System and method for predicting quality of printed circuit board assembly

Abstract

The invention relates to a system and method for predicting the quality of a printed circuit board assembly, the method comprising the steps of obtaining production data related to the production of a printed circuit board, PCB, assembly, mapping the production data onto a latent vector of a latent space of a trained adaptive algorithm, preferably based on a trained regression algorithm, wherein the trained adaptive algorithm is trained on and / or used to generate X-ray images of PCB assemblies, and determining a subspace of the latent space related to the latent vector, the subspace being indicative of the quality of the PCB assembly, in particular of a region of interest of the PCB assembly, and / or generating an X-ray image of the PCB assembly, in particular of the region of interest, based on the latent vector by the trained adaptive algorithm in order to determine the quality of the PCB assembly.

Description

Technical Field

[0001] This invention relates to the production of electronic products, and more particularly to the production and inspection of printed circuit board assemblies. Background Technology

[0002] The manufacturing of electronic products is prone to numerous potential failures during the production process. Therefore, the manufacturing process for surface-mount electronic devices (SMD) includes quality inspection processes for defect detection. The detection of certain error patterns, such as solder voids and pillow effect defects, requires radiographic inspection. These high-end inspection machines (such as X-ray inspection) rely on static inspection routines manually programmed by expert users of the machine to verify quality. Quality can be assessed using the tacit knowledge of domain experts based on soldering guidelines. This unique reliance on individual qualifications significantly impacts the error call rate of built-in computer vision routines.

[0003] According to existing technology, solder joint quality assessment is performed using X-ray inspection based on grayscale images. The intensity of certain areas in the X-ray image varies depending on the material being irradiated. This causes areas where metal material accumulates to darken. Automatic assessment of predefined image areas is performed based on these intensity differences. If a certain number of pixels within an area exceeds a set threshold, the area is marked as deficient. Operator confirmation is required. These product-specific test routines are manually created and adjusted. The analytical methods used in currently commercially available machines are based on deterministic, rule-based systems. In industrial applications, distinctions are made between different X-ray technologies. Vertical radiography of components is referred to as 2D X-ray. X-rays are a cost-effective solution for shadowless components with relatively short processing times. Because radiation penetrates matter, grayscale images (0-255) can be generated using (digital) sensors based on the material and volume penetrated by the radiation. Deviations within the grayscale value distribution can be identified using integrated image recognition or human visual inspection.

[0004] Due to higher packaging density or double-sided assembly, more complex systems are required to ensure adequate imaging. Unsuitable aspect ratios of PCBs necessitate separate 3D processes. To some extent, shadowing effects can be reduced by moving components between the X-ray source and detector and reconstructing quasi-3D images via sliced ​​images; this is known as 2.5D inspection. To further reduce shadowing, in X-ray tomography, the X-ray unit and detector move at an angle around the component on a coplanar trajectory, where the axis of rotation is tilted to the beam direction. A 3D image is formed through multiple image acquisitions and reconstruction algorithms (e.g., tomographic X-ray combination). High magnification is possible due to the specific setup and the small distance between the system and the component. Summary of the Invention

[0005] This invention aims to improve the inspection of PCB components, and in particular to reduce the number of inspection steps required. Therefore, the objective of this invention is to reduce the workload in product manufacturing.

[0006] During production, quality control or inspection is typically performed after product manufacturing and / or after each production step. Quality control is performed using imaging processes, thereby determining whether the manufactured product meets quality requirements based on images. This is especially useful for products where external imaging is insufficient, as X-rays can be used to create images.

[0007] Quality control is often time-consuming and sometimes requires high safety standards. In particular, the use of X-rays is detrimental to quality control. Furthermore, quality control is a significant cost factor.

[0008] According to the first aspect, this objective is achieved through a computer-implemented method for predicting the quality of a printed circuit board (PCB). The method includes the step of obtaining production data related to the production of the PCB assembly. The method further includes, preferably, the step of mapping the production data onto a latent vector in a latent space of a trained adaptive algorithm based on a trained regression algorithm, wherein the trained adaptive algorithm is trained on and / or used to generate X-ray images of the PCB assembly on real X-ray images of the PCB assembly. The method also includes the step of determining a subspace of the latent space associated with the latent vector, which indicates the quality of the PCB assembly, particularly the quality of the region of interest of the PCB assembly. Alternatively or additionally, the method further includes the step of generating X-ray images of the PCB assembly, particularly the region of interest, based on the latent vector using the trained adaptive algorithm, in order to determine the quality of the PCB assembly.

[0009] According to the second aspect, this objective is achieved by an apparatus, preferably a check station, which preferably includes a processor and a memory, operable to perform the method steps according to the first aspect.

[0010] According to a third aspect, this objective is achieved by a computer-implemented method for obtaining an adaptive algorithm for predicting the quality of one or more PCB assemblies. The method includes, for example, obtaining a real X-ray image of one or more PCB assemblies from an X-ray inspection system. The method also includes training an autoencoder capable of reconstructing the real X-ray image input. Furthermore, the method includes obtaining the adaptive algorithm for predicting the quality of one or more PCB assemblies by identifying a decoder portion of the autoencoder, for example, by removing the encoder portion of the autoencoder, wherein the decoder portion serves as a latent spatial interpreter.

