Parallel seam welding full-automatic production process method

Abstract

The application provides a parallel seam welding full-automatic production process method, and relates to the technical field of microelectronic packaging manufacturing. The parallel seam welding full-automatic production process method comprises the following steps: firstly, the cover plate and the tube shell to be welded are sent into a protective gas sealed environment after vacuum baking; secondly, visual alignment and pre-welding fixation of the cover plate and the tube shell are completed in the sealed environment; thirdly, parallel seam welding is performed on two opposite edges of the pre-welded cover plate and tube shell, and welding energy is adjusted based on real-time electrode pressure and welding loop parameters; finally, the welded cover plate and tube shell are subjected to single-station time-sharing imaging, bright-field images and dark-field images of the weld are obtained, weld defects are identified through image processing, and the welded cover plate and tube shell are sorted based on the defect identification result. The method is automatically executed throughout the whole process, realizes full-process closed-loop production from baking, alignment, pre-welding, seam welding, defect detection to sorting, and greatly improves production efficiency and product packaging quality stability.

Description

Technical Field

[0001] This invention relates to the field of microelectronic packaging manufacturing technology, specifically to a fully automated parallel seam welding production process. Background Technology

[0002] Parallel seam welding is a key process for the hermetic packaging of microelectronic devices. It is widely used in the metal-cased packaging of hybrid integrated circuits, sensors, and optoelectronic devices. Its welding reliability directly determines the long-term operational stability of the device in harsh environments such as humidity, heat, and salt spray. As electronic components become increasingly miniaturized and denser, the industry has placed higher demands on the consistency of weld nuggets, the hermeticity of the packaging, and the accuracy of post-weld inspection, urgently requiring stable and efficient automated production solutions.

[0003] In related technologies, existing parallel seam welding production mostly adopts a segmented processing mode. Typically, the pipe shell and cover plate are first manually or semi-automatically loaded, and alignment, pre-welding, and parallel seam welding are completed in an open or semi-open environment. Post-weld quality assessment relies on manual microscopic inspection or conventional machine vision testing. The entire process depends on manual intervention and the independent operation of multiple devices, failing to form an integrated, fully automated production system. Regarding welding control, open-loop / semi-closed-loop control methods using constant current, constant voltage, or constant power are commonly employed. Alignment is mostly achieved manually or through simple positioning mechanisms. Inspection often uses single lighting conditions and traditional grayscale and edge detection algorithms for defect judgment.

[0004] However, the existing production model has obvious defects: First, the welding energy control does not take into account the dynamic changes in the welding circuit load, and the nonlinear fluctuation of contact resistance can easily lead to poor soldering and overheating, making it difficult to guarantee the hermeticity of the package; Second, the alignment accuracy between the cover plate and the shell is low, and misalignment is prone to occur, affecting the quality of the weld formation; Third, the welding motion control is rigid, and vibration, overheating and trajectory deviation are prone to occur at corners, resulting in uneven energy density of the weld; Finally, the high reflectivity of the metal weld and the strong interference of lattice texture make it easy for conventional visual imaging and algorithms to miss microcracks and misdetect scratches and water stains, making it impossible to achieve stable and reliable automated detection. Overall, it is difficult to meet the fully automated, highly consistent and highly reliable packaging production requirements of high-density microelectronic devices. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a fully automated parallel seam welding production process, which solves the technical problems of unstable welding energy control, low alignment accuracy between the cover plate and the shell, easy interference from high reflectivity and texture in weld defect detection, and the inability to achieve fully automated closed-loop production in existing parallel seam welding processes.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a fully automated production method for parallel seam welding, comprising: The cover plate and the pipe shell to be welded are vacuum baked and then sent into a protective gas sealed environment. Complete the visual alignment and pre-welding fixation of the cover plate and the pipe shell in a sealed environment; Parallel seam welding is performed on the two opposite edges of the pre-welded cover plate and the shell. Pressure sensors are used to collect the pressure applied to the electrodes in real time. Combining the real-time electrode pressure with the welding circuit parameters, the welding energy is adjusted in a closed loop using an incremental PID control algorithm. Single-station time-division imaging is performed on the welded cover plate and pipe shell to obtain bright field and dark field images of the weld. Weld defects are identified through image processing, and the welded cover plate and pipe shell are sorted based on the defect identification results.

[0007] Preferably, the visual alignment refers to using a visual imaging component to identify the pose deviation between the tube shell and the cover plate, and driving the robotic arm to perform alignment compensation, including: A downward-looking camera assembly is used to photograph the tube shell to obtain the geometric center and deflection angle of the tube shell; The top-view camera assembly is used to photograph the cover plate being sucked up by the suction nozzle to obtain the geometric center and deflection angle of the cover plate. Based on the pre-calibrated nozzle rotation center, the angular deviation between the cover plate and the tube shell, as well as the coupled displacement caused by the rotation, are calculated to generate motion commands that include translational compensation and rotational compensation to drive the robot to complete the alignment.

[0008] Preferably, the cover plate and the tube shell are pre-welded and fixed using spot welding electrodes, and the energy control method for the pre-welding and fixing is as follows: The instantaneous welding current and the voltage across the electrodes in the welding circuit are synchronously acquired using a high-frequency sampling rate, and the dynamic resistance is calculated in real time. Based on the Joule heating principle, the product of the square of the instantaneous welding current and the real-time dynamic resistance is integrated over the total discharge time, and the integral result is multiplied by the thermal efficiency coefficient of the system to obtain the effective energy. By dynamically adjusting the power output duty cycle, the effective energy can meet the preset target value.

[0009] Preferably, a roller electrode is used to perform parallel seam welding on the two opposite edges of the pre-welded cover plate and the tube shell; The motion control of the parallel seam welding adopts an S-shaped velocity curve planning strategy. The jerk is limited by constraining the third derivative of the position command. At the same time, a real-time coupling mechanism between velocity and welding energy is established. The pulse triggering frequency of the welding power source is dynamically mapped based on the real-time linear velocity of the roller electrode to maintain a constant weld point density per unit length of weld.

[0010] Preferably, in the closed-loop adjustment of welding energy based on real-time electrode pressure and welding circuit parameters, the welding power source adopts a high-frequency inverter AC square wave power supply. The high-frequency inverter AC square wave power supply uses high-frequency inverter AC square wave output and inserts dead time at the zero crossing point of the positive and negative half-wave switching; and the high-frequency inverter AC square wave power supply calculates the actual current effective value of the current in the current cycle based on the true RMS algorithm, and uses incremental proportional and integral control algorithms to adjust the pulse width modulation duty cycle of the next cycle.

[0011] Preferably, a composite optical acquisition component is used to perform single-station time-division imaging of the welded cover plate and tube shell. The composite optical acquisition component adopts a dual-sided telecentric lens and an industrial camera equipped with a global shutter image sensor. The object-side optical resolution of the composite optical acquisition component is determined by multiplying the ratio of the physical size of the image sensor pixel to the optical magnification of the telecentric lens by a sampling factor. The sampling factor is determined based on the color array structure of the camera sensor to ensure that the image edge sharpness meets the requirements of the detection algorithm for the extraction accuracy and signal-to-noise ratio of edge gradient information.

[0012] Preferably, the composite optical acquisition component further includes a coaxial parallel light source and a low-angle ring light source, and the specific logic of the single-station time-division imaging includes: The coaxial parallel light source is lit up at the first hardware trigger timing to acquire the bright field image of the weld seam. After a preset time interval, the low-angle ring light source is switched to acquire the dark field image of the weld seam. The light source intensity is adaptively adjusted based on the Weber contrast model. The optimal lighting parameters are determined by calculating the average gray level difference ratio between the region of interest and the background region.

[0013] Preferably, before identifying weld defects through image processing, anisotropic bilateral filtering is performed on the region of interest in the image, and the pixel value update logic includes: The neighboring pixels within the convolution window are summed using a weighted summation. The weighting coefficients of this summation are jointly determined by a spatial domain Gaussian function based on geometric distance and a pixel domain Gaussian function based on gray-level difference, in order to preserve high-frequency edge features while smoothing the background texture.

[0014] Preferably, in the process of identifying weld defects through image processing, the defect identification employs cascaded logic, including: A Hessian matrix is ​​constructed by convolving the image using a Gaussian second-order differential kernel. Based on the geometric topological relationship, ratio, and sum of squares norm of the two eigenvalues ​​of the Hessian matrix, a crack probability response function is constructed to extract suspected defect areas. The suspected defect region is used as the input to the convolutional neural network classifier. The Softmax function is used to calculate the posterior probability of the current region belonging to various types of defects. The existence of a defect is determined only when the predicted probability exceeds a preset confidence threshold. The types of defects include at least microcracks.

[0015] A second aspect of the present invention provides a fully automated parallel seam welding production system for implementing the above-described process. The system includes: The glove box module is filled with inert protective gas to create a sealed process environment, and a handling module is arranged in the central area inside. The baking module is sealed on one side of the glove box module, with an openable and closable automatic sealing door between them, for performing vacuum baking on the cover plate and tube shell to be welded. The transfer module is sealed on the other side of the glove box module, forming a sealed passage for the glove box module to interact with the external environment. The pre-welding module, located inside the glove box module, integrates a vision imaging component for aligning and positioning the cover plate and the tube shell, as well as a spot welding electrode component for completing the pre-welding fixation. The seam welding module, located inside the glove box module, is equipped with roller electrodes and motion drive components for performing parallel seam welding along the two opposite edges of the cover plate and the tube shell. The detection module, located inside the glove box module, is equipped with a composite optical acquisition component for time-division acquisition of bright-field and dark-field images of the weld at a single workstation. The main control module is electrically connected to the baking module, glove box module, handling module, pre-welding module, seam welding module, detection module, and transfer module, respectively. The main control module is configured to: schedule and transport the cover plate and the shell between each workstation, coordinate and control the action sequence of each module, regulate the welding electrode pressure and welding circuit energy output in a closed loop, complete the image processing and recognition of weld defects, and control the sorting and circulation of finished products based on the defect judgment results.

[0016] (III) Beneficial Effects This invention provides a fully automated production process for parallel seam welding. Compared with existing technologies, it has the following advantages: This invention effectively removes adsorbed moisture and impurities from the workpiece by vacuum baking the cover plate and shell to be welded, thus avoiding welding oxidation and porosity defects. Visual alignment and pre-welding are performed within the sealed environment, significantly improving the assembly accuracy of the cover plate and shell, ensuring alignment consistency, and solving the problem of low alignment accuracy in traditional processes. Parallel seam welding is performed on the two opposite edges of the pre-welded cover plate and shell, and the welding energy is dynamically adjusted based on real-time electrode pressure and welding circuit parameters in a closed loop. This dynamically compensates for contact resistance fluctuations, avoiding incomplete welds and overheating, and achieving stable and controllable welding energy. Single-station time-division imaging acquires bright-field and dark-field images, suppressing high metal reflectivity and lattice texture interference, improving the accuracy and reliability of weld defect identification. This method is fully automated, realizing a closed-loop production process from baking, alignment, pre-welding, seam welding, defect detection, and sorting, significantly improving production efficiency and product packaging quality stability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of the steps of a fully automated parallel seam welding production process provided by the present invention; Figure 2 A structural block diagram of a fully automated parallel seam welding production process system provided by the present invention; Figure 3 A process flow diagram of a fully automated parallel seam welding production process provided by the present invention; Figure 4 A comparison diagram of the electrical characteristics of a dynamic resistance energy closed-loop control mode and a traditional constant current control mode provided by the present invention; Figure 5 A comparison diagram of the effects of an S-shaped velocity planning and position-energy synchronization (PSO) mechanism provided by the present invention; Figure 6 This invention provides a schematic diagram of a microcrack defect feature extraction principle based on the Hessian matrix. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solutions in the embodiments of the present invention are intended to solve the above-mentioned technical problems, and the overall approach is as follows: First, the embodiments of the present invention aim to realize an integrated closed-loop operation of vacuum baking, atmosphere protection, alignment pre-welding, parallel seam welding and weld inspection, so as to solve the problems of traditional process segmented operation, high dependence on manual labor and difficulty in achieving fully automated continuous production.

