Chip welding method based on machine vision and control

Abstract

The invention provides a chip welding method based on machine vision and control, and the method comprises the steps: obtaining a three-dimensional model of a chip and a substrate, obtaining an initial relative position of a chip welding spot and a substrate welding pad, and generating a mechanical arm initial motion instruction; images of the chip and the substrate are shot, pose deviation is calculated through image registration, if the deviation exceeds a threshold value, a compensation instruction is generated to be fused with an initial instruction, and the track of a mechanical arm is adjusted to be aligned with a bonding pad; a temperature sensor array is constructed to monitor the temperature of the welding area, the average temperature is calculated, and if a difference exists between the average temperature and the ideal welding temperature interval, the heater power is adjusted through a proportional integral differential algorithm; a piezoelectric sensor is used for obtaining a mechanical arm joint vibration signal, the dominant vibration frequency and amplitude are extracted through fast Fourier transform, and a piezoelectric actuator is used for generating an anti-phase driving signal to counteract vibration.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to a chip welding method based on machine vision and control. Background Art

[0002] The introduction of robotic automation technology in the manufacture of micro-nano electronic components faces the core contradiction between precision and cost. Take the precision welding of microchips and substrates as an example. The chip size is only a few hundred microns, and the pad spacing is also at the micron level, which places extremely high demands on positioning accuracy and process stability.

[0003] First, chips and substrates are extremely sensitive to temperature. Slight temperature fluctuations may cause size changes of several microns, which directly affects welding accuracy. However, there is currently a lack of precise temperature control technology in microscopic areas.

[0004] Secondly, the tiny vibrations of the robotic arm during grasping, placing and welding can become a significant source of error. The vibration amplitude at the nanometer level is enough to cause misalignment between the chip and the pad, but existing robots find it difficult to achieve sub-nanometer vibration control.

[0005] These problems reflect the stringent requirements of micro-nano manufacturing on equipment performance and are key technical barriers that hinder the application of automation technology. It is necessary to develop new precision control technologies based on process details in order to promote the application of automated assembly in this field. Summary of the invention

[0006] The present invention provides a chip welding method based on machine vision and control, which mainly includes: The three-dimensional model of the chip and substrate is obtained to obtain the initial relative position of the chip solder joint and the substrate pad, and the initial motion command of the robot arm is generated; the chip and substrate images are captured, and the posture deviation is calculated through image registration. If the deviation exceeds the threshold, a compensation command is generated and merged with the initial command to adjust the robot arm trajectory to align with the pad; a temperature sensor array is constructed to monitor the temperature of the welding area and calculate the average temperature. If there is a difference between the average temperature and the ideal welding temperature range, the heater power is adjusted through the proportional integral differential algorithm; the piezoelectric sensor is used to obtain the vibration signal of the robot arm joint, the dominant vibration frequency and amplitude are extracted through fast Fourier transform, and the piezoelectric actuator is used to generate an anti-phase drive signal to offset the vibration; a thermal-mechanical coupling model is established, with a welding process parameter set as input and a welding quality evaluation index as output. The Gaussian process regression algorithm is used to extract the key parameter association rules, and the parameters are optimized according to real-time data using the gradient descent method to generate optimized process parameters. Further, the three-dimensional model of the chip and the substrate is obtained to obtain the initial relative position of the chip solder joint and the substrate pad, including: obtaining a set of discrete coordinate points of the outer boundary of the chip and the substrate, using a B-spline surface fitting algorithm to obtain a continuous surface model, constructing a three-dimensional grid model, extracting the coordinate information of the solder joint and the pad, simulating an ideal welding state, calculating the target coordinates of the solder joint and the target pose matrix of the end effector of the robot, establishing a forward kinematics model of the robot according to the DH parameters of the robot, and obtaining a numerical solution of the joint angle by using an inverse kinematics solution algorithm of the robot. Further, the chip and substrate images are photographed and the pose deviation is calculated by image registration, including: obtaining the images of the chip and the substrate through a high-precision visual sensor and preprocessing them, extracting a set of feature points from the images, obtaining the positions of the feature points on the chip image and the substrate image, using a robust method of anti-error estimation to calculate the homography matrix to obtain the parameters of the spatial pose deviation matrix, using an inverse solution algorithm of the kinematics of the robot to generate a joint angle compensation value, weightedly fusing the joint angle compensation value with the initial motion control instruction of the robot to obtain the adjusted motion control instruction, and guiding the chip pad to align with the substrate pad through vision. Furthermore, the construction of the temperature sensor array to monitor the temperature of the welding area and calculate the average temperature includes: dividing the welding area into sub-areas, deploying a micro temperature sensor array at the center of each sub-area to form an array, obtaining the real-time temperature value of each sensor in the array, summing the measured values ​​of all temperature sensors and calculating the average temperature value of the welding area, using the traversal optimization method to obtain the optimal welding temperature threshold, calculating the deviation between the actual temperature and the midpoint of the ideal temperature range, and using the welding process parameter set as the input condition of the finite element model.Further, the method of obtaining the vibration signal of the joint of the mechanical arm by using the piezoelectric sensor and extracting the dominant vibration frequency and amplitude by fast Fourier transform includes: obtaining the vibration signal of each joint of the mechanical arm according to the piezoelectric sensor, calculating the spectrum of the vibration signal of each joint by fast Fourier transform, extracting the frequency with the largest amplitude in each joint spectrum as the dominant vibration frequency, accumulating all amplitudes to obtain the amplitude corresponding to the dominant vibration frequency, and constructing the anti-phase drive signal according to the extracted dominant vibration frequency and amplitude. Further, the method of generating the anti-phase drive signal by the piezoelectric actuator to offset the vibration includes: real-time monitoring of the vibration signal at the joint of the mechanical arm, judging whether the vibration amplitude exceeds the threshold, if the vibration amplitude exceeds the threshold, re-obtaining the vibration signal and generating a new drive signal, using the adaptive control algorithm to continuously optimize the parameters of the drive signal according to the feedback signal of the piezoelectric sensor and the piezoelectric actuator, adjusting the heating power in real time to stabilize the welding temperature within the ideal range, and sending the control instruction to the power control module of the heater. Further, the establishment of a thermal-mechanical coupling model, taking the welding process parameter set as input and outputting welding quality evaluation indicators, includes: establishing a three-dimensional finite element model according to the material properties and geometric dimensions of the chip and substrate, simulating the temperature field, stress field and deformation field distribution under different parameter combinations, extracting the maximum temperature, temperature gradient, cooling rate, and equivalent plastic strain characteristic parameters of the solder joint area according to the simulation results, and using a finite element model calculation method based on data consistency convergence to obtain key parameters of heat conduction, heat convection, and heat radiation. Further, the Gaussian process regression algorithm is used to extract key parameter association rules, and the parameters are optimized using the gradient descent method according to real-time data, including: using the Gaussian process regression algorithm, taking the welding process parameters as independent variables, and taking the void rate and interface bonding strength as dependent variables, training the key parameter association model to obtain the influence weight and sensitivity of each process parameter on the welding quality, optimizing the kernel function type and hyperparameters of the Gaussian process regression model by cross-validation and grid search, and comparing the predicted welding quality with the actual measured welding quality to obtain the mean square error between the two as the objective function of model optimization.

