Gold finger plate inner lead laser processing method and system based on intelligent feedback

Abstract

The application relates to the technical field of laser precision machining, and discloses a gold finger plate inner lead laser processing method and system based on intelligent feedback, which comprises the following steps: collecting molten pool shapes and heat affected zone images in real time through high-frame-rate coaxial visual sensing, extracting key feature parameters through image preprocessing; combining multi-physical field coupling modeling to solve heat-flow-solid equations in real time, and outputting temperature gradient and solidification rate physical quantities; dynamically adjusting laser power, scanning speed and focusing position by an adaptive neural network decision engine; and ensuring control stability and accurate connection of paths through a delay compensation and steady state judgment module; through a closed-loop intelligent feedback mechanism, the gold finger plate inner lead laser processing method and system can compress the lead width tolerance to ±0.8 microns, reduce the heat affected zone to within 4 microns, reduce the oxidation layer thickness to below 5 nanometers, improve the yield to 99.5%, reduce energy consumption by 40%, and significantly improve the machining consistency and industrial benefits.

Description

Technical Field

[0001] This invention belongs to the field of laser precision machining technology, specifically relating to a laser machining method and system for internal leads of gold finger boards based on intelligent feedback. Background Technology

[0002] With the widespread application of high-density interconnect circuit boards in 5G communications, artificial intelligence servers, and high-performance computing devices, the requirements for precision and yield in the gold finger area are becoming increasingly stringent. Traditional mechanical milling and chemical etching processes rely on fixed tool paths and preset etching parameters. Their processing is based on static design drawings and empirical thresholds, lacking real-time perception capabilities of material heterogeneity, thermal stress deformation, and surface oxide layer distribution within the board. In complex board types with multi-layer stacking, micron-level linewidths, and asymmetrical layouts, such static processing strategies are prone to causing burrs on lead edges, uncontrolled linewidth consistency, or localized gold layer peeling, leading to signal integrity degradation and contact impedance fluctuations, severely restricting the electrical reliability of high-speed interfaces. Furthermore, production lines are continuously upgrading their demands for processing efficiency, defect self-repair, and dynamic adaptation of process parameters. Traditional methods, lacking a closed-loop feedback mechanism, struggle to achieve coordinated optimization of defect identification, parameter self-correction, and dynamic trajectory compensation within a single processing flow.

[0003] Among these technologies, laser-based non-contact wire bonding has become a mainstream alternative due to its small heat-affected zone and strong path programmability. Its core principle is to selectively ablate non-conductive layers with a high-energy beam to expose the metal areas of the gold fingers. However, existing laser systems still use an open-loop control architecture. Processing parameters such as power density, scanning speed, and focusing focal length are set once before operation, making it impossible to dynamically adjust them at the millisecond level based on changes in board reflectivity, local copper foil thickness deviations, or fluctuations in ambient temperature and humidity. This results in significant variations in processing quality across different areas of the same board. Especially in mass production, micro-morphological variations caused by batch differences in substrates or residual stress from previous processes often lead to a sharp drop in laser energy coupling efficiency or uncontrolled thermal accumulation effects, resulting in jagged lead profiles, thermal damage to the gold layer, or accidental ablation of adjacent pads.

[0004] Existing technologies generally suffer from three major defects: a broken perception-decision-execution chain, rigid adaptation of processing parameters, and lag in defect response. This makes it difficult to simultaneously ensure lead geometry accuracy, surface integrity, and process robustness when dealing with highly complex board shapes, multivariate interference scenarios, and zero-fault-tolerant electrical performance requirements. There is an urgent need to build a new paradigm for laser processing with intelligent feedback closed loop. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a laser processing method and system for in-plate leads based on intelligent feedback. By constructing a closed-loop dynamic sensing and real-time control mechanism, it achieves precise control over the entire process of morphological evolution, thermal stress distribution, and material phase transformation behavior of the microstructure in the gold finger region during laser processing. This system integrates high-resolution in-situ visual sensing, multi-physics coupling modeling, and adaptive laser parameter control modules. During processing, it can dynamically adjust the laser power density, scanning speed, pulse frequency, and focusing position based on real-time acquired molten pool morphology, edge heat-affected zone width, and surface reflectivity changes. This ensures lead conductivity while minimizing the negative effects of thermally induced warping, metal recrystallization coarsening, and interface oxide layer thickening, achieving synergistic optimization of the three objectives: micron-level precision, submicron-level surface roughness, and high yield consistency.

[0006] According to one aspect of this application, a method for laser processing of internal leads in a gold finger board based on intelligent feedback and a control method mentioned in the system are provided, comprising:

[0007] The high-frame-rate coaxial vision sensing unit acquires real-time images of the molten pool contour and gray-scale gradient distribution of the edge heat-affected zone in the current laser-acting area. The vision sensing unit is set in the optical path inside the laser processing head, and its imaging plane has a zero-degree angle with the normal to the workpiece surface to ensure no perspective distortion. The image resolution reaches more than 200 pixels per millimeter, and the frame rate is no less than 1,000 frames per second. The light source uses a near-infrared light-emitting diode with a center wavelength of 850 nanometers, a radiation intensity of 850 milliwatts per spherical degree, and a half-intensity angle of ±25 degrees to penetrate the plasma plume and obtain the true morphology of the molten pool.

[0008] The image preprocessing module performs background subtraction, median filtering, and edge sharpening on the acquired raw image. It extracts four key feature parameters: the long axis dimension of the molten pool, the short axis dimension, the width of the edge heat-affected zone, and the mean surface reflectance. The image preprocessing module uses a two-dimensional convolution kernel with a fixed kernel size of 3x3 to perform spatial domain filtering. Edge detection uses the Sobel operator combined with the non-maximum suppression algorithm. Reflectance is calculated based on the integral area ratio of the image gray-level histogram within the preset gray-level range.

[0009] The multiphysics coupling modeling unit receives four key characteristic parameters and, combined with the current laser processing parameters, solves in real time the coupled equation set consisting of heat conduction equation, fluid dynamics equation, and phase transition dynamics equation. The heat conduction equation adopts a three-dimensional unsteady-state form, and the thermal conductivity of the material changes as a piecewise linear function with temperature. The fluid dynamics equation uses the Navier-Stokes equation to describe the convection behavior inside the molten pool, and the phase transition dynamics equation uses the Johnson-Mell equation to describe the propagation rate of the solid-liquid interface. All equations are discretized using the finite volume method, with a time step set to ten nanoseconds, a spatial grid size of 0.5 micrometers, and a computational domain covering a 50-micrometer cubic region around the laser point of action.

[0010] The model output interface obtains four physical quantities: the temperature gradient at the bottom of the molten pool, the solidification rate of the sidewall, the principal component of thermal stress, and the predicted value of the oxide layer thickness at the current moment. These physical quantities are used as feedback control variables and input to the adaptive control decision engine. The decision engine adopts a two-layer feedforward neural network structure with four nodes in the input layer, eight nodes in the hidden layer, and four nodes in the output layer. The activation function is a modified linear unit, and the weight matrix is ​​obtained through offline training. The training dataset contains 10,000 sets of simulation results and measured yield data of the processing process under different materials and different combinations of initial parameters. The loss function is a weighted sum of mean square error and yield.

[0011] The control instruction generation module converts the four control quantities output by the neural network into laser drive signals and scanning galvanometer control signals. The laser drive signal adjusts the output power density, with an adjustment range of 10 watts to 500 watts per square millimeter and an adjustment resolution of 1 watt per square millimeter. The scanning galvanometer control signal adjusts the scanning speed and path curvature, with the scanning speed adjustment range of 10 millimeters to 500 millimeters per second and an adjustment resolution of 1 millimeter per second. The path curvature is reconstructed in real time through cubic spline interpolation to ensure that the trajectory continuity and acceleration constraints meet the physical limits of the mechanical system.

