Method and device for optimizing an encoder model, for training a quality-predicting model, and for monitoring a process

Abstract

A method for optimizing an encoder model for an artificial intelligence-based quality predicting model (142) for generating a prediction of the quality of a connection between two components (104, 106) produced by means of a reactive multilayer system (108) comprises reading in raw data characterizing physical properties of a process for producing the connection and training the encoder model to encode the raw data into a dimensionally reduced feature vector and a decoder model to decode the feature vector into decoded raw data in order to obtain a trained encoder model in which a deviation between the raw data and the decoded raw data corresponds to a predetermined criterion.

Description

[0001] The present invention relates to a method and a device for optimizing an encoder model for an artificial intelligence-based quality-predicting model for generating a prediction of the quality of a connection between two components produced by means of a reactive multilayer system, a method and a device for training a quality-predicting model, and a method and a device for monitoring a process for producing a connection between two components by means of a reaction of a reactive multilayer system.

[0002] A reactive multilayer system, for example a reactive multilayer film (RMS), can be used to connect elements.

[0003] Against this background, the present invention provides an improved method and an improved device for optimizing an encoder model for an artificial intelligence-based quality-predicting model for generating a prediction of the quality of a connection between two components produced by means of a reactive multilayer system, an improved method and an improved device for training an artificial intelligence-based quality-predicting model for generating a prediction of the quality of a connection between two components produced by means of a reaction of a reactive multilayer system, and an improved method and an improved device for monitoring a process for producing a connection between two components by means of a reaction of a reactive multilayer system according to the main claims.Advantageous embodiments result from the dependent claims and the following description.

[0004] If an encoder is appropriately trained to generate a feature vector for an artificial intelligence-based quality predicting model, the size of the feature vector can be reduced, thereby simplifying the training and use of the quality predicting model.

[0005] A method for optimizing an encoder model for an artificial intelligence-based quality-predicting model to generate a prediction of the quality of a connection between two components produced using a reactive multilayer system comprises the following steps: Reading in raw data that characterizes the physical properties of a process for creating the compound; and Training an encoder model to encode the raw data into a dimensionally reduced feature vector and a decoder model to decode the feature vector into decoded raw data, in order to obtain a trained encoder model where a deviation between the raw data and the decoded raw data corresponds to a predetermined criterion.

[0006] The raw data can be pre-existing or currently acquired data. It can comprise single values ​​and, additionally or alternatively, multiple temporal sequences of values. Features can be derived from the raw data through encoding. This can utilize established feature extraction rules used in artificial intelligence models. For example, a feature can represent the mean or maximum value of a temporal sequence of values. The raw data could, for instance, depict the temporal progression of pressure or temperature during the manufacturing process, or a property of a component, such as the reactive multilayer system or one of the building elements. The dimensionally reduced feature vector can contain a smaller data volume than the raw data.Using the decoder model, the goal is to recover the original raw data from the dimensionally reduced feature vector. By training the encoder model, it can be modified so that the decoded raw data corresponds as closely as possible to the original raw data. For example, the predetermined criterion can include a threshold for the deviation between the original raw data and the decoded raw data. This deviation could, for instance, correspond to a so-called loss of reconstruction or cycle consistency. Known methods can be used to minimize such deviations. The quality-predicting model could, for example, be based on an artificial neural network.

[0007] Through activation and the resulting reaction of the reactive multilayer system, a stack of components and the reactive multilayer system can be assembled into a component. The component can thus be an assembly of components. The first component can be an electronic component or a sensor. For example, the first component could be a semiconductor device, a chip, or a wafer. For example, the component could be a transistor, a diode, or a force sensor. The second component can be a substrate for the first component. For example, the second component could be a substrate for electronic components, such as a printed circuit board or a substrate, such as a copper or ceramic substrate.The second component can also be a machine element, such as a connecting element or a force or motion transmission element, such as a shaft. The reactive multilayer system can be a reactive multilayer film (RMS) capable of providing instant heat for a wide variety of applications across many industries. The reactive multilayer system can be fabricated by vapor deposition of thousands of alternating nanoscale layers, for example, of aluminum and nickel. The reactive multilayer system can be arranged as a film between the components or comprise reactive layers that have been sequentially deposited directly onto one or both components to form the reactive multilayer system. The reactive multilayer system can perform the reaction in response to activation, such as ignition.For this purpose, the reactive multilayer system can be exposed to a laser beam or a spark, an ignition current can be passed through the reactive multilayer system, or the reactive multilayer system can be ignited using ultrasonic energy. This generates sufficient thermal energy to reach the temperature required to ignite the reactive multilayer system. Ignition can occur, for example, at one or more predefined positions, at a random position, or across the entire contact area between the reactive multilayer system and one of the components. After the reaction, the reacted reactive multilayer system can form a metallurgical bond between the components. This bond can be mechanically stable and electrically conductive and is also referred to as multilayer bonding.

[0008] During the data acquisition step, raw application data can be imported, characterizing the properties of the manufacturing device, the reactive multilayer system, and / or the components. A property of the manufacturing device could be, for example, the mass of a pressure plunger used to apply pressure to the reactive multilayer system, or the type of activation device used to activate the reactive multilayer system. A property of the reactive multilayer system could be, for example, its thickness or material composition. A property of the components could be, for example, its surface finish or material composition. Such properties can influence the quality of the resulting compound. Information about the relevant properties can be obtained, for example, through sensing or other means.