[0021] Based on this, the embodiments of the present invention also aim to improve the alignment accuracy of the cover plate and the shell, and ensure assembly consistency through visual guidance and posture compensation to avoid welding misalignment.

[0022] Meanwhile, the embodiments of the present invention further aim to improve the stability of welding energy control. By monitoring welding circuit parameters in real time and dynamically adjusting energy input, defects such as poor soldering and overheating caused by contact resistance fluctuations are overcome, thereby improving the hermeticity of the packaging.

[0023] Furthermore, the embodiments of the present invention also aim to improve the detection capability of metal weld defects. By optimizing the optical imaging method and defect recognition logic, high reflectivity and texture interference are suppressed, the false alarm rate and missed detection rate of microcracks are reduced, and the quality of the final product is guaranteed to be stable and reliable.

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] like Figure 1 As shown, this embodiment of the invention provides a fully automated parallel seam welding production process, including: Step 1: After vacuum baking, the cover plate and the pipe shell to be welded are placed into a protective gas sealed environment. Step 2: Complete the visual alignment and pre-welding fixation of the cover plate and the pipe shell in a sealed environment; Step 3: Perform parallel seam welding on the two opposite edges of the pre-welded cover plate and the pipe shell. Use a pressure sensor to collect the pressure applied to the electrode in real time. Combine the real-time electrode pressure with the welding circuit parameters and adjust the welding energy in a closed loop using an incremental PID control algorithm. Step 4: Perform single-station time-division imaging on the welded cover plate and pipe shell to obtain bright field and dark field images of the weld. Identify weld defects through image processing and sort the welded cover plate and pipe shell based on the defect identification results.

[0026] The process method provided in this invention is fully automated, realizing a closed-loop production process from baking, alignment, pre-soldering, seam welding, defect detection and sorting, which greatly improves production efficiency and product packaging quality stability.

[0027] In some embodiments, the visual alignment in step two refers to using a visual imaging component to identify the pose deviation between the tube shell and the cover plate, and driving the robotic arm to perform alignment compensation, including: A downward-looking camera assembly is used to photograph the tube shell to obtain the geometric center and deflection angle of the tube shell; The top-view camera assembly is used to photograph the cover plate being sucked up by the suction nozzle to obtain the geometric center and deflection angle of the cover plate. Based on the pre-calibrated nozzle rotation center, the angular deviation between the cover plate and the tube shell, as well as the coupled displacement caused by the rotation, are calculated to generate motion commands that include translational compensation and rotational compensation to drive the robot to complete the alignment.

[0028] This invention uses upper and lower cameras to acquire the position and pose information of the tube shell and the cover plate respectively, and calculates the angular deviation and coupling displacement by combining the rotation center of the suction nozzle, which significantly improves the alignment accuracy of the cover plate and the tube shell, avoids assembly offset, and ensures accurate and consistent welding positions.

[0029] In some embodiments, in step two, the cover plate and the tube shell are pre-welded and fixed using spot welding electrodes. The energy control method for the pre-welding and fixing is as follows: The instantaneous welding current and the voltage across the electrodes in the welding circuit are synchronously acquired using a high-frequency sampling rate, and the dynamic resistance is calculated in real time. Based on the Joule heating principle, the product of the square of the instantaneous welding current and the real-time dynamic resistance is integrated over the total discharge time, and the integral result is multiplied by the thermal efficiency coefficient of the system to obtain the effective energy. By dynamically adjusting the power output duty cycle, the effective energy can meet the preset target value.

[0030] The embodiments of the present invention achieve precise control of pre-welding energy based on dynamic resistance and Joule heat integral, automatically compensate for contact resistance fluctuations, ensure the stability of the pre-welding weld nugget, avoid incomplete welding, explosion, and over-welding, and improve the reliability of pre-welding.

[0031] In some embodiments, in step three, a roller electrode is used to perform parallel seam welding on the two opposite edges of the pre-welded cover plate and the tube shell. The motion control of the parallel seam welding adopts an S-shaped velocity curve planning strategy. The jerk is limited by constraining the third derivative of the position command. At the same time, a real-time coupling mechanism between velocity and welding energy is established. The pulse triggering frequency of the welding power source is dynamically mapped based on the real-time linear velocity of the roller electrode to maintain a constant weld point density per unit length of weld.

[0032] The embodiments of the present invention employ an S-shaped velocity curve and velocity... Real-time energy coupling suppresses mechanical vibration, reduces trajectory deviation, ensures uniform weld density and heat input per unit length of weld seam, and avoids overheating at corners.

[0033] Furthermore, in order to balance the heating of both electrodes, prevent short circuits, improve the accuracy of current control, and further improve the stability of welding energy, in the above embodiment, in the closed-loop adjustment of welding energy based on real-time electrode pressure and welding circuit parameters, the welding power supply adopts a high-frequency inverter AC square wave power supply. The high-frequency inverter AC square wave power supply uses high-frequency inverter AC square wave output and inserts dead time at the zero crossing point of the positive and negative half-wave switching; and the high-frequency inverter AC square wave power supply calculates the actual current effective value of the current in the current cycle based on the true RMS algorithm, and uses incremental proportional and integral control algorithms to adjust the pulse width modulation duty cycle of the next cycle.

[0034] In some embodiments, in step four, a composite optical acquisition component is used to perform single-station time-division imaging of the welded cover plate and tube shell. The composite optical acquisition component uses a dual-sided telecentric lens and an industrial camera equipped with a global shutter image sensor. The object-side optical resolution of the composite optical acquisition component is determined by multiplying the ratio of the physical size of the image sensor pixel to the optical magnification of the telecentric lens by a sampling factor. The sampling factor is determined based on the color array structure of the camera sensor to ensure that the image edge sharpness meets the requirements of the detection algorithm for the extraction accuracy and signal-to-noise ratio of edge gradient information.

[0035] The embodiments of the present invention employ a dual-sided telecentric lens and a global shutter camera, with the resolution scientifically configured according to pixels, magnification, and sampling factor to eliminate imaging distortion and motion blur, and improve image clarity and edge gradient quality.

[0036] Furthermore, in order to effectively suppress interference from high metal reflectivity and improve the contrast of defect features and image signal-to-noise ratio, in the above embodiments, the composite optical acquisition component further includes a coaxial parallel light source and a low-angle ring light source. The specific logic of the single-station time-division imaging includes: The coaxial parallel light source is lit up at the first hardware trigger timing to acquire the bright field image of the weld seam. After a preset time interval, the low-angle ring light source is switched to acquire the dark field image of the weld seam. The light source intensity is adaptively adjusted based on the Weber contrast model. The optimal lighting parameters are determined by calculating the average gray level difference ratio between the region of interest and the background region.

[0037] In some embodiments, before identifying weld defects through image processing in step four, anisotropic bilateral filtering is performed on the region of interest in the image, and the pixel value update logic includes: The neighboring pixels within the convolution window are summed using a weighted summation. The weighting coefficients of this summation are jointly determined by a spatial domain Gaussian function based on geometric distance and a pixel domain Gaussian function based on gray-level difference, in order to preserve high-frequency edge features while smoothing the background texture.

[0038] This invention, in its embodiments, preserves the high-frequency edges of cracks while smoothing the metallic background texture, reducing noise interference, highlighting defect features, and improving the accuracy of subsequent defect identification. In some embodiments, step four involves identifying weld defects through image processing. The defect identification employs cascaded logic, including: First, the image is convolved using a Gaussian second-order differential kernel to construct a Hessian matrix. Based on the geometric topological relationship, ratio, and sum of squares norm of the two eigenvalues ​​of the Hessian matrix, a crack probability response function is constructed to extract suspected defect areas. Subsequently, the suspected defect area is used as the input to the convolutional neural network classifier. The Softmax function is used to calculate the posterior probability of the current area belonging to various types of defects. The existence of a defect is determined only when the predicted probability exceeds a preset confidence threshold. The types of defects include at least microcracks.

[0039] This invention significantly improves the detection rate of microcracks and reduces the false alarm rate by using Hessian matrix coarse detection and convolutional neural network (CNN) fine judgment, thereby achieving accurate identification of defects in highly reflective metal welds.

[0040] It should be noted that the core innovation of this invention is the fully automated parallel seam welding production process. It achieves closed-loop control from baking, alignment, welding to inspection through an integrated process. At the same time, the following dedicated hardware system is designed to stably support the operation of this process.

[0041] like Figure 2 As shown, this embodiment of the invention provides a fully automated parallel seam welding production system, mainly including a baking module 100, a glove box module 200, a handling module 300, a pre-welding module 400, a seam welding module 500, a detection module 600, a main control module 700, and a transfer module 800. Wherein: The glove box module 200 is filled with inert protective gas to form a sealed process environment, and the handling module 300 is arranged in the central area inside it. The baking module 100 is sealed on one side of the glove box module 200, and an openable and closable automatic sealing door is provided between the two for vacuum baking of the cover plate and the tube shell to be welded. The transfer module 800 is sealed on the other side of the glove box module 200, forming a sealed interaction channel between the glove box module 200 and the external environment. The pre-welding module 400 is located inside the glove box module 200 and integrates a visual imaging component for aligning and positioning the cover plate and the tube shell, as well as a spot welding electrode component for completing the pre-welding fixation. The seam welding module 500 is located inside the glove box module 200 and is equipped with roller electrodes and motion drive components for performing parallel seam welding along the two opposite edges of the cover plate and the tube shell. The detection module 600 is located inside the glove box module 200 and is equipped with a composite optical acquisition component for acquiring bright field and dark field images of the weld seam at a single station in a time-division manner. The main control module 700 is electrically connected to the baking module 100, glove box module 200, handling module 300, pre-welding module 400, seam welding module 500, detection module 600 and transfer module 800 respectively. The main control module 700 is configured to: schedule and transport module 300 to automatically transfer cover plates and tube shells between each workstation (e.g., pre-welding module 400, seam welding module 500, and detection module 600 can be arranged in a ring around transport module 300), coordinate and control the action sequence of each module, control the welding electrode pressure and welding circuit energy output in a closed loop, complete the image processing and recognition of weld defects, and control the sorting and circulation of finished products based on the defect judgment results (e.g., glove box module 200 can be arranged to include a left buffer station and a right buffer station, which are used to temporarily store components to be processed or already processed, respectively).

