Printed circuit board welding spot three-dimensional shape high-precision analysis method
By combining multimodal sensing and digital twin models with dynamic stress simulation, the problem of multidimensional information gaps in solder joint detection is solved, enabling high-precision solder joint quality assessment and early defect identification, and supporting reliability assessment throughout the entire lifecycle.
Patent Information
- Application Number
- CN202610100693.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies struggle to simultaneously capture the geometric evolution and thermodynamic state transition of solder joints, cannot accurately identify defects at the microscale, lack stability in complex environments, and lack the capability for full lifecycle reliability assessment.
Multimodal sensing modules are used to synchronously acquire the three-dimensional morphology and thermal radiation data of the weld points. Combined with an adaptive variable resolution strategy and a digital twin model, a dynamic stress simulation module is embedded. Convolutional neural networks are used for defect prediction and graded alarms. Environmental compensation calibration and cross-production line data collaborative optimization are also performed.
It achieves high precision and stability in solder joint quality inspection and reliability assessment, can identify potential defects at an early stage, provide full life cycle reliability prediction and process improvement suggestions, and improve inspection efficiency and accuracy.
Smart Images

Figure CN122046937A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of solder joint analysis technology, specifically a high-precision method for analyzing the three-dimensional morphology of solder joints on printed circuit boards. Background Technology
[0002] The quality of solder joints on printed circuit boards (PCBs) is a core element for the long-term stable operation of electronic products, and the limitations of their detection and evaluation technologies directly affect the level of product reliability control. Current mainstream detection methods generally suffer from multi-dimensional information gaps, making it difficult to simultaneously capture the geometric morphological evolution and thermodynamic state migration of solder joints. Existing technical solutions, represented by patent CN120298605A, mainly rely on two-dimensional optical imaging combined with static thermal field distribution for quality judgment. The planar analysis mode used cannot accurately reconstruct the micron-level three-dimensional surface features of solder joints. Its single-point laser scanning mechanism is too time-consuming during data acquisition, especially in scenarios involving dense arrays of solder joints, resulting in significant inefficiency and making it difficult to effectively identify key geometric defects at the microscale, such as solder collapse and porosity distribution.
[0003] At the reliability assessment level, existing technologies for simulating the stress state of solder joints mostly remain at the static load analysis stage, failing to fully integrate the continuous mechanical vibration spectrum and temperature cycle fluctuations in the actual operating environment of the circuit board. The simplified processing methods used result in an underestimation of the creep damage accumulation process of the solder alloy, ultimately causing a large deviation between the fatigue life prediction results and the actual failure cycle.
[0004] Design flaws in the defect identification model also restrict early warning capabilities. Due to the lack of an effective correlation mechanism between deformation data and thermal radiation characteristics, the system cannot keenly perceive the local stress concentration and abnormal heat dissipation phenomena that accompany the initial initiation of microcracks inside the weld joint, resulting in the delay of the critical early warning window period.
[0005] In industrial field applications, existing methods are clearly not adaptable to environmental interference. Measurement drift caused by temperature and humidity fluctuations and electromagnetic noise from the production line will reduce the stability of the test results, and traditional calibration methods are difficult to achieve real-time dynamic compensation.
[0006] More importantly, the massive amounts of testing data generated by different production lines have long been isolated, lacking secure and reliable sharing mechanisms and collaborative optimization channels. This has prevented quality knowledge from flowing effectively within the manufacturing system, hindering the overall improvement of quality control efficiency in large-scale production.
[0007] In response to the trend of miniaturization and high density in electronic products, there is an urgent need to build a comprehensive analysis method that integrates submicron-level three-dimensional reconstruction, multi-physics dynamic coupling simulation and intelligent early warning decision-making, so as to break through the technical bottleneck of reliability assessment of solder joints throughout the entire life cycle from manufacturing process to service stage.
[0008] To address this, a high-precision method for analyzing the three-dimensional morphology of solder joints on printed circuit boards is proposed. Summary of the Invention
[0009] One of the technical problems this application aims to solve is the urgent need to construct a comprehensive analysis method that integrates submicron-level three-dimensional reconstruction, multi-physics dynamic coupling simulation, and intelligent early warning decision-making, so as to break through the technical bottleneck of reliability assessment of solder joints throughout their entire life cycle from manufacturing process to service stage.
[0010] To address the aforementioned technical problems, this application provides a high-precision analysis method for the three-dimensional morphology of solder joints on printed circuit boards, comprising the following steps: S1: Simultaneously acquire three-dimensional topographic data and thermal radiation data of the solder joint area through a multimodal sensing module; S2: Based on the material properties of the solder joint, the scanning path is dynamically planned, and an adaptive variable resolution strategy is used for laser scanning: the central area of the solder joint is acquired with ultra-high density point cloud, and the edge transition area is sampled with gradient downsampling. S3: Constructing a digital twin model of the weld joint by integrating 3D point cloud and thermodynamic features; S4: Embed a dynamic stress simulation module in the digital twin model to simulate the mechanical vibration and thermal cycling loads on the solder joints during circuit board operation; S5: Generate a potential fatigue failure index based on simulation results, and combine it with a convolutional neural network to achieve early prediction and graded alarm of defects.
[0011] In some embodiments, the multimodal sensing module in step S1 includes: a coaxial confocal laser sensor with a resolution ≤0.1μm; a short-wave infrared thermal imager with a wavelength of 3-5μm; and a miniature hyperspectral camera with a spectral range of 400-1000nm.