[0011] According to the fourth aspect, this objective is achieved by a method for obtaining a computer-implemented regression algorithm for predicting the quality of one or more PCB components. The method includes the step of obtaining generated data and latent vectors of a latent space from a trained encoder, wherein the trained encoder is used to compress X-ray images. The method also includes the step of training a regression algorithm (e.g., an X-Tree Boost algorithm) capable of mapping the generated data onto the latent vectors, thereby obtaining the trained regression algorithm.

[0012] These and other objects, features and advantages of this disclosure will become apparent when reading the following detailed description in conjunction with the accompanying drawings, which are briefly described below. Attached Figure Description

[0013] Figure 1 An illustration shows a production area used for manufacturing and inspecting printed circuit board assemblies.

[0014] Figure 2 A first exemplary embodiment is shown, in which production data is mapped onto a potential vector.

[0015] Figure 3 A second exemplary embodiment is shown, according to which an X-ray image is generated based on product data.

[0016] Figure 4 A third exemplary embodiment is shown, according to which the quality of a PCB assembly is determined based on a potential vector.

[0017] Figure 5 A fourth exemplary embodiment is shown, according to which the quality of a PCB assembly is determined based on the generated X-ray image.

[0018] Figure 6 Several real X-ray images are shown.

[0019] Figure 7 Several regions of interest for the PCB assembly are shown.

[0020] Figure 8 A PCB assembly including multiple components mounted on a PCB is shown.

[0021] Figure 9 An autoencoder for generating X-ray images according to one embodiment is shown.

[0022] Figure 10 The potential space of the trained adaptive algorithm is shown.

[0023] Figure 11 The determination of quality indicators based on X-ray images is shown.

[0024] Figure 12 Method steps according to one embodiment are shown, wherein a subspace indicating the quality of a PCB assembly is determined.

[0025] Figure 13 The method steps according to another embodiment are shown, wherein quality is determined based on the generated X-ray image.

[0026] Figure 14 The method steps according to yet another embodiment are shown, wherein a generated X-ray image is output.

[0027] Figure 15 The method steps according to yet another embodiment are shown, wherein quality indicators of a PCB assembly are determined.

[0028] Figure 16 The method steps according to yet another embodiment are shown, wherein a trained adaptive algorithm is selected based on the region of interest of the PCB assembly.

[0029] Figure 17 Method steps according to yet another embodiment are shown, wherein the inspection of the generated X-ray image is performed based on a static, rule-based computer vision and / or a trained machine learning model.

[0030] Figure 18 The method steps according to yet another embodiment are shown, wherein a trained adaptive algorithm is obtained.

[0031] Figure 19 The method steps according to yet another embodiment are shown, wherein a trained regression algorithm is obtained.

[0032] Figure 20 A device including a processor and memory is shown. Detailed Implementation

[0033] In electronic products, electrical components are electrically and mechanically connected to printed circuit boards, for example, using through-hole technology (THT). Figure 1 and Figure 8As shown. In most cases, at placement station 11b, such as a manual placement station or by means of a placement robot, electrical components are inserted into mounting holes on the printed circuit board. Before placement station 11b, solder paste can be applied to the printed circuit board via solder paste printing. Afterward, the solder paste applied to the printed circuit board can be inspected at solder paste inspection station 11a. The printed circuit board, along with the inserted electrical components, is then driven into wave soldering machine 12, where the printed circuit board is pulled onto a soldering shaft, where solder electrically creates a conductive connection between the circuit board and the electrical components. After the soldering process, the PCB assembly is obtained, and the contacts (hereinafter referred to as solder joints) and their characteristics are inspected at inspection station 10 for quality control. For this purpose, optical inspection (AOI) and / or X-ray inspection (AXI) are typically used, such as... Figure 1 As shown. Inspection station 10 includes hardware components for image processing, such as an X-ray source, digital sensors, a computer, etc. For example, Omron offers such an inspection station called VT-X750. The VT-X750 is designed using Computed Tomography (CT) to provide high-precision X-ray imaging for accurate and reliable inspection of welded areas during production. The VT-X750 inspects for weld defects such as pillow effects and voids within BGA, LGA, THT, and other discrete components. As a result, realistic X-ray images are obtained.

[0034] The possible defects that inspection station 10 should detect are missing parts, bent parts, insufficiently wetted solder joints (cold or open solder joints) or solder scratches, i.e., the electrical contact of the solder joint, which should have an electrical potential.

[0035] X-ray examinations capture real X-ray images through image acquisition. For example, digital X-ray detectors can be used to measure the flux, spatial distribution, spectrum, and / or other properties of X-rays.

[0036] If image processing detects one or more errors on a printed circuit board assembly, the PCB assembly is fed back to diagnostic station 13 via a return route for manual inspection. Field operators now perform manual (visual) inspections, for example, via display D on the diagnostic dashboard.