[0042] Based on the above system layout, such as Figure 3 As shown, the process method provided in this embodiment of the invention specifically employs the following steps for parallel seam welding: S10, the baking module 100 performs a vacuum baking procedure on the tooling tray loaded with the cover plate and tube shell assembly to be welded, and the main control module 700 monitors the baking operation status throughout the process. S20, after the baking process is completed, the automatic sealing door between the baking module 100 and the glove box module 200 is opened, and the transport module 300 takes out the first batch of components from the baking module 100 and transfers them to the pre-welding module 400 station inside the glove box. S30, the pre-welding module 400 performs visual alignment and pre-welding fixing operations on the first batch of components within the protective gas-sealed environment of the glove box; during this operation, the handling module 300 removes the remaining batch tooling trays from the baking module and transfers them to the left buffer station inside the glove box for temporary storage; after the material is picked up, the automatic sealing door closes, and the baking module 100 can start the next round of vacuum baking process. S40, after the first batch of component pre-welding is completed, the transport module 300 transfers the pre-welded cover plate and tube shell of this batch to the seam welding module 500 station; S50, the seam welding module 500 performs parallel seam welding on the two opposite edges of the first batch of cover plate and tube shell. During the welding process, the welding energy is adjusted in real time based on the electrode pressure and welding circuit parameters in a closed loop. At the same time as the seam welding operation, the transport module 300 transfers the second batch of components temporarily stored in the left buffer station to the pre-welding module 400. The pre-welding module simultaneously completes the visual alignment and pre-welding fixation of the second batch. S60, after the first batch of seam welding is completed, the transport module 300 transfers it to the detection module 600; the detection module 600 performs single-station time-division bright and dark field imaging on the welded components and identifies weld defects through image processing; at the same time, the transport module 300 transfers the second batch of pre-welded components to the seam welding module 500 and starts the parallel seam welding process for this batch. S70: After the first batch of component defects are detected, the handling module 300 will transfer the qualified / sortable components to the right buffer station for temporary storage. The system will iteratively execute steps S50-S60 until all batches of components temporarily stored in the left buffer station have completed welding and defect detection and have been transferred to the right buffer station. S80, the interaction channel between the glove box and the external transfer module 800 is opened. The transport module 300 transfers all the finished cover plates and tube shells in the right buffer station to the transfer module 800. Based on the defect identification results, the final sorting and unloading of the welded components is completed, and a single fully automatic closed-loop production process is completed.

[0043] Understandably, Figure 3 The complete process flow shown is Figure 1 The core method steps are matched one by one, specifically: Processes S10-S20 complete the vacuum baking and impurity removal of the workpiece and send it into the glove box inert protective sealed cavity, matching the technical content of step one above, "sending the cover plate and tube shell to be welded into the protective gas sealed environment after vacuum baking", that is, the vacuum baking and environmental replacement steps.

[0044] The pre-welding process in process S30 and batch cycle relies on visual imaging to achieve high-precision alignment and reliable pre-welding fixation, matching the technical content of step two above, "to complete the visual alignment and pre-welding fixation of the cover plate and the shell in a sealed environment," which is the visual-guided pre-welding step.

[0045] The bilateral parallel seam welding operation in process S50-S60 involves real-time closed-loop control of the welding electrode pressure and circuit energy output. This matches the technical content of step three above, which involves "performing parallel seam welding on the two opposite edges of the pre-welded cover plate and the shell, using pressure sensors to collect the pressure applied to the electrodes in real time, and combining the real-time electrode pressure with welding circuit parameters to adjust the welding energy in a closed loop using an incremental PID control algorithm." This is the technical content of the dynamic control of the parallel seam welding step. The time-division bright and dark field imaging, cascaded defect identification logic, and finished product sorting action in process S60-S80 match the technical content of step four above, which is "to perform single-station time-division imaging on the welded cover plate and pipe shell, acquire bright field and dark field images of the weld, identify weld defects through image processing, and sort the welded cover plate and pipe shell based on the defect identification results," that is, the multi-channel defect detection step.

[0046] Based on this, the following further introduction Figure 3 Technical details of the illustrated embodiments: In this embodiment, the main structure of the baking module 100 is a vacuum chamber with high airtightness. Its material is preferably 304 or 316L stainless steel. The inner wall is precision electrolytically polished to obtain a low roughness surface (e.g., Ra≤0.4μm), thereby minimizing the adsorption and desorption effects of gas molecules on the chamber wall and ensuring the rapid establishment of a high vacuum.

[0047] The vacuum chamber contains multiple independent heating plate assemblies arranged parallel to each other along a vertical direction. Considering that heat transfer in a vacuum environment mainly relies on thermal radiation and contact heat conduction rather than gas convection, the heating plates in this embodiment are made of aerospace-grade aluminum alloy or copper with high thermal conductivity, and their surfaces undergo hard anodizing treatment to improve infrared emissivity and enhance wear resistance. The heating plate surface has extremely high flatness (e.g., flatness tolerance controlled within 0.05mm) to ensure full contact with the bottom surface of the tooling tray placed on it, guaranteeing efficient and uniform contact heat conduction.

[0048] To address the interlayer temperature interference and edge heat dissipation effects that occur during the baking process of multi-layer components, the system employs a multi-temperature zone independent closed-loop control strategy. Each heating plate is equipped with an embedded or mounted high-precision temperature sensor (such as a Pt100 platinum resistance thermometer) to collect the actual temperature of that heating plate in real time. Meanwhile, the heating power of each heating plate is driven by an independent solid-state relay or thyristor power regulator. The main control module 700, after calculating using a PID algorithm, outputs a pulse width modulation (PWM) signal to adjust the on / off duty cycle of each heating element, thereby achieving precise temperature control for each layer. This control method effectively compensates for differences in heat loss caused by varying loading amounts or location.

[0049] To ensure temperature stability during the baking process, the system monitors and adjusts in real time according to the following logic. (Setting the...) The target process temperature of the heating plate is (Common temperature range is 80℃~150℃), the main control module 700 periodically collects the actual temperature of each layer. And calculate the instantaneous absolute temperature deviation. : In the formula, Indicator of heating plate layer number index , For the first Measured temperature of the heating plate.

[0050] When the calculation yields Within a preset precision temperature control range (e.g.) When the temperature is ≤2℃, the system maintains the current PID parameters for constant temperature regulation; if The abnormal increase exceeds the safety threshold (e.g.) If the temperature exceeds 10℃, the main control module 700 will immediately cut off the heating power supply to that layer and trigger an audible and visual alarm to prevent thermal damage to the device due to sensor detachment or uncontrolled heating.

[0051] In terms of gas path control, the baking module 100 is connected to an external dry vacuum pump group via a vacuum valve and to a high-purity nitrogen source (purity ≥99.999%) via an inflation valve. In this embodiment, in order to thoroughly remove water vapor deeply adsorbed on the device, a pulsed vacuum baking process is adopted, which uses periodic pressure changes to disrupt the adhesion balance of water molecules on the material surface.

[0052] During operation, the main control module 700 first confirms that the automatic sealing door is locked, then opens the vacuum valve and starts the pump unit. The vacuum pressure sensor monitors the internal pressure in real time. Once the internal pressure drops to a preset background vacuum level (e.g., background vacuum ≤ 10 Pa, this threshold can be set between 1 Pa and 50 Pa depending on the device packaging level), the main control module 700 controls the heating component to start heating at a preset heating rate (e.g., 5℃ / min). During the constant temperature baking stage, the system can periodically close the vacuum valve and slightly open the inflation valve to introduce hot nitrogen according to a set program, then evacuate again. Through the cyclical cleaning effect of inflation and evacuation, the removal efficiency of water vapor and volatile organic compounds is accelerated.

[0053] At the end of the baking process, the system stops heating and closes the vacuum valve, while the inflation valve is fully opened to refill the chamber with high-purity nitrogen. When the pressure inside the chamber rises to balance with the slightly positive pressure environment (e.g., relative to atmospheric pressure +50Pa to +200Pa) within the glove box module 200, the automatic sealing door unlocks and opens, allowing the transport module 300 to remove the components after high-temperature dehumidification. This pressure-balancing opening mechanism effectively prevents airflow impact or backflow of contaminated external atmosphere caused by pressure differences.

[0054] In this embodiment, the atmosphere-protected glove box module 200 serves as the core process environment carrier, providing a low-water-oxygen inert gas protective environment for the pre-welding and seam welding processes. This prevents oxidation of the metal casing and cover plate during high-temperature welding, or porosity defects in the weld caused by moisture. Its hardware architecture mainly includes a highly airtight sealed enclosure, a closed-loop gas purification and circulation component, an online atmosphere monitoring component, and a micro-pressure balance control component.

[0055] The high-airtightness sealed enclosure is fully welded from 304 stainless steel plate, and the observation window is made of tempered glass and sealed with an O-ring. Several standard KF interfaces are pre-installed on the inner wall of the enclosure for cable and pipe routing. The closed-loop gas purification and circulation assembly is connected to the high-airtightness sealed enclosure via inlet and outlet pipes. A high-flow-rate magnetohydrodynamic sealed circulation fan and a high-efficiency purification column are connected in series in the loop. In this embodiment, the purification column is filled with composite purification materials, specifically including a copper-based catalyst (or manganese-based catalyst) for chemical adsorption of oxygen and a 13X molecular sieve for physical adsorption of moisture. This dual-effect adsorption mechanism constitutes a deep water and oxygen purification unit, capable of continuously reducing the concentration of water and oxygen impurities in the circulating airflow to the ppm level (e.g., ≤1ppm). Furthermore, the closed-loop gas purification and circulation assembly is equipped with a regeneration heater. When the adsorption material is saturated, its adsorption activity can be restored through a high-temperature reduction process, ensuring the long-term stability of the closed-loop gas purification and circulation assembly.

[0056] The online atmosphere monitoring system consists of a trace water analyzer (dew point meter) and a trace oxygen analyzer installed at key flow field locations inside a highly airtight sealed enclosure. Sensor selection must possess rapid response characteristics; the trace water analyzer is preferably a corrosion-resistant phosphorus pentoxide electrolytic sensor, while the trace oxygen analyzer is preferably a long-life electrochemical fuel cell sensor. The sensors collect real-time data on the water content of the current environment. and oxygen content data The data is then transmitted to the main control module 700. To prevent external air infiltration and deformation of the high-pressure-tightness sealed enclosure, the system maintains a slightly positive pressure within the enclosure using a micro-pressure balancing control component. This component includes a high-precision differential pressure transmitter, a proportional regulating inlet valve, and a solenoid exhaust valve. Based on the differential pressure feedback from the differential pressure transmitter, the main control module 700 dynamically adjusts the opening of the proportional regulating inlet valve using a PID closed-loop control algorithm. When the pressure inside the high-pressure-tightness sealed enclosure fluctuates due to glove operation, the system can respond in milliseconds, maintaining the differential pressure within a preset range (e.g., 50 Pa to 200 Pa).