[0012] The control command is corrected for time delay by a closed-loop feedback delay compensation module. The delay includes four parts: image acquisition delay, data transmission delay, model calculation delay, and actuator response delay. The total delay time is calibrated to twelve milliseconds through a step response experiment. The compensation algorithm adopts a zero-order hold combined with a feedforward predictor structure. The predictor input is the trend of the control quantity change in the most recent three sampling cycles, and the output is the advance estimate of the control quantity in the next cycle, ensuring the closed-loop stability of the system.

[0013] The processing termination judgment module continuously monitors the convergence of the molten pool morphology and the stability of the heat-affected zone width. When the fluctuation of the long axis dimension of the molten pool is less than 0.5 micrometers and the change rate of the heat-affected zone width is less than 0.5 percent within five consecutive sampling periods, it is determined that the current lead processing segment has reached a steady state, triggering the processing path advancement step. The advancement step adopts incremental displacement control, and the advancement distance each time is 80% of the designed width of the lead, ensuring that there is a 20% overlap area between adjacent processing segments, avoiding conduction breaks caused by processing gaps.

[0014] The global path planning module generates a processing path sequence based on the lead topology in the circuit board design file. The path sequence is sorted according to the principle of minimum empty travel. Path turning points adopt arc transitions with an arc radius of not less than twice the lead width. During path execution, the processing parameters and feedback characteristics of each segment are recorded in real time to form a digital twin archive of the processing process for subsequent quality traceability and process optimization.

[0015] According to another aspect of this application, a method for laser processing of in-plate leads based on intelligent feedback and a control system mentioned in the system are provided, comprising:

[0016] The high frame rate coaxial vision sensing unit is used to acquire images of the molten pool morphology and heat-affected zone in real time during laser processing. Its optical system includes an objective lens, a beam splitter, a filter, and an image sensor. The objective lens has a numerical aperture of 0.8 and a working distance of 10 mm. The beam splitter transmits 90% of the laser energy to the workpiece surface and guides 10% of the reflected light to the image sensor. The filter has a center wavelength of 850 nm and a bandwidth of 20 nm. The image sensor uses a global shutter type CMOS device with a pixel size of 2.2 μm by 2.2 μm and a maximum frame rate of 1200 frames per second.

[0017] The image preprocessing module performs noise suppression and feature extraction operations on the original image. Its hardware carrier is a field-programmable gate array chip, and the internal image processing pipeline is embedded. The pipeline includes a background modeling unit, a spatial filtering unit, an edge detection unit, and a gray-level statistics unit. Background modeling adopts the sliding window mean method with a window size of 50 x 50 pixels. Spatial filtering adopts a 3 x 3 median filter. Edge detection adopts a parallel Sobel operator to calculate the gradient magnitude and direction. Gray-level statistics adopts a histogram accumulator to calculate the pixel ratio of a specified gray-level interval.

[0018] The multiphysics coupling modeling unit is used to solve the thermo-fluid-structure interaction equations in real time and output key physical quantities. Its computing core is a graphics processor array, which contains four computing nodes. Each node is equipped with three thousand stream processors and eight gigabytes of video memory. The computing tasks are distributed to each node using a spatial domain decomposition strategy. Communication adopts a point-to-point direct memory access mechanism to ensure that the data exchange latency is less than one microsecond. The equation solver uses the preprocessed conjugate gradient method to accelerate convergence. The convergence criterion is that the residual norm is less than 10 to the power of negative six.

[0019] The adaptive control decision engine generates laser parameter adjustment instructions based on physical quantity feedback. Its neural network model is stored in non-volatile memory. The inference engine is implemented using a tensor processing unit, supports eight-bit integer quantization inference, and the time for a single inference is less than 0.5 milliseconds. Input data preprocessing includes normalization and moving average filtering. The normalization coefficients are stored in a lookup table, and the sliding window length is five. Output data postprocessing includes dead zone suppression and rate limiting. The dead zone width is two percent of the control range, and the rate limit is that the adjustment amount per millisecond does not exceed five percent of the total range.

[0020] The control command generation module is used to convert digital control quantities into analog drive signals. It includes a digital-to-analog converter, a power amplifier, and a current loop controller. The digital-to-analog converter has a resolution of 16 bits and an update rate of one million times per second. The power amplifier has a bandwidth of 100 kHz and an output current range of 0 to 20 amperes. The current loop controller adopts a proportional-integral structure with a proportional gain of 0.5 and an integral time of 10 milliseconds to ensure that the laser power response has no overshoot and the steady-state error is less than 0.1%.

[0021] The closed-loop feedback delay compensation module is used to correct the impact of the inherent system delay on control stability. It includes a timestamp recorder, a delay measuring device, and a feedforward predictor. The timestamp recorder marks the system clock at the moment of image acquisition. The delay measuring device obtains the end-to-end delay through loop closure testing. The feedforward predictor uses a third-order polynomial to fit the trend of the control variable change in the most recent three cycles. The prediction step size is one sampling period. The prediction result is weighted and summed with the current feedback variable before being output. The weighting coefficients are obtained through offline identification.

[0022] The processing termination judgment module is used to determine whether the current processing segment has reached a steady state and trigger path advancement. Its logic judgment unit adopts a state machine structure, which includes five states: initialization state, acquisition state, calculation state, judgment state, and advancement state. The state transition condition is based on the characteristic parameter change rate threshold, which is stored in a register. The advancement command is sent to the motion control card through the pulse output interface. The pulse frequency is proportional to the advancement distance, and the pulse width is ten microseconds to ensure positioning accuracy better than 0.1 micrometer.

[0023] The global path planning module is used to generate the optimal processing path sequence and manage processing data. It runs on an industrial control computer with a real-time Linux kernel operating system. The path planning algorithm adopts an improved genetic algorithm with a population size of one hundred, a crossover probability of 0.8, and a mutation probability of 0.1. The fitness function includes a weighted sum of path length, number of turns, and heat accumulation. The processing data is stored in a solid-state drive and adopts a time-series database structure. Each record contains five types of fields: timestamp, location coordinates, laser parameters, feedback features, and yield label.

[0024] Compared with the prior art, the advantages and positive effects of this application are as follows:

[0025] The closed-loop intelligent feedback system constructed in this application fundamentally solves the three major technical bottlenecks faced by traditional open-loop laser processing methods in the manufacturing of gold finger lead wires: uncontrollable precision, irreversible thermal damage, and unstable yield. By introducing in-situ visual sensing and real-time multiphysics modeling, it achieves, for the first time, millisecond-level perception and micron-level quantification of the dynamic behavior of the molten pool during processing, transforming the process parameter setting, which previously relied on experience-based trial and error, into precise calculation and dynamic correction based on physical mechanisms. The introduction of the adaptive control engine enables the system to automatically maintain the consistency of processing quality under conditions of material batch fluctuations, environmental temperature changes, or equipment aging disturbances, compressing the lead wire width tolerance from ±3 micrometers in traditional methods to within ±0.8 micrometers, and reducing the surface roughness from 0.5 micrometers to below 0.08 micrometers.

[0026] In terms of thermal management, by real-time adjustment of the synergistic relationship between laser power density and scanning speed, the lateral expansion of the heat-affected zone (HAZ) was effectively suppressed, reducing the HAZ width from an average of twelve micrometers to less than four micrometers. Simultaneously, by controlling the molten pool solidification rate, excessive growth of columnar crystals was avoided, refining the recrystallized grain size from fifty micrometers to less than five micrometers, significantly improving the mechanical strength and fatigue resistance of the leads. Regarding oxidation control, by maintaining the molten pool surface temperature below the critical oxidation threshold and shortening the high-temperature exposure time, the thickness of the interface oxide layer was reduced from fifty nanometers in traditional processes to less than five nanometers, ensuring low-impedance ohmic contact between the leads and the pads.

