SMT patch optimization control method and system based on multi-source data

Abstract

The invention relates to the technical field of surface mounting control, and discloses an SMT surface mounting optimization control method and system based on multi-source data, and the method comprises the steps: collecting current surface mounting information, comparing the current surface mounting information with optimization judgment parameters, and judging whether to carry out surface mounting optimization control or not. When it is judged that optimization is carried out, historical SMT patch image data are collected, and defect feature vectors are obtained; and constructing an electrical attribute vector and a three-dimensional space distribution model of the soldering paste, executing thermal simulation calculation, and obtaining a thermal field distribution diagram of the component mounting area. And taking the defect feature vector, the electrical attribute vector and the thermal field distribution diagram as multi-source input, inputting the multi-source input into a multi-head attention neural network model, and outputting a patch optimization control parameter group. And driving the surface mounting equipment to execute surface mounting operation according to the surface mounting information and the surface mounting optimization control parameter group. According to the invention, the mounting reliability, customization capability and self-adaptability are improved, and precise and high-quality mounting control of complex special-shaped devices and high-temperature-sensitive elements can be realized.

Description

Technical Field

[0001] This invention relates to the field of surface mount control technology, and more specifically, to a method and system for SMT placement optimization control based on multi-source data. Background Technology

[0002] Surface Mount Technology (SMT) is widely used in the assembly of high-density electronic circuit boards. The accuracy and compatibility of its mounting process directly affect the subsequent soldering quality, electrical performance, and product consistency. With the widespread use of irregularly shaped devices, high-thermal-sensitive devices, and multi-package components, traditionally experience-based mounting parameters are no longer sufficient to meet the requirements of modern high-precision and high-reliability manufacturing.

[0003] Currently, some existing technologies incorporate image recognition feedback mechanisms to achieve optimized control of the surface mount technology (SMT) process. For example, patent document CN117202532A proposes an optimized control method and system for SMT processes. This method acquires mounting images, identifies typical defects such as mounting offset and bridging, and combines this with online detection data such as abnormal power rail voltage and missing solder in vias to extract mounting quality feature vectors. It then uses a neural network to output optimization suggestions. While these solutions have made progress in mounting defect identification and feature vector-based control strategy generation, they still have some shortcomings. For instance, they lack correlation modeling between electrical properties and mounting behavior. These solutions primarily analyze from the perspectives of appearance defects and structural parameters, without considering the impact of component electrical sensitivity (such as thermal sensitivity and capacitance level) on the mounting control strategy, making it difficult to achieve customized optimized control for different devices. Furthermore, they lack three-dimensional modeling of solder paste distribution and thermal simulation analysis, making it impossible to assess the impact of solder paste distribution on local thermal stress and mounting stability. They also lack the ability to perceive and control the potential failure mechanisms of thermally sensitive components in uneven thermal environments. Although image defects and process features were introduced, the data sources were relatively singular and did not integrate multi-dimensional information such as patch coordinates, component thermal parameters, and historical process logs. The input dimensions of the neural network model were limited, making it impossible to make a comprehensive judgment on multiple causal factors.

[0004] Therefore, it is necessary to design an SMT placement optimization control method and system based on multi-source data to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes an SMT placement optimization control method and system based on multi-source data, aiming to solve the current problem of lacking a complete optimization control scheme that unifies the modeling of image defects, component electrical properties and pad thermal field information, and combines deep learning mechanisms to output personalized placement parameters.

[0006] In one aspect, this invention proposes an SMT placement optimization control method based on multi-source data, comprising: Collect the current patch information and compare it with the optimization judgment parameters. Based on the comparison result, determine whether to perform patch optimization control. When it is determined that surface mount optimization control is to be performed, historical SMT surface mount image data is collected and preprocessed, and defect identification is performed on the image data to obtain defect feature vectors. The defect feature vectors include defect probability, defect offset and mounting posture deviation. Collect electrical attribute data of the components to be mounted, and construct an electrical attribute vector. The electrical attribute vector includes thermal sensitivity parameters, electrical tolerance level, package type information, and rated power value. Based on the target PCB pad diagram, component placement coordinate information, and historical solder paste distribution log data, a three-dimensional spatial distribution model of solder paste is constructed, and thermal simulation calculations are performed to obtain the thermal field distribution map of the component placement area. The defect feature vector, electrical attribute vector, and thermal field distribution map are used as multi-source inputs and fed into a multi-head attention neural network model, which outputs a set of patch optimization control parameters, including patch speed, bonding pressure, mounting angle, and path compensation. The placement equipment is driven to perform placement operations based on the placement information and the placement optimization control parameter set.

[0007] Furthermore, when comparing the patch information with the optimization judgment parameters and determining whether to perform patch optimization control based on the comparison result, the process includes: The current placement information includes at least one placement batch's target component model, package type, PCB process reference number, equipment placement speed, placement pressure, and historical image detection defect rate; The optimized judgment parameters include thermal sensitivity threshold, defect rate threshold, component packaging complexity level, and mounting accuracy requirement level. The corresponding fields in the current patch information are compared with the optimization judgment parameters at the field level, including: determining whether the target component is of a high thermal sensitivity level; determining whether the frequency of defects of the target component in historical images exceeds a set threshold; determining whether the package type of the target component is fine pitch or irregular structure; and determining whether the mounting position is at the edge of the PCB or a heat accumulation area. If any comparison result is determined to meet the optimization conditions, then the placement optimization control is performed; otherwise, the optimization step is skipped, and the default placement parameters are called for placement control.

[0008] Furthermore, when collecting and preprocessing historical SMT patch image data, and performing defect identification on the image data to obtain defect feature vectors, the process includes: The image data is subjected to image enhancement processing, which includes brightness normalization, noise filtering, edge enhancement, and geometric correction. Based on convolutional neural networks, the image is segmented and classified into multiple categories to identify defect types including mounting offset, warping, bridging, cold solder joints and missing parts, and output the corresponding defect labels and location box coordinates. The identification results are mapped to the corresponding mounting device reference numbers to form component-level defect structured data.

[0009] Furthermore, when collecting and preprocessing historical SMT patch image data, and performing defect identification on the image data to obtain defect feature vectors, the process also includes: For each component to be evaluated, the frequency of defect occurrence is statistically analyzed based on the component-level defect structured data, and the defect probability is calculated as the first feature dimension. Based on the offset distance of the defect location box in the image relative to the theoretical mounting center of the device, the offset value and rotation angle offset are calculated to form the defect offset as the second feature dimension. The degree of contour tilt of the device in the identified image is quantified as a posture deviation factor as the third feature dimension. The first feature dimension, the second feature dimension, and the third feature dimension are integrated to form the defect feature vector.

[0010] Furthermore, when constructing a three-dimensional spatial distribution model of solder paste based on the target PCB pad diagram, component placement coordinate information, and historical solder paste distribution log data, the following steps are included: The target PCB pad diagram is analyzed to extract the pad position, size and arrangement structure corresponding to each mounted component, and a pad geometric model is established in a spatial coordinate system. Alignment and registration are performed based on the coordinate information of the surface mount components and the positions of the pads on the pad diagram to generate a mounting alignment reference area; Read historical SMT solder paste distribution log data, including solder paste injection path, injection thickness, injection speed, and applied pressure; The log data is reconstructed into a point cloud in a three-dimensional coordinate space to construct a volume distribution model after solder paste injection, which is then superimposed on the corresponding pad structure to form a component-level three-dimensional volume distribution model of solder paste.

[0011] Furthermore, when performing thermal simulation calculations to obtain the thermal field distribution map of the component mounting area, the following is included: On the three-dimensional distribution volume model of the solder paste, thermophysical simulation boundary conditions are constructed based on the component thermal parameters and reflow soldering process parameters; the component thermal parameters include thermal conductivity, specific heat capacity, and packaging material; the reflow soldering process parameters include temperature profile, heating stage, and duration; Transient heat conduction simulation of solder paste-pad-component is performed based on the finite difference method to obtain the temperature field evolution results at each time point; The thermal field distribution map is extracted from the local area where the target component is located. The thermal field distribution map is used to generate a two-dimensional thermal map through interpolation and mapped according to the component reference number.

[0012] Furthermore, when the defect feature vector, electrical attribute vector, and thermal field distribution map are used as multi-source inputs to a multi-head attention neural network model, and the patch optimization control parameter set is output, it includes: The defect feature vector and electrical attribute vector are respectively subjected to numerical normalization. Based on an independent embedding network, the defect feature vector and the electrical attribute vector are linearly mapped to generate a first embedding feature and a second embedding feature, respectively. Two-dimensional convolutional encoding is performed on the thermal field distribution map to extract spatial thermal feature maps; The first embedded feature, the second embedded feature, and the spatial heat map are concatenated under a unified feature dimension to form a multi-source fusion input tensor, which is then input to the input layer of the multi-head attention neural network model.