[0037] Therefore, the production of PCB assemblies typically requires one or more, i.e., multiple production steps. The quality of the PCB assembly is determined by these production steps and is only inspected by an X-ray machine, i.e., without being affected or modified. Thus, the quality of the PCB assembly is already determined before X-ray inspection. Therefore, it is recommended to utilize and benefit from the production data generated related to one or more production steps of the PCB assembly.

[0038] Regarding the production of printed circuit board assemblies, production data can be obtained from different production stations performing one or more production steps. Production data may include: the position of solder paste on the circuit board, the deviation of the applied solder paste from a target position or target path, the thickness of the applied solder paste layer, and / or the temperature or humidity of the production environment. Production data 101 can be obtained by one or more sensors at each production station. Sensors are preferably cameras, temperature sensors, position sensors, distance sensors, vibration sensors, or flow sensors. Alternatively or additionally, production data can be obtained from a database in memory, which can be located at an inspection station, on-site at a production station, or off-site via a remote connection to the memory, such as in a storage facility from a component manufacturer. Production data 101 can be obtained from the respective production station via a data bus. For this purpose, the production station can be communicatively connected to the data bus.

[0039] Production data may include or be based on one or more of the following:

[0040] - Solder paste information, such as the formation, size, and / or volume of one or more solder libraries applied to the PCB, obtained from a solder paste inspection station.

[0041] - Component information, such as information obtained from the component placement machine (11b) or from the component supplier, such as the coplanarity information of the component's pins.

[0042] - For example, information about one or more characteristics of the PCB obtained from the barcode (B) on the PCB, such as the PCB substrate material, solder resist application process type, manufacturing location, and / or

[0043] - Residual oxygen level, temperature profile of reflow oven (12) and / or aging of the PCB and / or components mounted on the PCB.

[0044] Inspection station 10 can be operated to determine the quality of printed circuit board assembly 1. For this purpose, a quality index I1 can be used. The quality index preferably indicates whether the PCB assembly meets the quality requirements. The quality index can preferably be represented by "0" or "1" (in binary). "0" indicates, for example, that the corresponding PCB assembly does not meet the quality requirements, and "1" indicates that the corresponding PCB assembly meets the quality requirements.

[0045] For example, quality indicators show whether all connections between components have been adequately developed. Quality indicators can also indicate whether a PCB assembly functions correctly after its production. Preferably, quality indicators are calculated using a computer program installed and / or accessible, for example, at inspection station 10 or... Figure 20The calculation is performed on a computing unit, such as a processor, in the illustrated device. Quality metrics can now be calculated based on production data rather than actual X-ray images. Preferably, this calculation is performed using a (trained) regression algorithm 102 and / or a (trained) adaptive algorithm 104, as will be described below. Both the (trained) regression algorithm 102 and the (trained) adaptive algorithm 104 are (trained) machine learning models.

[0046] Figure 2 A first exemplary embodiment is shown, in which production data is mapped onto a potential vector.

[0047] Production data 101 is input into a (trained) regression algorithm 102. The training of regression algorithm 102 will be described later in this document. The regression algorithm maps the production data 101 input onto a latent vector 103. The latent vector resides in the latent space of the (trained) adaptive algorithm, as will be described here later. Thus, the trained regression algorithm 102 determines the latent vector to predict the quality of the printed circuit board assembly. Therefore, the latent vector is used to predict the quality of the PCB assembly.

[0048] In the current context, the production data is mapped to the latent space of a trained adaptive algorithm, such as the latent space of a trained autoencoder. An autoencoder (AE) is an artificial neural network with an encoder-decoder architecture. The encoder part maps the input to the latent space, and the decoder part maps the latent space to a reconstruction of the input. Therefore, the decoder part can interpret the latent vectors in the latent space determined based on the production data (by a (trained) regression algorithm).

[0049] A further development of the classic autoencoder represents the architecture of the Variational Autoencoder (VAE). This model achieves regularization of the distribution of the latent representation, ensuring that the learned latent space is continuous, and therefore each point in the latent space can be mapped by the decoder to a meaningful data point. A key difference between VAEs and other types of autoencoders is that VAEs represent latent representations as latent variables with their own preferential distribution. This gives them Bayesian interpretation. Therefore, the Variational Autoencoder is a generative model with both prior and posterior data distributions.

[0050] Alternatively, a shrinkage autoencoder CAE based on ResNet18 can be used to automatically generate X-ray images from digital production data obtained from previous production steps. In this case, the encoder portion can be based on ResNet18, and the decoder can be based on inverse ResNet18.

[0051] Figure 3A second exemplary embodiment is shown, according to which X-ray images are generated based on production data. Similar to... Figure 2 In this embodiment, production data is input into a (trained) regression algorithm. The regression algorithm maps production data related to the production of PCB components onto latent vectors. These latent vectors belong to a latent space. This latent space is created using one or more real X-ray images. For example, an autoencoder can be trained on real X-ray images, and the latent space can be extracted from the autoencoder after training. Therefore, the decoder portion of the autoencoder, also known as a (trained) adaptive algorithm, can be used to generate the X-ray image. Thus, based on latent vector 102, an artificially generated X-ray image can be obtained by decoding, decompressing, or reconstructing latent vector 103. As a result, an (artificially) generated X-ray image is obtained, which can be used to determine the quality of one or more PCB components. Therefore, an X-ray imaging process is no longer needed, as the generated X-ray image can be used for inspection.