[0057] To address the outgassing generated during welding and the impurities introduced during material transfer, this embodiment designs an intelligent atmosphere management strategy based on a comprehensive contamination index. The main control module 700 periodically reads monitoring data and combines it with preset process standard thresholds. (For example, according to the GJB548B-2021 standard for test methods and procedures for microelectronic devices, the value is set to 100 ppm, or according to higher-level process requirements, the value is set to 50 ppm), calculate the comprehensive pollution index. : In the formula, and These are real-time collected data on water and oxygen content, respectively. This index can normalize and reflect the degree of deviation of the current environment from the process red line.

[0058] based on Based on the calculation results, the main control module 700 adaptively switches its operating mode: in the normal state, i.e. When the value is ≤1, the system executes a low-power cycle maintenance mode, with the circulating fan operating at low frequency to maintain airflow uniformity within the highly airtight sealed enclosure; when slight contamination is detected, i.e. ( When the preset engineering tolerance factor (usually 3-5) is reached, the system switches to full-power purification mode, and the circulating fan speed increases to the rated value to accelerate the adsorption efficiency through the purification column; once severe pollution occurs (such as accidental opening of the transfer door), the system switches to full-power purification mode. At this time, the system triggers the emergency replacement cleaning mode. The main control module 700 simultaneously opens the proportional regulating intake valve to introduce high-purity nitrogen (purity ≥99.999%) and opens the electromagnetic exhaust valve, using a large flow rate of nitrogen (e.g., flow rate ≥100L) to purge and flush the highly airtight sealed chamber until… It quickly drops to a safe range and then automatically resets to cycle mode.

[0059] Furthermore, in scenarios involving material interaction between the baking module 100 and the glove box module 200, the system executes strict pressure interlock logic. Before opening the automatic sealing door at the connection point, the main control module 700 calculates the absolute value of the pressure difference between the two in real time. : In the formula, The chamber pressure after the baking module 100 is recharged. This refers to the current pressure within the glove box module 200. Only when... Less than the preset safe opening differential pressure threshold (For example The system only releases the safety lock and allows the automatic door to open when the pressure difference is ≤200Pa. This interlock mechanism physically eliminates the risk of airflow impact caused by severe pressure differences, ensuring that the precision dust control level within the glove box module 200 is not affected by the material flow process.

[0060] The vision pre-welding module 400 mainly consists of an automatic cover and a vision alignment device. In terms of hardware, the device integrates precision mechanical motion components, pneumatic actuation components and vision imaging components, aiming to achieve high-precision automatic picking and alignment of the cover plate with the tube shell.

[0061] Specifically, the precision mechanical motion component includes a three-axis Cartesian coordinate robot and a rotary axis module installed at the end of its Z-axis; the pneumatic actuation component is specifically manifested as a vacuum nozzle mechanism installed at the lower end of the rotary axis module; and the visual imaging component includes a downward-looking camera component and an upward-looking camera component.

[0062] The three-axis Cartesian coordinate robot is mounted on a precision marble base and adopts a high-rigidity aluminum alloy gantry structure. It is driven by an AC servo motor and a ground ball screw to achieve spatial movement in the X, Y, and Z directions. To meet the stringent angular accuracy requirements of microelectronic devices, the end-rotor module uses a hollow direct-drive torque motor to directly drive the load without backlash error from a reducer, achieving a repeatability accuracy better than ±2 arcseconds.

[0063] The vacuum nozzle mechanism, acting as an end effector, is connected to the negative pressure generator via an internal air passage. Considering that the casing and cover are mostly made of brittle ceramic or glass, this embodiment incorporates a flexible floating buffer mechanism between the nozzle rod and the rotating shaft. This mechanism consists of a precision miniature linear guide and a preloaded spring, providing the nozzle with an elastic stroke of 1mm to 3mm in the Z-axis direction. When the nozzle contacts the workpiece surface, the spring compression triggers the torque limiting function of the photoelectric sensor or the Z-axis motor, achieving zero-impact soft-land pickup and placement.

[0064] The visual imaging component is the core sensing unit for achieving precise alignment. The downward-facing camera component is mounted vertically downwards on the Z-axis side of the robot arm, moving with it to image the tube shell placed on the tooling table. The upward-facing camera component is fixedly mounted on the machine frame, with its lens pointing vertically upwards, to image the cover plate being sucked up by the suction nozzle. To eliminate magnification variations (i.e., perspective errors) caused by changes in workpiece height or focusing errors, both the upward and downward vision systems in this embodiment utilize dual telecentric lenses. The dual telecentric lenses achieve parallel principal rays on both the object and image sides, ensuring that measurement accuracy is unaffected by changes in object distance within the depth of field (e.g., ±5mm). Combined with a red coaxial light source and a white low-angle ring light source, the coaxial light source illuminates the smooth, gold-plated surface of the cover plate or tube shell, making it highly bright, while the low-angle ring light source highlights the physical edges of the device, thus obtaining a high-contrast binarized image.

[0065] To avoid coordinate system deviations during vision operations, the main control module 700 runs a vision alignment algorithm. The core logic of this algorithm is to eliminate deviations between the robot's coordinate system, the image coordinate system, and the workpiece coordinate system. The specific alignment process is as follows: In step S301, the downward-facing camera assembly moves above the bottom shell fixture to capture an image of the tube shell positioned on the fixture. The main control module 700 first performs image filtering, noise reduction, and distortion correction. Then, it uses a sub-pixel edge extraction algorithm to identify the contour features of the tube shell and calculates the coordinates of the tube shell's geometric center in the robot's coordinate system. and the deflection angle of the tube shell relative to the X-axis .

[0066] In step S302, the three-axis Cartesian coordinate robot drives the vacuum nozzle mechanism to pick up the cover plate from the material tray and move it above the center of the field of view of the upward-looking camera assembly. The upward-looking camera assembly captures an image of the bottom surface of the cover plate, identifies the edge features of the cover plate, and calculates the geometric center coordinates of the cover plate in the current pose of the nozzle. and cover plate deflection angle .

[0067] In step S303, the main control module 700 calculates the motion compensation amount based on the measured cover plate deflection angle data. The system's alignment target is to align the center of the cover plate with the center of the tube shell, ensuring angular alignment. During this calculation, rotation center calibration parameters must be introduced. Due to machining errors, the physical rotation axis of the nozzle typically does not coincide with the image center and the cover plate center. In this embodiment, the projection coordinates of the rotation axis in the image coordinate system are pre-calibrated using a multi-point circular approximation method. This involves controlling the suction nozzle to pick up the calibration object and rotate it at three different angles, extracting the coordinates of feature points, and fitting a circle center, which becomes the rotation center. .

[0068] First, the system calculates the required angular deviation of the cover plate. : Subsequently, based on the principle of rigid body transformation, the compensation angle rotation and position deviation are calculated, and the comprehensive translational compensation amount that the robot arm needs to perform is determined. ,in These represent the translational compensation amounts in the X and Y directions, respectively. This calculation model considers the cover plate's rotation around its center of rotation. Rotation The coupling displacement generated at that time: in, For two-dimensional rotation transformation matrix: In the above formula, This represents the initial eccentricity vector of the cover plate center relative to the axis of rotation; Represents a two-dimensional rotation transformation matrix; finally, add... Obtain the theoretical coordinates of the center of the rotated cover plate. These theoretical coordinates are related to the center of the tube shell. The difference is the compensation stroke of the X-axis and Y-axis motors.

[0069] In step S304, the main control module 700 sends the calculated compensation stroke to the servo driver to drive the X-axis and Y-axis motors to move. , Distance, driving simultaneously Shaft motor rotation Angle. After the movement is in place, the Z-axis descends, and under the protection of the flexible buffer mechanism, the cover plate is precisely fitted and placed on the tube shell. Through the above algorithm compensation, the system can ensure that the edge alignment error between the cover plate and the tube shell converges to within ±0.003mm, and the angle error is less than 0.1°.

[0070] Given that the surfaces of the components to be welded (shells and covers) are typically gold-plated or nickel-plated, exhibiting strong specular reflectivity and featuring stamped chamfers at the edges, the main control module 700 incorporates a visual recognition and correction algorithm unit. This unit employs sub-pixel-level geometric contour extraction logic based on gradient analysis to ensure stable measurement data even under complex lighting conditions. The visual recognition and correction process specifically includes the following steps: S310, Image Acquisition and Adaptive Preprocessing. After the visual imaging component acquires the original grayscale image, the algorithm unit first defines the region of interest based on the preset material size and masks background noise. To address random specular noise and texture interference on the metal surface, the system uses a 3×3 or 5×5 median filter for nonlinear smoothing, removing isolated noise points while preserving edge details. Subsequently, to achieve effective separation of the target and background, the system automatically calculates the optimal segmentation threshold using the maximum inter-class variance method. This algorithm traverses 0–255 grayscale levels to find the threshold that maximizes the variance between the foreground and background pixel sets, thereby dynamically adapting to overall brightness fluctuations caused by differences in oxidation levels in different batches of devices.

[0071] S320, Subpixel Edge Feature Extraction. After initial segmentation, the system uses the Canny operator to extract pixel-level edge contours. To overcome the limitations of the physical resolution of the imaging sensor and support the system to achieve micron-level alignment accuracy, this embodiment introduces a subpixel interpolation algorithm based on gray-level moments. This algorithm no longer treats the edge as a line, but rather as a sloping region where gray-level values ​​change abruptly. For pixels in the edge's neighborhood, the system calculates their zeroth, first, and second moments. By fitting an ideal gray-level distribution model of the edge, the precise location of the gray-level transition center is calculated, thereby improving the edge positioning resolution to 0.1–0.05 pixel equivalents. Based on the extracted subpixel edge point set, the algorithm uses the least squares method to fit the four edge lines of the shell or cover plate and calculates the intersection of the lines as corner points, thus determining the geometric center of the component. and the angle relative to the image coordinate system .

[0072] S330, Image coordinate to physical coordinate mapping transformation. This is to transform the pixel coordinates obtained from image processing... Convert to the physical world coordinates required for the robotic arm to perform the operation. The system employs an affine transformation model for spatial mapping. This model relies on a pre-executed calibration procedure, which establishes the correspondence between the image plane and the robot's motion plane by photographing a high-precision calibration board. For any feature point within the field of view, its physical coordinates are calculated using the following transformation formula: In the formula, These are the pixel coordinates of the feature points in the image coordinate system (unit: pixel). These are the physical calibration equivalents of the image sensor in the horizontal and vertical directions (unit: mm / pixel), respectively, typically ranging from 0.005 to 0.02. The rotational deviation angle (in rad) of the camera imaging array coordinate system relative to the XY motion axis system of the robot. This represents the translational offset (unit: mm) of the image coordinate origin relative to the robot's physical coordinate origin. The above parameters... All of these are constants that are calibrated and stored in the system's non-volatile memory.

[0073] S340, a compliance inspection based on weld area avoidance. While obtaining physical coordinates, the algorithm performs a weld area clearance check. The system determines the weld width based on the process settings. (For example, 0.5mm) A virtual welding exclusion zone is generated inward along the fitted edge of the pipe shell in the image. The algorithm calculates the Euclidean distance between the fitted line and the actual extracted edge contour points. If the Euclidean distance exceeds the allowable tolerance (e.g., 0.05mm) and the defect enters the welding exclusion zone, the system determines that there is a burr or defect in the material, immediately outputs an alarm signal and stops the operation to prevent sparking or airtightness failure during parallel seam welding due to foreign objects.