[0013] Furthermore, the multi-head attention neural network model includes: It includes at least three attention heads, each of which independently learns an attention pattern for a specific feature category, wherein one attention head is used to focus on defect offset anomalies, another attention head is used to focus on thermal anomaly regions, and yet another attention head is used to identify highly sensitive characteristics of electrical properties. Each attention head employs a scaled dot-product attention mechanism, performing a weighted product between the key, query, and value vectors, and outputting the feature weighting result. The outputs of each attention head are concatenated and then fed into a fully connected layer and a ReLU activation layer via a residual connection and normalization module to form an intermediate feature representation generated by the optimized control strategy.

[0014] Furthermore, when optimizing the output patch control parameter set, it includes: The intermediate feature representation is input into the regression prediction module, which contains several parallel fully connected output layers. Each output layer is used to predict the patch speed, bonding pressure, mounting angle and path compensation amount. The output of each control parameter is a continuous value in the corresponding physical unit, and the output is constrained by the Sigmoid or Tanh function; The control parameter set is compared with the process constraint parameters of the current mounting task. If any parameter exceeds the limit, the parameter re-constraint module is triggered to perform range rollback and automatic correction.

[0015] Compared with existing technologies, the advantages of this invention are as follows: By introducing a multi-source data fusion mechanism, it overcomes the limitations of existing SMT placement control methods that rely primarily on single image information. Before placement control, a comparison and judgment logic between the current placement information and optimized judgment parameters is introduced, avoiding ineffective optimization and improving resource utilization efficiency. By performing image enhancement and defect identification on historical placement images, defect probability, offset, and posture deviation are quantified to form a structured defect feature vector, which helps to deeply explore quality fluctuation patterns during the placement process. The electrical attributes of components are fused to construct an electrical attribute vector, and combined with pad diagrams, coordinate information, and solder paste injection logs, a three-dimensional solder paste distribution model is established and thermal simulation is performed to generate a thermal field distribution map that realistically reflects thermal sensitivity risks, thereby achieving collaborative modeling of multiple physical factors including heat, force, and electricity. Through a multi-head attention neural network, the three types of information—defects, thermal field, and electrical attributes—are efficiently focused and weighted, outputting precise optimized control parameters including placement path, angle, and pressure, improving placement reliability, customization capabilities, and adaptability, which is beneficial for achieving precise and high-quality placement control of complex irregularly shaped devices and highly thermally sensitive components.

[0016] On the other hand, this application also provides an SMT placement optimization control system based on multi-source data, used to apply the above-mentioned SMT placement optimization control method based on multi-source data, including: The judgment module is configured to collect the current patch information, compare the patch information with the optimization judgment parameters, and determine whether to perform patch optimization control based on the comparison result; The identification module is configured to collect and preprocess historical SMT placement image data when it is determined that placement optimization control is to be performed, perform defect identification on the image data, and obtain a defect feature vector, wherein the defect feature vector includes defect probability, defect offset and placement posture deviation. The processing module is configured to collect electrical attribute data of the components to be mounted and construct an electrical attribute vector, which includes thermal sensitivity parameters, electrical tolerance levels, package type information and rated power values. The simulation module is configured to construct a three-dimensional spatial distribution model of solder paste based on the target PCB pad diagram, component placement coordinate information, and historical solder paste distribution log data, and to perform thermal simulation calculations to obtain the thermal field distribution map of the component placement area. The optimization module is configured to take the defect feature vector, electrical attribute vector and thermal field distribution map as multi-source inputs, input them into a multi-head attention neural network model, and output a set of patch optimization control parameters, which includes patch speed, bonding pressure, mounting angle and path compensation amount. The execution module is configured to drive the placement device to perform placement operations based on the placement information and the placement optimization control parameter group.

[0017] It is understandable that the above-mentioned SMT placement optimization control method and system based on multi-source data have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of an SMT chip placement optimization control method based on multi-source data provided in an embodiment of the present invention; Figure 2 This is a functional block diagram of an SMT placement optimization control system based on multi-source data, provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] In traditional surface mount technology (SMT) systems, mounting parameter settings primarily rely on empirical rules and single-dimensional defect feedback mechanisms, lacking a collaborative optimization model that integrates electrical properties, thermal field distribution, and multi-source data. The mounting process lacks dynamic sensing capabilities regarding component thermal sensitivity, package complexity, and solder paste 3D distribution, resulting in parameter adjustments failing to meet the customized needs of irregularly shaped devices and highly thermally sensitive components. Furthermore, the insufficient correlation between historical defect data and real-time process parameters during mounting control strategy generation makes it difficult to accurately predict potential failure risks in areas of thermal stress concentration, impacting mounting accuracy and soldering yield.

[0021] For example, in a high-density PCB assembly line, a batch contained a high-sensitivity thermistor in a fine-pitch QFN package and irregularly shaped connectors. Traditional methods adjust the mounting path based solely on image recognition offsets, without integrating component thermal conductivity and solder paste distribution data. During reflow soldering, temperature gradients arise in localized solder paste accumulation areas due to differences in thermal capacity, leading to cold solder joints on the sensor pins. Irregularly shaped connectors experience package warping due to uneven mounting pressure distribution. Production line monitoring data shows that the mounting defect rate of similar components fluctuates beyond the process tolerance, and the rework rate rises to the production line bottleneck threshold.

[0022] If the above issues are not addressed, the adaptability of the mounting process will be limited by a single data source and static parameter settings, making it impossible to achieve optimal control of complex components. Potential failures in areas of concentrated thermal stress will directly lead to an increased early failure rate of products, affecting the long-term reliability of electronic components. In multi-package mixed mounting scenarios, the lack of an optimization model that integrates multi-dimensional data will exacerbate parameter conflicts, reduce overall production line efficiency, and increase manufacturing costs and the difficulty of quality control.

[0023] For this, please refer to Figure 1 As shown, this application proposes an SMT placement optimization control method based on multi-source data, including: S100: Collect the current patch information and compare it with the optimization judgment parameters. Based on the comparison result, determine whether to perform patch optimization control.

[0024] S200: When it is determined that SMT placement optimization control is to be performed, historical SMT placement image data is collected and preprocessed, defects are identified in the image data, and defect feature vectors are obtained. The defect feature vectors include defect probability, defect offset and placement posture deviation.

[0025] S300: Collects electrical attribute data of the components to be mounted and constructs an electrical attribute vector, which includes thermal sensitivity parameters, electrical tolerance level, package type information and rated power value.

[0026] S400: Based on the target PCB pad diagram, component placement coordinate information, and historical solder paste distribution log data, construct a three-dimensional spatial distribution model of solder paste and perform thermal simulation calculations to obtain the thermal field distribution map of the component placement area.

[0027] S500: The defect feature vector, electrical attribute vector and thermal field distribution map are used as multi-source inputs to the multi-head attention neural network model, and the output is a set of patch optimization control parameters, which includes patch speed, bonding pressure, mounting angle and path compensation.

[0028] S600: Drives the placement equipment to perform placement operations based on the placement information and the placement optimization control parameter group.

[0029] Specifically, the component placement information refers to the component model, package type, process reference number, equipment operating parameters, and historical defect rate data involved in the current placement task. This can be achieved through real-time sensor acquisition or by retrieving placement batch records from the production management system, and is used to trigger conditional judgments for optimized control. The defect feature vector refers to a quantitative indicator of component quality anomalies extracted through image processing and classification algorithms. This can be achieved by using convolutional neural networks to segment and identify regions in historical images, and statistically analyzing defect probabilities, offsets, and orientation deviations, reflecting potential risk factors in the component placement process. The electrical attribute vector refers to a set of parameters characterizing the electrical performance and thermal sensitivity of components. This can be achieved by extracting thermal sensitivity parameters, tolerance levels, package types, and rated power data from the component datasheet, guiding differentiated adjustments to placement parameters to avoid damage to electrical performance. The solder paste three-dimensional spatial distribution model refers to a volume distribution model constructed based on pad structure, component coordinates, and historical solder paste injection data. This can be achieved by using point cloud reconstruction technology to map solder paste injection parameters to a three-dimensional coordinate system and superimpose it onto the pad geometry model, used to predict the heat conduction behavior of solder paste during reflow soldering. The thermal field distribution map refers to the visualized temperature field obtained through thermophysical simulation calculations. Specifically, it can be implemented using the finite difference method combined with component thermal parameters and reflow soldering process parameters to perform transient heat conduction simulation, used to evaluate the impact of local thermal stress on mounting stability. The multi-head attention neural network model refers to a deep learning architecture capable of parallel processing of multi-source heterogeneous data. Specifically, it can be implemented by using independently embedded networks to linearly map different features, and then using multiple attention heads to focus on the pattern association of defects, thermal fields, and electrical characteristics, used to generate optimized control parameters that take into account the coupled effects of multiple factors. The chip placement optimization control parameter set refers to the set of process parameters that drive the dynamic adjustment of the chip placement equipment. Specifically, it can be implemented by outputting continuous values ​​of placement speed, pressure, angle, and path compensation amount through a regression prediction module, and then correcting the parameters in conjunction with process constraints, used to improve placement accuracy and reliability.