[0007] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: 1. This method first constructs a three-dimensional digital model of the chip and substrate, uses finite element analysis to calculate the relative position relationship under the ideal welding state, and generates the initial motion control instructions of the robot arm. The position deviation between the chip and the substrate is detected in real time by the visual sensor, and the motion trajectory of the robot arm is dynamically adjusted to achieve precise alignment. During the welding process, a temperature sensor array is used to monitor the temperature distribution in the welding area, and the heater power is adjusted in real time through the PID algorithm to ensure that the welding temperature is in the ideal range.

[0008] 2. At the same time, piezoelectric sensors and actuators are used to suppress the vibration of the mechanical arm and improve the welding stability. In addition, the present invention also establishes a thermal-mechanical coupling model, combines the Gaussian process regression algorithm to extract key process parameters, and continuously optimizes the welding process parameter model through the gradient descent method to achieve continuous improvement of welding quality. This method can significantly improve the accuracy, stability and quality consistency of chip-to-substrate welding, and is suitable for the field of high-end electronic product manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flow chart of the chip and substrate welding quality control of the present invention. The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0010] like Figure 1 In this embodiment, the chip and substrate welding quality control may specifically include: Step S101, constructing a three-dimensional digital model of the chip and the substrate, wherein the outer boundary coordinate set of the chip model is denoted as {(x ci , y ci , z ci ) | i=1,2,...,N c}, the outer boundary coordinate set of the substrate model is recorded as {(x si , y si , z si ) | i=1,2,...,N s},N c and N s Represent the number of discrete points on the chip and substrate models respectively. The relative position relationship between the chip solder joint and the substrate pad under the ideal welding state is calculated by finite element analysis, and the coordinates of the solder joint on the three-dimensional model (x f , y f , z f ) and the initial pose matrix T of the end effector of the robot arm init By solving the inverse kinematics equations, the joint angles θ1, θ2, ..., θ of the robot arm are calculated. n , and then generate the initial motion control instructions containing the angle information of each joint.

[0011] According to the three-dimensional digital model of the chip and substrate, the discrete coordinate point sets of the outer boundaries of the chip and substrate are obtained, which are recorded as Sc and Ss respectively. The B-spline surface fitting algorithm is used to convert the discrete coordinate point sets Sc and Ss into continuous surface models. Based on the surface model, the Delaunay triangulation algorithm is used to construct the three-dimensional mesh models Mc and Ms of the chip and substrate. By extracting features from the three-dimensional mesh models Mc and Ms, the solder joints on the chip and the corresponding pads on the substrate are identified, and their coordinate information is obtained. Using the obtained solder joint and pad coordinates, the finite element analysis method is used to simulate the ideal welding state of the chip solder joint and the substrate pad, and the target coordinates (xf, yf, zf) of the solder joint in three-dimensional space and the target pose matrix Tgoal of the end effector of the robot arm are calculated. According to the DH parameters of the robot arm, the forward kinematics model of the robot arm is established to describe the mathematical relationship between the angles of each joint of the robot arm and the end pose. On this basis, the inverse kinematics solution algorithm of the manipulator, such as the analytical method or the numerical iteration method, is used to take the target posture matrix Tgoal as input to solve the numerical solution of each joint angle (θ1, θ2, ..., θn) of the manipulator, so as to realize the mapping from the end posture to the joint angle. According to the numerical solution of the joint angle (θ1, θ2, ..., θn) obtained by solving, the manipulator motion control instruction is generated. The control instruction is expressed in the form of G code, which contains parameter information such as the target angle, movement speed, acceleration, etc. of each joint of the manipulator. The generated G code instruction sequence is transmitted to the manipulator controller through the communication interface. The manipulator controller integrates a G code interpreter and a motion planning module. After the controller receives the G code instruction, the interpreter parses the instruction and extracts parameters such as the joint target angle. The motion planning module plans the motion trajectory of each joint of the manipulator according to the joint target angle, and controls the servo motor to drive the manipulator to move. Finally, the manipulator moves to the initial posture of the welding operation according to the planned motion trajectory, and prepares for the subsequent welding operation.

[0012] Specifically, the three-dimensional digital model of the chip and substrate is usually obtained by laser scanning or industrial camera. Taking a specific chip as an example, its outer size is ten millimeters by ten millimeters. The surface discrete point cloud data of about ten thousand points is obtained by scanning. These point cloud data constitute a set of discrete coordinate points of the outer boundary of the chip. When using the cubic B-spline surface fitting algorithm to process the discrete point set, it is first necessary to determine the size of the control point grid. For the above chip model, a twenty-by-twenty control point grid can be selected, which can better balance the fitting accuracy and calculation efficiency. The continuous surface model obtained by fitting can be used for subsequent meshing. When performing three-dimensional meshing, the maximum side length of the grid unit is set to 0.5 mm. For the above chip model, about five thousand triangular grid units can be generated by the Delaunay algorithm. This density of grid can ensure the calculation accuracy without causing excessive calculation. In the feature extraction process, the solder joints usually appear as convex features. Taking a typical ball grid array package chip as an example, the diameter of its solder joints is about 0.3 mm and the height is about 0.2 mm. These feature points can be identified by curvature analysis and their precise coordinates can be obtained. In finite element analysis, the material properties of solder joints need to be considered. Taking tin-lead solder as an example, its elastic modulus is thirty GPa and Poisson's ratio is zero point four. The ideal contact state between the solder joint and the pad can be obtained through analysis and calculation, so as to determine the target coordinates of the welding position. The kinematic model of the robot arm is established based on the DH parameters. Taking the six-axis robot arm as an example, the position and posture of its end effector can be represented by a four-by-four homogeneous transformation matrix. Through inverse kinematics solution, the target position and posture can be converted into the angle value of each joint. The control instructions finally generated are in the general G code format. For example, a typical motion instruction may include parameters such as feed speed of one hundred millimeters per minute and angle values ​​of each joint. These instructions are transmitted to the controller via industrial Ethernet to realize precise positioning control of the robot arm. After receiving the instructions, the controller will generate a suitable acceleration curve according to the dynamic characteristics of each joint. Trapezoidal velocity planning, that is, a motion planning method with constant acceleration, is usually adopted, which not only ensures the smoothness of motion, but also improves positioning accuracy.