[0012] In some embodiments, the adaptive variable resolution strategy in step S2 specifically involves: establishing a polar coordinate system with the center of the solder joint as the origin, and the scan point density satisfying the formula: ,in The distance from the center of the circle. Peak density, ≥200 points / mm² This is the correlation coefficient for the thermal conductivity of the material.
[0013] In some embodiments, the dynamic stress simulation of step S4 includes: importing the circuit board working environment parameters: temperature cycling range, -40℃~125℃, vibration spectrum 10-2000Hz; and calculating the cumulative amount of creep-fatigue damage based on the constitutive equation of the solder alloy.
[0014] In some embodiments, the early prediction in step S5 includes: extracting microcrack initiation tendency factors through a twin model. , Peak stress, For plastic strain, For Young's modulus; when A level 1 warning is triggered when the value is greater than 0.25.
[0015] In some embodiments, a solder wettability quantitative analysis is added after step S3: the solid-liquid-gas three-phase contact line profile is extracted; the time-series change rate of the wetting angle is calculated. ,like ° / s indicates a risk of poor soldering.
[0016] In some embodiments, the training of the convolutional neural network adopts a multi-task learning architecture, with the main branch outputting defect classifications: cold solder joints, bridging, and oxidation; and the auxiliary branch outputting reliability lifetime predictions in the form of thermal cycles.
[0017] In some embodiments, environmental compensation calibration is performed before step S1: real-time monitoring of ambient temperature, humidity and electromagnetic interference intensity; and generation of a noise cancellation model using a generative adversarial network (GAN).
[0018] In some embodiments, the method supports cross-production line data collaboration: the test results of different production lines are uploaded to a blockchain evidence storage platform, and the global defect identification model is updated based on federated learning.
[0019] In some embodiments, the final generated 3D inspection report includes: holographic projection coordinates of the solder joint defect location; an interactive dynamic demonstration model of the failure process; and a list of process improvement suggestions based on the ISO 9454 standard.
[0020] The present invention has at least the following beneficial effects: 1. By simultaneously acquiring three-dimensional topographic data and thermal radiation data of the solder joint area, a multi-dimensional complementary information foundation is laid for subsequent analysis, improving the overall efficiency of printed circuit board solder joint quality inspection and reliability assessment.
[0021] 2. The application of dynamic planning scanning path and adaptive variable resolution strategy enables ultra-high density point cloud acquisition in the core area of the solder joint, while gradient downsampling is performed in the edge area, which improves data acquisition efficiency, avoids resource waste, and ensures the integrity of information in key areas.
[0022] 3. The fusion of 3D point cloud data and thermodynamic features constructs a digital twin model of the solder joint, realizing a more comprehensive and accurate digital representation of the physical state of the solder joint. The embedded dynamic stress simulation module can effectively simulate the complex mechanical vibration and thermal cycling loads borne by the solder joint in actual operation of the circuit board, providing strong support for evaluating the mechanical behavior of the solder joint under actual working conditions.
[0023] 4. Based on the potential fatigue failure index generated from the simulation results, combined with the analytical capabilities of the convolutional neural network, early identification of weld defects and effective classification of risk levels are achieved, which helps to take early intervention measures to prevent potential failures.
[0024] 5. By quantifying the temporal change rate of the wetting angle to identify the risk of cold solder joints, the detection capability of process defects is enhanced. The convolutional neural network used, through a multi-task learning architecture, can simultaneously and accurately complete the classification of defect types and the prediction of solder joint reliability and lifespan, thereby improving the analytical efficiency and practicality of the model.
[0025] 6. An environmental compensation calibration step is performed, and an adversarial generative network is used to counteract environmental noise interference, ensuring the accuracy and stability of the data acquisition process. Especially in complex industrial environments, it supports cross-production line data collaboration, uses blockchain technology to ensure the credible storage of test results, and continuously optimizes the performance of the global defect identification model through a federated learning mechanism, giving it good adaptability and continuous evolution capabilities.
[0026] 7. The final generated 3D inspection report not only contains accurate weld defect location information, but also provides an interactive dynamic demonstration of the failure process and process improvement suggestions based on international standards, providing intuitive and operational guidance for production decisions and quality control. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of the method steps of the present invention; Detailed Implementation The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] Example 1, see Figure 1 This invention provides a technical solution: a high-precision analysis method for the three-dimensional morphology of solder joints on printed circuit boards, comprising the following steps: S1: Simultaneously acquire three-dimensional topographic data and thermal radiation data of the solder joint area through a multimodal sensing module; S2: Based on the material properties of the solder joint, the scanning path is dynamically planned, and an adaptive variable resolution strategy is used for laser scanning: the central area of the solder joint is acquired with ultra-high density point cloud, and the edge transition area is sampled with gradient downsampling. S3: Constructing a digital twin model of the weld joint by integrating 3D point cloud and thermodynamic features; S4: Embed a dynamic stress simulation module in the digital twin model to simulate the mechanical vibration and thermal cycling loads on the solder joints during circuit board operation; S5: Generate a potential fatigue failure index based on simulation results, and combine it with a convolutional neural network to achieve early prediction and graded alarm of defects.
[0029] Specifically, by synchronously acquiring the three-dimensional morphology and thermal radiation data of the solder joint through the multimodal sensing module, the aim is to establish a correlation mapping between physical state and thermodynamic behavior, overcome the analysis limitations of a single data source, capture the coupling effect of solder joint deformation and temperature field, and provide more comprehensive basic information for subsequent analysis.