[0009] Additionally or alternatively, sensor data can be read in as raw data, representing values ​​sensed during a reaction of the reactive multilayer system. Such values ​​can influence the quality of the connection to be produced. The sensor values ​​can represent values ​​acquired using a sensor device. The sensor device can have one or more suitable sensors for acquiring the sensor values. The sensor values ​​can represent or depict a measured quantity, a property, or a state of the reactive multilayer system, or at least one of its components, before, during, or immediately after activation.

[0010] The process may include a step of sensing sensor values ​​using a sensor device. This allows real-world data to be provided. Such a sensor device may comprise one or more sensors. The sensor device may be part of a manufacturing apparatus for establishing the connection.

[0011] The encoder model can be designed to extract features in the statistical, temporal, and / or spectral domains from the raw data. This allows for the use of established methods.

[0012] The training step can be repeated, for example, until the discrepancy between the raw data and the decoded raw data is minimal. Repeated training allows the trained encoder model to be continuously optimized.

[0013] A method for training an artificial intelligence-based quality-predicting model to generate a prediction of the quality of a connection between two components produced by means of a reaction of a reactive multilayer system comprises the following steps: Reading in raw data that characterizes the physical properties of a process for making the connection and quality data that characterizes the quality of the connection; Encoding the raw data into a dimensionally reduced feature vector using a trained encoder model representing an encoder model optimized according to an embodiment of a said method for optimizing an encoder model, in order to obtain input data for the quality-predicting model; and Training the quality predictor model using the input data and the quality data to obtain a trained quality predictor model.

[0014] The raw data can be the same as, or similar to, the data used to optimize the encoder model. The dimensionally reduced feature vector generated from the raw data using the trained encoder model can be fed into the quality predictor model as input. The quality predictor model can be trained to generate an estimate of the quality of a joint produced using the manufacturing device, based on the dimensionally reduced feature vector. Thus, the quality of a produced joint can be predicted using the quality predictor model. In this way, a statement about the quality can be made without having to analyze the manufactured component itself. As output data, the quality predictor model can, for example, indicate either good or poor quality of the joint.In the case of poor quality, the component may be considered scrap. Quality data can be used to train the quality predictor model. Such quality data can be obtained, for example, through the analysis of manufactured joints. This data can be derived from shear tests that destroy the joint, X-ray images, electrical functional tests, an analysis of the injection molding behavior, lifetime tests, and / or from a detected deviation from an optimal position. The quality data can be used as reference data for training the quality predictor model. Corresponding sensor data, representing process parameters that prevailed during the manufacturing of the joint analyzed to obtain the quality data, can be associated with the quality data.By using the dimensionally reduced feature vector, rather than the raw data itself or a feature vector created using an untrained encoder model, to train the quality predicting model, the training process can be accelerated.

[0015] A method for monitoring a process for creating a connection between two components by means of a reaction of a reactive multilayer system comprises the following steps: Reading in current process data that characterizes the physical properties of the process for establishing the connection; Encoding the process data into a dimensionally reduced feature vector using a trained encoder model representing an encoder model optimized according to an embodiment of a said method for optimizing an encoder model, in order to obtain input data for the quality predicting model; and Generating a prediction of the quality of the connection produced by the reaction of a reactive multilayer system between the two components using a trained quality-predicting model that represents a quality-predicting model trained according to an embodiment of said method for training.

[0016] Advantageously, the current process data can include sensor data acquired directly during the process using a sensor device. Additionally, the process data can include application data, which, for example, can be pre-known and provided accordingly. This allows the prediction of the connection quality to be generated simultaneously with the connection being made or, for example, immediately afterward. This enables a manufactured component to be sorted out or reworked directly after production if the trained quality predictor model indicates low or insufficient connection quality. No direct analysis of the manufactured component is required to assess the quality. Instead, the quality can be evaluated solely based on process data related to the current manufacturing process.Advantageously, a quality-predicting model trained according to a previously mentioned procedure can be used.

[0017] During the data acquisition step, application data can be imported along with the current process data. This data characterizes the properties of the manufacturing device, the reactive multilayer system, and / or the components, as previously described. In this way, the quality of the connection can be assessed based on the current properties of the manufacturing device used and the stack to be joined.

[0018] Additionally or alternatively, sensor data can be read in alongside the current process data, representing values ​​sensed before, during, and / or after the reaction of the reactive multilayer system. This allows a sensor-monitored reaction progression to be incorporated into the quality assessment.

[0019] The sensor data can be acquired using a sensor device. This allows current sensor data to be provided for each connection.

[0020] In the sensing step, for example, the position of at least one of the components, the movement of at least one of the components, the acceleration of at least one of the components, a force acting on at least one of the components, a temperature, an ignition voltage to ignite the reactive multilayer system, a current to ignite the reactive multilayer system, and / or a vibration of at least one of the components can be sensed.

[0021] The process can include a step of creating the connection between the two components using the reaction of the reactive multilayer system. A suitable manufacturing device can be used for this purpose. The manufacturing step can be repeated to produce a plurality of components in succession. For each component produced, a prediction of the connection quality can be generated. Thus, the described approach can be advantageously used in the mass production of components.

[0022] The approach presented here further creates a device designed to perform, control, and implement the steps of a variant of the method presented here in appropriate facilities. This embodiment of the invention in the form of a device also allows the problem underlying the invention to be solved quickly and efficiently.