[0030] This application integrates multi-source data on defect characteristics, electrical properties, and thermal field distribution, and uses a multi-head attention neural network model to perform multi-factor coupling analysis to generate optimized control parameters that take into account mounting quality, thermal stability, and electrical performance. This solves the problem of traditional methods having a single data source and lacking multi-dimensional comprehensive judgment.

[0031] The working process and principle of this application are as follows: Current patch information is collected and compared with optimization judgment parameters. Based on the comparison results, it is determined whether patch optimization control should be performed. By comparing the current patch status with preset optimization conditions, it is decided whether optimization control is needed, thus avoiding unnecessary optimization operations.

[0032] When it is determined that surface mount optimization control is needed, historical SMT surface mount image data is collected and preprocessed. Defects are identified in the image data to obtain defect feature vectors. The defect feature vectors include defect probability, defect offset, and mounting posture deviation. These data reflect the problems and trends in the historical surface mount process.

[0033] Collect electrical property data of the components to be mounted and construct an electrical property vector. The electrical property vector includes thermal sensitivity parameters, electrical tolerance class, package type information, and rated power value. These data reflect the electrical characteristics of the components and their sensitivity to mounting conditions.

[0034] Then, based on the target PCB pad diagram, component placement coordinates, and historical solder paste distribution log data, a three-dimensional spatial distribution model of the solder paste is constructed, and thermal simulation calculations are performed to obtain the thermal field distribution map of the component placement area. By simulating the solder paste distribution and heat conduction process, potential thermal stress problems during the placement process are predicted.

[0035] Defect feature vectors, electrical property vectors, and thermal field distribution maps are used as multi-source inputs and fed into a multi-head attention neural network model. This model can comprehensively consider multiple factors and output a set of optimized placement control parameters. The optimized placement control parameters include placement speed, bonding pressure, placement angle, and path compensation, which directly affect placement quality and accuracy.

[0036] Finally, the placement equipment is driven to perform the placement operation based on the placement information and the placement optimization control parameter set. In this way, placement parameters can be dynamically adjusted based on multi-source data, achieving precise control and optimization.

[0037] As a preferred embodiment, the solution of this application is specifically implemented as follows: On the SMT production line, the current placement information is first collected, including placement batch, target component model, package type, PCB process reference number, equipment placement speed, placement pressure, and historical image inspection defect rate. This information is then compared with preset optimized judgment parameters, including thermal sensitivity threshold, defect rate threshold, component package complexity level, and placement accuracy requirement level.

[0038] If it is determined that surface mount optimization control is required, the image processing module is invoked to preprocess historical SMT surface mount image data. Preprocessing steps include image enhancement, noise filtering, and edge enhancement. Then, a convolutional neural network is used to identify defects in the image, recognizing defect types such as mount offset, warping, bridging, cold solder joints, and missing components. Based on the recognition results, the defect probability, defect offset, and mount posture deviation are calculated to form a defect feature vector.

[0039] Extract electrical attribute data of the components to be mounted from the component database, including thermal sensitivity parameters, electrical tolerance levels, package type information and rated power values, and construct an electrical attribute vector.

[0040] Then, the 3D modeling module is invoked to construct a 3D spatial distribution model of the solder paste based on the target PCB pad diagram, component placement coordinates, and historical solder paste distribution log data. On this basis, thermal simulation calculations are performed to simulate the temperature field evolution during the reflow soldering process and obtain the thermal field distribution map of the component placement area.

[0041] The defect feature vector, electrical property vector, and thermal field distribution map are input into a pre-trained multi-head attention neural network model. This model contains multiple attention heads, each focusing on the highly sensitive characteristics of defect offset, thermal anomaly regions, and electrical properties. By weighted fusion of multi-source information, the model outputs a set of optimized patch control parameters, including patch speed, bonding pressure, mounting angle, and path compensation.

[0042] Finally, the optimized control parameter set is transmitted to the placement equipment control system, driving the placement equipment to perform the placement operation according to the optimized parameters. During the placement process, the placement quality is monitored in real time, and the data is fed back to the optimization control system, forming a closed-loop optimization mechanism.

[0043] Through the above-described scheme, this application enables SMT placement optimization control based on multi-source data. By integrating historical defect data, component electrical properties, and thermal distribution information, placement risks can be assessed more comprehensively, and placement parameters can be adjusted accordingly. This improves placement accuracy and consistency, especially for high-thermal-sensitive components and irregularly shaped packages. Dynamic optimization of placement parameters reduces solder joint defects and warpage caused by uneven thermal stress, improving product reliability. Simultaneously, the multi-head attention mechanism adaptively focuses on key features from different data sources, enhancing adaptability to complex placement scenarios. The multi-source data-based optimization method improves placement yield and reduces rework rates and production costs.

[0044] In some of the solutions mentioned above in this application, the process of determining whether to perform surface mount optimization control lacks multi-dimensional correlation analysis of component characteristics and process conditions, resulting in a single optimization trigger mechanism that cannot accurately identify the scenarios that need optimization, which may lead to unnecessary optimization calculations or omission of key optimization requirements.

[0045] This application further proposes to collect the target component model, package type, PCB process reference number, equipment placement speed, placement pressure, and historical image detection defect rate from the current placement information. The above information is compared at the field level with the thermal sensitivity threshold, defect rate threshold, component package complexity level, and placement accuracy requirement level in the optimization judgment parameters. The optimization control is triggered when any condition is met, such as whether the target component is of a high thermal sensitivity level, whether the historical defect frequency exceeds the threshold, whether the package type is fine pitch or irregular structure, or whether the placement position is at the edge of the PCB or a heat accumulation area.

[0046] The current placement information includes equipment operating parameters and component attribute data, while the optimization judgment parameters include process constraints and device characteristic thresholds. Field-level comparisons employ multi-dimensional logical judgments, with each comparison condition corresponding to different process risk factors. For example, thermal sensitivity thresholds are associated with component models, defect rate thresholds with historical inspection data, and package complexity levels with device physical structures. The judgment logic uses a parallel conditional detection mechanism; the optimization process is triggered when any condition is met.

[0047] Specifically, when a placement task starts, the system automatically extracts the model attributes and equipment operating parameters of the current component, and retrieves the component's historical defect statistics. By comparing the component's thermal sensitivity level with a preset threshold, highly thermally sensitive components can be identified as requiring optimized thermal field adaptation. By comparing historical defect rates with thresholds, high-failure-risk components can be identified as requiring adjustments to placement parameters. By analyzing whether the package type is fine-pitch or irregularly shaped, it is determined whether placement accuracy needs to be improved. By detecting whether the placement position is at the PCB edge or in a high-temperature area, the thermal stress risk is assessed. When any of the above conditions are met, the optimization control process is automatically initiated; otherwise, placement is performed using default parameters. Precise optimization triggering is achieved through multi-dimensional condition detection, avoiding misjudgments caused by single judgment conditions, and reducing unnecessary computational resource consumption. For example, when processing ordinary resistor components, if their historical defect rate is below the threshold and they are not thermally sensitive devices, the optimization step will be skipped, and the standard placement process will be executed directly.

[0048] As a preferred embodiment, the solution of this application is specifically implemented as follows: The current placement information includes the target component model, package type, PCB process reference number, equipment placement speed, placement pressure, and historical image inspection defect rate for at least one placement batch.

[0049] The optimized judgment parameters include thermal sensitivity threshold, defect rate threshold, component package complexity level, and mounting accuracy requirement level.

[0050] The corresponding fields in the current component placement information are compared with the optimized judgment parameters at the field level, including: determining whether the target component is of a high thermal sensitivity level; determining whether the frequency of defects appearing on the target component in historical images exceeds a set threshold; determining whether the package type of the target component is fine-pitch or irregularly shaped; and determining whether the placement position is at the edge of the PCB or in a heat-concentrating area.

[0051] If any comparison result determines that the optimization conditions are met, then the placement optimization control is performed. Otherwise, the optimization step is skipped, and the default placement parameters are used for placement control.

[0052] Specifically, the placement information may include the target component model being an MLCC capacitor, the package type being 0402, the PCB process reference number being R1, the equipment placement speed being 50mm / s, the placement pressure being 0.5N, and the historical image detection defect rate being 2%. Optimization judgment parameters may include a thermal sensitivity threshold of 80℃, a defect rate threshold of 1%, a component package complexity level of 3, and a placement accuracy requirement level of 2.

[0053] When performing field-level comparisons, first determine if the thermistor rating of the MLCC capacitor exceeds 80℃. Second, check if the defect rate of this capacitor model in historical images exceeds 1%. Then, confirm whether the 0402 package is a fine-pitch or irregularly shaped structure. Finally, determine whether it is located at the edge of the PCB or in a heat-concentrated area based on the PCB process reference R1.

[0054] If any of the above conditions are met, the placement optimization control process is triggered. Otherwise, the default placement parameters are used directly for placement.