[0030] An adaptive variable resolution scanning strategy based on material properties is employed. This strategy involves differentiated data acquisition according to the mechanical importance of different regions of the weld joint. The central region, as a stress concentration area, utilizes ultra-high density point cloud acquisition to ensure accurate capture of key mechanical features. Gradient downsampling is applied to the edge regions to balance detection efficiency and avoid data redundancy, thus ensuring analytical accuracy in the core area while significantly improving overall detection efficiency.
[0031] A digital twin model integrating 3D point cloud and thermodynamic features is constructed to create a digital twin of the physical state of the solder joint. This model breaks through the limitations of traditional static modeling, providing a high-fidelity foundation platform for dynamic simulation. The core concept of embedding the dynamic stress simulation module is to simulate real-world working conditions and loads. By importing temperature cycling parameters and vibration spectrum data, and combining them with the constitutive equations of the solder alloy, the cumulative creep-fatigue damage is calculated. This extends solder joint reliability assessment from the laboratory environment to actual working conditions, significantly improving the engineering value of the predictions.
[0032] Based on simulation results, a fatigue failure index is generated and coupled with a convolutional neural network to achieve a closed-loop technology from physical model to intelligent judgment. This method establishes a precise early warning threshold by quantifying microcrack initiation tendency factors and simultaneously completes defect classification and lifespan prediction using a multi-task learning architecture. This design not only meets the needs of immediate quality inspection but also provides reliability prediction throughout the product's entire lifecycle.
[0033] The overall technical approach includes environmental compensation calibration and cross-production line collaboration mechanisms. It eliminates environmental interference through generative adversarial networks and builds a distributed quality database using blockchain notarization and federated learning. The final generated 3D inspection report integrates holographic coordinate positioning, dynamic demonstration of the failure process, and standardized improvement suggestions, forming a complete solution from problem diagnosis to process optimization. Through the organic integration of multiphysics coupling modeling, operating condition simulation, and intelligent early warning, it improves inspection accuracy while achieving a strategic transformation in quality control from post-event judgment to pre-event prediction.
[0034] The process for generating the potential fatigue failure index is as follows: The first step is to calculate the cumulative material damage. Based on transient thermo-mechanical coupling simulation data, key physical quantities of the weld joint are extracted. Equivalent plastic strain amplitude, Creep strain energy density, in MJ / m³. Stress triaxiality is the principal stress ratio; single-cycle damage is calculated using a modified Coffin-Manson-Basquin model. ,in The fatigue ductility coefficient, The fatigue ductility index, The fatigue strength coefficient is determined by decomposing the random vibration spectrum into k load blocks using the rainflow counting method, combined with Miner's linear cumulative damage rule. This value quantifies the tendency of solder alloys to develop microcracks. >0.3 indicates accelerated grain boundary slip.
[0035] Next is the modeling of critical region feature factors, which defines spatial distribution parameters based on the geometric characteristics of BGA solder joints. This represents the maximum principal stress gradient, in MPa / μm. The coefficient of variation for the thickness of the interface IMC layer is the ratio of standard deviation to mean. The displacement coordination between adjacent weld points is represented in μm. High-risk areas are located and calculated using finite element nodal topology analysis. ,in To exceed the material's yield strength 85% of the area, This represents the total cross-sectional area of the solder balls. When this factor exceeds 0.15, the risk of interface peeling increases by 5 times.
[0036] Next is the construction of failure path prediction values, and the establishment of a crack propagation model based on fracture mechanics: Calculate the J integral at the crack tip: Where W represents strain energy density, that is, the deformation energy stored per unit volume of the material. It is the stress vector on the integration path Γ. Let Γ be the displacement field vector, ds be the arc length infinitesimal element along the integration path Γ, and dy be the coordinate differential of the path normal.
[0037] The Paris formula describes the expansion rate: ,in C is the crack propagation rate (the increment of crack length per unit cycle), usually measured in mm / cycle or m / cycle. C and m are material constants (Paris constants), determined experimentally. C reflects the material's inherent resistance to crack propagation, and m is an exponent of the sensitivity of the crack propagation rate to the stress intensity factor amplitude. ΔK is the stress intensity factor amplitude (ΔK = K_max - K_min), measured in MPa·√m, characterizing the range of stress intensity variation near the crack tip and being the main mechanical parameter driving fatigue crack propagation.
[0038] A microstructure sensitivity parameter β (tin grain orientation factor) is introduced to describe the amplification effect of the crystallographic orientation of tin grains inside the solder joint on the local stress / strain field. The value range is: β ≥ 1. Mechanism of action: Grains with specific orientations exacerbate local deformation, making the actual crack driving force ΔK_effective higher than the macroscopically calculated ΔK, i.e., ΔK_effective = βΔK.
[0039] Final predicted failure cycle number: ,in Predicted number of failure cycles (total lifetime from initial crack propagation to critical size). The integral of crack length *a* (from initial size to critical size), *a*0 is the initial crack size (the detected defect or assumed microcrack size), *a*0 is the integral of crack length *a*0. c Critical crack size (the size at which unstable propagation leads to complete fracture). The effective stress intensity factor amplitude is a function of the macroscopic stress field and crack geometry. The effective alternating stress amplitude acting on the crack plane.
[0040] Spatiotemporal feature fusion and threshold setting Multi-scale fusion is achieved through a 3D convolutional neural network: Input layer: Finite element mesh data is converted into a 64×64×8 tensor (spatial coordinates × physical quantity); Convolutional layer: Damage feature map is extracted using a 5×5×3 kernel; Fully connected layer: Weighted integration. , , Generate composite index The warning thresholds are set based on accelerated testing: a value less than 0.65 indicates a safe state. A value of 0.65 or higher triggers a warning state, corresponding to a load cycle count not exceeding 100,000. A value of 0.85 or higher indicates a high-risk state, corresponding to a load cycle count not exceeding 10,000.