[0023] A device can be an electrical appliance that processes electrical signals, such as sensor signals, and outputs control signals accordingly. The device can have one or more suitable interfaces, which can be implemented in hardware and / or software. In the case of a hardware implementation, the interfaces can, for example, be part of an integrated circuit in which the device's functions are implemented. The interfaces can also be separate integrated circuits or consist at least partially of discrete components. In the case of a software implementation, the interfaces can be software modules that are present, for example, on a microcontroller alongside other software modules.

[0024] It is also advantageous to have a computer program product with program code that can be stored on a machine-readable medium such as semiconductor memory, hard disk memory or optical memory and is used to carry out the method according to one of the embodiments described above when the program is executed on a computer or device.

[0025] The invention is explained in more detail by way of example with reference to the accompanying drawings. These show: Fig. 1. A representation of an exemplary embodiment of a manufacturing device; Fig. 2 a schematic representation of the sequence of an exemplary embodiment of a reaction of a reactive multilayer system; Fig. 3a to 3e Representations of an exemplary embodiment of a reaction process of a reactive multilayer system; Fig. 4 sectional views of an exemplary embodiment of a low-quality connection; Fig. 5 sectional views of an exemplary embodiment of a high-quality connection; Fig. 6. A representation of an exemplary embodiment of a method for optimizing an encoder model; Fig. 7. A representation of an exemplary embodiment of a method for training a quality-predicting model; Fig. 8 a flowchart of an exemplary implementation of a method for optimizing an encoder model; Fig. 9 a flowchart of an exemplary implementation of a method for training a quality-predicting model; and Fig. 10 a flowchart of an exemplary implementation of a method for monitoring a process.

[0026] In the following description of preferred embodiments of the present invention, the same or similar reference numerals are used for the elements shown in the various figures and having a similar effect, without repeating these elements.

[0027] Fig. Figure 1 shows an embodiment of a manufacturing device 100 for joining a stack 102 of a first component 104, a second component 106 and a reactive multilayer system 108 arranged between the first component 104 and the second component 106 to form a component.

[0028] The device 100 comprises an activation device 110 configured to activate the reactive multilayer system 108 in order to establish a connection between the first component 104 and the second component 106. The activation device 110 is also referred to as an ignition device.

[0029] The activation device 110 is designed, for example, to apply a spark to the reactive multilayer system 108 or to direct a laser beam at the reactive multilayer system 108. This allows sufficient energy to be supplied to reach an ignition temperature required to activate the reactive multilayer system 108.

[0030] Optionally, the device 100 is designed to subject the stack of the first component 104, the second component 106 and the reactive multilayer system 108 to pressure 120 during the reaction of the reactive multilayer system 108 in order to increase the quality of the compound to be produced by the reactive multilayer system 108.

[0031] As an example, a pressure device 111 is used, comprising a first mass 112, also referred to as m1, and a movable second mass 114, also referred to as m2. The masses 112 and 114 are coupled to each other via a spring-damper mechanism with a damper 122 and a spring 124. Pressure is exerted on the stack via the movable mass 114, specifically on the side of the first component 104 facing away from the reactive multilayer system 108. According to one embodiment, the pressure device 111 is a spring plunger, whose first mass 112 can perform a movement y1 and whose second mass 114 can perform a movement y2.

[0032] According to one embodiment, the manufacturing device 100 comprises a sensor device 130 configured to acquire and provide sensor data 132. The sensor data 132, for example, depict process variables prevailing during the manufacture of the connection between components 104 and 106. The sensor device 130 is configured, for example, to acquire data on the temporal profile of various process variables, such as pressure 120, the temperature of the stack 102, or the occurrence of splashes 132. To acquire the sensor data 132, the sensor device 130 comprises one or more suitable sensors or measuring devices. The sensor data 132 may include, for example, several time series of sensor values ​​and / or process data, which can be used, for example, based on the Fig. The procedures described in sections 6 to 10 can be used.

[0033] If several components are manufactured using the manufacturing device 100, for example in series production, current sensor data 132 can be acquired and provided during each successive manufacturing process using the sensor device 130. Thus, sensor data 132 can be provided for each manufactured component and therefore for each connection produced.

[0034] According to one embodiment, a device 140 with an artificial intelligence-based quality-predicting model 142 is used to monitor a process for creating the connection between the two components 104, 106 by means of the reaction of the reactive multilayer system 108. Optionally, the device 140 is part of the manufacturing device 100.

[0035] According to one embodiment, the device 140 is configured to read in sensor data 132 provided by the sensor device 130 or stored sensor data 132 as current process data and, using the current process data, to create a dimensionally reduced feature vector as input data for the quality-predicting model 142. Using the quality-predicting model 142 fed with the input data, the device 140 is configured to generate a prediction of the quality of the compound produced by the reaction of the reactive multilayer system 108 and to provide it, for example, in the form of a quality signal 144. For example, the quality signal 144 indicates good or poor quality of a compound produced in a current process.According to one embodiment, the device 140 is configured to additionally or alternatively read in application data 146 as current process data and use it to create the dimensionally reduced feature vector. The application data 146 can, for example, also be acquired using the sensor device 130 or read out from a storage device.

[0036] Thus, the device 140 can, for example, be used to determine the value based on Fig. 10 described procedures for monitoring a process to establish a connection.