[0013] Step S102, using a high-precision visual sensor deployed at the end of the robot arm to capture the first image of the chip and the substrate, and using a feature point-based image registration algorithm to calculate the spatial posture deviation between the first images. The position set of the feature points in the image coordinate system is represented by P c = {(u pci , v pci ) | i=1,2,...,N pc} and P s = {(u psi , v psi ) | i=1,2,...,N ps},N pc and Nps Respectively represent the number of feature points detected on the chip and substrate images, and u and v represent the image coordinates. The deviation is judged by the threshold value. If it exceeds the preset deviation threshold δ, the compensation control instruction generation module is triggered to generate compensation control instructions, which are integrated with the initial motion control instructions to adjust the robot arm motion trajectory in real time to align the pads.

[0014] The images of the chip and substrate are acquired by a high-precision visual sensor, and the images are preprocessed, including denoising by using a median filter and image enhancement by using a histogram equalization, to obtain the preprocessed chip image and substrate image. Feature points are extracted from the preprocessed images, and two sets representing the positions of feature points on the chip and substrate images are obtained by using the SIFT algorithm. At the same time, the position information of the pads on the chip and substrate is extracted by edge detection and Hough transform. According to the extracted feature point set, the homography matrix between the chip image and the substrate image is calculated by using the RANSAC algorithm to obtain the matrix parameters representing the spatial posture deviation between the two images. The calculated spatial posture deviation matrix is ​​compared with the preset deviation threshold to determine whether the deviation exceeds the allowable range. If the deviation is small, the next step is directly performed; if the deviation is large, the compensation control instruction generation module is triggered. The compensation control instruction generation module first determines the type and degree of the deviation according to the deviation matrix, and then uses the robot kinematics inverse solution algorithm, such as the CCD algorithm, to generate the joint angle compensation value used to compensate the robot motion trajectory. The generated joint angle compensation value is weighted fused with the initial motion control instruction of the robot to obtain the adjusted motion control instruction. The adjusted motion control instructions are sent to the control system of the robot arm. The control system adjusts the motion of each joint of the robot arm in real time according to the received instructions and corrects the motion trajectory of the robot arm. Through visual guidance, the pads on the chip are accurately aligned with the pads at the corresponding positions on the substrate to complete the chip assembly.

[0015] Specifically, to obtain images of the chip and substrate through high-precision visual sensors, it is necessary to first select a suitable sensor, such as a CMOS or CCD camera, with a resolution of usually more than 5 megapixels to ensure the clarity and details of the image. Image preprocessing is a key step. Median filtering can effectively remove noise. Assuming that there is salt and pepper noise in the original image, the noise points are replaced by the mean of the surrounding pixels after median filtering, and the image is smoother. Histogram equalization adjusts the grayscale distribution of the image to enhance the contrast and make the pad and the background clearer. The feature point extraction uses the SIFT algorithm, which is robust to scale and rotation changes. Taking a specific scene as an example, there are 100 feature points on the chip surface and 80 feature points on the substrate surface. The SIFT algorithm ensures that each feature point has a unique description vector through key point detection and descriptor generation. The Canny operator is commonly used for edge detection, combined with the Hough transform to extract straight lines. Assuming that the edge of the pad is a rectangle, the straight line equation of the edge is identified through the Hough transform, and then the center coordinates of the pad are determined. The RANSAC algorithm is used to calculate the homography matrix. Assuming that 50 pairs of corresponding points are matched from the SIFT feature points, RANSAC eliminates the mismatched points through iteration and finally obtains a stable homography matrix, which describes the spatial transformation relationship between the chip image and the substrate image. The calculated matrix parameters are compared with the preset deviation threshold (such as 0.1 mm). If the deviation is less than the threshold, it means that the posture error is within the acceptable range and can be assembled directly. If the deviation is large, the compensation control instruction generation module is triggered. Assuming that the deviation matrix shows that the chip is offset by 0.2 mm in the X direction and 0.15 mm in the Y direction, the deviation type is first determined to be a translation deviation. The CCD algorithm is used to solve the inverse kinematics and generate the joint angle compensation value, such as θ1 needs to be increased by 2 degrees and θ2 needs to be reduced by 1.5 degrees. The compensation value is weighted and fused with the initial motion control instruction. The weight is set according to the deviation size and the dynamic characteristics of the robot arm to ensure the best compensation effect. The adjusted motion control instruction is sent to the robot arm control system. The control system adjusts the movement of each joint in real time according to the instruction and corrects the trajectory. Assuming that the robot arm has 6 joints, the control system uses the PID control algorithm to ensure that each joint moves accurately according to the instructions, and finally the alignment error between the chip pad and the substrate pad is less than 0.05 mm. Visual guidance plays a key role in the assembly process. Through real-time image feedback, the robot arm posture is continuously corrected to ensure assembly accuracy. Assuming that the ambient light changes during the assembly process, the visual system needs to dynamically adjust the exposure parameters to maintain the image quality and ensure accurate feature point extraction and posture calculation. The whole process is closely linked, and each step provides accurate data support for the next step, ultimately achieving high-precision assembly of the chip and the substrate. Median filtering and histogram equalization improve image quality, the SIFT algorithm ensures the robustness of feature point extraction, the RANSAC algorithm ensures the stability of the homography matrix, and the CCD algorithm and PID control algorithm achieve accurate correction of the motion trajectory.Through the comprehensive application of these technologies, the efficiency and accuracy of chip assembly have been significantly improved, and the error and rework rate in the assembly process have been reduced.