[0041] The core mechanism of this index quantifies the failure mechanism of material damage accumulation through physical models, identifies key vulnerable areas by combining geometric features, and predicts crack propagation paths. Its spatiotemporal convolution fusion technology overcomes the limitations of traditional single-point assessment, enabling early warning three thousand temperature cycles earlier than the strain energy density method. Real-world application examples show that in the automotive electronics field, when the index exceeds 0.7, scanning electron microscopy verification results show a void aggregation rate of over 18% in the solder joints. In contrast, simultaneous X-ray inspection can only identify void rates exceeding 30%, confirming that this method possesses early micro-damage identification capabilities.
[0042] The multimodal sensing module in step S1 includes: a coaxial confocal laser sensor with a resolution ≤0.1μm; a short-wave infrared thermal imager with a wavelength of 3-5μm; and a miniature hyperspectral camera with a spectral range of 400-1000nm.
[0043] Specifically, the design of the multimodal sensing module follows the multi-dimensional physical characteristic characterization requirements of solder joint quality inspection. A coaxial confocal laser sensor provides sub-micron-level three-dimensional topographic data, with a resolution of ≤0.1μm, accurately capturing microscopic undulations and microcrack features of the solder. A short-wave infrared thermal imager operates in the 3-5μm band, exhibiting high sensitivity to solder temperature changes and accurately reflecting the heat distribution and cooling anomalies during the welding process. A miniature hyperspectral camera covers the 400-1000nm spectral range, identifying optical features such as oxide film formation, flux residue, and changes in intermetallic compound composition on the solder joint surface. The simultaneous acquisition of these three sensing technologies forms a complementary data source: three-dimensional topographic data characterizes geometric structural integrity, thermal radiation data reflects thermodynamic stability, and spectral data reveals changes in material chemical properties. This multi-physics collaborative detection mechanism overcomes the limitations of single-sensor modes; for example, topographic data alone cannot determine microscopic thermal anomalies in cold solder joints, and simple thermal imaging struggles to identify microcracks that do not affect heat conduction. By using hardware-level time synchronization technology, we ensure that data from different modes have strict spatiotemporal consistency, providing accurate input for the subsequent construction of digital twin models of solder joints. This enhances the ability to detect latent defects in solder joints, especially for failure modes that are difficult to detect by traditional methods, such as early material degradation caused by thermal cycling and microstructural damage caused by mechanical stress. This lays a solid data foundation for subsequent reliability analysis.
[0044] The adaptive variable resolution strategy in step S2 is as follows: A polar coordinate system is established with the center of the solder joint as the origin, and the scan point density satisfies the formula: ,in The distance from the center of the circle. Peak density, ≥200 points / mm² This is the correlation coefficient for the thermal conductivity of the material.
[0045] Specifically, this design principle stems from the spatial non-uniformity of solder joint failure mechanisms. The central region of the solder joint bears the main thermomechanical stress during circuit board operation, and micro-cracks typically initiate preferentially in this region. Therefore, the strategy sets the peak density in the central region to be no less than 200 sampling points per square millimeter, a density level sufficient to accurately capture micrometer-level deformation characteristics.
[0046] The stress gradient in the edge transition region exhibits a decreasing trend, and the strategy employs an exponential decay formula to dynamically control the point cloud density. The decay rate is adjusted in real time by the material thermal conductivity correlation coefficient λ, establishing a direct correlation between the scanning strategy and the material's physical properties. For high thermal conductivity materials such as silver paste, maintaining high edge region resolution is necessary to monitor rapid thermal diffusion effects; while for low thermal conductivity materials such as tin-lead alloys, a faster sampling rate decay is permissible.
[0047] This design achieves three optimization goals. The primary goal is to ensure the integrity of geometric data in stress concentration areas, identifying micron-level voids and cracks that are easily missed by traditional methods through ultra-high-density sampling. The secondary goal is to significantly improve scanning efficiency and avoid resource waste caused by excessive data acquisition in low-failure-risk areas. The ultimate goal is to enable the system to automatically adapt to the failure behavior characteristics of different solder alloys through physical property-driven dynamic resource allocation.
[0048] Practical applications have verified that this strategy can improve the overall system performance. In typical ball grid array solder joint detection, it reduces the amount of invalid data by 40% compared to traditional uniform scanning, while increasing the resolution of key stress areas to more than three times that of traditional methods. The exponential function form ensures the smooth transition of point cloud density, effectively avoiding the data abruptness problem caused by traditional partitioned scanning modes in boundary regions. This continuity provides a stable and reliable geometric input foundation for subsequent stress field simulation calculations of digital twin models.
[0049] Step S4 dynamic stress simulation includes: importing circuit board working environment parameters: temperature cycling range, -40℃~125℃, vibration spectrum 10-2000Hz; calculating creep-fatigue damage accumulation based on solder alloy constitutive equation.
[0050] Specifically, this design closely matches the actual service conditions of electronic products, with a temperature cycling range covering industrial-grade standards from -40℃ to 125℃, simulating the complete thermal shock process from extreme cold to high temperature environments. The vibration spectrum covers the 10-2000Hz frequency band, characterizing transportation vibrations and broadband mechanical excitations during equipment operation, thus constructing load boundary conditions close to real-world scenarios.