[0027] In terms of system robustness, the closed-loop delay compensation mechanism effectively overcomes the inherent time lag in the sensing, computing, and execution stages, enabling the system to maintain stable convergence even with a total delay of twelve milliseconds, avoiding oscillations or divergences caused by control lag. The combination of global path planning and digital twin archives not only optimizes processing efficiency and reduces idle travel time by more than 30%, but also provides complete data support for subsequent quality traceability and process iteration, shortening the new product introduction cycle from two weeks in traditional methods to less than three days. Ultimately, this system achieved a processing yield of over 99.5% in a mass production environment, an improvement of more than 10 percentage points compared to existing technologies, while reducing unit energy consumption by 40% and increasing overall equipment efficiency by 50%, demonstrating significant economic benefits and industrial promotion value. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall technical solution architecture of the laser processing method and system for the inner lead wire of the gold finger plate based on intelligent feedback proposed in this invention.

[0029] Figure 2 This is a schematic diagram of the core principle framework of the adaptive laser parameter control and multiphysics closed-loop feedback mechanism in this invention. Detailed Implementation

[0030] This application provides a laser processing method and system for in-plate leads based on intelligent feedback. Its core lies in constructing a closed-loop dynamic sensing and real-time control mechanism to achieve precise control over the entire process of morphological evolution, thermal stress distribution, and material phase transformation behavior of the microstructure in the gold finger region during laser processing. This method utilizes the collaborative operation of high-resolution in-situ visual sensing, multi-physics coupling modeling, and an adaptive laser parameter control module. During processing, based on real-time acquired data on molten pool morphology, edge heat-affected zone width, and surface reflectivity changes, it dynamically adjusts the laser power density, scanning speed, pulse frequency, and focusing position. This ensures lead conductivity while minimizing the negative effects of thermally induced warping, metal recrystallization coarsening, and interface oxide layer thickening, achieving a synergistic optimization of the three objectives: micron-level precision, submicron-level surface roughness, and high yield consistency.

[0031] Please refer to Figure 1 and Figure 2The method includes the following steps: Real-time acquisition of images of the molten pool contour and grayscale gradient distribution of the edge heat-affected zone in the current laser-affected area using a high-frame-rate coaxial vision sensing unit; background subtraction, median filtering, and edge sharpening operations performed on the acquired raw images using an image preprocessing module to extract four key feature parameters: the molten pool's major axis dimension, minor axis dimension, edge heat-affected zone width, and average surface reflectivity; receiving the four key feature parameters through a multi-physics coupling modeling unit and, combined with the current laser processing parameters, solving the coupled equation set consisting of the heat conduction equation, fluid dynamics equation, and phase transition dynamics equation in real time; and obtaining the current... The system monitors four physical quantities: the temperature gradient at the bottom of the molten pool, the solidification rate of the sidewalls, the principal component of thermal stress, and the predicted value of the oxide layer thickness. An adaptive control decision engine receives these four quantities and generates laser parameter adjustment commands. A control command generation module converts these commands into laser drive signals and scanning galvanometer control signals. A closed-loop feedback delay compensation module applies time delay corrections to the control commands. A processing termination judgment module continuously monitors the convergence of the molten pool morphology and the stability of the heat-affected zone width, triggering processing path advancement steps. A global path planning module generates processing path sequences based on the lead topology in the circuit board design file and manages the processing data.

[0032] In this step, a high-frame-rate coaxial vision sensing unit is used to acquire images of the molten pool contour and gray-scale gradient distribution of the edge heat-affected zone in real time within the laser-acting area. The vision sensing unit is located in the optical path inside the laser processing head, with its imaging plane having a zero-degree angle with the normal to the workpiece surface to ensure no perspective distortion. The image resolution reaches more than 200 pixels per millimeter, and the frame rate is no less than 1,000 frames per second. The light source uses a near-infrared light-emitting diode with a center wavelength of 850 nanometers, a radiation intensity of 850 milliwatts per steradian, and a half-intensity angle of ±25 degrees to penetrate the plasma plume and obtain the true morphology of the molten pool. The execution process of this step includes the following sub-steps: First, before the laser is started, the vision sensing unit completes self-testing and calibration, including light source intensity stability testing, image sensor dark current compensation, and optical path coaxiality verification. Second, simultaneously with the laser pulse triggering, the image sensor synchronously starts global shutter exposure, with the exposure time set to 0.8 milliseconds to ensure complete image acquisition during a single laser pulse. Third, the acquired raw image data is transmitted to the image preprocessing module via a high-speed serial interface, using a low-voltage differential signal standard and a data transmission rate set to 5 gigabits per second to ensure no image loss between frames. Finally, after each image acquisition, the vision sensing unit automatically performs lens contamination detection. The detection method is to analyze the ratio of the average grayscale value of the four corner areas to the average grayscale value of the center area. If the ratio deviates from a preset threshold by more than 5%, a cleaning alarm is triggered and the processing flow is paused. The data flow in this step is as follows: the laser pulse trigger signal serves as the image acquisition synchronization source; the image data flows through the optical imaging system, photoelectric conversion device, and high-speed data interface before entering the preprocessing pipeline. Simultaneously, the system records the current laser parameters and spatial coordinates as image metadata.

[0033] In the steps, the image preprocessing module performs background subtraction, median filtering, and edge sharpening operations on the acquired raw image, extracting four key feature parameters: the long axis dimension of the molten pool, the short axis dimension, the width of the edge heat-affected zone, and the mean surface reflectance. The image preprocessing module uses a two-dimensional convolution kernel with a fixed kernel size of 3x3 to perform spatial domain filtering. Edge detection uses the Sobel operator combined with the non-maximum suppression algorithm. Reflectance is calculated based on the integral area ratio of the image gray-level histogram within the preset gray-level range. The execution process of this step includes the following sub-steps: First, the background subtraction unit reads the current frame image and the pre-stored background template image. The background template image is generated by averaging fifty frames of images acquired under laser-free conditions. The subtraction operation uses pixel-by-pixel subtraction, and the result is truncated to ensure that the grayscale value is non-negative. Second, the median filtering unit performs a 3x3 window median filter on the background-subtracted image. The filtering window slides through the entire image in row-major order, and the median of the nine pixel values ​​in each window is taken as the output pixel value after sorting. This operation effectively suppresses salt-and-pepper noise while preserving edge sharpness. Third, the edge sharpening unit uses the Laplacian operator to perform secondary differential enhancement on the filtered image. The enhancement coefficient is set to 0.5. The enhanced image and the original image are weighted and superimposed to generate the sharpened result. Then, the edge detection unit calculates the Sobel gradient of the image in the horizontal and vertical directions in parallel. The gradient magnitude is calculated as the square root of the sum of the squares of the horizontal and vertical gradients, and the gradient direction is calculated as the arctangent of the vertical gradient divided by the horizontal gradient. The large value suppression algorithm compares the gradient magnitude of the center pixel with that of its neighboring pixels along the gradient direction, retaining only local maxima as edge candidates. Next, the edge connection algorithm uses a dual-threshold method, setting the high threshold to 70% of the maximum gradient magnitude and the low threshold to 30% of the high threshold. Starting from the high-threshold edge point, it tracks the low-threshold edge point along the gradient direction to complete edge closure. Subsequently, the melt pool contour extraction unit calculates the minimum bounding ellipse based on the closed edges, outputting the major and minor axis lengths of the ellipse as the melt pool's major and minor axis dimensions, respectively. The edge heat-affected zone width extraction unit searches outwards along the melt pool contour normal direction for the first time the grayscale gradient falls below a preset threshold. The distance between this location and the contour boundary is used as the heat-affected zone width. The search direction covers one hundred uniformly distributed sampling points on the contour, and the final output width is the arithmetic mean of the distances between all sampling points. The surface reflectance mean calculation unit counts the number of pixels in the image with grayscale values ​​between one hundred and two hundred. This number is divided by the total number of pixels in the image and multiplied by one hundred to output the mean reflectance. The data verification mechanism in this step includes: after each feature extraction, the system automatically verifies whether the long axis dimension of the molten pool is within the preset reasonable range of five to fifty micrometers. If it exceeds the range, the current frame is marked as an abnormal frame and resampling is triggered. At the same time, it verifies whether the width of the heat-affected zone is less than fifty percent of the short axis dimension of the molten pool. If it does not meet the requirements, the processing is paused and an alarm is triggered.The data flow in this step is as follows: the original image data is input to the background subtraction unit, output to the median filtering unit, then passed to the edge sharpening unit, the sharpening result is sent to the edge detection unit, the edge data is sent to the contour extraction and thermally affected area calculation unit, the grayscale data is sent to the reflectance statistics unit, and finally the four feature parameters are packaged into a structured data package and output to the multiphysics coupling modeling unit.