[0037] According to one embodiment, the device 140 or a corresponding device is additionally or alternatively used as a device for training the quality-predicting model 142. When the device 140 is used as a training device, it is configured, for example, to read in the sensor data 132 and additionally or alternatively the application data 146 as raw data and to create a dimensionally reduced feature vector from the raw data as input data for the quality-predicting model 142. The sensor data 132 can be currently sensed data or stored data.To train the quality-predicting model 142, additional quality data 148 are used, according to one embodiment. These data may have been determined in advance and may represent, for example, different quality characteristics of a connection between the components 104 and 106 created using the manufacturing device 100. The trained quality-predicting model 142 can then be used to monitor a process for creating a connection between the two components 104 and 106 using the reactive multilayer system.

[0038] Thus, the device 140 can be used, for example, to perform a calculation based on the Fig. 7 and Fig. 9 to implement the described procedure for training the quality-predicting model 142.

[0039] According to one embodiment, the device 140 or a corresponding device is additionally or alternatively used as a device for optimizing an encoder model with which the aforementioned dimensionally reduced feature vector can be created.

[0040] When the device 140 is used as a device for optimizing the encoder model, the device 140 is configured, according to one embodiment, for example, to read in the sensor data 132 and, additionally or alternatively, the application data 146 as raw data, wherein the sensor data 132 and the application data 146 can be currently sensed data or stored data. For this purpose, the device 140 comprises, for example, an encoder 150 and a decoder 152. The encoder 150 is configured to create a preliminary dimensionally reduced feature vector as input data for the quality-predicting model 142 using the raw data. The decoder 152 is configured to recover the raw data from the preliminary dimensionally reduced feature vector as far as possible.This process is repeated, according to one embodiment, until a deviation between the original raw data and the recovered decoded raw data meets a predetermined criterion, for example, falling below a threshold that represents a predetermined deviation. In this process, an encoder model used by the encoder 150 is modified for each repetition, according to one embodiment, to minimize the deviation. When the predetermined criterion is met, the underlying encoder model is made available for further use, for example, to provide a dimensionally reduced feature vector for a method to train the quality predictor model or to monitor a process for making a connection.

[0041] Thus, the device 140 can be used, for example, to perform a calculation based on the Fig. 6 and Fig. 8. To implement the described procedure for optimizing an encoder model.

[0042] According to one embodiment, the first component 104 is an electronic component, for example a semiconductor component, and the second component 106 is a support for the first component 104. For example, the second component 106 is a printed circuit board or a substrate, for example a copper substrate. For example, the first component 104 is a chip, or, by way of example, a transistor or a diode. According to one embodiment, at least one electrical connection of the first component 104 is permanently electrically and mechanically connected to an electrical contact of the second component 106 after ignition of the reactive multilayer system 108.

[0043] According to an alternative embodiment, the first component 104 is a sensor and the second component 106 is a machine element, for example, a shaft for an electric drive or a gearbox. For example, the first component 104 is a force sensor which, after ignition of the reactive multilayer system 108, is rigidly connected to the second component 106 in order to detect a deformation of the second component 106.

[0044] According to one embodiment, the reactive multilayer system 108 comprises a plurality of alternating nanoscale layers and is designed to react exothermically upon ignition, thereby establishing a metallurgical bond between the components 104 and 106. For example, the reactive multilayer system 108 is configured as a reactive multilayer film placed between the components 104 and 106. Alternatively, the layers of the reactive multilayer system 108 are grown, for example, on a surface of one of the components 104 and 106.

[0045] The ignition temperature required for the activation of the reactive multilayer system 108 is achieved, for example, using the activation device 110 by means of a laser beam.

[0046] A reactive multilayer film (RMS), which according to one embodiment can be used as a reactive multilayer system 108, is a film that provides instant heat for a wide variety of applications in many industries. This reactive multilayer film is produced by vapor deposition of thousands of alternating nanoscale layers, for example, of aluminum (Al) and nickel (Ni). Activation is triggered by a small local energy pulse, for example, from an electrical, optical, or thermal source. The reactive multilayer system 108 reacts exothermically and releases precise local heat up to 1500 °C in fractions of a second, for example, within a few thousandths of a second.

[0047] To provide a small pulse of local energy to trigger the activation of the reactive multilayer system 108, the following can be used, for example: By applying an electrical voltage directly to the reactive multilayer system 108.

[0048] Using a laser that is directed at the reactive multilayer system 108.

[0049] By applying a direct heat source to the reactive multilayer system 108.

[0050] The concepts for monitoring the reaction can of course also be applied to an RMS system in which the reactive multilayers of the reactive multilayer system 108 are grown directly on one of the components, for example chip or substrate, and thus no separate RMS film is used as the reactive multilayer system 108.

[0051] The pressure 120 required for a high-quality RMS bonding process, for example to press the chip, the RMS foil and the substrate together, is typically applied using spring plungers. These spring plungers have a moving mass, here the second mass 114, of over 100 g and can consist of several spring-damper mass systems.

[0052] However, the concept described here can also be transferred without the pressure device 111 or with a pressure device 111 in which the second mass 114 is pressed against the stack 102 in a different way.

[0053] The rapid exothermic reaction, e.g. ~7 m / s during the activation / bonding process, leads to a rapid expansion of the reactive multilayer system 108.

[0054] This rapid exothermic reaction, or the associated RMS system changes, can be monitored by sensors. However, the relationship between the measured RMS system changes and other input variables, such as film thickness, chip size, etc., is not trivial. Nevertheless, using the approach described here, conclusions can be drawn about the RMS bond quality.