[0055] Through the above technical solution, this application achieves surface mount optimization control judgment based on multi-dimensional information. This allows for flexible decision-making regarding whether optimization control is needed for components of different types and locations, avoiding redundant processing of components that do not require optimization, and improving the efficiency and accuracy of the surface mount process. Simultaneously, by comprehensively considering multiple factors such as thermal sensitivity, historical defect rate, and package complexity, the targeting and effectiveness of optimization control are enhanced, better adapting to the mounting requirements of different components and PCB layouts.

[0056] In some of the solutions described above in this application, a method is proposed to collect and preprocess historical SMT patch image data and perform defect identification on the image data to obtain defect feature vectors. However, in this process, the original image data may suffer from uneven lighting, noise interference, or geometric deformation, resulting in insufficient defect identification accuracy and thus affecting the accuracy of the defect feature vectors.

[0057] This application further proposes a method for acquiring and preprocessing historical SMT placement image data. When identifying defects in the image data and obtaining defect feature vectors, the method includes: image enhancement processing, which includes brightness normalization, noise filtering, edge enhancement, and geometric correction. Based on a convolutional neural network, the image is segmented into regions and classified into multiple categories to identify defect types including placement offset, warping, bridging, cold solder joints, and missing components, and outputs corresponding defect labels and location box coordinates. The identification results are mapped to the corresponding component reference numbers to form component-level defect structured data.

[0058] In the image enhancement process, brightness normalization adjusts the overall brightness distribution of the image through histogram equalization, eliminating differences in brightness caused by light source fluctuations. Noise filtering uses a median filter to remove salt-and-pepper noise while preserving image edge details. Edge enhancement strengthens the gradient information of defect contours using the Sobel operator. Geometric correction compensates for image rotation and scaling based on affine transformation, correcting deformation caused by shooting angle deviations. For region segmentation and multi-class classification based on convolutional neural networks, a pre-trained ResNet architecture is used as the backbone network. Transfer learning is used for fine-tuning on an SMT patch defect dataset. The output layer uses a Softmax function to generate a defect type probability distribution, and a non-maximum suppression algorithm is used to filter defect localization boxes with confidence levels higher than a set threshold. When mapping the recognition results to component reference numbers, the component reference number text is identified by parsing the QR code in the image or by OCR recognition, establishing a correlation between defect labels and the physical location of the component.

[0059] Specifically, brightness normalization adjusts images acquired in different batches to a uniform brightness range, preventing the same defect from exhibiting different appearances in different images due to variations in lighting conditions, thus ensuring the stability of subsequent processing. Noise filtering eliminates random noise points in the image, reducing the probability of false detections, especially when detecting minute defects, avoiding noise interference. Edge enhancement highlights the contour features of the defect area, enhancing the sensitivity of the convolutional neural network to defect edges and improving classification accuracy. Geometric correction performs spatial transformation on the image using an affine transformation matrix, eliminating image distortion caused by camera tilt or placement equipment vibration, making the defect location box coordinates more accurately reflect the actual placement position. Region segmentation and multi-class classification based on convolutional neural networks utilize the feature extraction capabilities of deep learning to automatically identify various complex defect types, exhibiting higher generalization ability compared to traditional threshold segmentation methods. The mapping relationship between defect location box coordinates and component reference numbers is established, enabling defect data to be accurately associated with specific components, providing component-level data support for subsequent optimization control. For example, when a placement offset is detected for a component with a certain reference number, the placement path compensation can be adjusted for the PCB location of that component, preventing the recurrence of the same type of defect.

[0060] As a preferred embodiment, the solution of this application is specifically implemented as follows: When collecting and preprocessing historical SMT patch image data, and performing defect identification on the image data to obtain defect feature vectors, the following steps are included: Image enhancement processing is performed on the image data. This includes brightness normalization, noise filtering, edge enhancement, and geometric correction. Specifically, histogram equalization is used for brightness normalization, a Gaussian filter is used to remove image noise, the Sobel operator is applied for edge enhancement, and affine transformation is used to correct image geometric distortion.

[0061] Image region segmentation and multi-class classification are performed based on convolutional neural networks. The U-Net network architecture is used to achieve pixel-level semantic segmentation, dividing the image into regions such as background, PCB board, components, and solder pads. Then, ResNet50 is used as the backbone network to construct a multi-class classifier to identify defect types including mount offset, warping, bridging, cold solder joints, and missing components. The corresponding defect labels and bounding box coordinates are output.

[0062] Finally, the identification results are mapped to the corresponding component reference numbers to form component-level defect structured data. By using OCR technology to identify PCB silkscreen markings and combining them with mounting coordinate information, a correspondence between defects and components is established, generating a structured data table containing information such as component number, defect type, and location coordinates.

[0063] The above technical solution enables automated defect identification and structured processing of SMT patch images. Image enhancement improves the accuracy of subsequent analysis, while the application of convolutional neural networks makes defect identification more precise and efficient. The generation of component-level defect data provides a reliable data foundation for subsequent optimization and control. This overcomes the subjectivity and inefficiency of traditional manual inspection, improving the accuracy and consistency of defect identification.

[0064] In some of the solutions described above in this application, although the defect types can be identified and structured data can be formed after collecting and preprocessing historical SMT placement image data, the defect data lacks multi-dimensional quantitative indicators, making it impossible to accurately assess the dynamic impact of defects on placement quality and difficult to generate optimized control parameters that adapt to different defect modes.

[0065] This application further proposes, for each component to be evaluated, to calculate the defect probability as the first feature dimension based on the statistical frequency of defect occurrence using structured data of component-level defects. Based on the offset distance of the defect location box in the image relative to the theoretical mounting center of the device, the offset value and rotation angle offset are calculated to form the defect offset as the second feature dimension. Based on the degree of contour tilt of the device in the identified image, it is quantified as a posture deviation factor as the third feature dimension. The first, second, and third feature dimensions are integrated to form a defect feature vector.

[0066] The defect probability is obtained by statistically analyzing the ratio of the number of defects occurring in historical batches of the same component to the total number of placements. The defect offset is determined by calculating the Euclidean distance and angle difference between the center point of the defect positioning frame and the theoretical placement center point. The attitude deviation factor is obtained by detecting the tilt angle of the device contour edge through Hough transform and normalizing it to a value between 0 and 1. The first feature dimension reflects the defect risk level of the component, the second feature dimension characterizes the direction and magnitude of the placement position deviation, and the third feature dimension describes the degree of attitude abnormality during device placement.

[0067] Specifically, the defect probability is calculated using a sliding window method, taking the most recent 50 placement batches as samples to calculate the dynamic defect rate. The defect offset calculation incorporates pad coordinate system transformation, mapping image pixel coordinates to actual physical coordinates before offset conversion. The posture deviation factor extracts the device contour through edge detection, fits the minimum bounding rectangle, calculates the angle between its major axis and the standard placement direction, and divides the angle by 90 degrees to obtain the normalized deviation value. These three feature dimensions are linearly weighted and combined into a three-dimensional vector, with weights dynamically adjusted according to the component type; for example, thermal components are given higher posture deviation weights. When the integrated defect feature vector is input into the neural network model, it can transform image defect information into a quantitative indicator directly related to placement control parameters. The defect probability guides the adjustment of placement speed, the defect offset drives the calculation of path compensation, and the posture deviation factor affects placement angle correction, thereby improving the adaptability of optimized control parameters and defect patterns.

[0068] As a preferred embodiment, the solution of this application is specifically implemented as follows: For each component to be evaluated, the frequency of defects is statistically analyzed based on the component-level defect structured data, and the defect probability is calculated as the first feature dimension. Specifically, the defect probability of a component can be obtained by dividing the number of times a specific component has a defect in historical placement data by the total number of placements. For example, if a certain type of capacitor has a defect 50 times in 1000 placements, its defect probability is 5%.

[0069] Based on the offset distance of the defect location box in the image relative to the theoretical mounting center of the component, the offset value and rotation angle offset are calculated to form the defect offset as the second feature dimension. Further, the centroid coordinates of the defect region can be extracted using image processing algorithms, and the Euclidean distance between these coordinates and the theoretical component center coordinates can be calculated to obtain the offset value. The rotation angle offset is obtained by comparing the angle between the principal axis direction of the defect region and the theoretical direction.

[0070] The tilt degree of the device's contour in the identified image is quantified as a pose deviation factor as the third feature dimension. Specifically, the component contour can be extracted using an edge detection algorithm, a minimum bounding rectangle can be fitted, and the angle between the rectangle's major axis and the horizontal line can be calculated as the tilt angle. This angle is then divided by a preset maximum allowable tilt angle to obtain the normalized pose deviation factor.

[0071] The first, second, and third feature dimensions are integrated to form a defect feature vector. This results in a multi-dimensional feature vector containing defect probability, offset, and attitude deviation, which is used for subsequent patch optimization control.