[0034] In this step, four key characteristic parameters are received through a multiphysics coupling modeling unit, and combined with the current laser processing parameters, the coupled equation set consisting of heat conduction equation, fluid dynamics equation and phase transformation dynamics equation is solved in real time. The heat conduction equation adopts a three-dimensional unsteady state form, and the thermal conductivity of the material changes as a piecewise linear function with temperature. The fluid dynamics equation uses the Navier-Stokes equation to describe the convection behavior inside the molten pool, and the phase transformation dynamics equation uses the Johnson-Mell equation to describe the propagation rate of the solid-liquid interface. All equations are discretized using the finite volume method, with a time step set to ten nanoseconds, a spatial grid size of 0.5 micrometers, and a computational domain covering a 50-micrometer cubic region around the laser point of action. The execution process of this step includes the following sub-steps: First, the model initialization unit loads thermophysical parameters from the material database according to the current processing material type, including solid-phase thermal conductivity, liquid-phase thermal conductivity, melting point temperature, latent heat value, density, viscosity, and surface tension coefficient. The material database is stored in non-volatile memory and supports online updates. Second, the mesh generation unit generates a structured hexahedral mesh based on the current laser focus position and the computational domain size. The total number of mesh nodes is one million, and the mesh generation adopts an equal-spacing strategy. The boundary condition is set to a constant temperature of 300 Kelvin on the outer surface of the computational domain. Third, the initial condition setting unit maps the current molten pool contour shape to the computational domain. The mapping method is to convert the visually extracted elliptical contour of the molten pool into an initial liquid phase region within the computational domain. The initial temperature field is set to the melting point temperature, and the initial velocity field is set to zero. Then, the equation coupling solver starts iterative calculation. The discretized form of the heat conduction equation is:

[0035] ;

[0036] in For density, For specific heat capacity, For temperature, For time, Thermal conductivity is temperature-dependent. For laser heat source items, This represents the latent heat of phase transition. The fluid dynamics equations employ the incompressible Navier-Stokes equations, which include the continuity equation and the momentum equation. The volume force term in the momentum equation includes gravity and surface tension, with the surface tension calculated using a continuous surface force model. The phase transition dynamics equations adopt the Johnson-Mell equations, which are exponential functions of the solid fraction with temperature. The phase transition rate is determined by both the temperature gradient and the cooling rate. The solver employs the preprocessed conjugate gradient method to solve the pressure Poisson equation and uses the explicit Runge-Kutta method to update the velocity and temperature fields, performing a full-field update every 10 nanoseconds. Next, the convergence judgment unit calculates the residual norm after each time step. The residual norm is defined as the L2 norm of the difference between the temperature, velocity, and pressure variables in the current and previous time steps. If the residual norm is less than 10 to the power of -6, convergence is determined; otherwise, iteration continues. Subsequently, the physical quantity extraction unit extracts the temperature gradient at the bottom of the molten pool, the solidification rate of the sidewalls, the principal components of thermal stress, and the predicted oxide layer thickness after convergence. The temperature gradient is calculated by fitting a temperature distribution line along the depth direction in the bottom region of the molten pool, with the slope of the line serving as the temperature gradient output. The solidification rate is calculated by tracking the change in the solid-liquid interface position, with the distance the interface moves per unit time serving as the solidification rate. Thermal stress is calculated based on the thermoelastic constitutive equation, inputting the temperature field and displacement boundary conditions, and outputting the principal stress direction components. The oxide layer thickness prediction is based on the Arrhenius equation, inputting the surface temperature and exposure time, and outputting the oxide layer growth thickness. The data verification mechanism in this step includes: verifying the completeness of the input feature parameters before each solution; if any parameter is missing, the value from the previous cycle is used to replace it and a data interpolation flag is marked; monitoring computational stability during the solution process; if a temperature value exceeds the material boiling point or a velocity value exceeds the speed of sound (physically unreasonable), the time step is automatically reduced and the calculation is recalculated; and verifying whether the physical quantities are within a preset safe range before output; if they are exceeded, a safety protection mechanism is triggered. The data flow in this step is as follows: the four key feature parameters and the current laser parameters are used as input data to enter the model initialization unit; the generated mesh and initial conditions are sent to the solver; the solution results are sent to the physical quantity extraction unit; and the four physical quantities are output to the adaptive control decision engine.

[0037] In this step, four physical quantities are obtained through the model output interface: the temperature gradient at the bottom of the molten pool, the solidification rate of the sidewall, the principal component of thermal stress, and the predicted value of the oxide layer thickness at the current moment. These physical quantities are used as feedback control variables and input to the adaptive control decision engine. The decision engine adopts a two-layer feedforward neural network structure with four nodes in the input layer, eight nodes in the hidden layer, and four nodes in the output layer. The activation function is a modified linear unit, and the weight matrix is ​​obtained through offline training. The training dataset contains 10,000 sets of simulation results and measured yield data of the processing process under different materials and different combinations of initial parameters. The loss function is a weighted sum of mean square error and yield. The execution process of this step includes the following sub-steps: First, the data normalization unit maps the four physical quantities to the interval between zero and one, and the mapping coefficients are stored in a lookup table, which is dynamically selected according to the material type; Second, the moving average filtering unit performs a five-point moving average on the normalized data. The filtering window includes the current point and the previous four historical points, and the output is the arithmetic mean of the data within the window. This operation smooths out instantaneous fluctuations and improves control stability; Third, the neural network inference engine loads the pre-trained weight matrix, which is stored in the dedicated storage area of ​​the tensor processing unit. The inference process uses eight-bit integer quantization for calculation. The input data is normalized and filtered before being sent to the input layer, and after nonlinear transformation by the hidden layer, it is output to the output layer. The output layer is processed by a post-processing unit that performs dead-zone suppression and rate limiting on the original output of the neural network. Dead-zone suppression maintains the original value if the change in output value is less than 2% of the control range, while rate limiting truncates the output value to the limit if the change in output value per unit time exceeds 5% of the total range. Next, the four nodes of the output layer correspond to laser power density adjustment, scanning speed adjustment, pulse frequency adjustment, and focus position adjustment, respectively. These adjustments are incremental values ​​relative to the current setpoint. Subsequently, the adjustments are packaged into control command data packets containing a timestamp, adjustment value, material identifier, and data interpolation flag, and transmitted to the control command generation module via a high-speed bus. The data verification mechanism in this step includes: verifying the integrity of the input data before inference, replacing any null values ​​with default safe values; verifying whether the output adjustment is within the allowed adjustment range after inference, trunculating to the boundary value if it exceeds the range; and recording the confidence index for each inference iteration, triggering a model retraining request when the confidence level falls below a preset threshold. The data flow in this step is as follows: the input data of the four physical quantities is normalized by the unit, the output is sent to the moving average filtering unit, the filtering result is sent to the neural network inference engine, the inference result is processed by the post-processing unit to generate the adjustment amount, and the adjustment amount data packet is output to the control instruction generation module.