[0016] Step S103, the welding area is divided into m×n sub-areas, and a micro temperature sensor is deployed at the center of each sub-area to form an m×n temperature sensor array. The measured value of the temperature sensor in the i-th row and j-th column at time t is recorded as T ij (t), in degrees Celsius, where i=1,2,...,m, j=1,2,...,n. Calculate the average temperature of the welding area at time t T(t) = (Σ i=1 m Σ j=1 n T ij (t)) / (m×n). The ideal welding temperature range is defined as [T melt -5, T melt +5], where T melt Indicates the melting point of the solder. The actual average temperature T(t) is compared with the ideal soldering temperature range in real time, and the proportional integral derivative control algorithm is used to adjust the heater power according to the difference between the two.

[0017] According to the size of the welding area, it is divided into m×n sub-areas of equal size. The values ​​of m and n can be determined according to the length and width of the welding area and the required temperature measurement accuracy. A micro temperature sensor is installed at the center of each sub-area. The model and specifications of the sensor are selected according to the requirements of temperature measurement range, accuracy, etc. Common models include thermocouples, thermal resistors and infrared sensors. All temperature sensors are arranged in the form of m rows and n columns to form an m×n two-dimensional sensor array. The spatial position coordinates of the sensor are (i, j), where i is the row number, j is the column number, the value range of i is 1 to m, and the value range of j is 1 to n. The sensor array can be installed above the welding area through a special bracket or fixture, and the sensor probe is kept at a certain distance from the welding surface to avoid affecting the welding process. During the welding process, the real-time temperature measurement value Tij(t) of each sensor in the sensor array is obtained at a certain time interval Δt, where t is the current moment. The temperature data is transmitted to the central control unit in real time through a data acquisition card or a wireless transmission module. The transmission frequency should be determined according to the rate of temperature change and the response speed of the control system. Generally, it can be selected from tens to hundreds of times per second. The central control unit calculates the average temperature T(t) of the welding area at time t according to the formula T(t)=(ΣΣTij(t)) / (m×n), where ΣΣ represents the sum of all i and j. Compare T(t) with the preset ideal welding temperature range [Tmelt-ΔT0, Tmelt+ΔT0], where Tmelt is the melting point temperature of the welding material, and ΔT0 is the allowable temperature deviation range, which can generally be set according to the requirements of the welding process, such as ±5°C. Calculate the deviation ΔT(t)=T(t)-Tmelt between the actual temperature and the midpoint of the ideal temperature range. Take the temperature deviation ΔT(t) as input and use the PID control algorithm to calculate the adjustment amount ΔP(t) of the heater power. The basic principle of the PID control algorithm is to output the optimal control amount based on the size and change trend of the deviation, taking into account the three factors of proportion, integration and differentiation. Among them, the proportional term is proportional to the size of the deviation, the integral term is proportional to the cumulative value of the deviation, and the differential term is proportional to the rate of change of the deviation. The weight coefficients Kp, Ki, and Kd of the three items can be adjusted according to the dynamic characteristics and control requirements of the system. Generally, the trial and error method or self-tuning method can be used to obtain the best parameter combination. The power adjustment amount ΔP(t) output by the PID control algorithm is superimposed on the reference power P0 of the heater to obtain the actual control power P(t)=P0+ΔP(t). The control command is sent to the power control module of the heater to adjust the heating power in real time so that the welding temperature is stable within the ideal range. In order to ensure the accuracy and stability of the control, appropriate adjustment cycles and dead zones can be set to avoid frequent power fluctuations. At the same time, necessary safety protection measures such as over-temperature alarms and burn-out protection are also required to prevent abnormal conditions caused by sensor or heater failures.Through the above method, real-time monitoring and automatic control of the temperature in the welding area can be achieved to ensure welding quality and efficiency. According to different welding objects and process requirements, the arrangement of the sensor array, the parameters of the control algorithm and the target range of the temperature can be flexibly adjusted to obtain the best control effect.

[0018] Specifically, the division of the welding area needs to consider the size and distribution characteristics of the solder joints. Taking chip welding as an example, assuming that the welding area is a square area with a length and width of 40 mm, it can be divided into eight rows and eight columns, a total of 64 sub-areas, and each sub-area is 5 mm by 5 mm in size. Such a division can not only ensure the accuracy of temperature measurement, but also facilitate the arrangement of sensors. The selection of temperature sensors requires weighing multiple factors. For example, copper-nickel thermocouples have high sensitivity and fast response characteristics, and the temperature measurement range can reach minus 40 degrees to 400 degrees, with an accuracy of 0.5 degrees, which is suitable for welding temperature monitoring. Thermocouple sensors are installed in the center of each sub-area to form an eight-by-eight sensor array, and the sensor spacing is maintained at five millimeters to ensure uniform temperature measurement coverage. The real-time temperature acquisition system uses multiplexing to complete full array data acquisition every 0.1 second. Assuming that the melting point of solder paste is 217 degrees, the ideal welding temperature range is set to 212 degrees to 222 degrees. In actual application, the highest temperature in the array was measured to be 225 degrees, the lowest temperature was 210 degrees, and the average temperature was 220 degrees at a certain moment, indicating that the temperature distribution in the welding area was relatively uniform. The temperature control adopts a three-stage adjustment strategy. When the temperature deviation is greater than ten degrees, high-power heating is used for rapid adjustment, medium power is used when the deviation is between five and ten degrees, and low-power fine adjustment is used when the deviation is less than five degrees. For example, when the average temperature is detected to be 205 degrees, it is 12 degrees different from the ideal temperature of 217 degrees, and the controller immediately starts the high-power heating mode. The proportional coefficient is selected as 3.5, the integral time constant is selected as 2 seconds, and the differential time constant is selected as 0.5 seconds. These parameters are determined according to the welding process requirements and the system response characteristics, which can make the temperature quickly reach the target value with good stability. When it is detected that the temperature rises too fast, the differential action can suppress overshoot in time to avoid excessive temperature damage to components. In actual application, the temperature control accuracy can reach plus or minus two degrees, and the heating rate can be controlled at about three degrees per second. This precise temperature control ensures the quality of the solder joint. The system also has a data logging function that can trace the temperature curve of each solder point, facilitating quality analysis and process optimization. The dense arrangement of the temperature sensor array not only achieves accurate temperature measurement, but also monitors local overheating and prevents welding defects in a timely manner.

