Method for detecting a faulty connection between electronic components during robot-guided assembly

Abstract

The invention relates to a method for detecting a faulty connection between electronic components during robot-guided assembly, wherein the method comprises receiving (S100) first process data (101) from a robot performing the connection and second process data (102) from the electronic components, and providing (S200) the first and second process data (101, 102) to a trained machine learning model (201), wherein the trained machine learning model (201) is trained to recognize a connection state of the electronic components. The method further comprises determining (S300) the connection state of the electronic components by analyzing the first and second process data (101, 102) using the trained machine learning model (201), and verifying (S700) the connection as a function of the determined connection state.The invention further relates to a method for training a machine learning model for the inventive method for detecting the faulty connection. The invention further relates to a system for training the inventive machine learning model for the inventive method for detecting the faulty connection, as well as a system for detecting a faulty connection comprising the inventively trained machine learning model (201).

Description

[0001] The invention relates to a method for detecting a faulty connection between electronic components during robot-guided assembly and a method for training a machine learning model for the inventive method for detecting the faulty connection between the electronic components during robot-guided assembly. The invention further relates to a system for training the inventive machine learning model for the inventive method for detecting the faulty connection between the electronic components, as well as a system for detecting a faulty connection between electronic components during robot-guided assembly comprising the inventively trained machine learning model for the inventive method for detecting the faulty connection between the electronic components during robot-guided assembly.

[0002] When assembling complex modules containing numerous electronic components, a variety of connectors are used to link these components. The connectors are plugged into each other and usually mechanically locked. However, partial insertion often occurs, resulting in unreliable electrical contact between the connectors.

[0003] Gloves equipped with sensors are known that can monitor connections during manual assembly, i.e., assembly carried out by a person.

[0004] WO 2021 / 211 547 A1 describes a system for error detection and correction in robotic assembly processes using sensors. The sensors monitor the assembly process, detect deviations, and a fusion controller then creates a recovery plan based on this information, such as resetting a component and restarting the process while avoiding the original error. Historical data supports the error analysis. Therefore, this method is not robust and depends on a large amount of historical data.

[0005] The invention is based on the objective of providing a method that at least partially overcomes the aforementioned disadvantages of the prior art and makes it possible to reliably and safely detect a faulty connection between electronic components during robot-guided assembly.

[0006] The problem according to the invention is solved by a method for detecting a faulty connection between electronic components during robot-guided assembly. Preferred embodiments are the subject of the respective dependent claims.

[0007] A first aspect of the invention relates to a method for detecting a faulty connection between electronic components during robot-guided assembly. The method comprises receiving first process data from a robot performing the connection and second process data from the electronic components, and providing the first and second process data to a trained machine learning model. The trained machine learning model is trained to recognize the connection state of the electronic components. The method further comprises determining the connection state of the electronic components by analyzing the first and second process data using the trained machine learning model and verifying the connection based on the determined connection state.

[0008] Receiving first and second process data preferably involves directly reading the process data, for example from appropriately designed sensors, or receiving the process data from a control unit or a server. This includes, among other things, sensors for continuous data acquisition from the robot performing the plugging operation, the electronic components, and the processes involved.

[0009] The first and second process data preferably include first and second real-time sensor data, first and second physical state variables and / or first and second process parameters.

[0010] The first and second process data preferably correlate with each other in time, meaning they contain information from the same points in time. This allows for particularly accurate inferences, as there is a high degree of comparability and causality between the process data.

[0011] The initial process data preferably includes the position of a robot performing the assembly in space, its travel speed, its acceleration, but also physical state variables including motor currents and torques of the robot, temperatures such as a working environment such as ambient air or an operating temperature of the robot, humidity of the working environment such as ambient air, and / or information on wear or age and required maintenance of the robot.

[0012] The second process data preferably includes information on the electronic components relevant for the plugging process, such as their respective position and orientation in space as well as their relative position to each other and to the robot, their installation location, their complexity, for example their respective geometric dimensions and shapes and interaction with each other, but also information on the plug itself.