[0038] In this step, the four control quantities output by the neural network are converted into laser drive signals and scanning galvanometer control signals by the control instruction generation module. The laser drive signal adjusts the output power density, with an adjustment range of 10 watts to 500 watts per square millimeter and an adjustment resolution of 1 watt per square millimeter. The scanning galvanometer control signal adjusts the scanning speed and path curvature, with the scanning speed adjustment range of 10 millimeters to 500 millimeters per second and an adjustment resolution of 1 millimeter per second. The path curvature is reconstructed in real time through cubic spline interpolation to ensure that the trajectory continuity and acceleration constraints meet the physical limits of the mechanical system. The execution process of this step includes the following sub-steps: First, the instruction parsing unit receives the control instruction data packet, parses out four adjustment values ​​and a timestamp, and judges the timeliness of the instruction based on the timestamp. If the instruction delay exceeds 20 milliseconds, the instruction is discarded and the previous valid instruction is used. Second, the power density adjustment unit converts the laser power density adjustment amount into the laser drive current setpoint. The conversion relationship is a linear mapping, and the mapping coefficient is preset according to the laser model. The drive current setpoint is sent to the digital-to-analog converter (DAC). The DAC has a resolution of 16 bits and an update rate of one million times per second, outputting an analog voltage signal. Third, the power amplifier receives the analog voltage signal, amplifies it proportionally, and then drives the laser diode. The power amplifier has a bandwidth of 100 kHz to ensure dynamic response without phase lag. Then, the current loop controller monitors the laser output current, adopting a proportional-integral control structure with a proportional gain of 0.5 and an integral time of 10 milliseconds. The feedback current is compared with the set current to generate an error signal. The error signal is then processed by the controller to correct the error. A positive drive voltage is applied to ensure a steady-state error of less than 0.1%. Next, the scanning speed adjustment unit adds the scanning speed adjustment to the current scanning speed setting to generate a new speed setting, which is then sent to the scanning galvanometer controller. The path curvature reconstruction unit, based on the geometry of the current processing path and the new speed setting, uses a cubic spline interpolation algorithm to calculate the radius of curvature at each point on the path in real time. This radius of curvature is used to calculate the angular acceleration required by the galvanometer, ensuring that the mechanical system acceleration does not exceed the physical limit of five radians per second. Subsequently, the galvanometer control signal generation unit converts the speed setting and curvature data into a drive pulse sequence for the galvanometer motor. The pulse frequency is proportional to the scanning speed, and the pulse duty cycle is inversely proportional to the radius of curvature. The drive pulses are amplified by the power drive circuit and sent to the galvanometer motor. Finally, the focusing position adjustment unit converts the focusing position adjustment into a displacement command for the objective lens drive motor. This displacement command is executed by the stepper motor controller, with each step of the stepper motor displacing 0.1 micrometers to ensure focusing accuracy. The data verification mechanism in this step includes: verifying the validity of the adjustment value before generating the drive signal, and using the safe default value if the value is abnormal; monitoring the actual execution parameters after output, and triggering an alarm and recording the deviation log if the actual value deviates from the set value by more than 5%; and monitoring the laser temperature and galvanometer current, and automatically reducing the power protection if they exceed the safety threshold.The data flow in this step is as follows: the control command data packet is input to the command parsing unit, the power density adjustment amount is sent to the power adjustment subsystem, the scanning speed and path curvature adjustment amount is sent to the galvanometer control subsystem, the focus position adjustment amount is sent to the focus adjustment subsystem, and the drive signals generated by each subsystem are output to the laser, the galvanometer motor and the objective lens drive motor respectively.

[0039] In this step, a time delay correction is applied to the control command through a closed-loop feedback delay compensation module. The delay includes four parts: image acquisition delay, data transmission delay, model calculation delay, and actuator response delay. The total delay time is calibrated to twelve milliseconds through a step response experiment. The compensation algorithm adopts a zero-order hold combined with a feedforward predictor structure. The predictor input is the trend of the control quantity change in the most recent three sampling cycles, and the output is the advance estimate of the control quantity in the next cycle, ensuring the closed-loop stability of the system. The execution process of this step includes the following sub-steps: First, the timestamp recorder marks the system's high-precision clock at the moment of image acquisition triggering. The clock resolution is one nanosecond, and the marked value is used as the start time of the current control cycle. Second, the delay measuring device performs a loop closure test during the system initialization phase. The test method is to send a known step signal and measure the time difference from sending it to the actuator response. The test is repeated one hundred times, and the average value is taken as the total delay time, which is stored in a register. Third, the feedforward predictor reads the historical values ​​of the control quantity in the most recent three cycles at the beginning of each control cycle. The historical values ​​are stored in a circular buffer with a depth of five. Then, the predictor samples... The historical value variation trend is fitted using a third-order polynomial. The fitting formula is a cubic polynomial function of the control amount and time, and the polynomial coefficients are calculated using the least squares method. Next, the predictor calculates the advance estimate of the control amount for the next cycle. The estimate equals the control amount at the current time plus the increment of the fitted curve at the next cycle time point. Subsequently, the zero-order hold holds the control amount for the current cycle until the start of the next cycle. The weighted sum of the hold output and the predictor output generates the final compensated control amount. The weighting coefficients are obtained through offline system identification, which minimizes the overshoot and settling time of the closed-loop system step response. Finally, the compensated control amount is sent to the control command generation module for execution. The data verification mechanism in this step includes: verifying the integrity of historical data before each prediction, and using linear extrapolation to replace missing data; verifying the rationality of the estimated value after prediction, and truncating to the boundary value if the estimated value exceeds the physical allowable range; and monitoring the prediction error. If the prediction error exceeds 10% for five consecutive cycles, the predictor parameter recalibration is triggered. The data flow in this step is as follows: the timestamp signal is input to the delay measuring device, the historical value of the control quantity is input to the feedforward predictor, the prediction result is weighted and summed with the output of the hold device to generate the compensated control quantity, and the compensated control quantity is output to the control command generation module.