[0055] The measurement data from sensor unit 130 can be used to determine correlations and draw conclusions about the quality of the RMS bond. For example, it can indirectly detect whether the bond is clean and free of voids, whether the bond has caused excessive spatter, whether the strength is sufficient, whether the service life is adequate, and much more.

[0056] This lack of information regarding quality has a positive impact on quality control, as defective components can be identified and thus sorted out or reworked. This prevents errors from being discovered later and resulting in costly rework or complaints.

[0057] AI-based quality prediction is advantageous. Since heuristic feature extraction can prove difficult to implement in multidimensional relationships, the approach presented here allows for the use of AI-supported feature extraction to manage and represent this complexity.

[0058] The approach described here, exemplified by the first component 104, for example a part, e.g., a bare die, which is bonded to the second component 106, for example a substrate, via an RMS film in the form of the reactive multilayer system 108, consists of sensorially monitoring the rapid exothermic reaction or related RMS system changes and using AI (artificial intelligence) to account for the multidimensional dependence of various process parameters on the RMS bond quality. This results in the following advantages: Highly complex relationships between diverse input variables can be abstracted by AI, subsequently enabling conclusions to be drawn about the quality of the RMS bond. Using this novel feature vector, a neural network can detect even more precisely whether the bond is clean and free of voids, whether the bond has caused excessive spatter, whether the strength is sufficient, whether the service life is adequate, and much more.

[0059] Quality control can be significantly improved, as defective components can be identified and therefore sorted out or reworked. Furthermore, costly subsequent errors are avoided and the number of complaints can be reduced.

[0060] Advantageously, iterative model optimization and adjustment of input variables are no longer necessary, which can reduce training time and effort. Nevertheless, it is still possible to implicitly determine which sensor parameters significantly influence the quality.

[0061] Fig. Figure 2 shows a schematic representation of the process of an embodiment of a reaction of a reactive multilayer system 108, such as that used for manufacturing a component. The reactive multilayer system 108 is, for example, composed of a plurality of first layers 270 and a plurality of second layers 272, which are arranged alternately, with optionally mixed regions 274 located between adjacent layers 270, 272.

[0062] A schematic representation of an activation position, exemplified by a spark 276, is shown. Starting from the activation position, the material of the reactive multi-layer system 108 reacts to form reacted material of a reacted multi-layer system 280. A corresponding direction of propagation 278 is indicated by an arrow.

[0063] According to one embodiment, the reactive multilayer system 108 is a reactive multilayer film. This film provides instant heat for a wide variety of applications in many industries. This reactive multilayer film is produced by vapor deposition of thousands of alternating nanoscale layers, which are in Fig. The layers are schematically represented by 270 and 272 and consist, for example, of aluminum (Al) and nickel (Ni). Activation is triggered by a small local energy pulse from an electrical source in the form of a capacitor. The reactive multi-layer system 108 reacts exothermically to generate precise local heat up to 1500 °C in fractions of a second.

[0064] In this way, a connection that is both electrically and thermally conductive can be established, e.g., between a bare chip in electronics and a leadframe, a packaged chip and a printed circuit board, etc. - i.e., generally a connection between two parts, which are referred to here by way of example as components.

[0065] Fig. Figure 3a shows an embodiment of a stack 102 comprising a first component 104, a second component 106, and a reactive multilayer system 108 arranged between the first component 104 and the second component 106. The stack 102 is exemplary according to the design shown in Figure 3a. Fig. 1 described stack executed.

[0066] To activate the reactive multilayer system 108, an activation device is used, as exemplified by the following: Fig. Figure 1 describes the following. A first electrical contact tip 341 is shown, which is brought close to the reactive multilayer system 108, and a second electrical contact tip 342, which contacts the reactive multilayer system 108. An ignition voltage is applied between the contact tips 341 and 342.

[0067] Fig. 3b shows the using Fig. 3a shown stack 102 at a time when an ignition spark caused by the ignition voltage jumps between the electrical contact tip 341 and the reactive multilayer system 108.

[0068] Fig. 3c shows the example of Fig. 3b shown stack 102 at a time when a reaction of the reactive multilayer system 108 begins due to the energy introduced into the reactive multilayer system 108 by the ignition spark.

[0069] Fig. 3D shows the based on Fig. 3c shown stack 102 at a time when the reaction has spread to the entire reactive multilayer system 108.

[0070] Fig. 3e shows a component 380, which is made from the one based on the Fig. The stack shown in Figures 3a to 3d is formed by the reaction of the reactive multilayer system to a reacted multilayer system 280. The quality of the metallurgical bond produced by the reacted multilayer system 280 between the components 104 and 106 can be determined using the aforementioned monitoring device.

[0071] Fig. Figure 4 shows sectional views of an embodiment of a connection between two components realized by a reacted multilayer system 280. The connection was produced, for example, using a manufacturing device such as that shown in Fig. 1 is described. The compound is of low quality, as many large voids 490 are present.

[0072] Depicted are in Fig. 4 four sectional views, starting at the top left and proceeding clockwise, showing sections top left, top right, bottom right and bottom left.

[0073] Fig. Figure 5 shows sectional views of an embodiment of a connection between two components realized by a reacted multilayer system 280. The connection was produced, for example, using a manufacturing device such as that shown in Figure 5. Fig. As described in section 1. The connection is of good quality, as there are few very small voids (590).