[0019] Step S104, the vibration signal V of each joint is measured by the piezoelectric sensor installed at the joint of the robot arm. k (t), where k represents the joint number. The frequency spectrum of the vibration signal is obtained by fast Fourier transform, with a frequency range of 0-1000 Hz, from which the dominant vibration frequency f is extracted.d and the corresponding amplitude A d The piezoelectric actuator generates a k (t) The driving signal V with a phase difference of 180 degrees drive (t) = -A d sin(2πf d t + φ), where φ is the initial phase. The piezoelectric actuator is driven to move to offset the vibration at the joint and control the vibration amplitude within the threshold β.

[0020] Step 1: Obtain the vibration signals at each joint of the robot arm. The vibration signals are measured by piezoelectric sensors installed at the joints. Step 2: Perform fast Fourier transform on the obtained vibration signal to obtain the spectrum of the vibration signal. The frequency range of the spectrum is 0-1000Hz. Step 3: Extract the dominant vibration frequency and the corresponding amplitude from the spectrum. The dominant vibration frequency is the frequency component with the largest amplitude in the spectrum. Step 4: Based on the extracted dominant vibration frequency and amplitude, construct a drive signal that is 180 degrees out of phase with the original vibration signal. The expression of the drive signal is V drive (t)=-A d sin(2πf d t+φ), where A d is the amplitude corresponding to the dominant vibration frequency, f d is the dominant vibration frequency, and φ is the initial phase. Step 5: Input the constructed driving signal to the piezoelectric actuator, and the piezoelectric actuator generates corresponding movement according to the driving signal to offset the vibration at the joint of the robotic arm. Step 6: Monitor the vibration signal at the joint of the robotic arm in real time to determine whether the vibration amplitude is within the preset threshold β. If the vibration amplitude exceeds the threshold, return to step 1 to re-acquire the vibration signal and generate a new driving signal until the vibration amplitude is controlled within the threshold. Step 7: According to the feedback signals of the piezoelectric sensor and the piezoelectric actuator, an adaptive control algorithm is used to continuously optimize the parameters of the driving signal, such as amplitude and phase, to achieve the optimal effect of vibration suppression and improve the smoothness and accuracy of the robotic arm movement.

[0021] Specifically, in the vibration control process of the robot arm, the vibration signal is first obtained through the piezoelectric sensor installed at the joint. The piezoelectric sensor uses the piezoelectric effect to convert mechanical vibration into an electrical signal, thereby achieving high-sensitivity measurement of vibration. For example, assuming that when the robot arm performs a high-speed handling task, the vibration frequency generated at the joint is mainly concentrated in the range of 200-500Hz, and the sensor can accurately capture these vibration signals. Next, the acquired vibration signal is subjected to a fast Fourier transform (FFT) to convert it from the time domain to the frequency domain to obtain the spectrum of the vibration signal. The frequency range of the spectrum is set to 0-1000Hz, which can cover all the main vibration frequencies that may occur in the robot arm. Through FFT analysis, the amplitude distribution of different frequency components can be clearly seen. Assume that in the spectrum, the amplitude at 300Hz is the largest, indicating that this frequency is the dominant vibration frequency. After extracting the dominant vibration frequency and its corresponding amplitude from the spectrum, a drive signal with a phase difference of 180 degrees from the original vibration signal is constructed. The expression of the drive signal is V drive (t)=-A d sin(2πf d t+φ), where A d is the amplitude corresponding to the dominant vibration frequency, f d is the dominant vibration frequency, and φ is the initial phase. For example, if the dominant vibration frequency is 300Hz and the amplitude is 0.5V, the drive signal is V drive(t)=-0.5sin(2π×300t+φ). This signal with opposite phase can effectively offset the original vibration signal. The constructed driving signal is input to the piezoelectric actuator, and the piezoelectric actuator generates corresponding reverse motion according to the driving signal to offset the vibration at the joint of the manipulator. The piezoelectric actuator has the characteristics of fast response speed and high precision, and can quickly generate mechanical vibration corresponding to the driving signal. For example, when the driving signal is -0.5sin(2π×300t+φ), the actuator will generate corresponding reverse vibration, thereby effectively suppressing the dominant vibration of 300Hz. The vibration signal at the joint of the manipulator is monitored in real time to determine whether the vibration amplitude is within the preset threshold β. Assuming that the threshold β is set to 0.1V, if the vibration amplitude is monitored to exceed 0.1V, it means that the vibration suppression effect is not ideal, and it is necessary to re-acquire the vibration signal and generate a new driving signal. This closed-loop control strategy ensures the real-time and effectiveness of vibration suppression. According to the feedback signals of the piezoelectric sensor and the piezoelectric actuator, the adaptive control algorithm is used to continuously optimize the parameters of the driving signal, such as amplitude and phase. The adaptive control algorithm can dynamically adjust the drive signal based on the real-time monitoring data to more accurately offset the vibration. For example, if the initial phase φ is set to 0, but the actual effect is not good, the algorithm will automatically adjust the φ value until the optimal phase is found to achieve the best vibration suppression effect. This vibration control method can not only improve the smoothness and accuracy of the robot movement, but also extend the service life of the robot. Through precise vibration suppression, the wear of the robot joints can be reduced, and mechanical failures caused by excessive vibration can be avoided. In addition, the data recording function can trace the vibration curve of each joint, which is convenient for quality analysis and process optimization, and further improves the overall performance of the robot. In a specific application, it is assumed that when the robot is performing a precision assembly task, the vibration of the joint will cause the assembly accuracy to decrease. Through the above vibration control method, the vibration amplitude of the joint can be controlled within 0.05V, which significantly improves the assembly accuracy. For example, in high-speed handling tasks, vibration control can reduce the jitter of the robot arm and improve handling efficiency and safety. In short, through the high-sensitivity measurement of piezoelectric sensors, FFT spectrum analysis, construction of drive signals with opposite phases, precise response of piezoelectric actuators, and dynamic optimization of adaptive control algorithms, a complete vibration control system has been formed to ensure that the robotic arm can maintain high-precision and high-stability motion performance under various working conditions.