[0040] In this step, the processing termination judgment module continuously monitors the convergence of the molten pool morphology and the stability of the heat-affected zone width. When the fluctuation of the long axis dimension of the molten pool is less than 0.5 micrometers and the change rate of the heat-affected zone width is less than 0.5 percent within five consecutive sampling cycles, it is determined that the current lead processing segment has reached a steady state, triggering the processing path advancement step. The advancement step adopts incremental displacement control, and the advancement distance each time is 80% of the designed width of the lead, ensuring that there is a 20% overlap area between adjacent processing segments, avoiding continuity breaks caused by processing gaps. The execution process of this step includes the following sub-steps: First, the state machine initialization unit enters the initialization state at the beginning of the processing section. The initialization content includes clearing the historical data buffer, resetting the counter, and loading the current lead wire design width parameters. Second, in the acquisition state, the system continuously receives data on the long axis dimension of the molten pool and the width of the heat-affected zone. The data is stored in a first-in-first-out queue with a queue depth of ten. Third, in the calculation state, the system calculates the standard deviation of the long axis dimension of the molten pool and the relative rate of change of the width of the heat-affected zone within the last five sampling periods. The standard deviation is calculated by dividing the sum of the squares of the differences between the dimensions of each period and the mean by four and then taking the square root. The relative rate of change is calculated by dividing the difference between the width of the latest period and the width of the earliest period by the width of the earliest period. Finally, in the judgment state, the system compares whether the standard deviation is... If the distance is less than 0.5 micrometers and the relative rate of change is less than 0.5 percent, the system enters the forward state; otherwise, it returns to the acquisition state to continue monitoring. Next, in the forward state, the system generates a path forward step command, which includes the forward distance and direction. The forward distance is the current lead wire design width multiplied by 0.8, and the direction is along the lead wire extension direction. Subsequently, the pulse output interface converts the forward command into a pulse signal recognizable by the motion control card. The pulse frequency is proportional to the forward distance, the pulse width is fixed at 10 microseconds, and the number of pulses equals the forward distance divided by the system resolution of 0.1 micrometers. Finally, after receiving the pulse signal, the motion control card drives the linear motor to perform displacement. After the displacement is completed, a completion signal is sent to the state machine, which returns to the initialization state to prepare for the next processing segment. The data verification mechanism for this step includes: verifying the validity of the data before calculation; if the data is abnormal, the monitoring cycle is extended; if the conditions are not met during judgment but the maximum processing time threshold has been exceeded, forward movement is forcibly triggered and the segment is marked as suspicious; after executing the forward movement, the actual displacement accuracy is verified; if the deviation exceeds 0.1 micrometers, the error is recorded and the parameters for the next forward movement are adjusted. The data flow for this step is as follows: the size of the molten pool and the width of the heat-affected zone are input into the state machine. After the judgment conditions are met, a forward command is generated. The forward command is converted into a pulse signal and output to the motion control card. The motion control card drives the actuator to complete the displacement.

[0041] In this step, the global path planning module generates a processing path sequence based on the lead topology in the circuit board design file. The path sequence is sorted according to the principle of minimum empty travel. Path turning points are transitioned by arcs with an arc radius of not less than twice the lead width. During path execution, the processing parameters and feedback characteristics of each segment are recorded in real time to form a digital twin archive of the processing process, which is used for subsequent quality traceability and process optimization. The execution process of this step includes the following sub-steps: First, the design file parsing unit reads the circuit board design file in extended data format and extracts the start coordinates, end coordinates, width, and layer number information of all gold finger guide wires; second, the path generation unit generates a processing path for each lead wire, the path consisting of a series of discrete points with a point spacing of 0.5 micrometers, and the path direction from the start to the end of the lead wire; third, the path sorting unit uses an improved genetic algorithm to optimize the processing order, with a population size of one hundred, a crossover probability of 0.8, a mutation probability of 0.1, and a fitness function that is the weighted sum of the total path length, the number of turns, and the heat accumulation, with weight coefficients of 0.4, 0.3, and 0.3 respectively. After one hundred iterations, the algorithm outputs the optimal path sequence; finally, the path smoothing unit performs arc transition processing on the path turning points. The radius of the arc is taken as twice the larger of the widths of the two adjacent lead wire segments. Arc interpolation is calculated using parametric equations to ensure curvature continuity. Next, the path execution unit performs processing segment by segment sequentially. Before processing each segment, the corresponding preset laser parameter values ​​are loaded. These preset values ​​are retrieved from the process database based on the lead wire width and material type. Subsequently, the data recording unit collects and stores timestamps, position coordinates, laser parameters, feedback features, and yield tags in real time during each processing segment. The storage format is a time-series database record, and each record contains the above five types of fields. Finally, a digital twin file is generated after the entire board is processed. The file contains a compressed package of all records and a processing process statistical summary. The summary includes the total processing time, average yield, maximum heat-affected zone width, and minimum surface roughness index. The file is stored on a solid-state drive and uploaded to a cloud server for backup. The data verification mechanism in this step includes: verifying that the path does not self-intersect or collide after path generation; if so, replanning is performed; verifying the integrity of fields during data recording; marking records as incomplete if missing; and calculating the data consistency checksum after file generation; triggering data retransmission if the checksum does not match. The data flow in this step is as follows: the design file is input into the parsing unit, the generated path data is sent to the sorting and smoothing unit, the optimized path sequence is sent to the execution unit, the data collected during the execution process is sent to the recording unit, and finally, a digital twin archive is generated and stored.

[0042] The system includes a high frame rate coaxial vision sensing unit, an image preprocessing module, a multiphysics coupling modeling unit, an adaptive control decision engine, a control command generation module, a closed-loop feedback delay compensation module, a processing termination judgment module, and a global path planning module. The high frame rate coaxial vision sensing unit is used to acquire images of the molten pool morphology and heat-affected zone in real time during laser processing. Its optical system includes an objective lens, a beam splitter, a filter, and an image sensor. The objective lens has a numerical aperture of 0.8 mm and a working distance of 10 mm. The beam splitter transmits 90% of the laser energy to the workpiece surface and guides 10% of the reflected light to the image sensor. The filter has a center wavelength of 850 nm and a bandwidth of 20 nm. The image sensor uses a global shutter type CMOS device with a pixel size of 2.2 μm x 2.2 μm and a maximum frame rate of 1200 frames per second. The image preprocessing module performs noise suppression and feature extraction operations on the original image. Its hardware is a field-programmable gate array (FPGA) chip, internally containing an image processing pipeline. This pipeline includes a background modeling unit, a spatial filtering unit, an edge detection unit, and a grayscale statistics unit. Background modeling uses a sliding window mean method with a window size of 50 x 50 pixels. Spatial filtering uses a 3 x 3 median filter. Edge detection uses a parallel Sobel operator to calculate gradient magnitude and direction. Grayscale statistics use a histogram accumulator to calculate the pixel percentage within a specified grayscale range. The multiphysics coupling modeling unit solves the thermo-fluid-structure interaction equations in real time and outputs key physical quantities. Its computational core is a graphics processing unit (GPU) array containing four computing nodes, each equipped with 3,000 stream processors and 8 gigabytes of video memory. Computational tasks are distributed to each node using a spatial domain decomposition strategy. Communication uses a point-to-point direct memory access mechanism to ensure data exchange latency is less than one microsecond. The equation solver uses a preprocessed conjugate gradient method to accelerate convergence, with the convergence criterion being a residual norm less than 10 to the power of -6. The adaptive control decision engine generates laser parameter adjustment commands based on physical quantity feedback. Its neural network model is stored in non-volatile memory. The inference engine is implemented using a tensor processing unit, supporting 8-bit integer quantization inference with a single inference time of less than 0.5 milliseconds. Input data preprocessing includes normalization and moving average filtering, with normalization coefficients stored in a lookup table and a sliding window length of five. Output data post-processing includes dead-zone suppression and rate limiting. The dead-zone width is 2% of the control range, and the rate limit is that the adjustment per millisecond does not exceed 5% of the total range. The control command generation module converts digital control quantities into analog drive signals. It includes a digital-to-analog converter (DAC), a power amplifier, and a current loop controller. The DAC has a 16-bit resolution and an update rate of one million times per second. The power amplifier has a bandwidth of 100 kHz and an output current range of 0 to 20 amperes. The current loop controller uses a proportional-integral (PI) structure with a proportional gain of 0.5 and an integral time of 10 milliseconds, ensuring no overshoot in the laser power response and a steady-state error of less than 0.1%.The closed-loop feedback delay compensation module is used to correct the impact of inherent system delay on control stability. It includes a timestamp recorder, a delay measuring device, and a feedforward predictor. The timestamp recorder marks the system clock at the moment of image acquisition triggering. The delay measuring device obtains the end-to-end delay through loop closure testing. The feedforward predictor uses a third-order polynomial to fit the trend of the control variable changes in the most recent three cycles, with a prediction step size of one sampling period. The prediction result is output after a weighted summation with the current feedback variable, and the weighting coefficients are obtained through offline identification. The processing termination judgment module is used to determine whether the current processing segment has reached a steady state and trigger path advancement. Its logic judgment unit adopts a state machine structure, including five states: initialization state, acquisition state, calculation state, judgment state, and advancement state. The state transition condition is based on the characteristic parameter change rate threshold, which is stored in a register. The advancement command is sent to the motion control card through the pulse output interface. The pulse frequency is proportional to the advancement distance, and the pulse width is ten microseconds to ensure positioning accuracy better than 0.1 micrometers. The global path planning module is used to generate the optimal processing path sequence and manage processing data. It runs on an industrial control computer with a real-time Linux kernel operating system. The path planning algorithm adopts an improved genetic algorithm with a population size of one hundred, a crossover probability of 0.8, and a mutation probability of 0.1. The fitness function includes a weighted sum of path length, number of turns, and heat accumulation. The processing data is stored in a solid-state drive and adopts a time-series database structure. Each record contains five types of fields: timestamp, location coordinates, laser parameters, feedback features, and yield label.