[0072] Through the above technical solution, this application achieves comprehensive quantification and characterization of defect features in the SMT placement process. By integrating defect probability, position offset, and posture deviation into a unified feature vector, multi-dimensional input information is provided for subsequent placement optimization control. It can capture different types of placement defects and transform them into quantifiable numerical features, which helps improve the accuracy and adaptability of placement control. In particular, by introducing a posture deviation factor, it can better identify and handle defects such as component tilt, which are difficult to describe by simple position offset, thereby improving the comprehensiveness of placement quality control.

[0073] In some of the solutions described above in this application, when constructing the three-dimensional spatial distribution model of solder paste, there are problems such as insufficient alignment accuracy between the pad structure and the mounting coordinates, and the failure to effectively restore the three-dimensional shape of historical solder paste distribution data. As a result, the solder paste volume distribution model cannot accurately reflect the actual process state, affecting the reliability of subsequent thermal simulation analysis.

[0074] This application further proposes a method for constructing a three-dimensional spatial distribution model of solder paste based on the target PCB pad diagram, component placement coordinate information, and historical solder paste distribution log data.

[0075] The process involves analyzing the target PCB pad diagram, extracting the pad position, size, and arrangement structure for each mounted component, and establishing a pad geometric model in a spatial coordinate system. Alignment and registration are performed based on the component coordinates and pad positions to generate a mounting alignment reference area. Historical SMT solder paste distribution log data is read, including solder paste injection path, injection thickness, injection speed, and applied pressure. The log data is then used to reconstruct a point cloud in three-dimensional coordinate space, constructing a volume distribution model of the solder paste after injection, which is then superimposed on the corresponding pad structure to form a component-level three-dimensional solder paste distribution volume model.

[0076] Specifically, the pad geometry model reconstructs the physical shape and spatial layout of the pads in a 3D mesh by parsing the coordinate data in the pad drawing file. The mounting alignment reference area ensures precise spatial matching between the mounting position and the pad structure by transforming the component placement coordinates and the pad position. The injection path in the historical solder paste distribution log data is used to generate a continuous path point cloud through 3D trajectory interpolation. Injection thickness and velocity data are used to calculate the solder paste volume deposition amount at each path point, and applied pressure data is used to adjust the volume distribution density. During the point cloud reconstruction process, a spatial interpolation algorithm is used to convert the discrete log data into a continuous 3D volume distribution model. This model is then superimposed with the pad geometry model through Boolean operations to generate a component-level 3D solder paste distribution volume model that includes pad structure constraints. Through these steps, the solder paste distribution model can accurately reflect the influence of actual process parameters on the solder paste morphology, providing accurate physical field input data for thermal simulation.

[0077] As a preferred embodiment, the solution of this application is specifically implemented as follows: When constructing a 3D spatial distribution model of solder paste based on the target PCB pad diagram, component placement coordinates, and historical solder paste distribution log data, the following steps are included: The target PCB pad image is analyzed to extract the pad position, size, and arrangement structure for each mounted component. Specifically, image processing algorithms are used to perform edge detection and contour extraction on the PCB pad image to identify the geometric shape and position coordinates of each pad. Then, a geometric model of the pads is established in a three-dimensional coordinate system, representing each pad as a three-dimensional structure with a specific shape and height.

[0078] Alignment and registration are performed based on the coordinate information of the surface mount components and their positions on the PCB pad map. A coordinate transformation algorithm is used to align the component's surface mount coordinate system with the PCB pad coordinate system, determining the precise placement of the component on the PCB. This generates a mounting alignment reference area, providing a spatial positioning benchmark for subsequent solder paste distribution modeling.

[0079] Read historical SMT solder paste distribution log data. This log data includes parameters such as solder paste injection path, injection thickness, injection speed, and applied pressure. By parsing the log file, extract detailed information for each solder paste application operation.

[0080] Log data is reconstructed into a point cloud in three-dimensional coordinate space. Specifically, based on the solder paste injection path and thickness information, a large number of discrete points are generated in space, each representing a tiny volume unit of solder paste. A point cloud density control algorithm is used to adjust the distribution of points to accurately reflect the volume distribution of the solder paste. This constructs a volume distribution model after solder paste injection. This model is then superimposed on the corresponding pad structure to form a component-level three-dimensional volume distribution model of the solder paste, achieving accurate characterization of the solder paste distribution.

[0081] Through the above technical solution, this application can accurately construct a three-dimensional spatial distribution model of solder paste. This model comprehensively considers the influence of PCB pad structure, component placement location, and the actual solder paste application process, providing a precise geometric basis for subsequent thermal simulation analysis. It helps to more accurately assess the impact of solder paste distribution on local thermal stress and placement stability, thereby optimizing placement parameters and improving placement quality and reliability. Especially for heat-sensitive components, this model can help predict potential thermal non-uniformity distribution, thus enabling the development of more targeted placement strategies and reducing the risk of component failure.

[0082] In some of the solutions described above in this application, after constructing a three-dimensional spatial distribution model of solder paste, the thermal field distribution of the mounting area could not be effectively predicted, resulting in the inability to assess the dynamic coupling effect between solder paste distribution and component thermal parameters, and making it difficult to avoid mounting stability problems caused by local thermal stress concentration.

[0083] This application further proposes constructing thermophysical simulation boundary conditions based on the component thermal parameters and reflow soldering process parameters on a three-dimensional volumetric model of solder paste distribution. The component thermal parameters include thermal conductivity, specific heat capacity, and packaging material, while the reflow soldering process parameters include temperature profiles, heating stages, and duration. Transient heat conduction simulation of the solder paste-pad-component model is performed using the finite difference method to obtain the temperature field evolution results at each time point. A thermal field distribution map is extracted from the local region where the target component is located. This thermal field distribution map is then used to generate a two-dimensional thermal map through interpolation and mapped according to the component reference designator.

[0084] The thermophysical simulation boundary conditions establish the initial and constraint conditions of the heat conduction equation by analyzing the material properties and process parameters of the components. The finite difference method discretizes the three-dimensional structure of solder paste-pad-component into mesh elements, and iteratively calculates the temperature change of each mesh node. The transient heat conduction simulation outputs temperature field data at different time steps, reflecting the dynamic heat distribution of the solder paste during the reflow soldering process. The interpolation method converts the discrete three-dimensional temperature data into a continuous two-dimensional thermal map, which is convenient for the mounting control system to read directly.

[0085] Specifically, among the component's thermal parameters, thermal conductivity determines the rate of heat transfer within the material, specific heat capacity affects the hysteresis of temperature changes, and the packaging material determines the coefficient of thermal expansion. The temperature profile in the reflow soldering process parameters includes temperature gradients during preheating, wetting, reflow, and cooling stages, and the duration of the heating stage affects the heat accumulation effect. The finite difference method discretizes the spatial and temporal dimensions, solves partial differential equations, and calculates the rate of temperature change over time for each grid cell. Transient simulation results include the distribution of the temperature field in the spatial and temporal dimensions. The interpolation process uses bilinear or cubic spline algorithms to project the three-dimensional temperature field onto a two-dimensional plane, generating a thermal map corresponding to the component's mounting point. The thermal field distribution map can identify localized high-temperature regions, predict the impact of thermal stress on mounting accuracy, and thus adjust the placement speed and path compensation to avoid misalignment defects caused by thermal deformation.

[0086] As a preferred embodiment, the solution of this application is specifically implemented as follows: On a three-dimensional volumetric model of solder paste distribution, thermophysical simulation boundary conditions are constructed based on component thermal parameters and reflow soldering process parameters. Component thermal parameters include thermal conductivity, specific heat capacity, and package material. Reflow soldering process parameters include temperature profiles, heating stages, and duration.

[0087] Specifically, the simulation boundary conditions are first set. For component thermal parameters, common thermophysical parameters of electronic components are selected, such as the thermal conductivity of copper wires being 400 W / (m·K) and the specific heat capacity of epoxy resin encapsulation being 1000 J / (kg·K). For reflow soldering process parameters, a typical four-stage temperature profile is adopted: preheating stage (150-200℃, 60-120s), activation stage (150-200℃, 60-120s), reflow stage (peak temperature 230-250℃, 10-30s), and cooling stage (cooling to room temperature below 6℃ / s).

[0088] Furthermore, transient heat conduction simulation of solder paste-pad-component based on the finite difference method was performed to obtain the temperature field evolution results at each time point. In specific implementation, the entire PCB board was divided into mesh cells, and the heat conduction equation was applied to each cell for iterative calculation. The time step was set to 0.1s, the spatial step was 0.1mm, and the total simulation time was 300s, covering the entire reflow soldering process.

[0089] Therefore, a thermal distribution map is extracted from the local area where the target component is located. The thermal distribution map is used to generate a two-dimensional thermal image and mapped according to the component reference designator. For example, for a 10mm×10mm BGA package chip, the temperature data of 100×100 grid points at its bottom are extracted, and a thermal image with a resolution of 1000×1000 pixels is generated by bilinear interpolation, with the color from blue to red representing the change of temperature from low to high.