[0052] Figure 4 A third exemplary embodiment is shown, according to which the quality of a PCB assembly is determined based on a potential vector.

[0053] The latent space can be one-dimensional or multi-dimensional, such as 2D, 3D, or n-dimensional, based on which the quality of a PCB assembly, including, for example, one or more regions of interest including one or more solder joints, can be determined. This latent space may include one or more subspaces characterized by the quality of the PCB assembly. These subspaces are determined by training an autoencoder. The boundaries of the subspaces can be determined by the autoencoder itself. Each subspace may correspond to a fault type of the PCB assembly, such as solder bridges or open solder joints. If a latent vector lies in a certain subspace, the PCB assembly may therefore have solder bridges or open solder joints. Thus, latent vector 103 can be used to determine the quality 107 of the PCB assembly.

[0054] Figure 5 A fourth exemplary embodiment is shown, according to which the quality of a PCB assembly is determined based on a generated X-ray image. The generated X-ray image can be used in place of a real X-ray image to determine the quality of the PCB assembly. For this purpose, the generated X-ray image is classified. For example, an image classification algorithm can be used, which determines the quality of the PCB assembly, for example, by classifying the image into one or more categories corresponding to the quality 107 of the PCB assembly. For example, the method described in the background section above can be used to determine the quality 107 of the PCB assembly.

[0055] Figure 6Multiple real X-ray images are shown. Multiple X-ray images 20 can be obtained as a result of an imaging process that irradiates one or more PCB components with X-rays. Here, each image corresponds to a different PCB component being manufactured. The real X-ray images 20 are generated by electromagnetic radiation (X-rays) penetrating the corresponding PCB component. After the PCB component is scanned, images 21, 22, 23, 24, 25, and 27 can be stored. The X-ray images 20 can be stored, for example, as a 3D-X-ray matrix or vector in the form of a file. The X-ray images 20 can have a resolution of height, width, and depth, for example, 483 × 623 × 65, and are preferably stored as metadata, for example, in the form of a separate file. Therefore, the real X-ray images 20 can be used, in whole or in part, to train adaptive algorithms.

[0056] Figure 7 Several regions of interest (ROIs) of the PCB assembly are shown. Depending on the specific context, an ROI may correspond to a region of the PCB assembly and / or a region of an image of the PCB assembly. In any case, an ROI includes one or more solder joints. Therefore, ROIs can vary in terms of the number of solder joints, their location, their size, and the function of the attached components. Figure 7 As shown, the first region of interest (ROI) R1 may include a single but large solder joint P1, while the second ROI R2 includes two solder joints P2 and P3. The third ROI R3 may include a large number of small solder joints, and the fourth ROI R4 includes a smaller number of fairly large solder joints. Therefore, it may be necessary to provide a corresponding (trained) adaptive algorithm for each ROI R1, R2, R3, and R4. Consequently, a corresponding regression algorithm may also be required.

[0057] Figure 8 A PCB assembly L is shown, comprising multiple components K1, K22, and K3 mounted on PCB L. A printed circuit board (PCB) uses conductive tracks, solder areas, and other features etched into one or more copper sheets laminated to and / or between non-conductive substrates to mechanically support and electrically connect electronic components. Components are typically soldered to the PCB to electrically connect and mechanically secure them to the PCB. Printed circuit boards are used in almost all electronic products and some electrical products, such as passive switch boxes.

[0058] A barcode B1 is applied to the PCB L. It contains component indicators, switch settings, test points, and other indications that aid in assembly, testing, repair, and sometimes use of the circuit board. The barcode B1 can be applied, for example, using silkscreen epoxy ink, liquid photoimaging, or inkjet printing.

[0059] In addition, one or more of the components K1 and K2 mounted on the PCB may include Data Matrix Codes (DMCs) B21 and B22, which may also be printed thereon. The DMC may include a unique component ID that is read during the assembly process. Production data generated during assembly may be output as a *.xml file. Furthermore, the production data may include metadata related to the component's assembly and logistics data (e.g., lifespan).

[0060] Component suppliers can also provide the coplanarity values ​​of the corresponding components in the form of *.csv files. By reading the DMC, the coplanarity values ​​can be explicitly assigned to the corresponding components.

[0061] Figure 9 An autoencoder for generating X-ray images according to one embodiment is shown. The autoencoder AE has two main parts: an encoder 104b that maps the input to a latent space C, and a decoder 104a that reconstructs the latent space C to the input.