[0022] Step S105, establish a thermal-mechanical coupling model including the chip and the substrate, the model is based on the welding process parameter set S = {P, T, t}, where P represents welding pressure, T represents welding temperature, and t represents welding time. The output is the evaluation index Q of welding quality, including the void rate V void , interface bonding strength σ bondFinite element analysis software is used to perform numerical simulation on parameter set S to determine welding process parameters. The key parameter association rule extraction algorithm based on Gaussian process regression is used on the first data set obtained from the numerical simulation process to extract the key process parameters affecting welding quality and the quantitative relationship between them.

[0023] According to the material properties and geometric dimensions of the chip and substrate, a three-dimensional finite element model including multiple physical fields such as heat conduction, heat convection, and heat radiation is established, and the mechanical deformation and residual stress in the welding process are considered. The welding process parameter set S is used as the input condition of the finite element model, and the temperature field, stress field, and deformation field distribution under different parameter combinations are obtained through numerical simulation. According to the numerical simulation results, the characteristic parameters such as the maximum temperature, temperature gradient, cooling rate, and equivalent plastic strain of the solder joint area are extracted, and the quantitative relationship with the void rate and interface bonding strength is established. The Gaussian process regression algorithm is used to train the key parameter association model with the welding process parameters as independent variables and the void rate and interface bonding strength as dependent variables to obtain the influence weight and sensitivity of each process parameter on the welding quality. Through cross-validation and grid search methods, the kernel function type and hyperparameters of the Gaussian process regression model are optimized to improve the fitting accuracy and generalization ability of the association rules. The trained association model is used to predict the void rate and interface bonding strength under different welding process parameters, and the parameter combination that meets the quality requirements is screened out and experimentally verified. Based on the obtained key parameter association rules, an optimization model of welding process parameters is established. With void ratio and interface bonding strength as objective functions and welding pressure, temperature and time as decision variables, the optimal parameter combination is solved to guide actual welding production.

[0024] Specifically, according to the material properties and geometric dimensions of the chip and substrate, a three-dimensional finite element model containing multiple physical fields such as heat conduction, heat convection, and heat radiation is first established. For example, assume that the chip material is silicon, the substrate material is ceramic, and the geometric dimensions are 10mm×10mm and 20mm×20mm, respectively. The model considers the mechanical deformation and residual stress during the welding process to ensure the accuracy of the simulation results. The welding process parameter set S (such as welding temperature, welding time, and welding pressure) is used as the input condition of the finite element model, and the temperature field, stress field, and deformation field distribution under different parameter combinations are obtained through numerical simulation. In the numerical simulation process, it is assumed that the welding temperatures are 250℃, 300℃, and 350℃, respectively, the welding times are 10 seconds, 20 seconds, and 30 seconds, respectively, and the welding pressures are 0.5MPa, 1MPa, and 1.5MPa, respectively. Through simulation, the temperature field distribution diagram, stress field distribution diagram, and deformation field distribution diagram under each parameter combination can be obtained. For example, under the conditions of 300℃, 20 seconds, and 1MPa, the maximum temperature of the solder joint area may reach 320℃, the temperature gradient is 50℃ / mm, the cooling rate is 10℃ / s, and the equivalent plastic strain is 0.02. According to the numerical simulation results, the characteristic parameters such as the maximum temperature, temperature gradient, cooling rate, and equivalent plastic strain of the solder joint area are extracted to establish a quantitative relationship with the void rate and interface bonding strength. For example, through data analysis, it is found that the higher the maximum temperature, the greater the void rate; the greater the temperature gradient, the lower the interface bonding strength. Specifically, when the maximum temperature increases from 300℃ to 350℃, the void rate increases from 2% to 5%; when the temperature gradient increases from 50℃ / mm to 60℃ / mm, the interface bonding strength decreases from 50MPa to 40MPa. The Gaussian process regression algorithm is used to train the key parameter association model with welding process parameters as independent variables and void rate and interface bonding strength as dependent variables. For example, the input parameters are welding temperature 300℃, welding time 20 seconds, welding pressure 1MPa, and the output is void rate 3% and interface bonding strength 45MPa. Through multiple trainings, the weight and sensitivity of each process parameter on welding quality are obtained. It is assumed that the influence weight of welding temperature on void rate is the largest, and the influence weight of welding time on interface bonding strength is second. The kernel function type and hyperparameters of the Gaussian process regression model are optimized through cross-validation and grid search. For example, the radial basis function (RBF) is selected as the kernel function, and the optimal hyperparameter combination is determined through grid search, such as the length scale parameter is 1.5, the signal variance is 0.1, and the noise variance is 0.01. The fitting accuracy and generalization ability of the optimized model are significantly improved, and the prediction error is reduced from 5% to 2%. The trained association model is used to predict the void rate and interface bonding strength under different welding process parameters, and the parameter combination that meets the quality requirements is screened out. For example, it is predicted that when the welding temperature is 280°C, the welding time is 15 seconds, and the welding pressure is 1.2MPa, the void rate is 1.5% and the interface bonding strength is 55MPa, which meets the quality requirements.Experimental verification was carried out, and the actual measurement results were consistent with the predicted results, verifying the accuracy of the model. Based on the key parameter association rules obtained, an optimization model for welding process parameters was established, with void rate and interface bonding strength as the objective function, and welding pressure, temperature and time as decision variables to solve the optimal parameter combination. For example, the objective function is set to minimize the void rate and maximize the interface bonding strength, and the optimal parameter combination is obtained by solving the optimization algorithm (such as genetic algorithm) as welding temperature 290℃, welding time 18 seconds, and welding pressure 1.3MPa. This parameter combination is applied in actual welding production, which significantly improves the welding quality, reduces the void rate, enhances the interface bonding strength, and extends the service life of the product. Through the above steps, a complete set of welding process parameter optimization system is formed to ensure that high-quality welding effects can be obtained under different working conditions. The implementation of this method not only improves production efficiency, but also reduces the rework rate caused by welding quality problems, and improves the reliability and market competitiveness of products.