[0090] Through the above technical solution, this application achieves accurate simulation of the local thermal field distribution during SMT placement. By constructing a thermal simulation model considering component characteristics and process parameters, high spatiotemporal resolution temperature field evolution data was obtained. This provides a reliable data foundation for subsequent placement optimization based on thermal stress analysis, helps identify potential failure risks of heat-sensitive components, and improves placement quality and reliability. Simultaneously, the generation of the thermal field distribution map provides key input for multi-source data fusion, enhancing the optimization control model's ability to perceive thermally related factors.

[0091] In some of the solutions described above in this application, when the defect feature vector, electrical attribute vector and thermal field distribution map are input into the multi-head attention neural network model, the differences in dimensions and modes of different data sources make feature fusion difficult, affecting the model's prediction accuracy for optimizing control parameters.

[0092] This application further proposes numerical normalization processing for the defect feature vector and electrical attribute vector. A linear mapping is performed on the defect feature vector and electrical attribute vector based on an independent embedding network to generate a first embedding feature and a second embedding feature, respectively. Two-dimensional convolutional encoding is applied to the thermal field distribution map to extract a spatial thermal feature map. The first embedding feature, the second embedding feature, and the spatial thermal feature map are concatenated under a unified feature dimension to form a multi-source fusion input tensor, which is then input to the input layer of a multi-head attention neural network model.

[0093] The numerical normalization process employs Z-score standardization, mapping numerical values ​​of different dimensions, such as defect probability, defect offset, and thermal sensitivity parameters, to a distribution range with a mean of 0 and a variance of 1. The independent embedding network consists of fully connected layers, with an input dimension consistent with the length of the defect feature vector or electrical attribute vector, and an output dimension of 64, mapping vectors of different structures to a latent space of the same dimension. The two-dimensional convolutional encoding uses three convolutional layers with kernel sizes of 3×3, 5×5, and 3×3, and a stride of 2. Max pooling layers compress the feature map size, ultimately outputting a 256-dimensional spatial thermal feature map with 64 channels. The concatenation operation stacks the first and second embedded features along the feature dimension with the spatial thermal feature map, forming a 256×64-dimensional multi-source fusion input tensor.

[0094] Specifically, the defect feature vector includes defect probability, defect offset, and mounting posture deviation, with values ​​ranging from 0 to 1, and millimeter-level offset and angular deviation. The thermal sensitivity parameter in the electrical attribute vector differs by orders of magnitude from the rated power value; directly inputting it into the model would lead to gradient imbalance. Z-score normalization maps the defect probability to the [-1,1] interval, the defect offset is scaled proportionally to the maximum historical offset, and the thermal sensitivity parameter has a normalization threshold based on component type. The independent embedding network uses a linear transformation matrix to map the 3D input of the defect feature vector to a 64D vector, and the 4D input of the electrical attribute vector to a 64D vector, eliminating the dimensional differences of the original features. The 2D convolutional encoding of the thermal field distribution map extracts the spatial distribution patterns of temperature gradients and hotspot regions through local receptive fields, generating a spatial thermal feature map that matches the dimensions of the embedded features. The multi-source fusion input tensor retains complete information on defect statistical features, electrical attribute features, and thermal field spatial features in a unified dimension, providing structurally aligned input data for the multi-head attention mechanism. For example, after normalization, the standard deviation of the defect probability is controlled within 0.2, and the mean of the thermal sensitivity parameter is normalized to 0.3. The resolution of the spatial thermal feature map output by convolutional coding is reduced to 1 / 8 of the original image. As a result, the model can simultaneously capture the statistical regularity of defects, electrical sensitivity characteristics, and spatial patterns of thermal distribution, improving the prediction accuracy of control parameters such as patching speed and bonding pressure.

[0095] As a preferred embodiment, the solution of this application is specifically implemented as follows: The defect feature vector and electrical attribute vector are numerically normalized. The normalization process uses the maximum-minimum normalization method to map each feature value to the interval [0,1].

[0096] The normalized defect feature vector and electrical attribute vector are linearly mapped using independent embedding networks. Each embedding network contains a fully connected layer with an output dimension of 128. This step generates the first and second embedding features.

[0097] Two-dimensional convolutional encoding is performed on the thermal distribution map. Specifically, three convolutional layers are used, each with a 3x3 kernel size, a stride of 1, and padding of 1. The first convolutional layer outputs 32 channels, the second 64, and the third 128. Each convolutional layer is followed by a BatchNorm and ReLU activation function. The output of the final convolutional layer is the extracted spatial thermal feature map.

[0098] The first and second embedded features are concatenated with the spatial heat map under a unified feature dimension. Specifically, the spatial heat map is globally averaged to obtain a 128-dimensional vector, which is then concatenated with the two 128-dimensional embedded features to form a 384-dimensional multi-source fusion input tensor.

[0099] The multi-source fusion input tensor is fed into the input layer of a multi-head attention neural network model. This model contains three attention heads, each with a hidden dimension of 128. Attention computation employs a scaled dot-product attention mechanism. The attention output, after passing through residual connections and layer normalization, is fed into two fully connected layers with a hidden dimension of 512 and ReLU activation function.

[0100] Through the above technical solution, this application achieves effective fusion of multi-source heterogeneous data such as defect features, electrical properties, and thermal field distribution. By using independent feature embedding and convolutional coding, the feature representation capabilities of each data source are preserved. The multi-head attention mechanism adaptively focuses on the correlation between different features, improving the model's understanding of complex patching scenarios. This combination of multi-source data fusion and attention mechanism provides a more comprehensive and accurate feature representation for the subsequent generation of patching optimization control parameters, helping to improve the adaptability and accuracy of the SMT patching process.

[0101] In some of the solutions described above in this application, when multi-head attention neural network models fuse multi-source input data, due to the large differences in the dimensions of the input features and the existence of complex correlations, a single attention mechanism is difficult to effectively capture the key information between different feature categories. This results in insufficient attention to defect offset, thermal anomaly and electrical sensitivity characteristics by the model, affecting the accuracy of generating optimized control parameters.

[0102] This application further proposes a multi-head attention neural network model comprising at least three attention heads, each independently learning an attention pattern for a specific feature category. One attention head focuses on defect offset anomalies, another on thermal anomaly regions, and yet another on identifying highly sensitive electrical properties. Each attention head employs a scaled dot-product attention mechanism, performing a weighted product between the key, query, and value vectors and outputting a weighted feature result. The outputs of each attention head are concatenated and passed through a residual connection and normalization module before being input into a fully connected layer and a ReLU activation layer to form an intermediate feature representation generated by an optimized control strategy.

[0103] The three attention heads are designed with independent parameter matrices for defect offset, thermal field distribution, and electrical properties, respectively. They calculate the correlation weights between different features using key and query vectors. The key vector is generated by a linear transformation of the input features, while the query vector is dynamically generated based on the task objective of the current attention head. The defect offset attention head generates attention weights by calculating the difference between the defect offset and the theoretical position. The thermal anomaly attention head extracts high thermal stress regions based on temperature gradient distribution, and the electrical property attention head filters highly sensitive components based on thermal sensitivity parameters and tolerance levels. The weighted feature outputs from each attention head are concatenated, and the original feature information is preserved through residual connections. Layer normalization eliminates feature scale differences, and a fully connected layer performs nonlinear mapping to form an intermediate feature representation.

[0104] Specifically, the defect offset attention head generates an attention weight matrix for the defect region by calculating the cosine similarity between the defect offset features and the mounting coordinates, thus strengthening the representation of the offset anomaly region. The thermal anomaly attention head performs convolutional encoding on the temperature gradient of the thermal field distribution map, extracts the spatial features of local thermal stress concentration areas, and calculates the correlation between the thermal field and the component position through a scaling dot product mechanism. The electrical attribute attention head maps thermal sensitivity parameters and tolerance levels to query vectors, interacts with key vectors, and identifies the electrical characteristics of highly sensitive components. The outputs of each attention head are concatenated along the channel dimension, added to the input features, and normalized to eliminate scale differences between multi-source data. The fused features are then compressed into a low-dimensional intermediate representation through a fully connected layer, providing highly discriminative feature input for the subsequent regression prediction module.

[0105] As a preferred embodiment, the solution of this application is implemented as follows: Three independent attention heads are set in the multi-head attention neural network model. The first attention head is configured to receive defect offset data and establish a correlation mapping between defect offset anomalies and mounting path compensation amounts by calculating the spatial distance between the center coordinates of the defect location box and the theoretical mounting center. The second attention head receives the two-dimensional encoded features of the thermal field distribution map, strengthens spatial attention for areas where the temperature gradient exceeds a set threshold, and generates a feature weight matrix for thermal anomaly areas. The third attention head is configured to parse the thermal sensitivity parameters and tolerance levels in the electrical attribute vector, and identifies the electrical characteristic patterns of highly sensitive components through a learnable parameter matrix. Each attention head adopts a scaled dot product attention mechanism, performs matrix multiplication on the linearly transformed query vector and key vector, adjusts the weight distribution through a temperature coefficient, and performs a weighted summation of the value vectors to obtain the output features of each head. The output features of the three attention heads are concatenated along the channel dimension, residually connected to the original input features, and then input into the fully connected network after layer normalization. The fully connected network consists of two linear transformation layers, with a nonlinearity introduced in between through the ReLU activation function, ultimately generating an intermediate representation that includes features related to patch speed, pressure, angle, and path compensation.