[0074] S350, Relative Deviation Calculation and Correction Execution. After confirming the material's appearance is acceptable, the system calculates the final correction vector of the cover plate relative to the casing. Let the physical center of the casing be... The current physical center of the cover plate is The system constructs a deviation vector. : The deviation vector After being decomposed into X-axis and Y-axis components, the results are sent as position commands to the servo controller. The servo controller then drives the robot to perform the corresponding displacement until... The alignment is converged to a preset alignment accuracy threshold (e.g., ≤0.01mm, which is set according to the device package density level), thereby achieving high-precision visual-guided alignment.

[0075] After visual alignment is completed, in order to prevent the cover plate from undergoing slight displacement during subsequent rapid transfer or roller seam welding, the system must perform a pre-welding fixing process. The pre-welding fixing device (or spot welding assembly) is symmetrically installed on both sides of the visual alignment platform and mainly consists of a high-frequency inverter DC welding power supply, a pressurizing mechanism, and alloy electrodes designed for the tube shell packaging structure.

[0076] The pressurization mechanism preferably uses a voice coil motor or a servo electric cylinder with a torque sensor as the drive source, and its output is connected to a high-sensitivity thin-film pressure sensor (e.g., range 0–50 N, resolution 0.01 N). This hardware configuration achieves millinewton-level pressure regulation accuracy, ensuring that a low-contact-resistance conductive path is established at the moment the electrode contacts the cover plate, while effectively preventing the brittle ceramic tube shell from chipping or the glass insulator from cracking due to over-impact.

[0077] The core of pre-welding process control lies in the timing coordination of welding energy release and electrode pressure application. The main control module 700 executes the automated pre-welding process according to the following logic: S360, Flexible Contact and Pressure Establishment. To balance operational efficiency and safety, the main control module 700 employs a segmented speed control strategy. The electrode initially approaches the workpiece at a high speed (e.g., 100 mm / s). When it reaches a preset safe distance (e.g., 2 mm) from the cover surface, the system automatically switches to a low-speed search mode (e.g., 2 mm / s). At this time, the system monitors the pressure sensor feedback value in real time. Once the feedback value exceeds the trigger threshold (e.g., 0.2 N), it determines that the electrode has contacted the workpiece, and the motor immediately switches to a constant torque control mode, smoothly increasing the pressure until the preset pre-pressure threshold (e.g., 5 N to 15 N) is reached. Under this pressure-holding state, the system applies a small detection voltage to measure the static contact resistance between the electrode and the workpiece. If the static contact resistance exceeds the normal range (e.g., >), the system will detect the contact resistance. This indicates the presence of oxide layer contamination or foreign matter in the insulation. The system will immediately lock the discharge circuit and trigger an abnormal alarm to prevent sparking or poor soldering from the source.

[0078] S361, Multi-Pulse Energy Release. Given the high sensitivity of microelectronic devices to thermal shock, this embodiment employs a three-stage multi-pulse discharge strategy: preheating, welding, and tempering. The main control module 700 sends timing waveform commands to the high-frequency inverter power supply: first, it outputs a low-amplitude preheating current to soften the plating metal at the contact points using resistive heat, increasing the microscopic contact area and stabilizing the contact resistance; then, it outputs a high-amplitude welding current, causing the plating metal on the cover plate and shell to reach its melting point in a very short time (milliseconds) and undergo interatomic diffusion to form a micro-melt nucleus; finally, it outputs a slowly decreasing tempering current to control the cooling rate of the melt nucleus region and eliminate internal lattice stress caused by rapid cooling.

[0079] S362 employs a closed-loop energy control system based on the Joule thermal model. During the discharge process, to ensure welding consistency across different batches of materials and variations in coating thickness, the system utilizes a constant power or constant energy closed-loop control mode. Its control principle is based on Joule's law, with the effective heat generated at the pre-soldering point as the controlled object. Effective energy... The integral calculation model is as follows: In the formula, Effective energy accumulated during the pre-welding process (unit: J); The thermal efficiency coefficient of the system (typically ranging from 0.6 to 0.8) reflects the proportion of heat actually used for nucleus formation after deducting electrode heat dissipation and bypass losses. Total discharge time (unit: ms); for Instantaneous welding current flowing through the circuit at all times (unit: A); for Dynamic resistance of the welding circuit at any time (unit: ).

[0080] The main control module 700 synchronously acquires data at a high sampling rate (e.g., 4kHz). and voltage across the electrodes Real-time calculation of dynamic resistance Furthermore, the PWM duty cycle of the power inverter circuit is dynamically adjusted using a PID algorithm to ensure the effective energy input. Target energy set by the process Deviations are strictly controlled within ±2%.

[0081] S363, Pressure Maintenance and Crystallization Cooling. After the current is cut off, the voice coil motor does not immediately remove the pressure, but enters the maintenance phase to continuously maintain the welding pressure. Duration (typically 50ms to 200ms). This step utilizes the high thermal conductivity of the electrodes to quickly remove excess heat, while simultaneously using the forging effect generated by mechanical pressure to prevent internal shrinkage cavities or cracks from forming in the melt nucleus during the liquid-solid phase transformation, thus promoting the densification of metal grains.

[0082] S364, Position Verification and Reset. After cooling, the electrode is flexibly lifted and reset. The system again calls the vision component to verify the position of the cover plate, confirming that no deviation exceeding the tolerance range occurred due to force during the pre-welding process (e.g., displacement ≤ 0.01mm), thus determining the pre-welding process to be qualified. Regarding the above process parameters, The current values ​​for each segment are typically determined in advance through orthogonal experiments and stored in an expert database, based on the resistivity, specific heat capacity, and coating thickness of the material being welded. For example, for commonly used gold-plated Kovar alloy tube shells, the preheating current amplitude is generally set to 30%–50% of the welding current, and the welding main pulse width is controlled between 5ms and 20ms.

[0083] The parallel seam welding module 500 includes two symmetrically arranged seam welding head assemblies, located on both sides of the workpiece transport platform, for simultaneously performing roll welding on two opposite edges of the tube shell. Each seam welding head assembly mainly consists of inclined conical roller electrodes, a rotary conductive and drive assembly, a linear direct-drive pressurization assembly, and a precision floating follower slide.

[0084] The conical roller electrode, used as a direct working tool, is preferably made of dispersion-strengthened copper (such as alumina copper) or copper-tungsten alloy in this embodiment to balance conductivity and high-temperature softening resistance. Its conical half-angle is designed to be 10°–25°. This angle is chosen to ensure that when the electrode contacts the edge of the casing cover, it avoids the casing sidewall or insulator protrusions, preventing mechanical interference.

[0085] To achieve stable high-current transmission of the roller electrode during continuous rotation, this embodiment employs a dedicated rotary conductive and driving assembly. An AC servo motor drives the conductive spindle to rotate via a high-insulation coupling. The conductive spindle integrates a rotary conductive slip ring with a high rated current (e.g., ≥500A). Considering the extremely high requirements for contact resistance stability during the welding process, this slip ring is preferably made of liquid metal (mercury) or a multi-contact high-silver carbon brush slip ring, thereby controlling the rate of change of dynamic contact resistance to the micro-ohm level and ensuring controllable heat generation during continuous current flow.

[0086] To address the flatness errors in the tube shell and cover plate along their length (e.g., saddle-shaped warping of the ceramic substrate), the system must ensure that the roller electrodes maintain a constant pressure against the cover plate edge. Therefore, this embodiment constructs a high-frequency response active flexible force control system based on direct drive of a linear motor and feedback from a force sensor. The precision floating follower slide is supported by low-friction coefficient cross roller guides or air-bearing guides, eliminating static friction hysteresis in mechanical transmission, allowing the thrust of the voice coil motor or coreless linear motor to be linearly converted into electrode contact pressure.

[0087] A high-precision S-type tension / compression sensor (or a miniature spoke-type load cell) is connected in series at the end of the linear direct-drive pressurization assembly to collect the pressure sensor feedback value applied to the electrodes in real time. This feedback is then sent to the main control module 700. The main control module 700 internally runs a pressure and speed coordinated control algorithm, with the specific control logic as follows: S410, Linear velocity synchronization matching calculation. Before starting welding, to prevent relative slippage caused by a mismatch between the tangential linear velocity of the roller electrode edge and the linear velocity of the workpiece stage movement, which could lead to scratches on the cover plate coating or electrode wear, the system calculates the welding speed based on the process settings. (e.g., 5m / s to 50m / s) and the effective contact radius of the current electrode Calculate the servo motor speed command : In the formula, This is the speed command for the servo motor (unit: degrees / second). Welding speed (unit: mm / s); The effective radius of the roller electrode contact point (unit: mm). Pi is a constant, and a coefficient is introduced. This is used to convert the radian angular velocity calculated based on linear velocity and radius into the angle units required by the servo driver. This conversion is necessary because electrodes wear down during use. To reduce this, the system updates this parameter through periodic tool calibration procedures or visual measurements to maintain synchronization accuracy.

[0088] S411, soft landing and initial pressure establishment. After speed synchronization is complete, the main control module 700 drives the linear motor to rapidly move the seam welding head towards the workpiece. Approaching the predetermined welding position (e.g., at a distance from the surface)... When the pressure sensor feedback value is (mm), the motor switches to torque control mode. When the contact threshold is exceeded (e.g., 0.3N to 0.5N), the system determines that the electrode has contacted the workpiece, and then linearly increases the pressure according to the preset pressure ramp rate until the set welding pressure is reached. (Typically 10N to 30N).

[0089] S412, high-frequency real-time pressure PID regulation. After entering the welding process, to address the slight changes in workpiece surface height caused by warping, the main control module 700 employs an incremental PID control algorithm based on pressure sensor feedback values. Calculate the thrust compensation amount of the motor for fluctuations. To maintain a constant contact force, the control law is as follows: In the formula, This is the current increment command output to the motor driver during the current sampling period; These are the proportional, integral, and differential gain coefficients, which are usually tuned based on the mass of the linear motor's mover and the system's damping characteristics, using the Ziegler-Nichols method or step response experiments. This represents the pressure deviation at the current moment. The update frequency of this closed-loop control circuit is set to 2kHz to 5kHz to ensure that the system can effectively suppress mechanical vibration interference within 100Hz and achieve dynamic pressure stability better than ±0.5N.

[0090] S413 features dual monitoring of position and pressure. In addition to force closed-loop control, the system also monitors the real-time position of the linear motor in parallel using a linear encoder. This serves as a safety redundancy. When When the rate of change or positional deviation exceeds the preset safety threshold (indicating drastic changes, steps, or foreign objects on the workpiece surface), the system will automatically reduce... The gain is used to soften the system stiffness or trigger emergency stop protection to prevent the brittle shell from being crushed due to an overreaction.

[0091] Furthermore, considering the uneven or asymmetrical thickness of the two edges during actual shell manufacturing, this embodiment is equipped with an independent dual-channel pressure control mechanism. Specifically, the pressure at the left head of the die... and right-side head pressure It can be set and adjusted independently. The system simultaneously monitors the positional deviation of both machine heads. This deviation value reflects the centering of the tube shell in the fixture. If If the workpiece continues to exceed the allowable range (e.g., 0.2 mm), the system will issue an alarm to prompt the operator to check the workpiece clamping status, thereby ensuring the consistency of the welds on both sides.