[0062] Real-world data, such as product data, voice signals, digital images, or MRI scans, often has high dimensionality. However, to process such real-world data, it is usually necessary to reduce its dimensionality. The term dimensionality reduction in machine learning is understood as transforming high-dimensional data into a meaningful representation with reduced dimensionality, thus preserving the original fundamental properties. Dimensionality reduction provides a reduced number of new features created based on the original dataset. Ideally, the simplified representation should have a dimensionality corresponding to the inherent dimensionality of the data. Dimensionality reduction algorithms typically consist of an encoder-decoder architecture. In the encoder, new features are created from the input data. The decoder performs the reverse process, reconstructing the input data based on the new features. Therefore, dimensionality reduction can be interpreted as data compression, where the encoder compresses data from the output space into an encoding space, also called the latent space or latent representation. The decoder then decompresses the data again in the original space. This architecture is used to create representations from input data of the original dimensionality, also called latent representations or latent spaces. The transition from input x to the latent space h is used as encoding, and returning the code of the latent representation h to the original space is also called decoding.

[0063] In the current context, for training purposes, real X-ray images are input into the autoencoder AE. The encoder section 104b encodes the X-ray image input and generates latent vectors. The decoder 104a decodes the latent vectors belonging to the input image 20, thereby generating the output image 105. When multiple X-ray images are input, a latent space C is generated. In the case of using a variational autoencoder, the latent space consists of a mixture of distributed vectors rather than fixed vectors. Since the quality of the X-ray images used for training is known, the latent vectors in the latent space and / or the clusters of the latent space can also be used to determine the quality of the PCB assembly.

[0064] An autoencoder can be trained in an unsupervised manner on historical real-world X-ray images. These images can contain both defective and error-free PCB assemblies. Regularization is necessary to ensure that the autoencoder can reconstruct meaningful images from all possible latent vectors in the latent space via the decoder. A regression algorithm then maps the production data to latent vectors in the latent space. Therefore, regularization ensures that each latent vector regressed by the regression algorithm can also generate meaningful defective or error-free images of the PCB assembly by the decoder. Regularization prevents the autoencoder from using a pure mapping function on the images learned from the training dataset. Due to the high similarity of images of individual solder joints, the autoencoder needs to be able to detect small differences and biases between images. Therefore, both the encoder and decoder have multi-layered convolutional neural networks. Meanwhile, X-ray images are pure grayscale images with low pixel counts and simple structures, making them highly susceptible to model overfitting, especially for deep convolutional neural networks. Therefore, in general, a trade-off must be struck between the autoencoder's ability to detect detailed images and the risk of overfitting after only a few training epochs.

[0065] Furthermore, the autoencoder allows for the reconstruction of latent vectors into X-ray images. On one hand, this makes it possible to evaluate the accuracy of the autoencoder, since, for a labeled X-ray image test set, the difference between the original image and the image generated by the autoencoder can be evaluated, for example, by mean squared error. Thus, a pixel-by-pixel comparison is performed between the real X-ray image input and the resulting generated X-ray image.

[0066] Furthermore, the optical visualization of the generated X-ray images increases the acceptability of the quality checks suggested by the user.

[0067] Figure 10The latent space of the trained adaptive algorithm is shown. This latent space comprises multiple clusters C1, C2, C3. Each cluster can correspond to one or more features created by the trained adaptive algorithm. For example, one or more clusters can be created by partitioning the latent space. This can be done, for example, manually. Alternatively, one or more clusters can be defined using, for example, a K-means clustering algorithm. Each cluster can correspond to a fault type and / or a fault-free PCB assembly, respectively. Therefore, each cluster can be assigned a quality metric. For example, each cluster can be assigned a quality metric indicating the fault type. In any case, one or more clusters can be assigned a quality metric corresponding to a fault-free PCB assembly.

[0068] To achieve optimized feature extraction and clustering, the autoencoder should be trained with as many training images as possible. Furthermore, the training dataset should be balanced, including both faulty and error-free images. In addition, sufficient images representing different fault categories, i.e., fault types, should be included in the training dataset.

[0069] Figure 11 The determination of quality metrics based on X-ray images is illustrated. Quality metrics can be determined based on real X-ray images or X-ray images generated by a trained adaptive algorithm. Image classification can be used for this purpose. As previously mentioned, quality metrics can take binary values, i.e., "0" or "1", which represent error-free and faulty PCB components, respectively. For each region of interest of the PCB component, a quality metric can be obtained. Image classification can be rule-based or, again, based on a trained machine learning model.

[0070] Figure 12 The method steps according to one embodiment are illustrated, wherein a subspace indicative of the quality of a PCB assembly is determined. In a first step S0, production data related to the production of the PCB assembly is obtained. As described above, the production data can be obtained based on one or more production steps at one or more production stations. That is, after each production step, production data related to the performed production step is obtained. For this purpose, the production data can be obtained via a shared medium (e.g., a (wired or wireless) data bus) and / or can be stored in memory. The production data can be related to the production of the PCB assembly. Therefore, different or new production data is obtained for each PCB assembly.

[0071] In step S1, the production data is mapped onto latent vectors in the latent space of a trained adaptive algorithm. For this purpose, a trained regression algorithm can be used. This regression algorithm takes the production data as input and provides latent vectors as output. The trained regression algorithm can be executed once, for example, production data related to a PCB component is received at a checkpoint where the regression algorithm is performed.