[0106] Through the above technical solutions, this application achieves differentiated focusing and dynamic correlation of key features in multi-source heterogeneous data, effectively solving the control parameter deviation problem caused by traditional single-dimensional feature analysis. By using independent attention heads to target defect offsets, thermal anomalies, and electrical susceptibility, the model's ability to identify high-risk factors is enhanced. Residual connectivity and normalization processing ensure the stability of feature fusion and avoid interference from multi-source data noise on the control strategy. This structure enables optimized control parameters to accurately match the electrical characteristics, thermal environment, and historical defect patterns of components, improving the adaptability of the surface mount technology (SMT) process to complex operating conditions.

[0107] In some of the solutions described above in this application, after the multi-head attention neural network model generates intermediate feature representations, directly outputting control parameters may cause the parameters to exceed the physical limitations of the equipment or the constraints of the process, resulting in placement failure or equipment malfunction.

[0108] This application further proposes inputting intermediate feature representations into a regression prediction module. The regression prediction module comprises several parallel fully connected output layers, each used to predict placement speed, bonding pressure, placement angle, and path compensation. The output of each control parameter is a continuous value in its corresponding physical unit, constrained by a Sigmoid or Tanh function before output. The control parameter set is compared with the process constraint parameters of the current placement task. If any parameter exceeds the limit, a parameter re-constraint module is triggered for range rollback and automatic correction.

[0109] The parallel fully connected output layer of the regression prediction module independently processes different control parameters to avoid interference between parameters. The Sigmoid or Tanh function limits the output value to a preset range; for example, the placement speed is constrained to 30%-100% of the equipment's rated speed. The parameter re-constraint module dynamically adjusts out-of-limit parameters through the process constraint parameter table; for example, when the path compensation exceeds the maximum allowable offset of the PCB process reference number, it automatically corrects it to the upper limit value.

[0110] Specifically, intermediate features are mapped to placement speed, bonding pressure, placement angle, and path compensation amount through independent fully connected layers. Each output layer uses a linear activation function to generate the original predicted value. After compressing the original values ​​to the 0-1 range using the Sigmoid function, a linear mapping is performed based on the equipment parameters; for example, 0.5 is mapped to the median value of the equipment's default speed. If the mapped placement angle exceeds the maximum allowable tilt angle of the component package, the parameter re-constraint module truncates the parameters according to the safety threshold in the historical process log, ensuring that the parameters conform to actual production conditions. This method ensures the accuracy of parameter prediction while avoiding equipment failures or placement failures caused by parameters exceeding limits.

[0111] As a preferred embodiment, the solution of this application is implemented as follows: The regression prediction module consists of four independent fully connected layers, each connected to the intermediate feature representation output by a multi-head attention neural network. Specifically, the first fully connected layer uses a linear activation function to predict the patch placement speed; the second fully connected layer constrains the output range of the bonding pressure using a hyperbolic tangent function; the third fully connected layer generates the mounting angle parameters using normalization processing; and the fourth fully connected layer uses a sigmoid function to map the feasible range of the path compensation amount. The parameter re-constraint module has a built-in process parameter database. When the predicted patch placement speed exceeds the rated speed of the equipment, the output value is automatically reset to 95% of the rated speed. When the path compensation amount exceeds the boundary of the mounting area, the compensation amount is adjusted to within a safe threshold based on a coordinate mapping algorithm.

[0112] Through the above technical solution, this application effectively solves the problem of placement control parameters exceeding the physical limits of the equipment or process specifications, ensuring that the optimized control parameter set is always within the executable range. The automated parameter correction mechanism avoids abnormal equipment operation caused by neural network prediction bias, while also ensuring the structural integrity of highly sensitive components during the placement process, thus improving the reliability of placement operations under complex conditions.

[0113] The above embodiments overcome the limitations of existing SMT placement control methods that rely primarily on single image information by introducing a multi-source data fusion mechanism. Before placement control, a comparison and judgment logic between the current placement information and optimized decision parameters is introduced to avoid ineffective optimization and improve resource utilization efficiency. By performing image enhancement and defect identification on historical placement images, defect probability, offset, and posture deviation are quantified to form a structured defect feature vector, which helps to deeply explore quality fluctuation patterns during the placement process. Electrical attribute vectors are constructed by fusing component electrical attributes, and combined with pad diagrams, coordinate information, and solder paste injection logs, a three-dimensional solder paste distribution model is established and thermal simulation is performed to generate a thermal field distribution map that realistically reflects thermal sensitivity risks, thereby achieving collaborative modeling of multiple physical factors including heat, force, and electricity. A multi-head attention neural network is used to efficiently focus and weight the three types of information—defects, thermal field, and electrical attributes—outputting precise optimized control parameters, including placement path, angle, and pressure. This improves placement reliability, customization capabilities, and adaptability, facilitating precise and high-quality placement control for complex irregularly shaped devices and highly thermally sensitive components.

[0114] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides an SMT placement optimization control system based on multi-source data, used to apply the above-mentioned SMT placement optimization control method based on multi-source data, including: The judgment module is configured to collect the current patch information, compare the patch information with the optimization judgment parameters, and determine whether to perform patch optimization control based on the comparison result.

[0115] The identification module is configured to collect and preprocess historical SMT placement image data when it is determined that placement optimization control is to be performed, identify defects in the image data, and obtain defect feature vectors. The defect feature vectors include defect probability, defect offset, and placement posture deviation.

[0116] The processing module is configured to collect electrical attribute data of the components to be mounted and construct an electrical attribute vector, which includes thermal sensitivity parameters, electrical tolerance levels, package type information, and rated power values.

[0117] The simulation module is configured to construct a three-dimensional spatial distribution model of solder paste based on the target PCB pad diagram, component placement coordinate information, and historical solder paste distribution log data, and to perform thermal simulation calculations to obtain the thermal field distribution map of the component placement area.

[0118] The optimization module is configured to take the defect feature vector, electrical attribute vector and thermal field distribution map as multi-source inputs, input them to the multi-head attention neural network model, and output a set of patch optimization control parameters, which includes patch speed, bonding pressure, mounting angle and path compensation amount.

[0119] The execution module is configured to drive the placement device to perform placement operations based on the placement information and the placement optimization control parameter set.

[0120] The judgment module triggers an optimization process by comparing component models, package complexity, and historical defect rates. The identification module performs brightness normalization and geometric correction on the image, then extracts defect feature vectors using region segmentation and classification algorithms. The processing module integrates the component's thermal sensitivity parameters and tolerance levels to form an electrical attribute vector. The simulation module reconstructs the solder paste injection path and pressure data into a 3D point cloud model, overlays the pad structure, and performs transient heat conduction simulation to generate a 2D thermal map. The optimization module concatenates the normalized feature vectors with the convolutionally encoded thermal map into a multi-source input tensor. It then focuses on defect offsets and thermal anomaly regions using an attention head, generating intermediate feature representations, which are then output by the regression prediction module to determine the placement speed and path compensation. The execution module adjusts the equipment's motion mechanism based on the parameter set, achieving dynamic control of the bonding pressure and angle. Through multi-module data collaboration, the system can comprehensively consider the thermal effects of solder paste distribution, device electrical sensitivity, and historical defect patterns to generate more adaptable control parameters, improving placement accuracy and soldering reliability.

[0121] As a preferred embodiment, the specific implementation of this application is as follows: The judgment module collects the component model, package type, equipment placement speed, and historical defect rate data of the placement equipment in real time through the industrial bus interface, and performs logical comparison with the preset thermal sensitivity threshold and defect rate threshold. When a packaged component is detected and the historical bridging defect rate exceeds 5%, an optimization control command is triggered. The identification module calls the placement images of the same type of components stored in the image system within three months, uses a bilateral filtering algorithm for noise reduction, segments the device contour through a U-Net network and identifies offset defects, and outputs a defect feature vector containing an average offset of 0.12mm. The processing module extracts the thermal resistance coefficient of QFN-48 packaged devices (1.2℃ / W) and tolerance level (±5%) from the component database and constructs an eight-dimensional electrical attribute matrix. The simulation module imports the 0.5mm pitch pad array from the PCB Gerber file, combines it with reflow oven temperature curve data, and uses ANSYS transient thermal analysis to calculate the temperature gradient distribution of the placement area, generating a thermal field map containing three high-temperature clusters. The optimization module inputs the aforementioned features into a Transformer network with four attention heads. The second attention head focuses on the 135°C high-temperature region in the thermal field map. Through a cross-attention mechanism, it generates a set of control parameters that reduce the placement speed to 85% and increase Z-axis compensation by 0.03mm. The execution module writes the optimized parameters into the placement machine's motion controller, driving a linear motor to complete the component placement along the corrected path.