[0051] The calculation of creep-fatigue damage accumulation using the constitutive equation of solder alloys hinges on establishing a quantitative correlation between microscopic deformation and macroscopic failure of the material. The formula for calculating the damage accumulation D is: Key parameters include: β represents the plastic strain amplitude generated by a single temperature cycle, which is obtained through thermo-mechanical coupled finite element analysis; c is the fatigue ductility coefficient related to the solder composition, for example, the typical value of SAC305 alloy is about 0.65; β is the fatigue index, which reflects the material's resistance to cyclic deformation. The formula represents the number of failure cycles. Based on Miner's linear cumulative damage theory, it quantifies the intrinsic relationship between plastic strain and material life.
[0052] The advantages of this computational method are threefold. First, by correlating the material's microscopic properties through constitutive equations, it avoids misjudgments caused by simple stress analysis; for example, high-stress, low-strain states and low-stress, high-strain states have completely different failure mechanisms. Second, it quantifies the damage accumulation process, enabling the identification of early-stage latent damage, which traditional methods require waiting for crack propagation to reach a visible stage before detection. Finally, the parametric model supports dynamic correction, updating the c-value in real time when fluctuations in solder composition are detected, improving prediction accuracy. This module ultimately outputs the damage evolution curve of the solder joint throughout its entire lifespan, providing precise timeline references for preventative maintenance.
[0053] The early prediction in step S5 includes: extracting microcrack initiation tendency factors through a twin model. , Peak stress, For plastic strain, For Young's modulus; when A level 1 warning is triggered when the value is greater than 0.25.
[0054] Specifically, this design is based on the physical mechanism of solder joint fatigue failure. Peak stress reflects the degree of local stress concentration, and plastic strain characterizes the cumulative amount of microscopic deformation in the material. The product of the two reflects the internal damage energy density of the material. Young's modulus is introduced as the denominator to eliminate the influence of differences in the alloy hardness of different solders on the evaluation results, making this factor a dimensionless and universal criterion. When this factor exceeds the threshold of 0.25, it indicates that the internal damage energy of the solder joint has reached a critical state, and microscopic dislocations begin to converge to form initial crack nuclei. This early warning mechanism has a clear theoretical basis in materials mechanics and conforms to the energy critical condition for crack nucleation set in dislocation theory.
[0055] The core purpose of using this factor is to detect signs of material degradation before physical cracks become visible. Traditional detection methods rely on optically visible macroscopic cracks, while this factor, by quantifying microscopic plastic deformation energy, can issue an early warning 3-5 thermal cycles before crack initiation, reserving a critical time window for production line process adjustments and preventing defective welds from flowing into subsequent assembly stages.
[0056] The advantage of this design is that it simplifies complex multiaxial stress states into single-valued criteria, significantly reducing the complexity of engineering applications. Calculations based on intrinsic physical quantities avoid reliance on black-box models, ensuring interpretability of the results. A threshold of 0.25 has been validated through accelerated life testing, corresponding to the failure threshold of approximately 1000 thermal cycles under actual operating conditions, with a warning accuracy exceeding 92%. This is superior to traditional methods relying on visual inspection or single stress analysis, and is particularly valuable for predicting early failures of invisible solder joints on the bottom of BGA packages.
[0057] After step S3, a quantitative analysis of solder wettability is added: the contact line profile of the solid-liquid-gas three phases is extracted; the time-series change rate of the wetting angle is calculated. ,like ° / s indicates a risk of poor soldering.
[0058] Specifically, the design of the quantitative analysis of solder wettability originates from the physicochemical mechanism of solder interface formation. The solid-liquid-gas three-phase contact line profile directly reflects the spreading dynamics of molten solder on the surface of the metal pad, and its morphological evolution is closely related to the solder surface tension, substrate adhesion, and flux activity. The wetting angle change rate is used as the core criterion; its physical essence is to quantify the rate of change of solid-liquid interfacial energy during the solidification process of liquid solder. A normal soldering process exhibits continuous and smooth contact line contraction, corresponding to a wetting angle change rate typically less than 0.5 degrees per second. When the change rate exceeds this threshold, it indicates abnormal resistance at the solder solidification front, usually caused by localized surface energy abrupt changes due to incomplete removal of the pad oxide layer, plating contamination, or uneven heat distribution.
[0059] The extraction of the three-phase contact line contour was achieved using a hyperspectral imaging system. During the holding phase above the solder liquidus temperature, the system acquired top-view images of the solder joint at a rate of 100 frames per second. The image processing workflow first utilized the difference in spectral reflectance between the molten solder zone and the solid substrate in the near-infrared band to perform region segmentation using a 900nm feature band. Subsequently, a sub-pixel edge detection algorithm was employed to extract the set of points with the maximum gradient in the transition region along the outer edge of the solder. A key step was applying morphological topology analysis to identify the contact line trajectory at the interface between the gaseous meniscus and the solid substrate, constructing a continuous contour curve through cubic spline interpolation. This process required simultaneous compensation for image distortion caused by thermal disturbance to ensure a spatial resolution of 5μm.
[0060] Setting a threshold of 0.5 degrees per second for the rate of change of the wetting angle has a clear engineering basis. When the rate of change of the wetting angle exceeds this value, the interfacial reaction energy barrier at the solder solidification front rises to a critical state, leading to discontinuities in the formation of the intermetallic compound layer. This microscopic defect will develop into micron-level interfacial separation during subsequent cooling, but the macroscopic solder joint morphology still meets the acceptance criteria. This criterion can identify the risk of poor solder joints 3-5 seconds before the solder joint solidification is complete, two process steps ahead of traditional X-ray inspection.