[0074] Depicted are in Fig. 5 four sectional views, starting at the top left and proceeding clockwise, showing sections top left, top right, bottom right and bottom left.

[0075] Fig. Figure 6 shows an embodiment of a method for optimizing an encoder model for an AI-based quality-predicting model for generating a prediction of the quality of a connection between two components produced by means of a reactive multilayer system, as exemplified by Fig. 1 is described.

[0076] To carry out this procedure as well as the ones described below, training data can be used that consists of sensor data plus application data, as exemplified by... Fig. As described in section 1. The training data can be available both during AI training and later application. Additionally, quality data is provided, which is collected specifically for training but is no longer available during the application of the AI ​​algorithm.

[0077] Typical sensor data include: position / motion measurements, acceleration measurements, force measurements, temperature measurements, ignition voltage and current measurements, vibration measurements, camera recordings, etc.

[0078] Typical application data include: component dimensions (such as chip size, RMS film thickness, substrate thickness), RMS film type, surface coating of the components, moving mass, spring constant of the spring plunger, etc.

[0079] A combination of several measurement principles and / or measured variables from several concepts for RMS tape quality monitoring is possible in order to obtain additional conclusions about the quality of the RMS bond.

[0080] Typical quality data include: shear forces / pressures from destructive shear tests, X-ray images including the percentage of voids (see Fig. 4a), electrical function tests, splashing behavior, lifetime tests, deviation from the optimal position, etc.

[0081] The algorithm is trained based on this data.

[0082] The following section will discuss in more detail the procedure for AI-based RMS bond quality monitoring with AI-based feature extraction.

[0083] The AI ​​(Artificial Intelligence)-based methods for RMS bond quality monitoring described here enable the detection of high-quality RMS bonds without potentially damaging components. An alternative feature determination and selection based on a library and heuristics is solved using a neural network, as described here. This allows for the simpler representation of more complex relationships between input variables and also accelerates the selection process.

[0084] For this purpose, raw data 630 are initially used, which characterize the physical properties of the process for establishing the connection. The raw data 630 represent, for example, time series of sensor values ​​or other process data. This original raw data 630 is encoded in an encoder 632 using an encoder model to obtain a dimensionally reduced feature vector 634, which is dimensionally reduced compared to the raw data 630. The dimensionally reduced feature vector 634 is then decoded in a decoder 636 using a decoder model to obtain the raw data 630 again, if possible. The encoding and decoding process is repeated until a deviation between the original raw data 630 and the raw data recovered from the dimensionally reduced feature vector 634 meets a predetermined criterion.The encoder model that generates this dimensionally reduced feature vector 634 is stored as a trained encoder model or provided or used for a subsequent procedure to train the quality-predicting model or to monitor a process. Suitable methods, including neural networks, can be used to train the encoder model.

[0085] According to one embodiment, the raw data 630 comprise a plurality of data S1, S2, S3 ..., which are converted into recovered data S1', S2', S3' ... using the encoder 632 and the downstream decoder 636, wherein the difference between the original data S1, S2, S3 and the recovered data Ŝ1', Ŝ2', Ŝ3' ... is formed as L(Ŝ1', S1), L(Ŝ2', S2), L(Ŝ3', S3).

[0086] The in Fig. The process for training the encoder 632, as shown in section 6, is described in detail below using an example implementation.

[0087] In a first step, an innovative method for obtaining the feature vector 634 is presented: the raw data 630, for example, raw, high-dimensional input and application data (here: meta-information), is dimensionally reduced using the encoder model of the encoder 632. Subsequently, the resulting feature vector 634 is mapped back to the original raw data, here input and application data, using a decoder model of the decoder 636. During training, this encoder-decoder process is repeated until the deviation, here the loss L, between the actual input data (S) and the input data (Ŝ) reconstructed by the architecture is minimal (min. L(Ŝ, S)). According to one embodiment, a variant of the Reconstruction Loss or Cycle Consistency Loss is chosen for more complex architectures.This initial training process is unsupervised, meaning it occurs without manually labeled data, and is therefore relatively inexpensive. This is particularly true because it eliminates the need for destructive labeling methods. This data is comparatively inexpensive and can be generated in large quantities.

[0088] The trained encoder 632 is then used in "frozen" mode for feature extraction to map the application data into a low-dimensional feature vector. This vector then serves as the input vector for the quality-predicting model θP.

[0089] Fig. Figure 7 shows an embodiment of a method for training a quality-predicting model 142, as implemented, for example, using Fig. 1 is described.

[0090] To train the quality-predicting model 142, the trained encoder 632 is used, which is configured according to the model defined by Fig. The approach described in section 6 was used for training. Furthermore, raw data 630 is used, which is based on... Fig. The raw data described in section 6 may correspond to or resemble each other.

[0091] The raw data 630, here exemplified by a plurality of data S1, S2, S3 ..., each representing time series, are first processed using the trained encoder 632 and thus using the trained encoder model in the following: Fig. The dimensionally reduced feature vector 634 described in section 6 is encoded and used to train model 142. For example, model 142 generates a quality signal 144, as described in section 6. Fig. As described in section 1. Furthermore, quality data, which is provided specifically for the training process, is used to train model 142. Once model 142 is sufficiently trained, it is stored as a trained quality predictor model 142, according to one embodiment, or made available or used for a subsequent process monitoring procedure.

[0092] The in Fig. The process shown in section 7 for training model 142 is described in detail below using an exemplary embodiment.