[0025] Based on the key parameter association rule extraction algorithm of Gaussian process regression, a welding process parameter optimization model is constructed. Combined with the real-time production data generated on the production line, the gradient descent method is used for iterative optimization. In the optimization process, the objective function J(θ) = Σ(Q pred -Q actual ) 2 , where θ represents the model parameters, Q pred is the predicted welding quality, Q actual The actual welding quality. By minimizing the objective function J(θ), the model parameters are continuously updated. The optimized welding process parameter optimization model is used to predict the welding quality under given process parameter conditions, guide the production process, and set process parameters.

[0026] According to the welding process parameters and corresponding welding quality data collected on the production line, a Gaussian process regression model is constructed to establish a mapping relationship between process parameters and welding quality. The welding process parameter data collected in real time on the production line is obtained and input into the Gaussian process regression model to predict the welding quality under the current process parameter setting. The predicted welding quality is compared with the actual measured welding quality, and the mean square error between the two is calculated as the objective function of model optimization. Using the objective function, the hyperparameters of the Gaussian process regression model are iteratively optimized using the gradient descent method. Hyperparameters include parameters of the kernel function and the variance of the noise term. According to the objective function, the partial derivative of the hyperparameter is calculated to obtain the gradient direction of the hyperparameter, and the hyperparameter is updated along the gradient direction. The iteration is repeated until the value of the objective function is less than the preset convergence threshold. The convergence threshold can be set according to the fitting error of the historical data, for example, it can be set to the mean square error of the historical optimal model. If the value of the objective function is continuously greater than the convergence threshold during the iteration process, the learning rate can be adjusted or the number of iterations can be increased. The welding process parameters that need to be set on the production line are input into the optimized Gaussian process regression model to predict the welding quality under the process parameter setting. According to the gap between the predicted welding quality and the target quality, it is judged whether the current process parameter setting meets the quality requirements. A quality error tolerance range can be set. If the error between the predicted quality and the target quality is within the tolerance range, it is considered to meet the requirements. If the predicted quality does not meet the requirements, the process parameters need to be adjusted. The probability prediction results of the Gaussian process regression model can be used to obtain the probability distribution of welding quality under different process parameters, and the process parameters with the highest quality probability are selected as the new parameter settings. The process of quality prediction and parameter adjustment is repeated until the optimal process parameters that meet the quality requirements are found.

[0027] Specifically, when collecting welding process parameters on the production line, the sampling frequency and quality of the data need to be considered. For example, for welding temperature, data can be collected ten times per second to ensure that rapid changes in temperature can be captured. For a production line, assuming that one thousand products are produced every day, three key parameters are collected during the welding process of each product: welding temperature, pressure and time, and the corresponding welding quality indicators such as void rate and interface bonding strength are recorded at the same time. When building a Gaussian process regression model, the collected data must first be preprocessed. For example, standardizing the data with welding temperatures between 250 and 350 degrees to a range of zero to one can improve the training effect of the model. For a specific welding workpiece, assuming that the collected temperature is 300 degrees, the pressure is 1 MPa, and the time is 20 seconds, the corresponding void rate is 3%. In the model training process, it is crucial to choose a suitable kernel function. The radial basis function is a common choice, in which the length scale parameter determines the correlation between data points. For example, when the length scale parameter is set to 0.5, it means that there is a strong correlation between adjacent data points, which is conducive to capturing the nonlinear relationship between welding parameters and quality. When predicting welding quality, the model not only gives a point estimate, but also provides the uncertainty of the prediction. For example, when the input welding temperature is 280 degrees, the model predicts a void rate of 2.5 percent, with a confidence interval of 2 to 3 percent. This probabilistic prediction result helps to optimize the process parameters. The model performance is evaluated by calculating the mean square error between the predicted value and the measured value. Assume that on the validation set, the predicted mean square error of the void rate is 0.02, indicating that the model has good prediction accuracy. If the error exceeds the preset threshold of 0.05, the hyperparameters of the model need to be adjusted. In the hyperparameter optimization process, the gradient descent method is used for iterative update. The initial learning rate is set to 0.01, and the hyperparameters are adjusted according to the gradient direction of the objective function in each iteration. If the improvement of the objective function is less than 0.001 for five consecutive iterations, the optimization process is considered to converge. Finally, the optimized model can be used to guide the parameter setting of the production line. When the welding temperature needs to be adjusted to 290 degrees, the model predicts a void rate of 2%, which is lower than the upper limit of the quality control of 3%, so this temperature parameter can be used. If the predicted void ratio exceeds the standard, it is necessary to look for other parameter combinations until the optimal parameter setting that meets the quality requirements is found.

[0028] The above embodiment is only one of the preferred implementation modes of the present invention and should not be used to limit the protection scope of the present invention. Any changes or modifications that are made to the main design concept and spirit of the present invention and have no substantive significance, and the technical problems they solve are still consistent with the present invention, should be included in the protection scope of the present invention.

Claims

1. A chip welding method based on machine vision and control, characterized in that: Acquire the three-dimensional model of the chip and substrate, get the initial relative position of the chip solder joint and the substrate pad, and generate the initial motion command of the robot arm; Capture the images of the chip and substrate, calculate the posture deviation through image registration, and if the deviation exceeds the threshold, generate a compensation command and merge it with the initial command to adjust the robot arm trajectory to align with the pad; Construct a temperature sensor array to monitor the temperature of the welding area and calculate the average temperature. If there is a difference between the average temperature and the ideal welding temperature range, adjust the heater power through the proportional integral differential algorithm; Use piezoelectric sensors to obtain the vibration signal of the robot arm joint, extract the dominant vibration frequency and amplitude through fast Fourier transform, and generate an anti-phase drive signal through the piezoelectric actuator to offset the vibration; A thermal-mechanical coupling model is established, which takes the welding process parameter set as input and outputs the welding quality evaluation index. The Gaussian process regression algorithm is used to extract the key parameter association rules. According to the real-time data, the gradient descent method is used to optimize the parameters and generate the optimized process parameters.

2. The chip welding method based on machine vision and control according to claim 1 is characterized in that: The method of obtaining the three-dimensional model of the chip and the substrate and obtaining the initial relative positions of the chip solder joints and the substrate pads includes: obtaining a set of discrete coordinate points of the outer boundaries of the chip and the substrate, using a B-spline surface fitting algorithm to obtain a continuous surface model, constructing a three-dimensional mesh model, extracting coordinate information of the solder joints and pads, simulating an ideal welding state, calculating the target coordinates of the solder joints and the target pose matrix of the end effector of the robot arm, establishing a forward kinematics model of the robot arm according to the DH parameters of the robot arm, and obtaining a numerical solution of the joint angles through an inverse kinematics solution algorithm of the robot arm.