[0061] This design transforms the difficult-to-quantify wetting process into measurable physical parameters. By capturing the dynamic behavior of the contact line, it effectively distinguishes between normal solidification dominated by surface tension and non-wetting phenomena caused by interface contamination. Compared to static wetting angle measurement, the time-series change rate index is more sensitive to process fluctuations such as flux activity decay and insufficient nitrogen protection. Actual production line verification shows that this solution reduces the rate of missed solder joints from 17% with traditional optical inspection to 3.2%, while avoiding misjudgments caused by over-reliance on indirect parameters such as thermal history curves. The detection effectiveness is significantly improved, especially for hidden issues in the center pads of QFN packages.
[0062] The training of the convolutional neural network adopts a multi-task learning architecture. The main branch outputs defect classification: poor solder joint, cold solder joint, bridging, and oxidation; the auxiliary branch outputs reliability lifetime prediction, in units of thermal cycles.
[0063] Specifically, the multi-task learning architecture is designed based on the intrinsic correlation between solder joint quality characteristics and reliability lifetime. The main branch performs defect classification, focusing on visual pattern recognition of four typical defects: cold solder joints, bridging, and oxidation. The auxiliary branch performs reliability lifetime prediction, with its output being the number of thermal cycles required for failure. Both branches share the front-end feature extraction layer of the convolutional neural network, enabling the model to simultaneously learn both the apparent features of defects and deep performance degradation features.
[0064] During training, the main branch drives the network to extract explicit features such as weld joint morphology and texture through a classification loss function (e.g., cross-entropy). The auxiliary branch guides the network to capture implicit features related to material fatigue, such as microcrack propagation trends and intermetallic compound distribution, through a regression loss function (e.g., mean squared error). The weight updates of the shared layers are simultaneously affected by both loss functions, forcing the network to discover common feature expressions for both types of tasks. For example, the local micropore features generated by cold welding defects are both the basis for the main branch to judge cold welding and a key indicator for the auxiliary branch to predict shortened lifespan.
[0065] The advantage of this architecture lies in the synergistic enhancement effect of feature learning. The accurate identification of defects by the main branch provides semantic constraints for lifetime prediction, preventing the regression branch from misjudging surface oxidation as severe failure. The lifetime prediction task of the auxiliary branch supplements the main branch with physical failure knowledge. For example, when the network learns that a specific shape of solder bulge corresponds to high stress concentration, it will improve the sensitivity of identifying cold solder joints. Experimental data shows that after introducing the lifetime prediction task, the confusion rate between cold solder joints and cold solder joints by the main branch decreases by about 17%, because although the two have similar surface shapes, there are quantifiable differences in their internal stress distribution.
[0066] Traditional methods require separate defect detection and accelerated aging tests, while this architecture directly outputs a physical mechanism-based lifetime prediction value through an auxiliary branch, allowing the production line to implement tiered control based on this value. For example, solder joints with a lifetime prediction value below 2000 cycles can trigger process parameter optimization in advance, even if the defect classification threshold has not yet been reached. This design elevates post-event detection to pre-event warning, significantly reducing the field failure rate of electronic components.
[0067] Before step S1, environmental compensation calibration is performed: real-time monitoring of ambient temperature, humidity and electromagnetic interference intensity; and a noise cancellation model is generated using a generative adversarial network (GAN).
[0068] Specifically, the environmental compensation calibration design addresses the impact mechanism of complex interference in industrial environments on precision detection, monitoring environmental temperature, humidity, and electromagnetic interference intensity in real time, stemming from three core physical effects. Temperature drift causes laser wavelength shift and infrared thermal radiation reference distortion; humidity changes cause fluctuations in air refractive index, affecting the optical measurement path; and electromagnetic interference directly couples into the sensor's analog circuitry, generating parasitic signals. These interferences introduce systematic errors in micron-level detection; for example, a temperature fluctuation of 0.5°C can cause a measurement deviation of 0.8μm in a laser displacement sensor.
[0069] A noise cancellation model is constructed using a generative adversarial network (GAN), whose training process involves two deep neural networks. The generator network learns to separate environmental noise features from the original mixed signal, while the discriminator network undergoes adversarial training by comparing the generated results with real noise samples. The network takes data from three environmental sensors as input, including temperature gradient distribution, humidity change rate, and electromagnetic spectrum characteristics, while also incorporating the raw output streams from multimodal sensors. After adversarial training, the generator constructs the environmental noise generation function N=f(T,H,E), where T represents the temperature field, H represents the humidity field, and E is the electromagnetic interference time-domain spectrum. During the inference phase, the model outputs noise estimates in real time, performing vector subtraction from the sensor data to achieve noise stripping.
[0070] The key advantage of this design lies in its ability to handle nonlinear interference. Traditional digital filtering has limited effectiveness in suppressing sudden electromagnetic noise from frequency converters, while the GAN model captures the complex mapping relationship between noise and the environment through adversarial training. Experimental data shows that during the start-up and shutdown of arc welding equipment, the system improves the signal-to-noise ratio of the hyperspectral camera from 15dB to 28dB. More importantly, the model has online adaptability; when environmental sensors detect new interference patterns, it can be fine-tuned within 200ms through incremental learning with small samples.
[0071] This compensation mechanism ensures the system maintains measurement stability in workshop environments ranging from -10℃ to 50℃, with the laser sensor repeatability error controlled within ±0.12μm. The direct benefit is a 43% reduction in the false positive rate, particularly preventing the misidentification of signal jitter caused by electromagnetic interference as microcracks. As a preliminary step in the detection process, its calibration accuracy directly determines the validity of subsequent twin model input data, laying a reliable foundation for the entire quality prediction chain.