[0043] The system modules are interconnected via a high-speed backplane bus. The bus protocol combines the physical layer of the Controller Area Network (CAN) bus with the application layer of the transmission control protocol to ensure reliable and real-time data transmission. Upon system startup, a self-test is performed, checking the hardware status of each module, communication link connectivity, and parameter configuration integrity. After passing the self-test, the system enters standby mode, awaiting processing instructions. During processing, the system monitors the operating status of each module in real time. If an anomaly is detected, it automatically switches to a safe mode. In safe mode, laser output stops and the system maintains its current position, while simultaneously uploading alarm information to the operation interface. The system supports remote diagnostics and parameter updates, communicating with an external server via an Ethernet interface using a Hypertext Transfer Protocol (HTTP) to ensure data security. The system power supply employs a redundant design, automatically switching between the main and backup power supplies in less than ten milliseconds to ensure uninterrupted processing. The system casing is made of electromagnetic shielding material with a shielding effectiveness greater than 60 Baud, preventing external electromagnetic interference from affecting control accuracy. The system is equipped with an ambient temperature and humidity sensor; sensor data is used to compensate for material thermophysical parameter drift, and the compensation algorithm is built into the multiphysics coupling modeling unit. The system's human-machine interface provides real-time processing status display, historical data query, parameter setting, and alarm information management functions. The interface refresh rate is no less than 30 frames per second to ensure timely operation response.

[0044] The method and system in this embodiment achieve intelligent closed-loop control of the laser processing of the lead wires inside the gold finger board through the coordinated work of the above steps and modules. While ensuring processing accuracy and surface quality, it significantly improves processing yield and system robustness, and has important industrial application value.

Claims

1. A laser processing method for internal leads of a gold finger board based on intelligent feedback, characterized in that, include: The high-frame-rate coaxial vision sensing unit acquires images of the molten pool contour and gray-scale gradient distribution of the edge heat-affected zone in real time within the current laser-affected area. The image preprocessing module performs background subtraction, median filtering and edge sharpening operations on the acquired raw image, and extracts four key feature parameters: the long axis dimension of the molten pool, the short axis dimension, the width of the edge heat-affected zone and the mean surface reflectivity. The multiphysics coupling modeling unit receives the four key feature parameters and, in conjunction with the current laser processing parameter input, solves in real time the coupled equation set consisting of the heat conduction equation, the fluid dynamics equation, and the phase transition dynamics equation. The model output interface obtains four physical quantities at the current moment: the temperature gradient at the bottom of the molten pool, the solidification rate of the sidewall, the principal component of thermal stress, and the predicted value of the oxide layer thickness. The adaptive control decision engine receives the four physical quantities and generates laser parameter adjustment instructions. The laser parameter adjustment command is converted into a laser drive signal and a scanning galvanometer control signal by the control command generation module. The control command is corrected for time delay by using a closed-loop feedback delay compensation module. The process termination judgment module continuously monitors the convergence of the molten pool morphology and the stability of the heat-affected zone width. The global path planning module generates a sequence of processing paths based on the lead topology in the circuit board design file.

2. The laser processing method for internal leads of a gold finger board based on intelligent feedback according to claim 1, characterized in that, The high-frame-rate coaxial vision sensing unit acquires real-time images of the molten pool contour and gray-scale gradient distribution of the edge heat-affected zone in the current laser-affected area, including: Simultaneously with the laser pulse triggering, the image sensor synchronously starts the global shutter exposure, with the exposure time set to 0.8 milliseconds; The acquired raw image data is transmitted to the image preprocessing module via a high-speed serial interface. The transmission protocol adopts the low-voltage differential signal standard, and the data transmission rate is set to five gigabits per second. After each image acquisition, the visual sensing unit automatically performs lens contamination detection. The detection method is to analyze the ratio of the average gray value of the four corner areas to the average gray value of the central area. If the ratio deviates from the preset threshold by more than 5%, a cleaning alarm is triggered and the processing flow is suspended.

3. The laser processing method for internal leads of a gold finger board based on intelligent feedback according to claim 2, characterized in that, The image preprocessing module performs background subtraction, median filtering, and edge sharpening operations on the acquired raw image, extracting four key feature parameters: the long axis dimension of the molten pool, the short axis dimension, the width of the edge heat-affected zone, and the mean surface reflectivity. The background subtraction unit reads the current frame image and the pre-stored background template image. The background template image is generated by averaging fifty frames of images acquired under conditions without laser action. The subtraction operation uses pixel-by-pixel subtraction. The median filtering unit performs a 3x3 window median filter on the background-subtracted image, and the filtering window slides through the entire image in row-major order. The edge sharpening unit uses the Laplacian operator to perform secondary differential enhancement on the filtered image, with the enhancement coefficient set to 0.

5. The edge detection unit calculates the Sobel gradient of the image in the horizontal and vertical directions in parallel, and uses a non-maximum suppression algorithm to preserve local maxima along the gradient direction; The molten pool contour extraction unit calculates the minimum bounding ellipse based on the closed edge, and outputs the lengths of the major and minor axes of the ellipse as the dimensions of the molten pool in the major and minor axis directions, respectively. The edge heat-affected zone width extraction unit searches outward along the normal direction of the molten pool contour for the position where the gray-level gradient first falls below the preset threshold. The search direction covers one hundred sampling points evenly distributed on the contour, and the final output width is the arithmetic mean of the distances between all sampling points. The surface reflectance mean calculation unit counts the number of pixels in the image whose grayscale value is between 100 and 200 in the preset range. The number of pixels is divided by the total number of pixels in the image and then multiplied by 100 to calculate the reflectance and output it.

4. The laser processing method for internal leads of a gold finger board based on intelligent feedback according to claim 3, characterized in that, The multiphysics coupling modeling unit receives the four key characteristic parameters and, in conjunction with the current laser processing parameter input, solves in real time the coupled equation set consisting of the heat conduction equation, the fluid dynamics equation, and the phase transition dynamics equation, including: The model initialization unit loads thermophysical parameters from the material database based on the current material type being processed, including solid phase thermal conductivity, liquid phase thermal conductivity, melting point temperature, latent heat value, density, viscosity, and surface tension coefficient. The mesh generation unit generates a structured hexahedral mesh based on the current laser focus position and the size of the computational domain. The total number of mesh nodes is one million, and the boundary conditions are set to a constant temperature of 300 Kelvin on the outer surface of the computational domain. The initial condition setting unit maps the current molten pool contour shape to the computational domain. The mapping method is to convert the visually extracted elliptical molten pool contour into an initial liquid phase region within the computational domain. The equation-coupled solver uses the preprocessed conjugate gradient method to solve the pressure Poisson equation and the explicit Runge-Kutta method to update the velocity and temperature fields. After convergence, the physical quantity extraction unit extracts the temperature gradient at the bottom of the molten pool, the solidification rate of the sidewall, the principal component of thermal stress, and the predicted value of the oxide layer thickness. The temperature gradient is calculated by fitting a straight line of temperature distribution along the depth direction in the bottom region of the molten pool. The solidification rate is calculated by tracking the change in the position of the solid-liquid interface. The thermal stress is calculated based on the thermoelastic constitutive equation, and the oxide layer thickness is predicted based on the Arrhenius equation.