[0072] Latent vectors can belong to a latent space. That is, the adaptive algorithm is trained on and / or used to generate X-ray images from real X-ray images of PCB components, and the latent space of the adaptive algorithm is identified, such as by combining... Figure 18 Explanation.

[0073] In step S2, a subspace of the latent space associated with the latent vector is determined, which indicates the quality of the PCB assembly. As shown, the latent space may include one or more clusters of subspaces forming the latent space. A latent vector may belong to a certain subspace, depending on the value of the latent vector obtained by the regression algorithm. Once the subspace is known, the quality of the PCB assembly can be determined. Therefore, the quality of the PCB assembly can be predicted without taking a real X-ray image of the PCB assembly or otherwise inspecting the PCB assembly. Different subspaces of the latent space may correspond to aspects of a faulty or error-free PCB assembly. In particular, the quality of one or more regions of interest of the PCB assembly can be determined in this way.

[0074] Figure 13 Method steps according to another embodiment are shown, wherein quality is determined based on the generated X-ray image. As previously described, in Figure 12 In steps S0 and S1, production data is obtained and mapped onto latent vectors. In step S3, X-ray images of the PCB assembly are generated based on the latent vectors using a trained adaptive algorithm to determine the quality of the PCB assembly. The generated X-ray images can then be displayed, for example, for manual inspection by a user, and / or can undergo image processing at an AXI inspection station, using only the now-generated images instead of taking actual X-ray images.

[0075] Figure 14 The method steps according to yet another embodiment are shown, wherein a generated X-ray image is output. After step S3, the generated X-ray image can be output on, for example, a display, to determine the quality of the PCB assembly. The display can be part of an inspection station or part of a handheld device. In either case, the generated image is output to determine the quality of the PCB assembly. The generated image can then be further processed by one or more other applications, which perform further image processing, for example, based on the generated X-ray image.

[0076] Figure 15Method steps according to yet another embodiment are shown, wherein a quality index of a PCB assembly is determined. The quality of the PCB assembly can be represented by a quality index. As mentioned above, the quality index can be given in binary form. The quality index can also be displayed via a display unit or a separate signaling device, such as in the form of a traffic light system. Quality index I1 can indicate a faulty or error-free state of the PCB assembly. In addition, the quality index can represent multiple error categories. For example, the error index can indicate several different fault states of the PCB assembly, i.e., errors. Possible errors during the production of a printed circuit board assembly include, for example: missing or incomplete solder paste printing, solder paste application, missing or incomplete adhesive printing, adhesive application, missing components, component misalignment or twisting, tombstone effect, component reversal, color detection of MELF components, confusion testing of IC / μC types, confusion testing of resistance values, component reversal (IC, diode), short circuits and solder bridging (IC, μC), missing solder joints, open THT solder joints, non-contact THT solder joints, misalignment of solder deposits, solder paste application, and bridging formation. The quality index can represent one or more of these error categories and can take different values ​​depending on the error category. The values ​​of quality metrics can consist of alphanumeric strings or numbers including alphanumeric strings. Furthermore, quality metrics can describe errors in plain text (see above) or contain errors in coded form.

[0077] In any case, the quality or quality index of the PCB assembly can be determined in step S5 based on one or more of steps S1, S2 and S3 as described above.

[0078] Once the latent vectors are obtained, for example, from a regression algorithm, the quality of the manufactured PCB assembly can be determined. Optionally, the latent vectors can be assigned to a subspace of the latent space, or further, the X-ray image can be determined by an adaptive algorithm.

[0079] Figure 16The method steps according to yet another embodiment are shown, wherein a trained adaptive algorithm is selected based on the region of interest (ROI) of the PCB assembly. As previously described, production data is received in step S0. Depending on the situation, the production data may be associated with one or more ROIs on the PCB assembly. Thus, a first portion of the production data may be associated with a first ROI, and a second portion of the production data may be associated with a second ROI. The first portion of the production data is now input into a first regression algorithm to determine a first latent vector for the first ROI. The second portion of the production data is input into a second regression algorithm to determine a second latent vector for the second ROI. The first latent vector belongs to a first latent space of the first trained adaptive algorithm. The second latent vector belongs to a second latent space of the second trained adaptive algorithm. The first and second trained adaptive algorithms are trained on real X-ray images including the first and second ROIs, respectively. Thus, multiple trained adaptive algorithms can be provided. In step S6, a corresponding trained adaptive algorithm is selected to process each latent vector, resulting in a quality (index) associated with the ROI.

[0080] Figure 17 Method steps according to yet another embodiment are shown, wherein the inspection of the generated X-ray image is performed based on a static, rule-based computer vision and / or a trained machine learning model.