[0072] The method supports cross-production line data collaboration: the test results of different production lines are uploaded to the blockchain evidence storage platform, and the global defect identification model is updated based on federated learning.
[0073] Specifically, the cross-production line data collaboration design aims to resolve the core contradiction in multi-site collaboration within the manufacturing industry. Different production lines are physically isolated and their data ownership is independent, but the sparse distribution of defect samples necessitates the integration of global data to train high-precision models. A blockchain-based evidence platform provides an immutable trust anchor for the data. The inspection results from each production line generate timestamped hash values and are uploaded to the blockchain, preserving data sovereignty while establishing a traceability mechanism. A federated learning framework enables collaborative model evolution. Each node trains its model locally using private data, exchanging only encrypted model parameter gradients with the central server for aggregation. This architecture ensures that the original inspection data never leaves its local area, complying with the ISO / IEC 27001 data security standard.
[0074] This design aims to build a distributed defect knowledge base. A single production line has a limited number of defect samples, especially for latent defects with an incidence rate below 0.1%. Traditional methods require months of data accumulation to effectively model these defects. Federated learning integrates the feature learning results from N production lines in real time, reducing the model update cycle from weeks to hours. For example, a newly emerging chip warpage defect feature on a production line can be synchronized to all production line models within 24 hours via federated learning, enabling production lines that haven't encountered the defect to acquire the ability to identify it.
[0075] The advantages of this design are severalfold. First, it overcomes the limitations of data silos, improving model recognition accuracy by 12-15 percentage points while protecting trade secrets. Second, blockchain-based evidence storage enables a complete quality traceability chain covering the supplier system, increasing the efficiency of defect root cause analysis by 40%. Finally, the edge computing characteristics of federated learning reduce raw data transmission by 90%, saving approximately 170,000 RMB in bandwidth costs per production line annually. This collaborative mechanism is particularly suitable for fields with zero tolerance for defects, such as automotive electronics. When a solder joint failure is found on a circuit board for a particular vehicle model at the final assembly plant, it is possible to trace back to the process fluctuations of the specific batch of solder paste supplier, achieving supply chain-level quality control.
[0076] The final 3D inspection report includes: holographic projection coordinates of the solder joint defect location; an interactive dynamic demonstration model of the failure process; and a list of process improvement suggestions based on the ISO 9454 standard.
[0077] Specifically, the design of the 3D inspection report follows the core requirements of the electronics manufacturing industry for quality data visualization and operability. The generation of holographic projection coordinates is derived from the spatial registration technology of multimodal sensor data. The laser 3D point cloud, thermal imaging temperature field and hyperspectral feature map are integrated and positioned in a unified coordinate system, so that maintenance personnel can accurately know the spatial distribution of defects without disassembling the package. In particular, the positioning efficiency of the bottom solder joint of BGA package is significantly improved.
[0078] The dynamic demonstration model is built based on time-series data from twin simulations. By extracting the evolution of physical quantities such as peak stress and plastic strain, a physics engine drives the visualization of the material failure process. The model supports rotation, scaling, and time-axis dragging operations, intuitively presenting the full-cycle behavior of microcracks from initiation to fracture. This allows process engineers to observe the changing patterns of stress concentration areas within weld joints and understand the root causes of failure.
[0079] The generation of the process improvement suggestion list uses a rule-based reasoning engine to establish a mapping relationship between detected defect characteristics and clauses of the ISO 9454 standard. For example, when solder balling is detected, it automatically associates the standard with clauses regarding insufficient solder paste activity or mismatched reflow soldering temperature profiles, and outputs specific parameter adjustment ranges. This standards-based diagnostic mechanism ensures the authority and enforceability of the suggestions.
[0080] Holographic coordinates overcome spatial cognitive barriers to defect localization, dynamic models reveal the temporal evolution of failure mechanisms, and a list of standard recommendations provides proven improvement paths. These three elements form a complete "localization-analysis-resolution" closed loop, upgrading quality analysis from traditional two-dimensional image interpretation to three-dimensional engineering decision support. Practical applications show that this design shortens process optimization cycles by approximately 40%, and in particular, provides a systematic solution for the synergistic improvement of composite defects in multilayer stacked packages.
[0081] Example 2: The operation process of this method is as follows: The process begins with the synchronous acquisition of multi-source sensor data. During the solder joint thermal cycling test, a high-frame-rate infrared thermal imager continuously records the temperature field distribution of the solder joint, a laser displacement sensor scans the surface morphology of the solder joint with a precision of 10 micrometers, and a distributed strain gauge array monitors the substrate deformation data. All sensor data are synchronized at the millisecond level through timestamps and spatial coordinates, constructing a sequence of snapshots of the solder joint status in three-dimensional space.
[0082] Sensing data is input into a digital twin model for physical field reconstruction. The stress-strain distribution within the weld joint is calculated using the finite element method, with a focus on extracting the peak stress region and the plastic strain accumulation region. At the end of each thermal cycle, the system automatically calculates a microcrack initiation tendency factor, which is a specific combination of peak stress, plastic strain, and Young's modulus. When this factor exceeds a threshold of 0.25 for three consecutive cycles, [the crack initiation tendency factor is determined]. The system triggered a Level 1 warning and initiated an in-depth analysis process.
[0083] During the defect confirmation phase, the system utilizes X-ray tomography to perform sub-micron level layered imaging of the warning solder joints. A 3D image segmentation algorithm identifies the initial location of the microcracks, and thermomechanical simulation data is used to determine the direction of the crack propagation vector. At this point, the spatial registration module begins operation, matching the crack coordinates with the pad positions and component layouts in the PCB design file to generate the defect spatial coordinate dataset required for holographic projection.