[0092] In order to provide stable and controllable welding energy to the parallel seam welding head, this embodiment is equipped with a high-performance welding power supply. The welding power supply is designed as a high-frequency inverter AC square wave power supply. Its hardware core mainly consists of a three-phase rectifier and filter module, an IGBT full-bridge inverter module, an intermediate frequency welding transformer, and a digital signal processing control unit based on a digital signal processor and field programmable gate array architecture.

[0093] In the specific circuit implementation, three-phase AC power (e.g., 380V / 50Hz) is rectified and filtered to convert into a smooth high-voltage DC bus voltage. The IGBT full-bridge inverter module, as the core stage of power conversion, adopts an H-bridge topology. Driven by a high-frequency pulse width modulation signal from a digital signal processor, it inverts the DC power into an AC pulse voltage with an adjustable frequency (preferably 1kHz to 10kHz in this embodiment). This high-frequency AC power is input to the primary winding of the intermediate-frequency welding transformer, and after being stepped down and boosted, it is output to the conductive spindle of the seam welding head through the secondary winding. To minimize transmission losses and reduce the impact of line inductance on the current waveform, a low-inductance water-cooled coaxial cable is used to connect the transformer secondary and the welding head. The distributed inductance of this cable is controlled within 0.5μH.

[0094] This embodiment specifically employs an AC square wave instead of a DC pulse as the output form. The fundamental reason for this is to overcome the Peltier effect in the parallel seam welding process. In the contact welding of micron-level coated metals, if DC current is used, the unidirectional current flow will trigger a thermoelectric effect, resulting in a severe asymmetry in heat generation at the contact surfaces of the positive and negative electrode rollers (for example, overheating of the positive electrode leads to electrode adhesion, while insufficient heat on the negative electrode leads to poor soldering). By outputting alternating positive and negative AC square waves, the system can force the current direction to reverse in each cycle, thereby achieving a dynamic balance of heat distribution on both sides of the weld joint over time integration, extending the service life of the roller electrodes and improving the consistency of weld width.

[0095] The digital signal processor (DSP) control unit internally runs a fully digital power control algorithm, ensuring the stability of heat input during the welding process through real-time closed-loop adjustment of output energy. The specific power control and adjustment logic is as follows: Step S420, Waveform Modulation and Dead-Time Protection. The system generates a reference waveform command based on preset process parameters. For an AC square wave, one complete fundamental frequency cycle... Includes positive half-wave time Dead time and negative half-wave time To prevent shoot-through short circuits in the upper and lower transistors of the inverter bridge arm, the DSP inserts a microsecond-level dead time at the zero-crossing point of the positive and negative half-wave switching. This time parameter is typically set to 2μs–5μs, with the specific value depending on the IGBT device's turn-off tail time. Initial PWM duty cycle. Based on the target current RMS value With the current DC bus voltage The ratio is used to set the feedforward prediction.

[0096] Step S421, high-speed synchronous sampling. A high-bandwidth Rogowski coil or Hall current sensor is installed at the welding power supply output to collect the instantaneous welding current. Simultaneously, the instantaneous voltage across the electrodes is acquired using a Kelvin four-wire voltage probe. To ensure the accuracy of power calculation, the current and voltage signals are synchronously sampled by a high-speed ADC, with the sampling frequency set to more than 10 times the switching frequency (e.g., 100kHz to 200kHz) to ensure accurate capture of the rising edge jitter and overshoot characteristics of the square wave current waveform.

[0097] Step S422, True RMS value calculation. Because square wave current contains abundant high-order harmonics, simple average value calculations cannot accurately reflect the thermal effect of the current. This embodiment uses a true RMS algorithm to calculate the actual RMS value of the current in the current cycle. The calculation formula is as follows: In the formula, For the first The actual effective value of the current for each PWM control cycle (unit: A); The fundamental period of alternating current (unit: s); The instantaneous welding current value obtained from sampling; This marks the start time of the current cycle. Subsequently, the system calculates the current valid value and the set target value. Current deviation between : Step S423, Incremental PI Regulation. To eliminate steady-state error and ensure rapid response to bus voltage fluctuations, the DSP uses an incremental proportional-integral control algorithm to calculate the duty cycle adjustment for the next cycle. The governing equations are: In the formula, This is the incremental adjustment value for the duty cycle; This is the proportional gain coefficient of the current loop, used to determine the system's response speed; This is the current loop integral gain coefficient, used to eliminate static errors; and These represent the current deviations for the current cycle and the previous cycle, respectively. Finally, the PWM duty cycle for the next cycle... Updated to: To prevent integral saturation and transformer overexcitation, the system performs calculations... Strict amplitude limiting is applied, and its value range is restricted to... Between (e.g., 0.05≤) ≤0.95).

[0098] Step S424, Constant Power Mode Switching. In some heat-sensitive packaged applications, simple constant current control can cause heat generation fluctuations due to changes in contact resistance. In this case, the system can switch to constant power control mode. In this mode, the algorithm simultaneously utilizes synchronously acquired instantaneous voltage... and instantaneous welding current Calculate instantaneous power The average power is obtained by periodic integration. As a controlled variable, the power feedback loop introduces a voltage variable, enabling this mode to automatically compensate for resistance drift caused by increased contact area due to electrode wear or workpiece surface oxidation, ensuring a constant energy density per unit length of weld. For , For parameter tuning, those skilled in the art typically connect a standard resistive load box to the load end and measure the rise time and overshoot of the system through a step response experiment to select a parameter combination that puts the system in a critically damped or slightly underdamped state.

[0099] In parallel seam welding, besides precise control of current and pressure, the stability of the welding speed (i.e., the relative tangential linear velocity between the roller electrode and the edge of the casing cover plate) directly determines the heat input density and forming quality of the weld. To address the mechanical shock and overheating issues that easily arise during the start-up and shutdown phases and at corners of traditional trapezoidal speed curves, this embodiment constructs a high-order motion planning model based on an S-curve. This model runs in a high-performance motion controller based on EtherCAT or RTEX bus, achieving smooth trajectory tracking by strictly constraining the third derivative of the position command (jerk).

[0100] The specific logic flow of the system executing programmable speed control is as follows: Step S430: Parametric trajectory analysis and look-ahead planning. The main control module analyzes the user-input shell geometry (length L, width W, corner radius R) and the target welding speed set in the process settings. The system automatically generates a geometric description of the welding path. To ensure trajectory accuracy under high-speed motion, a multi-step look-ahead algorithm is used to pre-read the path features of the subsequent N interpolation cycles (e.g., N=500), discretizing the continuous welding trajectory into a series of micro-segments. During this process, the system automatically identifies straight segments, circular transition segments, and acceleration / deceleration segments, and assigns initial dynamic constraint values ​​to each segment.

[0101] Step S431: S-shaped velocity planning based on jerk constraints. Given the brittle nature of the ceramic tube shell, any minute mechanical vibration can lead to edge chipping or microcracks. Therefore, this embodiment abandons the trapezoidal velocity curve with abrupt acceleration and adopts a seven-segment S-shaped velocity curve planning strategy: acceleration, uniform acceleration, deceleration, uniform speed, acceleration / deceleration, uniform deceleration, and deceleration. This algorithm introduces jerk. (i.e., the rate of change of acceleration) is used as the core control variable to ensure that the acceleration curve is continuous and without abrupt changes. At any moment during the acceleration phase... Instantaneous speed The calculation follows the following kinematic equations: In the formula, for Instantaneous velocity at a given moment (unit: mm / s); The initial velocity of the current interpolation cycle (unit: mm / s); The initial acceleration of the current period (unit: mm / s²) 2 ); Preset jerk limit value (unit: mm / s) 3 ).in, The specific value is usually set to the system's maximum acceleration. 10 to 50 times (e.g., 5000 mm / s) 3 ~20000mm / s 3 This parameter determines the smoothness of the inflection point of the speed curve. By inserting controlled transition zones at the beginning and end of acceleration and deceleration, the system can achieve smooth start and stop, effectively suppressing residual vibration of the machine tool.

[0102] Step S432, Adaptive Corner Velocity Suppression. In the rounded corner transition area of ​​the rectangular tube shell, maintaining high-speed welding on a straight segment can easily lead to trajectory deviation due to centrifugal force. Furthermore, since the heat dissipation area is relatively reduced at the rounded corner, constant-speed welding can cause heat accumulation and overheating. Therefore, the system applies a centripetal acceleration constraint model to automatically calculate the safe tangential velocity at the corner. The constraint relationship is as follows: In the formula, The safe tangential velocity at the corner (unit: mm / s); The radius of the fillet for the welding trajectory (unit: mm); The maximum permissible centripetal acceleration threshold for a mechanical structure (unit: mm / s²) 2 ). The value range is determined by the rigidity of the motion platform and the grip of the electrodes, and is usually set at... (i.e., 2000mm / s) 2 ~5000mm / s 2 When planning speed Greater than the calculated At that time, the look-ahead algorithm will plan the deceleration curve in advance to ensure a smooth reduction to [a certain value] before entering the rounded corner. After exiting the rounded corner, smoothly accelerate back to normal speed. .

[0103] Step S433: Position and energy are output synchronously via hardware. Because a variable speed process is introduced in step S432, triggering the welding pulse at a fixed time interval would result in excessive overlap (overheating) at slower speeds and sparse weld points (incomplete welds) at faster speeds. Therefore, this embodiment utilizes the hardware position comparison function of the motion controller to establish a real-time coupling mechanism between speed and welding energy. The system does not rely on a software timer but directly triggers the welding power supply based on the actual arc length position fed back by a grating ruler or encoder. The system outputs the welding energy based on the current real-time linear velocity. Dynamically mapped welding power supply pulse trigger frequency To maintain a constant solder joint density per unit length. Its synchronization control model satisfies: In the formula, The pulse trigger frequency (unit: Hz); The real-time linear velocity of the electrode (unit: mm / s); The expected diameter of a single weld nugget (unit: mm) for the process. This is the overlap coefficient between adjacent solder joints (typically ranging from 0.3 to 0.7). Denominator term This is the equivalent center-to-center distance of the weld joint. Through this PSO hardware synchronization mechanism, even if a sudden speed drop occurs due to a corner in step S432, the system can automatically reduce the pulse frequency with a microsecond-level response speed, ensuring the energy density of the weld. Maintaining uniformity along the geometric path enables equal-energy welding throughout complex trajectories. For the specific drive of the servo motor, those skilled in the art can employ a three-loop control strategy consisting of a position loop, a speed loop, and a current loop, which will not be elaborated upon here.

[0104] When performing precise defect identification on minute, highly reflective metal welds, conventional vision solutions often suffer from feature loss due to specular reflection and insufficient depth of field. To address this, this embodiment constructs a composite optical acquisition module consisting of a dual telecentric optical imaging component and a multi-channel hardware synchronous strobe illumination component. Specifically, the imaging front-end module is configured as an industrial camera equipped with a global shutter large-area CMOS image sensor and matching dual telecentric lenses.