[0081] After generating an X-ray image, for example, according to step S3, the quality of the PCB assembly, particularly the quality of one or more regions of interest, can be determined in step S7 using predetermined static criteria based on a first computer vision algorithm to examine the generated X-ray image of the PCB assembly, and / or in step S8 using a trained machine learning model based on a second computer algorithm to examine the generated X-ray image of the PCB assembly. One or more methods for examining PCB assemblies and / or images of PCB assemblies are described in European Patent Application No. EP 21190840, which is incorporated herein by reference.

[0082] The first computer vision algorithm corresponds to the algorithm described in the background section, wherein a grayscale threshold is compared pixel-by-pixel (i.e., per-pixel). The second computer algorithm may also use a machine learning model trained to determine quality based on generated X-ray images, for example, by classifying the generated X-ray image input. For this purpose, the machine learning model can be trained on real X-ray images.

[0083] Figure 18The method steps according to yet another embodiment are shown, wherein a trained adaptive algorithm is obtained. Specifically, in step S9, a real X-ray image of one or more PCB components is obtained, for example, from an X-ray machine at an inspection station. In step S10, an autoencoder capable of reconstructing the real X-ray image is trained. For this purpose, the real X-ray image is input into the input layer of the autoencoder. As a result, the image input is reconstructed by the autoencoder, and the reconstructed image is output via the output layer of the autoencoder. The autoencoder includes the features described above. Figure 9 The encoder and decoder portions are described above. In step S11, the encoder portion is removed. In step S12, a trained adaptive algorithm (corresponding to the remaining decoder portion) for predicting the quality of one or more PCB components is obtained. The trained adaptive algorithm corresponding to the decoder portion of the trained autoencoder serves as a latent space interpreter. Therefore, the trained adaptive algorithm is able to receive latent space vectors as input and reconstruct or generate X-ray images based on them.

[0084] Figure 19 The method steps according to yet another embodiment are shown, wherein a trained regression algorithm is obtained. In step S13, production data and latent vectors are obtained. Both can be stored in memory. Production data can be obtained from one or more production steps used to produce PCB assemblies. The latent vectors associated with the production data can be obtained from a combination of... Figure 18 The encoder portion of the autoencoder is obtained. For this purpose, one or more real X-ray images are acquired from the PCB assembly associated with the production data. These real X-ray images are input into the encoder, and latent spatial vectors are obtained as a result. Thus, as described above, production data for the PCB assembly and one or more latent vectors associated with the production data are obtained. This step can be repeated to obtain sufficient samples, i.e., production data and associated latent vectors. Now, a regression algorithm can be trained in step S14. The regression algorithm is capable of mapping the production data to one or more latent vectors. Therefore, the trained regression algorithm is obtained in step S15.

[0085] In regression algorithms, the target variable or label y is determined based on the attributes of covariates or features x1, ..., xn. The target variable is also called the dependent variable, and the covariates are called explanatory variables. In this case, production data is used as covariates and the latent vectors to be identified as the target variable. As a result, the target value y is a random variable that depends on the distribution of the explanatory variables. The main purpose of regression algorithms is to determine the influence of explanatory variables on the target variable. Below, we explain the application of decision trees in regression algorithms. In the case of regression trees, rules are formulated from the training data. Through these rules, the outcome space is defined; that is, the rules create the outcome space. By concatenating the rules, the outcome space is continuously reduced until a final assignment is made. At the node, the attribute of the input vector is queried, and then it is divided into branches of that attribute's output value. Through iterative processing of this process, the outcome space of the input vector is reduced by increasing the depth of the regression tree. If the branch cancellation criterion is met, the iteration ends, and the final node's category is assigned to the input vector, which is then called a leaf. During the training phase, the query for attributes in the tree is defined to obtain the maximum possible information gain of the query. The category assignment of the leaf is determined by the average value.

[0086] Random forests are a so-called ensemble method where decision trees are used as the basic regressors. Therefore, predictions are made by a majority of decisions from an ensemble of trees. This process is also known as bootstrap aggregation or simply packing. To generate individual decision trees, N (for N > n) training records n (bootstrap samples) of data points are randomly selected (by backtracking). The corresponding model is trained based on this training dataset. This process is repeated t times, thus obtaining different regressors from the same training dataset. In subsequent tests, the values ​​of the data points are determined by averaging the total values ​​of the t classifiers.

[0087] Besides packing, the boosting process can also be used to combine multiple regressors. The Boosting Tree process uses a few simple regressors with only a few branches, which have poor prediction accuracy but require only short computation time. These regressors are weighted based on the prediction error during training. This process is repeated until all regressors are weighted in an optimized manner. Known methods include Adaptive Boosting (AdaBoost), Gradient Boosting (GBoost), Extreme Gradient Boosting (XGboost), Light GBM, and Cat Boost.

[0088] Figure 20An apparatus 100 including a processor and memory is shown. This apparatus is preferably an inspection station or part of an inspection station. To perform the embodiments and method steps described above, the apparatus includes memory in which program code (non-transitory) can be stored. The program code can be executed by the processor. When the program code is executed, the method steps according to any of the above embodiments can be performed. In particular, the apparatus may be equipped with an interface for receiving production data. Furthermore, the apparatus may be equipped with an interface for outputting the quality (indicators) of one or more PCB assemblies produced. The processor and memory of the apparatus can be used to store and / or execute trained regression algorithms and / or trained adaptive algorithms. Furthermore, the apparatus is operable to store and / or execute the inspection routine via the memory and the processor to determine the quality (indicators) of the PCB assembly based on the generated X-ray image.