[0093] In this process, the raw data 630, here raw sensor data, are preprocessed by the encoder 632, and the quality data, here referred to as quality values, serve as labels (ỹ) for the optionally supervised training of the quality-predicting model 142 (θ). P ). Due to the comparatively small number of parameters in θ PAccordingly, only a small number of these generated quality values ​​are needed. During inference, quality values ​​can either be estimated as a regression problem (ŷ), or they can be abstracted so that the model, based on the input data, simply determines whether a sample meets the quality standards or is defective (classification problem).

[0094] This makes it possible to optimally map and abstract high-dimensional input data and dependencies so that the model 142 (θ P) can then be optimized and trained with a manageable input vector. This technical solution makes it possible to reduce a high-dimensional optimization problem to a low-dimensional one, in order to subsequently train and solve the low-dimensional problem in a data-driven manner, thus achieving optimal predictive accuracy. With this optimized prediction of bond quality based on the input variables, quality control can be significantly improved. Defective components can be detected, rejected, or reworked. Furthermore, costly consequential errors can be avoided, and the number of complaints can be reduced.

[0095] Fig. Figure 8 shows a flowchart of an exemplary implementation of a method for optimizing an encoder model. The method can be implemented, for example, using the method described in Fig. The 6 encoders and decoders shown will be executed.

[0096] The optimized encoder model can then be used as a trained encoder model for application with an artificial intelligence-based quality predicting model to generate a prediction of the quality of a connection between two components produced using a reactive multilayer system.

[0097] To obtain the optimized encoder model, raw data characterizing the physical properties of a process for creating the connection is read in step 801. For example, application data characterizing the properties of the manufacturing device, the reactive multilayer system, and / or the components is read in. Additionally or alternatively, sensor data representing values ​​sensed during a reaction of the reactive multilayer system is also read in.

[0098] Optionally, the procedure includes step 803 of acquiring the raw data using a suitable acquisition device, for example, a sensor device as described in Fig. 1 is described.

[0099] In step 805 of the training process, an encoder model is trained to encode the raw data into a dimensionally reduced feature vector. A corresponding decoder model is used to recover the raw data from the dimensionally reduced feature vector. According to one embodiment, step 805 is repeated until the encoder model is sufficiently trained that a deviation between the original raw data and the decoded raw data meets a predetermined criterion. For example, a deviation, also referred to as a loss, between the original raw data and the decoded raw data is determined and evaluated for each training run to assess whether the predetermined criterion has been met, for example, whether a predetermined deviation threshold has been reached or, for example, fallen below.For example, the encoder model and optionally the decoder model are trained until the deviation between the raw data and the decoded raw data is minimal.

[0100] After completion of the training process, the trained encoder model is optionally made available for further use in step 807.

[0101] For example, the trained encoder model is used to extract features in the statistical, temporal and / or spectral domain from the raw data input.

[0102] Fig. Figure 9 shows a flowchart of an exemplary implementation of a method for training a quality-predicting model. The method can be implemented, for example, using the model described in Fig. The 7 encoders and models shown will be executed.

[0103] To train the model, a trained encoder can be used, for example by executing the following based on... Fig. The model was created using the 8 described process steps. The trained model can then be used, for example, to predict the quality of a connection between two components created by a reaction of a reactive multilayer system.

[0104] To train the model, raw data is first read in step 901. This data characterizes the physical properties of a process for creating the connection, as well as quality data that characterizes the quality of the connection. For example, application data is read in that characterizes the properties of the manufacturing device, the reactive multilayer system, and / or the components. Additionally or alternatively, sensor data is read in that represents values ​​sensed during a reaction of the reactive multilayer system. Quality data that characterizes the quality of the connection is also read in.

[0105] Optionally, the procedure includes a step 903 in which the raw data are acquired using a suitable acquisition device, for example, using a sensor device such as those described in Fig. 1 is described.

[0106] In step 905, the raw data is encoded into a dimensionally reduced feature vector. The trained encoder model is used for this purpose. The dimensionally reduced feature vector is then used as input data for the quality-predicting model.

[0107] In step 907, the quality predicting model is trained using the input data and using the quality data to obtain the trained quality predicting model.

[0108] After completion of the training process, the trained quality-predicting model is optionally made available for further use in step 909.

[0109] Fig. Figure 10 shows a flowchart of an exemplary implementation of a method for monitoring a process, as it can be implemented, for example, using the following: Fig. 1. This can be done using a trained encoder and a trained quality-predicting model, as described, for example, in Fig. 7 is shown.

[0110] The method can be used to monitor a manufacturing process for creating a connection between two components using a reaction of a reactive multilayer system.

[0111] In one step, 1001 current process data are read in, characterizing the physical properties of the process for creating the compound. For example, application data is read in that characterizes the properties of the manufacturing device, the reactive multilayer system, and / or the components. Additionally or alternatively, sensor data is read in, representing values ​​sensed during a reaction of the reactive multilayer system.

[0112] Optionally, the procedure includes step 1003 of sensing the current process data using a suitable sensing device, for example a sensor device, as described in Fig. 1 is described. For example, the position of at least one of the components, the movement of at least one of the components, the acceleration of at least one of the components, a force acting on at least one of the components, a temperature, an ignition voltage for igniting the reactive multilayer system, a current for igniting the reactive multilayer system, and / or a vibration of at least one of the components is sensed.