[0105] Considering the warping, pin height difference, and weld edge sloping characteristics present on the welded surface, ordinary industrial lenses suffer from perspective distortion (i.e., objects appear larger when closer and smaller when farther away) due to parallax, severely affecting the quantification of micron-level defects. Therefore, this embodiment uses a dual-telecentric lens with aperture stops on both the object and image sides, allowing only principal rays parallel to the optical axis to pass through. This optical architecture not only strictly controls the telecentricity to within 0.1°, eliminating parallax shift, but also exhibits an optical shearing effect, effectively filtering out stray light interference from the curved metal surface. To ensure the ability to resolve micron-level cracks or pores, the optical system's resolution must satisfy the Nyquist sampling theorem. The required object-side optical resolution... With lens magnification and camera pixel size The following matching relationship applies between them: In the formula, Object-side optical resolution (unit: µm); The physical size of the image sensor's pixels (e.g., 3.45µm). The optical magnification of the telecentric lens (preferably 0.5X to 1.0X in this embodiment); This is the sampling factor. For a color camera using a Bayer array, The value is usually 2; for monochrome cameras, A value of 1.5 is chosen to ensure that the image edge sharpness meets the gradient extraction requirements of the detection algorithm. Based on this configuration, the physical precision of a single pixel of the component can be designed to be between 3µm and 7µm.

[0106] Because the weld seam after parallel seam welding exhibits a complex metallic texture, encompassing both specular reflection (the smooth center of the weld seam) and diffuse reflection (edge ​​oxidation marks), a single-angle light source cannot simultaneously capture all these features. This embodiment employs a combined lighting module, including a vertically incident coaxial parallel light source and a large-aperture, low-angle ring light source. These two light sources are driven by a digital lighting controller with FPGA logic, supporting microsecond-level high-speed flicker and overdrive modes, using instantaneous high-energy output to suppress ambient light interference and freeze motion blur.

[0107] The specific illumination imaging process and control logic are as follows: Step S510: Multi-channel time-division hard-triggered acquisition. When the workpiece carrier moves the welded tube shell to the inspection station, the position signal fed back by the grating ruler directly triggers the lighting controller and camera. The component adopts a single-station double-exposure timing logic, using the rising and falling edges of the hardware signal to synchronize different illumination modes. In the first trigger sequence, the controller only illuminates the coaxial parallel light source, and the light is perpendicularly incident on the workpiece surface through the beam splitter. At this time, the flat cover surface and smooth weld ridges will totally reflect the light back to the lens, presenting a high grayscale area in the image, while pits, scratches, or severely deformed areas will scatter the light outside the lens aperture, presenting a low grayscale dark area. This mode is bright-field imaging, mainly used to detect surface flatness and foreign objects on the cover.

[0108] Step S511, Dark-field texture enhancement acquisition. Within a very short time interval (e.g., 10ms to 30ms) after the first exposure, the component automatically performs a second exposure. At this time, the coaxial light is turned off, and the low-angle ring light source is instantly illuminated. The light grazes the weld area at a near-horizontal angle (e.g., 10° to 20° with the horizontal plane). Because the weld surface has a raised, fish-scale-like texture, the low-angle light can illuminate the illuminated side of the texture and form a shadow on the shaded side, thus transforming the subtle three-dimensional topography into high-contrast two-dimensional image features. This mode is called dark-field imaging, mainly used to identify topological defects such as weld breaks, cracks, and pinholes.

[0109] Step S512: Adaptive optimization of light intensity and contrast. To address the differences in reflectivity between different tube housing materials (such as gold-plated, nickel-plated, or bare ceramic), the system incorporates an adaptive brightness adjustment algorithm. This algorithm, based on the Weber contrast model, automatically adjusts the PWM duty cycle of each light source channel. The criteria for determining the optimal lighting intensity are as follows: In the formula, For the calculated Weber contrast: The average grayscale value of the region of interest (i.e., the region with defects or weld features); This represents the average grayscale value of the background area (i.e., the defect-free metal surface); To prevent the use of tiny positive constants with a denominator of zero; This is the preset minimum contrast threshold. The value is typically determined through calibration experiments on standard samples, generally ranging from 0.3 to 0.5, to ensure sufficient separation between the feature and the background. During the debugging phase, the system automatically iterates through different PWM duty cycle combinations and selects the one that best suits the background. The maximized parameter is used as the standard illumination parameter for this batch of products.

[0110] Step S513, Image Data Fusion and Transmission. The industrial camera transmits the two acquired bright-field and dark-field images to the image processing host via a high-speed bus (such as GigEVision or CoaXPress). To eliminate pixel displacement caused by mechanical micro-vibration between the two shots, the host performs sub-pixel-level rigid registration of the two frames based on feature points. Then, a weighted fusion algorithm is used to superimpose the surface information of the bright field and the texture information of the dark field to construct a composite image containing complete weld seam features, which is then used by the subsequent defect detection algorithm module. Regarding the specific selection of the aforementioned dual telecentric lenses and digital light source controller, those skilled in the art can make conventional configurations based on the actual field of view and detection speed requirements, which falls within the scope of known technology.

[0111] After receiving the composite image of the complete weld seam features transmitted by the composite optical acquisition component, the image processing host enters the core algorithm processing stage. Thanks to the aforementioned hardware-level illumination suppression and dual telecentric imaging, the raw data already possesses high signal-to-noise ratio characteristics physically. To address the random mounting tolerances of the tube shell in the carrier and to balance detection speed and accuracy, this embodiment employs a cascaded processing logic of global attitude correction, feature engineering coarse screening, and deep learning fine judgment.

[0112] Step S520: Global coordinate correction and region of interest (ROI) construction. Although step S513 eliminated inter-frame jitter during the shooting process, the tube shell still has slight degrees of freedom (e.g., ±0.1mm translation or 0.5° rotation) within the welding fixture. To achieve standardized detection, the algorithm first performs a Hough transform or sub-pixel edge search based on the bright-field image to extract the linear equations of the physical edges of the tube shell cover. The system calculates the affine transformation matrix between the current workpiece center and the theoretical reference center, mapping the image to a standard coordinate system. Subsequently, a morphological mask is used to extract the ROI region covering the theoretical trajectory of the weld. To completely preserve the high-frequency edge features of micron-level cracks while removing random lattice noise from the metal surface, this embodiment performs anisotropic bilateral filtering on the ROI region. The pixel value update follows a nonlinear combination of spatial proximity and pixel similarity, calculated using the following model: In the formula, and The center pixels before and after filtering are respectively grayscale value; For The convolution window centered on the center; Let Gaussian function in spatial domain have standard deviation The value is set to 1.5 to 2 times the expected pixel width of the crack (e.g., 3.0) to ensure that the filter kernel can cover the crack structure; Let Gaussian function be defined in the pixel domain, and its standard deviation be defined as follows: The value is set to 2 to 3 times the standard deviation of the gray level of the background area to achieve smoothing of the background texture without blurring strong edges; These are the normalized weighting coefficients.

[0113] Step S521: Microcrack feature extraction based on the Hessian matrix. For micron-sized cracks present in dark-field images, given that they appear as directional, bright ridge-like structures against a dark background, simple thresholding is easily interfered with by reflective points. This embodiment constructs a second-order derivative multi-scale filtering operator based on the Hessian matrix, utilizing the geometric characteristic of the crack having the maximum curvature in the cross-sectional direction for detection. For any pixel in the image... Calculate its Hessian matrix. : In the formula, , , These are the partial derivative responses of the image after convolving it with a Gaussian second-order differential kernel with a scale factor of σ. The Hessian matrix is ​​then solved. Two eigenvalues and (and sorted by absolute value) The system makes a judgment based on the geometric topology of the bright ridge structure: an ideal crack structure should satisfy the condition that the curvature along the crack direction is minimal ( The curvature perpendicular to the crack direction is extremely large and negative. Based on this, a crack probability response function is constructed. : In the formula, The speckled structure discrimination ratio is used to distinguish between circular holes (with similar eigenvalues) and linear cracks; This is the structural strength norm, used to distinguish noise from structure; This is the shape sensitivity threshold, typically set to 0.5; The structural strength threshold is set to 50% of the maximum value of the Hessian norm. This algorithm can suppress point noise and specifically enhance the grayscale response of linear cracks, thereby generating candidate regions for potential defects.

[0114] Step S522: Comprehensive determination by the deep learning classifier. To further eliminate false positives caused by water stains, dust, or metal scratches, this embodiment introduces a lightweight convolutional neural network (such as a pruned and optimized ResNet-18) as a secondary classification decision-maker. The system crops the suspected defect regions extracted in the above steps into image patches of a fixed size (e.g., 64×64 pixels) and inputs them into a pre-trained CNN network. The network output layer uses the Softmax function to calculate the posterior probability that the current image patch belongs to various types of defects (cracks, pores, ablation, normal texture). : In the formula, The first output of the fully connected layer Logits; This represents the total number of defect categories. The system only determines the existence of a defect when the predicted probability of a certain defect exceeds a stringent confidence threshold (e.g., 0.95). This hybrid architecture, combining traditional algorithm recall with deep learning filtering, leverages both the high sensitivity of the Hessian operator to weak cracks and the high specificity of CNNs to complex textures, effectively solving the technical challenge of balancing false negative and false positive rates in the detection of defects on highly reflective metal surfaces. Finally, the system integrates weld continuity data and defect classification results to generate a quality inspection report and sends qualified / scrap classification instructions to the sorting mechanism via an industrial bus.

[0115] Experimental example: To verify the effectiveness of the technical solution of the present invention, the applicant conducted a group comparison experiment under the conditions of the above embodiments.

[0116] Experiment 1: Comparison of Welding Energy Control Stability (Refer to the instruction manual appendix) Figure 4 ) Comparison objects: Control group: Adopting the traditional constant current control mode (set current constant at 450A, corresponding to the attached...) Figure 4 (The dashed line in neutron diagram B).

[0117] Experimental group: Employing the dynamic resistance energy closed-loop control mode of this invention (target energy set at 12 J / mm, corresponding to the attached...) Figure 4 (Solid line in neutron graph B).

[0118] Test conditions: A small amount of oxidized grease was manually applied to the surface of some of the tube cover plates to simulate contact resistance fluctuations (resistance fluctuation range 20mΩ~60mΩ, fluctuation details are shown in the appendix). Figure 3 Neutron diagram A).

[0119] The test results are shown in Table 1: Table 1 Experimental conclusion: The experimental results fully demonstrate that the dynamic resistance energy closed-loop control mode adopted in this invention is superior to the traditional constant current control mode. Under harsh operating conditions with drastic fluctuations in contact resistance (such as surface contamination), this invention can compensate for changes in the Joule heating effect in real time, reducing the dispersion of weld shear strength by approximately 85.9% (standard deviation from 8.5N to 1.2N), and increasing the fine leak detection pass rate to 99.8%. This indicates that the control scheme can effectively solve the problems of overheating or incomplete welding caused by nonlinear resistance changes, improving the adaptability of the parallel seam welding process to the surface condition of materials and the consistency of the welded products.

[0120] Experiment 2: Comparison of Corner Welding Quality and Heat Input (Refer to the instruction manual appendix) Figure 5 ) Comparison objects: Control group: Trapezoidal velocity planning + fixed-frequency pulse triggering (traditional solution, corresponding to the attached diagram) Figure 5 (The constant frequency curve of neutron diagram B).