Claims

1. A computer-implemented method for predicting the quality of a printed circuit board (PCB) assembly (1), characterized in that, The method includes the following steps: Obtain (S0) production data (101) related to the production of the PCB assembly (1). The production data is mapped (S1) to latent vectors (103) in the latent space of a trained adaptive algorithm (104) based on a trained regression algorithm (102), wherein the trained adaptive algorithm (104) is trained on real X-ray images (20) of PCB assemblies and used to generate X-ray images (105) of PCB assemblies, and (S2) Determine (S3) a subspace (106) of the potential space associated with the potential vector (103), the potential space being capable of including one or more subspaces characterized by the quality of the PCB assembly, the subspaces being determined by the trained adaptive algorithm, the boundaries of the subspaces being determined by the trained adaptive algorithm itself, each subspace corresponding to a fault type of the PCB assembly, and The trained adaptive algorithm (104) generates (S3) an X-ray image (105) of the PCB assembly (1) based on the latent vector (103), and determines the quality (107) of the PCB assembly (1) based on the latent vector (103), the subspace (106) of the trained adaptive algorithm (104), and the generated X-ray image (105).

2. The method according to claim 1, characterized in that, The generated X-ray image (105) is output (S4) on the display (D) by the trained adaptive algorithm (104) in order to determine the quality (107) of the PCB assembly (1).

3. The method according to claim 1 or 2, characterized in that, Based on the potential vector (103), the subspace (106) of the trained adaptive algorithm (104), and the generated X-ray image (105), a quality index (I1) for the PCB assembly (1) is determined (S5).

4. The method according to claim 1 or 2, characterized in that, The PCB assembly (1) includes multiple regions of interest (R1, R2), each region of interest (R1, R2) includes one or more solder joints (P1), the production data (101) is associated with each of the regions of interest (R1, R2), and the method further includes the following steps: For each region of interest (R1, R2), a trained adaptive algorithm (104) is selected (S6) to determine the quality of the PCB assembly (1).

5. The method according to claim 1 or 2, characterized in that, The production data (101) includes one or more of the following: - Solder paste information obtained from the solder paste inspection station (11a), including the formation, size, and / or volume of one or more solder libraries applied to the PCB. - Component information obtained from the component placement machine (11b) or from the component supplier, including the coplanarity information of the component's pins. - Information on one or more characteristics of the PCB (L) obtained from the barcode (B) on the PCB, including the substrate material of the PCB, the type of solder resist application process, the production location, and / or - Residual oxygen level, temperature profile of reflow oven (12), and / or aging degree of the PCB and / or aging degree of the components mounted on the PCB (L).

6. The method according to claim 1 or 2, characterized in that, The generated X-ray image (105) of the PCB assembly (1) is inspected (S7) using a predetermined static standard based on a first computer vision algorithm, and / or the generated X-ray image (105) of the PCB assembly (1) is inspected (S8) using a trained machine learning model based on a second computer algorithm.

7. The method according to claim 1, characterized in that, The quality of the PCB assembly is the quality of the region of interest of the PCB assembly.

8. The method according to claim 1, characterized in that, The X-ray image (105) of the PCB assembly (1) is an X-ray image of the region of interest (R1) of the PCB assembly (1).

9. The method according to claim 3, characterized in that, The quality index (I1) for the PCB assembly (1) is the quality index for the region of interest (R1) of the PCB assembly (1).

10. The method according to claim 1 or 2, characterized in that, The method further includes: Obtain real X-ray images (20) of one or more PCB components (1) from the X-ray inspection system (100). The training is able to reconstruct the input of the real X-ray image (20) into an autoencoder (AE). By identifying the decoder portion (104a) of the autoencoder (AE), an adaptive algorithm (104) for predicting the quality of one or more PCB components (1) is obtained, wherein the decoder portion (104a) serves as a latent space interpreter.

11. The method according to claim 10, characterized in that, The method further includes: The latent vector (103) of the latent space (C) and production data (101) are obtained from the trained encoder (104b), wherein the trained encoder (104b) is used to compress X-ray images. Training the regression algorithm (102) that maps the production data (101) to the latent vector (103), and Thus, the trained regression algorithm (102) is obtained.

12. The method according to claim 10, characterized in that, By removing the encoder portion (104b) of the automatic encoder (AE), an adaptive algorithm (104) for predicting the quality of one or more PCB components (1) is obtained.

13. The method according to claim 11, characterized in that, The regression algorithm is the X-Tree Boost algorithm.

14. An apparatus comprising a processor and a memory, characterized in that, The memory stores code that is executed by a processor, causing the device to perform the steps of the method according to any one of the preceding claims.

15. The apparatus according to claim 14, characterized in that, The device is an inspection station.

Images