5. The laser processing method for internal leads of a gold finger board based on intelligent feedback according to claim 4, characterized in that, The adaptive control decision engine receives the four physical quantities and generates laser parameter adjustment instructions, including: The data normalization unit maps the four physical quantities to the interval between zero and one, and the mapping coefficients are stored in a lookup table; The moving average filtering unit performs a five-point moving average on the normalized data, and the filtering window includes the current point and the previous four historical points. The neural network inference engine loads a pre-trained weight matrix, and the inference process uses eight-bit integer quantization for calculation. The output data post-processing unit performs dead zone suppression and rate limiting on the raw output of the neural network. The dead zone suppression method is to keep the original value unchanged if the change in the output value is less than 2% of the control range. The rate limiting method is to truncate to the rate limit if the change in the output value per unit time exceeds 5% of the total range. The adjustment amount is packaged into a control instruction data packet, which includes a timestamp, adjustment amount value, material identifier, and data interpolation flag.

6. The laser processing method for internal leads of a gold finger board based on intelligent feedback according to claim 5, characterized in that, The adjustment command is converted into laser drive signal and scanning galvanometer control signal by the control command generation module, including: The instruction parsing unit receives control instruction data packets, parses out four adjustment values ​​and a timestamp, and determines the timeliness of the instruction based on the timestamp. If the instruction delay exceeds 20 milliseconds, the timestamp is discarded to determine the instruction. The power density adjustment unit converts the laser power density adjustment amount into the laser drive current setting value, and the drive current setting value is sent to the digital-to-analog converter. The digital-to-analog converter has a resolution of sixteen bits and an update rate of one million times per second. The current loop controller monitors the laser output current and adopts a proportional-integral control structure with a proportional gain of 0.5 and an integral time of 10 milliseconds. The scanning speed adjustment unit adds the scanning speed adjustment amount to the current scanning speed setting value to generate a new speed setting value; The path curvature reconstruction unit calculates the radius of curvature at each point on the path in real time using a cubic spline interpolation algorithm based on the geometry of the current processing path and the new speed setting. The galvanometer control signal generation unit converts the speed setpoint and curvature data into a drive pulse sequence for the galvanometer motor. The pulse frequency is directly proportional to the scanning speed, and the pulse duty cycle is inversely proportional to the radius of curvature. The focusing position adjustment unit converts the focusing position adjustment amount into a displacement command for the objective lens drive motor. Each step of the stepper motor is 0.1 micrometers.

7. The laser processing method for internal leads of a gold finger board based on intelligent feedback according to claim 6, characterized in that, The control command is corrected for time delay by a closed-loop feedback delay compensation module, including: The timestamp recorder marks the system's high-precision clock at the moment image acquisition is triggered, with a clock resolution of one nanosecond. The delay measuring device performs a loop closure test during the system initialization phase. The test method is to send a known step signal and measure the time difference from the time of the signal to the actuator response. The test is repeated 100 times and the average value is taken as the total delay time. The feedforward predictor reads the historical values ​​of the control amount from the most recent three cycles at the beginning of each control cycle and uses a third-order polynomial to fit the trend of the historical values. The predictor calculates a leading estimate of the control amount for the next cycle, which is equal to the current control amount plus the increment of the fitted curve at the next cycle time point. The zero-order hold maintains the current cycle control amount until the start of the next cycle. The weighted sum of the hold output and the predictor output generates the final compensated control amount. The weighting coefficients are obtained through offline system identification.

8. The laser processing method for internal leads of a gold finger board based on intelligent feedback according to claim 7, characterized in that, The process termination judgment module continuously monitors the convergence of the molten pool morphology and the stability of the heat-affected zone width, including: The state machine initialization unit enters the initialization state at the beginning of the processing section, clearing the historical data buffer, resetting the counter, and loading the current lead wire design width parameters. In the acquisition state, the system continuously receives data on the long axis dimension of the molten pool and the width of the heat-affected zone. The data is stored in a first-in-first-out queue with a queue depth of 10. In computational mode, the system calculates the standard deviation of the major axis dimension of the molten pool and the relative rate of change of the width of the heat-affected zone over the most recent five sampling periods; In the judgment state, the system compares whether the standard deviation is less than 0.5 micrometers and whether the relative rate of change is less than 0.5 percent. If both conditions are met, it enters the forward state. In the forward state, the system generates path forward step instructions, which include forward distance and direction. The forward distance is the current lead design width multiplied by 0.

8. The pulse output interface converts the forward command into a pulse signal that can be recognized by the motion control card. The pulse frequency is proportional to the forward distance, and the pulse width is fixed at ten microseconds.

9. A laser processing system for internal leads of a gold finger board based on intelligent feedback, characterized in that, include: The high frame rate coaxial vision sensing unit is used to acquire images of the molten pool morphology and heat-affected zone in real time during laser processing. Its optical system includes an objective lens, a beam splitter, a filter, and an image sensor. The objective lens has a numerical aperture of 0.8 and a working distance of 10 mm. The beam splitter transmits 90% of the laser energy to the workpiece surface and guides 10% of the reflected light to the image sensor. The filter has a center wavelength of 850 nm and a bandwidth of 20 nm. The image sensor uses a global shutter type CMOS device with a pixel size of 2.2 μm x 2.2 μm and a maximum frame rate of 1200 frames per second. The image preprocessing module is used to perform noise suppression and feature extraction operations on the original image. Its hardware carrier is a field-programmable gate array chip, and the internal image processing pipeline includes a background modeling unit, a spatial filtering unit, an edge detection unit, and a grayscale statistics unit. The multiphysics coupling modeling unit is used to solve the thermo-fluid-structure interaction equations in real time and output key physical quantities. Its computing core is a graphics processor array, which contains four computing nodes, each equipped with three thousand stream processors and eight gigabytes of video memory. The adaptive control decision engine is used to generate laser parameter adjustment instructions based on physical quantity feedback. Its neural network model is stored in non-volatile memory. The inference engine is implemented using tensor processing units, supports eight-bit integer quantization inference, and the time taken for a single inference is less than 0.5 milliseconds. The control command generation module is used to convert digital control quantities into analog drive signals. It includes a digital-to-analog converter, a power amplifier, and a current loop controller. The digital-to-analog converter has a resolution of sixteen bits and an update rate of one million times per second. The closed-loop feedback delay compensation module is used to correct the impact of the inherent delay of the system on control stability. It includes a timestamp recorder, a delay measurer, and a feedforward predictor. The processing termination judgment module is used to determine whether the current processing segment has reached a steady state and trigger the path to move forward. Its logic judgment unit adopts a state machine structure, which includes five states: initialization state, acquisition state, calculation state, judgment state and forward state. The global path planning module is used to generate the optimal processing path sequence and manage processing data. It runs on an industrial control computer with a real-time Linux kernel operating system, and the path planning algorithm adopts an improved genetic algorithm.

10. The laser processing system for internal leads of a gold finger board based on intelligent feedback according to claim 9, characterized in that, The image preprocessing module is used for: Background modeling uses the sliding window mean method, with a window size of 50 x 50 pixels; Spatial filtering employs a 3x3 median filter; Edge detection employs a parallel Sobel operator to calculate gradient magnitude and direction; Gray-level statistics use a histogram accumulator to calculate the percentage of pixels in a specified gray-level range.

Images