[0122] Through the above technical solution, this application effectively solves the problem that electrical properties and thermodynamic factors are not involved in control decisions in traditional methods, and realizes dynamic optimization of the mounting process of heat-sensitive components. By integrating multi-dimensional data to establish a comprehensive control model, the mounting and positioning accuracy of irregularly shaped packaged devices is improved, and the risk of soldering failure due to thermal stress is reduced. The feature fusion method based on the attention mechanism can adaptively identify key influencing factors, making the control parameters more in line with actual process requirements, thereby improving product yield and reducing rework costs.

[0123] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0124] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for SMT placement optimization control based on multi-source data, characterized in that, include: Collect the current patch information and compare it with the optimization judgment parameters. Based on the comparison result, determine whether to perform patch optimization control. When it is determined that surface mount optimization control is to be performed, historical SMT surface mount image data is collected and preprocessed, and defect identification is performed on the image data to obtain defect feature vectors. The defect feature vectors include defect probability, defect offset and mounting posture deviation. Collect electrical attribute data of the components to be mounted, and construct an electrical attribute vector. The electrical attribute vector includes thermal sensitivity parameters, electrical tolerance level, package type information, and rated power value. Based on the target PCB pad diagram, component placement coordinate information, and historical solder paste distribution log data, a three-dimensional spatial distribution model of solder paste is constructed, and thermal simulation calculations are performed to obtain the thermal field distribution map of the component placement area. The defect feature vector, electrical attribute vector, and thermal field distribution map are used as multi-source inputs and fed into a multi-head attention neural network model, which outputs a set of patch optimization control parameters, including patch speed, bonding pressure, mounting angle, and path compensation. The placement equipment is driven to perform placement operations based on the placement information and the placement optimization control parameter set.

2. The SMT placement optimization control method based on multi-source data according to claim 1, characterized in that, When comparing the patch information with optimization judgment parameters and determining whether to perform patch optimization control based on the comparison result, the following steps are included: The current placement information includes at least one placement batch's target component model, package type, PCB process reference number, equipment placement speed, placement pressure, and historical image detection defect rate; The optimized judgment parameters include thermal sensitivity threshold, defect rate threshold, component packaging complexity level, and mounting accuracy requirement level. The corresponding fields in the current patch information are compared with the optimization judgment parameters at the field level, including: determining whether the target component is of a high thermal sensitivity level; determining whether the frequency of defects of the target component in historical images exceeds a set threshold; determining whether the package type of the target component is fine pitch or irregular structure; and determining whether the mounting position is at the edge of the PCB or a heat accumulation area. If any comparison result is determined to meet the optimization conditions, then the placement optimization control is performed; otherwise, the optimization step is skipped, and the default placement parameters are called for placement control.

3. The SMT placement optimization control method based on multi-source data according to claim 1, characterized in that, When collecting and preprocessing historical SMT patch image data, and performing defect identification on the image data to obtain defect feature vectors, the process includes: The image data is subjected to image enhancement processing, which includes brightness normalization, noise filtering, edge enhancement, and geometric correction. Based on convolutional neural networks, the image is segmented and classified into multiple categories to identify defect types including mounting offset, warping, bridging, cold solder joints and missing parts, and output the corresponding defect labels and location box coordinates. The identification results are mapped to the corresponding mounting device reference numbers to form component-level defect structured data.

4. The SMT placement optimization control method based on multi-source data according to claim 3, characterized in that, When collecting and preprocessing historical SMT patch image data, and performing defect identification on the image data to obtain defect feature vectors, the process further includes: For each component to be evaluated, the frequency of defect occurrence is statistically analyzed based on the component-level defect structured data, and the defect probability is calculated as the first feature dimension. Based on the offset distance of the defect location box in the image relative to the theoretical mounting center of the device, the offset value and rotation angle offset are calculated to form the defect offset as the second feature dimension. The degree of contour tilt of the device in the identified image is quantified as a posture deviation factor as the third feature dimension. The first feature dimension, the second feature dimension, and the third feature dimension are integrated to form the defect feature vector.

5. The SMT placement optimization control method based on multi-source data according to claim 1, characterized in that, When constructing a 3D spatial distribution model of solder paste based on the target PCB pad diagram, component placement coordinates, and historical solder paste distribution log data, the following are included: The target PCB pad diagram is analyzed to extract the pad position, size and arrangement structure corresponding to each mounted component, and a pad geometric model is established in a spatial coordinate system. Alignment and registration are performed based on the coordinate information of the surface mount components and the positions of the pads on the pad diagram to generate a mounting alignment reference area; Read historical SMT solder paste distribution log data, including solder paste injection path, injection thickness, injection speed, and applied pressure; The log data is reconstructed into a point cloud in a three-dimensional coordinate space to construct a volume distribution model after solder paste injection, which is then superimposed on the corresponding pad structure to form a component-level three-dimensional volume distribution model of solder paste.

6. The SMT placement optimization control method based on multi-source data according to claim 5, characterized in that, When performing thermal simulation calculations to obtain the thermal field distribution map of the component mounting area, the following should be included: On the three-dimensional distribution volume model of the solder paste, thermophysical simulation boundary conditions are constructed based on the component thermal parameters and reflow soldering process parameters; the component thermal parameters include thermal conductivity, specific heat capacity, and packaging material; the reflow soldering process parameters include temperature profile, heating stage, and duration; Transient heat conduction simulation of solder paste-pad-component is performed based on the finite difference method to obtain the temperature field evolution results at each time point; The thermal field distribution map is extracted from the local area where the target component is located. The thermal field distribution map is used to generate a two-dimensional thermal map through interpolation and mapped according to the component reference number.

7. The SMT placement optimization control method based on multi-source data according to claim 1, characterized in that, When the defect feature vector, electrical property vector, and thermal field distribution map are used as multi-source inputs to a multi-head attention neural network model, and the output patch optimization control parameter set is generated, it includes: The defect feature vector and electrical attribute vector are respectively subjected to numerical normalization. Based on an independent embedding network, the defect feature vector and the electrical attribute vector are linearly mapped to generate a first embedding feature and a second embedding feature, respectively. Two-dimensional convolutional encoding is performed on the thermal field distribution map to extract spatial thermal feature maps; The first embedded feature, the second embedded feature, and the spatial heat map are concatenated under a unified feature dimension to form a multi-source fusion input tensor, which is then input to the input layer of the multi-head attention neural network model.

8. The SMT placement optimization control method based on multi-source data according to claim 7, characterized in that, The multi-head attention neural network model includes: It includes at least three attention heads, each of which independently learns an attention pattern for a specific feature category, wherein one attention head is used to focus on defect offset anomalies, another attention head is used to focus on thermal anomaly regions, and yet another attention head is used to identify highly sensitive characteristics of electrical properties. Each attention head employs a scaled dot-product attention mechanism, performing a weighted product between the key, query, and value vectors, and outputting the feature weighting result. The outputs of each attention head are concatenated and then fed into a fully connected layer and a ReLU activation layer via a residual connection and normalization module to form an intermediate feature representation generated by the optimized control strategy.

9. The SMT placement optimization control method based on multi-source data according to claim 8, characterized in that, When outputting the patch optimization control parameter group, it includes: The intermediate feature representation is input into the regression prediction module, which contains several parallel fully connected output layers. Each output layer is used to predict the patch speed, bonding pressure, mounting angle and path compensation amount. The output of each control parameter is a continuous value in the corresponding physical unit, and the output is constrained by the Sigmoid or Tanh function; The control parameter set is compared with the process constraint parameters of the current mounting task. If any parameter exceeds the limit, the parameter re-constraint module is triggered to perform range rollback and automatic correction.

10. A multi-source data-based SMT placement optimization control system, used to apply the multi-source data-based SMT placement optimization control method as described in any one of claims 1-9, characterized in that, include: The judgment module is configured to collect the current patch information, compare the patch information with the optimization judgment parameters, and determine whether to perform patch optimization control based on the comparison result; The identification module is configured to collect and preprocess historical SMT placement image data when it is determined that placement optimization control is to be performed, perform defect identification on the image data, and obtain a defect feature vector, wherein the defect feature vector includes defect probability, defect offset and placement posture deviation. The processing module is configured to collect electrical attribute data of the components to be mounted and construct an electrical attribute vector, which includes thermal sensitivity parameters, electrical tolerance levels, package type information and rated power values. The simulation module is configured to construct a three-dimensional spatial distribution model of solder paste based on the target PCB pad diagram, component placement coordinate information, and historical solder paste distribution log data, and to perform thermal simulation calculations to obtain the thermal field distribution map of the component placement area. The optimization module is configured to take the defect feature vector, electrical attribute vector and thermal field distribution map as multi-source inputs, input them into a multi-head attention neural network model, and output a set of patch optimization control parameters, which includes patch speed, bonding pressure, mounting angle and path compensation amount. The execution module is configured to drive the placement device to perform placement operations based on the placement information and the placement optimization control parameter group.

Images