[0084] The dynamic demonstration model is constructed by coupling multiphysics data. Stress field evolution data for 20 thermal cycles before and after the warning time point are extracted, and the material point method is used to simulate the elastoplastic behavior of the solder alloy. The visualization engine transforms the material yielding process into color gradient changes, and the crack propagation path is presented as a dynamically growing red trajectory line. Users can adjust the viewing angle through the interactive interface or drag the timeline to view the internal stress distribution at a specific number of cycles.
[0085] The process improvement suggestion generation module activates the rule inference engine. The system matches the detected crack location characteristics (such as interface cracks / body cracks) and stress concentration factors (range 1.8-3.2) with the ISO 9454 standard clause library. For example, when solder-copper interface separation is detected, it automatically associates the "excessive thickness of intermetallic compound in solder" clause in the standard and recommends reducing the peak reflow temperature by 8-12°C based on thermal history data. All suggestion entries are accompanied by specific standard clause numbers and implementation priority assessments.
[0086] The final 3D inspection report is generated through a data fusion platform. A holographic coordinate system works in conjunction with augmented reality equipment to overlay and display the crack location projection onto the physical sample. A dynamic demonstration model is embedded with an interactive timeline controller, and the process recommendation list is linked to the process parameter database of the enterprise's MES system. Once the report is confirmed, the system automatically generates a quality traceability code containing spatial coordinates, failure animation links, and a sequence of standard clause references. This code is then written into a blockchain-based evidence storage system to complete the closed loop.
[0087] 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 process, method, article, or apparatus.
[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
Claims
1. A high-precision analysis method for the three-dimensional morphology of solder joints on a printed circuit board, characterized in that, Includes the following steps: S1: Simultaneously acquire three-dimensional topographic data and thermal radiation data of the solder joint area through a multimodal sensing module; S2: Based on the material properties of the solder joint, the scanning path is dynamically planned, and an adaptive variable resolution strategy is used for laser scanning: the central area of the solder joint is acquired with ultra-high density point cloud, and the edge transition area is sampled with gradient downsampling. S3: Constructing a digital twin model of the weld joint by integrating 3D point cloud and thermodynamic features; S4: Embed a dynamic stress simulation module in the digital twin model to simulate the mechanical vibration and thermal cycling loads on the solder joints during circuit board operation; S5: Generate a potential fatigue failure index based on simulation results, and combine it with a convolutional neural network to achieve early prediction and graded alarm of defects.
2. The method for high-precision analysis of three-dimensional morphology of solder joints on printed circuit boards according to claim 1, characterized in that: The multimodal sensing module in step S1 includes: Coaxial confocal laser sensor with a resolution ≤0.1μm; Shortwave infrared thermal imager, 3-5μm band; Miniature hyperspectral camera with a spectral range of 400-1000nm.
3. The method for high-precision analysis of three-dimensional morphology of solder joints on printed circuit boards according to claim 1, characterized in that: The adaptive variable resolution strategy in step S2 is as follows: Establish a polar coordinate system with the center of the solder joint as the origin, and the density of scanned points satisfies the formula: ,in The distance from the center of the circle. Peak density, ≥200 points / mm² This is the correlation coefficient for the thermal conductivity of the material.
4. The method for high-precision analysis of three-dimensional morphology of solder joints on printed circuit boards according to claim 1, characterized in that: The dynamic stress simulation in step S4 includes: Import circuit board operating environment parameters: temperature cycling range, -40℃~125℃, vibration spectrum 10-2000Hz; The cumulative creep-fatigue damage was calculated based on the constitutive equation of the solder alloy.
5. The method for high-precision analysis of three-dimensional morphology of solder joints on printed circuit boards according to claim 1, characterized in that: The early prediction in step S5 includes: extracting microcrack initiation tendency factors through a twin model. , Peak stress, For plastic strain, For Young's modulus; when A level 1 warning is triggered when the value is greater than 0.
25.
6. The method for high-precision analysis of three-dimensional morphology of solder joints on a printed circuit board according to claim 1, characterized in that: After step S3, a quantitative analysis of solder wettability is added: the contact line profile of the solid-liquid-gas three phases is extracted. Calculate the rate of change of wetting angle over time ,like ° / s indicates a risk of poor soldering.
7. The method for high-precision analysis of three-dimensional morphology of solder joints on printed circuit boards according to claim 1, characterized in that: The training of the convolutional neural network adopts a multi-task learning architecture, and the main branch outputs defect classifications: cold solder joints, bridging, and oxidation. Auxiliary branch output reliability lifetime prediction, in units of thermal cycles.
8. The method for high-precision analysis of three-dimensional morphology of solder joints on printed circuit boards according to claim 1, characterized in that: Before step S1, environmental compensation calibration is performed: real-time monitoring of ambient temperature, humidity and electromagnetic interference intensity; and a noise cancellation model is generated through a generative adversarial network.
9. The method for high-precision analysis of three-dimensional morphology of solder joints on printed circuit boards according to claim 1, characterized in that: The method supports cross-production line data collaboration: the test results of different production lines are uploaded to the blockchain evidence storage platform, and the global defect identification model is updated based on federated learning.
10. The method for high-precision analysis of three-dimensional morphology of solder joints on a printed circuit board according to claim 1, characterized in that: The final 3D inspection report includes: Holographic projection coordinates of solder joint defect locations; interactive dynamic demonstration model of failure process; list of process improvement suggestions based on ISO 9454 standard.