[0121] Experimental group: S-shaped velocity planning + position and energy hardware synchronization (PSO) (This invention's solution, corresponding to the appendix) Figure 5 The velocity curve of neutron diagram A and the frequency conversion curve of sub-diagram B.

[0122] Test area: R0.5mm rounded transition area of ​​the tube shell (involving acceleration and deceleration processes).

[0123] The test results are shown in Table 2: Table 2 Experimental conclusion: The experimental results demonstrate that by introducing an S-shaped velocity planning and position / energy hardware synchronization mechanism, this invention successfully solves the problems of mechanical vibration and heat accumulation in complex trajectory welding. Data shows that S-shaped curve planning reduces trajectory tracking error by 81.2%, achieving micron-level motion control accuracy. Simultaneously, the PSO mechanism ensures that the welding energy density deviation per unit length is controlled within ±2% during non-constant velocity motion phases (such as the rounded corner deceleration zone). The combination of these two mechanisms eliminates microcracks in the ceramic substrate caused by localized overheating, achieving equal-energy welding throughout the entire path and ensuring the mechanical integrity and hermetic reliability of the device packaging.

[0124] Experiment 3: Comparison of Detection Rates of Defects in High-Reflectivity Welds (Refer to the instruction manual appendix) Figure 6 ) Dataset: Contains 1000 weld seam images (including 50 microcrack samples, 200 scratch samples, and the rest are normal textures).

[0125] Comparison Algorithms: Control group: Canny edge detection + geometric threshold screening.

[0126] Experimental group: Anisotropic bilateral filtering, Hessian matrix and CNN classification (this invention).

[0127] The test results are shown in Table 3: Table 3 Experimental conclusion: Experimental data confirm that the cascaded detection algorithm based on anisotropic bilateral filtering, Hessian matrix, and CNN classification proposed in this invention has advantages in identifying minute defects on highly reflective metal surfaces. Compared to traditional edge detection algorithms, this scheme improves the detection rate of microcracks from 62.0% to 98.0% while maintaining the industrial production cycle time (<2s / pcs), and reduces the false alarm rate from 35.5% to 1.5%. These results demonstrate that this invention can effectively suppress metal lattice noise and specular reflection interference, accurately extract topological defect features, and meet the stringent requirements of fully automated online appearance inspection of high-reliability microelectronic devices.

[0128] In summary, compared with existing technologies, it has the following beneficial effects: 1. The embodiments of the present invention calculate the dynamic resistance by collecting the real-time voltage and current of the welding circuit, and introduce the thermal efficiency coefficient to correct the Joule heat integral model. This can automatically compensate for the nonlinear changes in contact resistance caused by differences in the oxidation degree of the tube shell surface or the pressure fluctuation of the tooling fixture, ensuring that the effective energy actually acting on the molten core in each discharge process meets the preset target, effectively avoiding the phenomenon of false welding or overburning caused by energy input deviation, and ensuring the airtightness of the tube shell packaging.

[0129] 2. In this embodiment of the invention, the single-station dual-exposure logic combined with dual telecentric lenses enables the system to acquire a bright-field image reflecting surface flatness and a dark-field image reflecting three-dimensional texture under the same field of view. This avoids positioning errors caused by mechanical movement. Combined with Weber contrast-based adaptive light source adjustment and sampling factor matching design, specular reflection interference is suppressed from the hardware source, ensuring high signal-to-noise ratio and edge sharpness of the image data, and providing a high-quality data foundation for the detection of minute defects.

[0130] 3. The embodiments of the present invention adopt a cascaded defect recognition strategy that combines geometric morphology and deep learning, which effectively solves the contradiction between false detection and false detection in the context of complex textures on metal surfaces. By using anisotropic bilateral filtering and Hessian matrix eigenvalue analysis, the algorithm can specifically extract weak linear crack structures while suppressing the inherent lattice noise of the metal. By using a convolutional neural network to perform secondary semantic judgment on suspected areas, it can accurately distinguish physical cracks from non-fatal surface scratches or water stains.

[0131] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0132] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fully automated production process for parallel seam welding, characterized in that, include: The cover plate and the pipe shell to be welded are vacuum baked and then sent into a protective gas sealed environment. Complete the visual alignment and pre-welding fixation of the cover plate and the pipe shell in a sealed environment; Parallel seam welding is performed on the two opposite edges of the pre-welded cover plate and the shell. Pressure sensors are used to collect the pressure applied to the electrodes in real time. Combining the real-time electrode pressure with the welding circuit parameters, the welding energy is adjusted in a closed loop using an incremental PID control algorithm. Single-station time-division imaging is performed on the welded cover plate and pipe shell to obtain bright field and dark field images of the weld. Weld defects are identified through image processing, and the welded cover plate and pipe shell are sorted based on the defect identification results.

2. The fully automated parallel seam welding production process method as described in claim 1, characterized in that, The aforementioned visual alignment refers to using a visual imaging component to identify the pose deviation between the tube shell and the cover plate, and driving a robotic arm to perform alignment compensation, including: A downward-looking camera assembly is used to photograph the tube shell to obtain the geometric center and deflection angle of the tube shell; The top-view camera assembly is used to photograph the cover plate being sucked up by the suction nozzle to obtain the geometric center and deflection angle of the cover plate. Based on the pre-calibrated nozzle rotation center, the angular deviation between the cover plate and the tube shell, as well as the coupled displacement caused by the rotation, are calculated to generate motion commands that include translational compensation and rotational compensation to drive the robot to complete the alignment.

3. The fully automated parallel seam welding production process method as described in claim 1, characterized in that, The cover plate and the tube shell are pre-welded and fixed using spot welding electrodes. The energy control method for the pre-welding and fixing is as follows: The instantaneous welding current and the voltage across the electrodes in the welding circuit are synchronously acquired using a high-frequency sampling rate, and the dynamic resistance is calculated in real time. Based on the Joule heating principle, the product of the square of the instantaneous welding current and the real-time dynamic resistance is integrated over the total discharge time, and the integral result is multiplied by the thermal efficiency coefficient of the system to obtain the effective energy. By dynamically adjusting the power output duty cycle, the effective energy can meet the preset target value.

4. The fully automated parallel seam welding production process method as described in claim 1, characterized in that, Parallel seam welding is performed on the two opposite edges of the pre-welded cover plate and the pipe shell using roller electrodes; The motion control of the parallel seam welding adopts an S-shaped velocity curve planning strategy. The jerk is limited by constraining the third derivative of the position command. At the same time, a real-time coupling mechanism between velocity and welding energy is established. The pulse triggering frequency of the welding power source is dynamically mapped based on the real-time linear velocity of the roller electrode to maintain a constant weld point density per unit length of weld.

5. The fully automated parallel seam welding production process method as described in claim 4, characterized in that, In the closed-loop adjustment of welding energy based on real-time electrode pressure and welding circuit parameters, the welding power source adopts a high-frequency inverter AC square wave power supply. The high-frequency inverter AC square wave power supply uses high-frequency inverter AC square wave output and inserts dead time at the zero crossing point of the positive and negative half-wave switching; and the high-frequency inverter AC square wave power supply calculates the actual current effective value of the current in the current cycle based on the true RMS algorithm, and uses incremental proportional and integral control algorithms to adjust the pulse width modulation duty cycle of the next cycle.

6. The fully automated parallel seam welding production process method as described in claim 1, characterized in that, A composite optical acquisition component is used to perform single-station time-division imaging of the welded cover plate and tube shell. The composite optical acquisition component adopts a dual-sided telecentric lens and an industrial camera equipped with a global shutter image sensor. The object-side optical resolution of the composite optical acquisition component is determined by multiplying the ratio of the physical size of the image sensor pixel to the optical magnification of the telecentric lens by a sampling factor. The sampling factor is determined based on the color array structure of the camera sensor to ensure that the image edge sharpness meets the requirements of the detection algorithm for the extraction accuracy and signal-to-noise ratio of edge gradient information.

7. The fully automated parallel seam welding production process method as described in claim 6, characterized in that, The composite optical acquisition component also includes a coaxial parallel light source and a low-angle ring light source. The specific logic of the single-station time-division imaging includes: The coaxial parallel light source is lit up at the first hardware trigger timing to acquire the bright field image of the weld seam. After a preset time interval, the low-angle ring light source is switched to acquire the dark field image of the weld seam. The light source intensity is adaptively adjusted based on the Weber contrast model. The optimal lighting parameters are determined by calculating the average gray level difference ratio between the region of interest and the background region.

8. The fully automated parallel seam welding production process method as described in claim 1, characterized in that, Before identifying weld defects through image processing, anisotropic bilateral filtering is performed on the region of interest in the image, and the pixel value update logic includes: The neighboring pixels within the convolution window are summed using a weighted summation. The weighting coefficients of this summation are jointly determined by a spatial domain Gaussian function based on geometric distance and a pixel domain Gaussian function based on gray-level difference, in order to preserve high-frequency edge features while smoothing the background texture.

9. The fully automated parallel seam welding production process method as described in claim 1, characterized in that, In the image processing-based weld defect identification, the defect identification employs cascaded logic, including: A Hessian matrix is ​​constructed by convolving the image using a Gaussian second-order differential kernel. Based on the geometric topological relationship, ratio, and sum of squares norm of the two eigenvalues ​​of the Hessian matrix, a crack probability response function is constructed to extract suspected defect areas. The suspected defect region is used as the input to the convolutional neural network classifier. The Softmax function is used to calculate the posterior probability of the current region belonging to various types of defects. The existence of a defect is determined only when the predicted probability exceeds a preset confidence threshold. The types of defects include at least microcracks.

10. A fully automated parallel seam welding production system, characterized in that, The method for implementing the fully automated parallel seam welding production process according to any one of claims 1-9 includes: The glove box module (200) is filled with inert protective gas to form a sealed process environment, and a handling module (300) is arranged in the central area inside. The baking module (100) is sealed on one side of the glove box module (200), and an openable and closable automatic sealing door is provided between the two for vacuum baking of the cover plate and the tube shell to be welded. The transfer module (800) is sealed on the other side of the glove box module (200), forming a sealed interaction channel between the glove box module (200) and the external environment. The pre-welding module (400) is located inside the glove box module (200) and integrates a visual imaging component for aligning and positioning the cover plate and the tube shell, as well as a spot welding electrode component for completing the pre-welding fixation. The seam welding module (500) is located inside the glove box module (200) and is equipped with roller electrodes and motion drive components for performing parallel seam welding along the two opposite edges of the cover plate and the tube shell. The detection module (600) is located inside the glove box module (200) and is equipped with a composite optical acquisition component for acquiring bright field and dark field images of the weld seam in a single station time-division manner. The main control module (700) is electrically connected to the baking module (100), glove box module (200), handling module (300), pre-welding module (400), seam welding module (500), detection module (600) and transfer module (800), respectively; The main control module (700) is configured to: automatically transfer the cover plate and the shell between each workstation, coordinate and control the action sequence of each module, regulate the welding electrode pressure and welding circuit energy output in a closed loop, complete the image processing and recognition of weld defects, and control the sorting and circulation of finished products based on the defect judgment results.

Images