically adjusts the Z-axis reference of the second side dispensing based on the predicted final shrinkage displacement.

[0230] Let the initial time after the flip be The measured wafer tilt angle compensation amount is (Calculated using the method in step three of Example 1). The system further predicts from the current moment... Additional wafer displacement caused by adhesive dot shrinkage during the curing process The final Z-axis compensation for the second surface dispensing is:

[0231]

[0232] The system updates every 10 seconds. .when When the shrinkage is less than 0.2μm (i.e. the shrinkage tends to stabilize) or the curing time exceeds the preset maximum waiting time (90 seconds in this embodiment), the system triggers the second dispensing process.

[0233] As an alternative, the system can also actively intervene in the curing process: when the curing level is detected to be too low ( When production cycles require rapid completion, the closed-loop control unit triggers an ultraviolet-assisted curing light source (center wavelength 365nm, power density 50mW / cm²) to irradiate the adhesive dot area for 30 seconds, accelerating the curing process. During ultraviolet irradiation, the system monitors the fluorescence ratio at a higher frequency (5Hz) to ensure a uniform increase in curing degree.

[0234] Step E: Closed-loop feedback and curing consistency control;

[0235] For each batch of continuously produced products, the system records the final degree of curing for each product. and the corresponding shrinkage compensation amount. When 10 consecutive products... When the average value deviates from the target value (0.95±0.03), the system determines that the adhesive batch or environmental conditions have changed and automatically triggers a recalibration procedure: production is paused, and a rapid calibration is completed within 15 minutes using the current adhesive sample (using DSC rapid mode or using a standard sample with known cure degree), and the mapping model parameters are updated. , and dynamic parameters #imgpt250#

[0236] This embodiment operates under the same production conditions as Example 1, with the additional introduction of adhesive batch change (switching from batch A to batch B, where the curing rate of batch B is 18% slower than that of batch A). Production data is collected after 24 hours of continuous operation.

[0237] The correlation coefficient between the real-time monitoring value of curing degree and the offline DSC measurement value reached 0.97, and the maximum absolute deviation did not exceed ±0.04.

[0238] Double-sided center overlap error: 0.72 μm (better than 1.05 μm in Example 1 and 1.08 μm in Example 3), because dynamic compensation eliminates residual stress caused by differences in curing rate.

[0239] The average curing time for the first adhesive dot is 52 seconds for batch A and 68 seconds for batch B (the system automatically extends the waiting time to ensure stable shrinkage). However, UV-assisted curing can shorten the waiting time for batch B to 44 seconds, balancing quality and efficiency.

[0240] Scrap rate: 0.06%, a further reduction compared to Example 1.

[0241] Automatic calibration during batch switching takes 12 minutes, and all indicators return to optimal levels starting from the second hour after calibration.

[0242] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0243] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A visual closed-loop control method for double-sided dispensing in crystal oscillator production, characterized in that, Includes the following steps: Step 1: Before dispensing, establish the coordinate transformation relationship between the coordinate system of the top vision acquisition component, the coordinate system of the bottom vision acquisition component, and the physical coordinate system of the multi-axis linear motion platform. Map the pixel coordinates captured by the heterogeneous vision sensors to a unified three-dimensional physical space coordinate system. During the dispensing process on the first side, continuously capture dynamic images of the adhesive detaching from the nozzle and contacting the pad surface at a sampling frequency of no less than 500 frames / second. Calculate the pixel area and pixel perimeter of the sprayed adhesive dots on the two-dimensional plane in real time, and calculate the shape factor based on the pixel area and pixel perimeter. The shape factor is equal to 4π multiplied by the pixel area divided by the square of the pixel perimeter. Identify dispensing anomalies by monitoring the dynamic evolution curve of the shape factor. Step 2: After the first side is dispensing, periodically collect the image sharpness evaluation values ​​of at least three reference points on the wafer surface. Calculate the height change of each reference point based on the change in the image sharpness evaluation values. Solve the tilt vector of the wafer plane based on the height change of each reference point and its coordinates on the wafer plane. Calculate the Z-axis compensation amount based on the tilt vector and the coordinates of the target dispensing position on the second side. Then, superimpose the Z-axis compensation amount into the Z-axis motion command for dispensing on the second side. During the dispensing process on the second side, based on the relative deviation between the actual glue dot volume and the target volume on the first side, a proportional-integral-derivative (PID) control algorithm is used to calculate the adjustment amount of the driving parameters, and the proportional coefficient, integral coefficient, and derivative coefficient of the PDD control algorithm are corrected in real time according to the ambient temperature.

2. The visual closed-loop control method for double-sided dispensing in crystal oscillator production according to claim 1, characterized in that, Establishing the coordinate transformation relationship in step 1 specifically includes: A high-precision calibration plate is used to perform multi-point sampling within the overlapping field of view of the top and bottom visual acquisition components. The Zhang calibration method is used to establish the coordinate transformation relationship between the top visual coordinate system, the bottom visual coordinate system, and the physical coordinate system of the multi-axis linear motion platform. The coordinate transformation relationship includes the camera intrinsic parameter matrix, distortion coefficients, and rotation and translation matrices obtained from hand-eye calibration.