3. The chip welding method based on machine vision and control according to claim 2 is characterized in that: The method of obtaining a discrete coordinate point set of the outer boundaries of the chip and the substrate, obtaining a continuous surface model by using a B-spline surface fitting algorithm, and constructing a three-dimensional mesh model includes: Denoted as Sc and Ss respectively, the discrete coordinate point sets Sc and Ss are transformed into continuous surface models using the B-spline surface fitting algorithm; Based on the surface model, the Delaunay triangulation algorithm is used to construct the three-dimensional mesh models Mc and Ms of the chip and substrate; By extracting features from the three-dimensional mesh models Mc and Ms, the solder joints on the chip and the corresponding pads on the substrate are identified, and their coordinate information is obtained; Using the acquired solder joint and pad coordinates, the finite element analysis method is used to simulate the ideal welding state between the chip solder joint and the substrate pad, and the target coordinates (xf, yf, zf) of the solder joint in three-dimensional space and the target pose matrix Tgoal of the end effector of the robot arm are calculated; According to the DH parameters of the robot arm, the forward kinematics model of the robot arm is established to describe the mathematical relationship between the joint angles of the robot arm and the end posture.

4. The chip welding method based on machine vision and control according to claim 1, characterized in that: The method of shooting chip and substrate images and calculating posture deviation through image registration includes: acquiring images of the chip and substrate through a high-precision visual sensor and preprocessing them, extracting a set of feature points from the images to obtain the positions of feature points on the chip image and the substrate image, using a robust method for anti-error estimation to calculate the homography matrix to obtain spatial posture deviation matrix parameters, using a robot arm kinematic inverse algorithm to generate joint angle compensation values, weighted fusion of the joint angle compensation values ​​and the initial motion control instructions of the robot arm to obtain adjusted motion control instructions, and aligning the chip pads with the substrate pads through visual guidance.

5. The chip welding method based on machine vision and control according to claim 1, characterized in that: The method of constructing a temperature sensor array to monitor the temperature of the welding area and calculate the average temperature includes: dividing the welding area into sub-areas, deploying a micro temperature sensor array in the center of each sub-area, obtaining the real-time temperature value of each sensor in the array, summing the measured values ​​of all temperature sensors and calculating the average temperature value of the welding area, using a traversal optimization method to obtain the optimal welding temperature threshold, calculating the deviation between the actual temperature and the midpoint of the ideal temperature range, and using the welding process parameter set as the input condition of the finite element model.

6. The chip welding method based on machine vision and control according to claim 5, characterized in that: The real-time temperature value of each sensor in the array is obtained, and the measured values ​​of all temperature sensors are summed up and the average temperature value of the welding area is calculated, including: The welding area is divided into m×n sub-areas, and a micro temperature sensor is deployed at the center of each sub-area to form an m×n temperature sensor array; The measured value of the temperature sensor in the i-th row and j-th column at time t is denoted as T ij (t), in degrees Celsius, where i = 1, 2, ..., m, j = 1, 2, ..., n; Calculate the average temperature of the welding area at time t T(t) = (Σ i=1 m Σ j=1 n T ij (t)) / (m×n); The ideal welding temperature range is defined as [T melt -5, T melt +5], where T melt Indicates the melting point of solder; The actual average temperature T(t) is compared with the ideal welding temperature range in real time, and the heater power is adjusted according to the difference between the two using the proportional integral derivative control algorithm.

7. The chip welding method based on machine vision and control according to claim 1, characterized in that: The method of obtaining the vibration signal of the joints of the robotic arm by using a piezoelectric sensor and extracting the dominant vibration frequency and amplitude by fast Fourier transform includes: obtaining the vibration signal of each joint of the robotic arm according to the piezoelectric sensor, calculating the frequency spectrum of each joint vibration signal by fast Fourier transform, extracting the frequency with the largest amplitude in the spectrum of each joint as the dominant vibration frequency, accumulating all amplitudes to obtain the amplitude corresponding to the dominant vibration frequency, and constructing an anti-phase drive signal according to the extracted dominant vibration frequency and amplitude.

8. The chip welding method based on machine vision and control according to claim 1, characterized in that: The method of generating an anti-phase drive signal by a piezoelectric actuator to offset vibration includes: real-time monitoring of the vibration signal at the joint of the robotic arm, determining whether the vibration amplitude exceeds a threshold value, if the vibration amplitude exceeds the threshold value, reacquiring the vibration signal and generating a new drive signal, using an adaptive control algorithm to continuously optimize the parameters of the drive signal based on the feedback signal of the piezoelectric sensor and the piezoelectric actuator, adjusting the heating power in real time to stabilize the welding temperature within an ideal range, and sending control instructions to the power control module of the heater.

9. The chip welding method based on machine vision and control according to claim 1, characterized in that: The thermal-mechanical coupling model is established, which takes the welding process parameter set as input and outputs welding quality evaluation indicators, including: establishing a three-dimensional finite element model according to the material properties and geometric dimensions of the chip and substrate, simulating the temperature field, stress field and deformation field distribution under different parameter combinations, extracting the maximum temperature, temperature gradient, cooling rate, and equivalent plastic strain characteristic parameters of the solder joint area according to the simulation results, and using a finite element model calculation method based on data consistency convergence to obtain key parameters of heat conduction, heat convection, and heat radiation.

10. The chip welding method based on machine vision and control according to claim 1, characterized in that: The method uses a Gaussian process regression algorithm to extract key parameter association rules, and optimizes parameters using a gradient descent method based on real-time data, including: using a Gaussian process regression algorithm, with welding process parameters as independent variables, and void ratio and interface bonding strength as dependent variables, training a key parameter association model to obtain the weight and sensitivity of each process parameter on welding quality, optimizing the kernel function type and hyperparameters of the Gaussian process regression model through methods such as cross-validation and grid search, comparing the predicted welding quality with the actually measured welding quality, and obtaining the mean square error between the two as the objective function for model optimization.

Images