[0113] In step 1005, the current process data is encoded into a dimensionally reduced feature vector. The trained encoder model is used for this purpose. The dimensionally reduced feature vector is then used as input data for the trained quality predictor model.

[0114] In step 1007, a prediction of the quality of the connection between the two components, created by the reaction of a reactive multilayer system, is generated. The trained quality-predicting model is used for this purpose.

[0115] Optionally, the process includes a step 1009 in which the connection between the two components is established by means of the reaction of the reactive multilayer system. For this purpose, a fabrication device such as that described in Fig. As described in section 1, steps 1001, 1003, 1005, 1007, and 1009 can be repeated for each manufacturing process of a component and thus of a joint. In this way, the quality of each manufactured joint can be assessed simultaneously with or after the respective manufacturing step 1009 using the trained quality predictor model. Reference sign 100 manufacturing device 102 stacks 104 first component 106 second component 108 reactive multilayer system 110 Activation device 111 Printing device 112 first mass 114 second mass 120 Print 122 dampers 124 springs 130 Sensor device 132 sensor data 134 splashes 140 Device 142 quality-predicting models 144 Quality signal 146 Application data 148 quality data 150 encoders 152 decoders 270 first layer 272 second layer 274 mixed regions 276 Ignition 278 Direction of propagation 280 reacted multilayer system 341 first electrical contact tip 342 second electrical contact tip 380 component 490 large voids 590 small voids 630 raw data 632 Encoders 634 Feature vector 636 decoders Step 801 of the reading process 803 Step of the capture process Step 805 of training Step 807 of deployment 901st step of the reading process 903 Step of the capture process 905th step of encoding Step 907 of training 909th step of deployment 1001 steps of the reading process 1003 Step of Sensing 1005th step of encoding 1007th step of generation 1009th step of the manufacturing process

Claims

[1] Method for optimizing an encoder model for an artificial intelligence-based quality predicting model (142) to generate a prediction of the quality of a connection between two components (104, 106) produced by means of a reactive multilayer system (108), wherein the method comprises the following steps: Reading (801) raw data (630) that characterize the physical properties of a process for making the connection; and Training (805) an encoder model to encode the raw data (630) into a dimensionally reduced feature vector (634) and a decoder model to decode the feature vector (634) into decoded raw data, in order to obtain a trained encoder model in which a deviation between the raw data (630) and the decoded raw data corresponds to a predetermined criterion. [2] Method according to claim 1, wherein in step (801) of reading in the raw data (630) application data (146) are read in which characterize the properties of a manufacturing device (100), the reactive multilayer system (108) and / or the components (104, 106), and / or sensor data (132) are read in which represent values ​​sensed during a reaction of the reactive multilayer system (108). [3] Method according to any of the preceding claims, wherein the encoder model is configured to extract features in the statistical and / or temporal and / or spectral domain from the raw data (630). [4] Method according to any of the preceding claims, wherein the training step (805) is repeated until the deviation between the raw data (630) and the decoded raw data is minimal. [5] Method for training an artificial intelligence-based quality-predicting model (142) to generate a prediction of the quality of a connection between two components (104, 106) produced by a reaction of a reactive multilayer system (108), wherein the method comprises the following steps: Reading (901) raw data (630) that characterize the physical properties of a process for making the connection and quality data (148) that characterize the quality of the connection; Encoding (905) the raw data (630) into a dimensionally reduced feature vector (634) using a trained encoder model representing an encoder model optimized according to a method according to one of the preceding claims to obtain input data for the quality predicting model (142); and Training (907) the quality predicting model (142) using the input data and the quality data (148) to obtain a trained quality predicting model (142). [6] Method for monitoring a process for creating a connection between two components (104, 106) by means of a reaction of a reactive multilayer system (108), the method comprising the following steps: Reading (1001) current process data that characterize the physical properties of the process for making the connection; Encoding (1005) the process data into a dimensionally reduced feature vector (634) using a trained encoder model representing an encoder model optimized according to a method according to one of the preceding claims to obtain input data for the quality predicting model (142); and Generating (1007) a prediction of the quality of the connection produced by the reaction of a reactive multilayer system (108) between the two components (104, 106) using a trained quality predictor model (142) that represents a quality predictor model (142) trained using a method according to claim 5. [7] Method according to claim 6, wherein in step (1001) of reading in application data (146) which characterize the properties of the manufacturing device (100), the reactive multilayer system (108) and / or the components (104, 106) are read in as the current process data, and / or sensor data (132) which represent values ​​sensed during the reaction of the reactive multilayer system (108) are read in as the current process data. [8] Method according to claim 7, comprising a step (1003) of sensing the sensor data (132) using a sensor device (130). [9] Method according to claim 8, wherein in step (1003) of sensing a position of at least one of the components (104, 106), a movement of at least one of the components (104, 106), an acceleration of at least one of the components (104, 106), a force acting on at least one of the components (104, 106), a temperature, an ignition voltage for igniting the reactive multilayer system (108), a current for igniting the reactive multilayer system (108), and / or a vibration of at least one of the components (104, 106) is sensed. [10] Method according to claim one of claims 6 to 9, comprising a step (1009) of making the connection between the two components (104, 106) by means of the reaction of the reactive multilayer system (108). [11] Device which is set up to perform and / or control the steps of the method according to any of the preceding claims in corresponding units. [12] Computer program configured to execute and / or control the steps of a method according to any one of claims 1 to 10.

Images