3. The visual closed-loop control method for double-sided dispensing in crystal oscillator production according to claim 1, characterized in that, The identification of dispensing abnormalities in step 1 specifically includes: When the shape factor shows a downward trend over multiple consecutive spraying cycles, it is determined that the needle of the piezoelectric jet dispensing actuator is stuck with glue, and the cleaning program is automatically triggered. The cleaning procedure includes a high-frequency, low-amplitude pulse vibration needle step, a high-pressure pulse jetting residual adhesive step, and a vacuum adsorption cleaning needle tip step, performed sequentially.

4. The visual closed-loop control method for double-sided dispensing in crystal oscillator production according to claim 1, characterized in that, Before calculating the shape factor in step 1, the method further includes a step of extracting the pad contour using a sub-pixel edge detection algorithm: The improved Canny edge detection operator is invoked to achieve coarse localization of pad edges through adaptive double threshold segmentation. In the neighborhood of the coarsely localized edge, the pixel gray-level gradient is reconstructed using a bicubic interpolation algorithm. Subpixel-level edge refinement is achieved by finding the zero point of the second derivative of the gray-level gradient. Based on the refined edge point set, the geometric center of the pad is fitted using the least squares method.

5. The visual closed-loop control method for double-sided dispensing in crystal oscillator production according to claim 1, characterized in that, The calculation of the height change at each reference point in step 2 is performed in the following way: Multiply the lens depth-of-field calibration factor by a logarithmic function base e, where the independent variable of the logarithmic function is 1 plus the ratio of the change in the sharpness evaluation value to the sharpness evaluation value of the reference point at the initial moment. The change in the sharpness evaluation value is equal to the sharpness evaluation value at the current moment minus the sharpness evaluation value at the initial moment.

6. The visual closed-loop control method for double-sided dispensing in crystal oscillator production according to claim 1, characterized in that, In step 2, the tilt vector of the wafer plane is calculated using the least squares plane fitting method, and the plane model is set as follows: The height change of any reference point is equal to the tilt angle of the wafer around the X-axis multiplied by the x-coordinate of the reference point, plus the tilt angle of the wafer around the Y-axis multiplied by the y-coordinate of the reference point, plus the overall height change of the wafer center; where both the x-coordinate and y-coordinate are based on the wafer center as the origin.

7. The visual closed-loop control method for double-sided dispensing in crystal oscillator production according to claim 1, characterized in that, The method for calculating the adjustment amount of the driving parameter using the incremental proportional-integral-derivative control algorithm in step 2 is as follows: The adjustment amount of the driving parameter is equal to the proportional coefficient multiplied by the difference between the two previous volume relative deviations, plus the integral coefficient multiplied by the current volume relative deviation, plus the derivative coefficient multiplied by the current volume relative deviation minus twice the volume relative deviation of the previous cycle plus the difference between the volume relative deviations of the two previous cycles. Wherein, the relative volume deviation is the ratio of the deviation between the actual glue dot volume and the target volume in the current spraying cycle to the target volume.

8. The visual closed-loop control method for double-sided dispensing in crystal oscillator production according to claim 1, characterized in that, It also includes an active inhibition step for fiber drawing: The process of the adhesive droplet detaching from the needle is captured by a lateral vision acquisition component at a frame rate of no less than 1000 frames per second. Multiple frames are continuously acquired in each spraying cycle. The connected region between the adhesive droplet and the needle is extracted and the skeleton length of the connected region is calculated as the string length. When the string length exceeds a preset threshold, a reverse pulse is applied immediately after the piezoelectric jet dispensing actuator finishes its drive pulse. The duration of the reverse pulse is set to 0.3 to 0.8 times the width of the forward pulse, and the voltage amplitude of the reverse pulse is set to 30% to 50% of the forward pulse voltage.

9. The visual closed-loop control method for double-sided dispensing in crystal oscillator production according to claim 1, characterized in that, It also includes a fluorescence tracer imaging step: A fluorescent tracer is added to the adhesive, and a narrow-band filter is disposed in front of the top visual acquisition component. The center wavelength of the narrow-band filter matches the excitation spectrum of the fluorescent tracer, and an ultraviolet laser light source is used as auxiliary illumination to acquire high-brightness fluorescent images of the adhesive dots under ultraviolet light excitation.

10. The visual closed-loop control method for double-sided dispensing in crystal oscillator production according to claim 1, characterized in that, It also includes a step for measuring the three-dimensional volume of adhesive dots based on multi-view photometric stereo method: At least four independently controllable ring light sources are evenly arranged around the top visual acquisition component. Each ring light source is lit in sequence and an image of the glue dot is acquired. For each pixel in the image, the surface normal vector is solved according to the Lambert reflection model. The relative height map is reconstructed from the surface normal field through a global integration algorithm. The height values ​​of all pixels in the glue dot area are integrated to obtain the true three-dimensional volume of the glue dot. The true three-dimensional volume is used to replace the two-dimensional area to convert the volume for calculating the volume deviation.

Images