[0013] Determining the connection state, in other words, involves providing the trained model, for example, by loading the trained model for machine learning. Checking the connection, in other words, involves determining whether the connection is faulty or faultless, depending on the determined connection state. For example, if an incomplete connection state is determined, a faulty connection is identified, and if a complete connection state is determined, a faultless connection is identified.

[0014] A plug connection is also called a connector. A plug connection is a mechanical and electrical connection created by connecting at least two electronic components. It serves to transmit electrical signals or energy and simultaneously fulfills mechanical purposes, such as fixing or connecting components. Electrical connectors include, for example, sockets and plugs for power supplies, cable connectors such as USB, HDMI, or RJ45, and socket-plug systems often used in electronic devices, such as jack plugs or Molex connectors. Connectors typically consist of a plug, the male electronic component with protruding contacts, and a socket, the female electronic component into which the plug is inserted.In other words, electronic components comprise at least one male and one female electronic component. Contact materials such as copper or gold-plated surfaces ensure conductivity and corrosion protection, while insulating materials such as plastic prevent short circuits. Connectors are easy to use, as they can often be assembled or disassembled without tools, and they are suitable for repeated disconnection and reconnection. They are widely used in fields such as electronics, automotive, construction, and household appliances. Well-known examples include USB ports, Europlugs or Schuko plugs for electricity, jack plugs for audio, and pneumatic connectors in compressed air systems. Their versatility and reliability make them an essential technology.

[0015] The electronic components are preferably part of a complex assembly. The complex assembly is preferably a vehicle. The electronic components are preferably wiring harnesses and control units that are connected to each other by plug connectors. The electronic components are also preferably electronic consumers of the vehicle.

[0016] The electronic components preferably comprise at least one connecting element, such that a plug connection exists between electronic components, for example, between a cable harness and a connecting element, or between the connecting element and a control unit. A plug connection between two cable harnesses, between a cable harness and an electronic device, or between an electronic device and a control unit is also conceivable, with a connecting element optionally being involved as a third electronic component in each case.

[0017] One of the electronic components involved in the connection process is therefore preferably a cable harness, also called a wiring bundle. A cable harness is an organized arrangement of cables, wires, and connectors, bundled together and organized by sheathing, cable ties, or protective sleeves. It serves to make electrical connections within a system efficient, safe, and clearly arranged. Cable harnesses bundle multiple electrical cables or wires into an organized bundle, thus reducing confusion and disorganization. At the same time, the sheathing protects the wires from mechanical damage, abrasion, moisture, or high temperatures. Furthermore, a cable harness connects various electrical or electronic components and ensures the transmission of power or signals. Since each system has different requirements, cable harnesses are usually custom-made.A wiring harness consists of wires or cables that carry electrical energy or signals, connectors or sockets for connecting to devices or other wiring harnesses, and protective materials such as insulating tape, heat-shrink tubing, or braided sleeving. Cable ties or clips are used to secure and organize the wires. Wiring harnesses have a wide range of applications. In the automotive industry, for example, they connect electrical systems such as engine control units, lighting, sensors, airbags, and infotainment systems. In aerospace, they ensure reliable connections where space and weight are critical. In electronics and household appliances, they bundle connections between components, for example, in washing machines or computers, while in mechanical engineering, they connect motors, controllers, and sensors.Cable harnesses offer numerous advantages: They save space by minimizing loose cables, are clearly arranged, which facilitates maintenance and repair, are safe because they reduce the risk of short circuits or cable breaks, and are reliable because they protect the wires and extend their service life. They are therefore a key component of many electrical systems and essential for modern technologies.

[0018] A faulty connection includes connections such as partial and incorrect connections, which are faulty because they are not fully mechanically and / or electrically connected. A connection is therefore characterized by the completeness of the connection between the components involved. A faulty connection thus encompasses a mechanical and electrical connection that deviates from a complete connection. This could, for example, be a geometric deviation. The determined connection status contains this information. In the case of an electrically or mechanically faulty connection, this can impair assembly, durability, and functionality.

[0019] The method according to the invention enables the precise control and sustainable optimization of manufacturing and assembly processes. In particular, robot-guided assembly can be improved by reliably detecting and correcting assembly errors. By using the trained machine learning model, it is possible to assess whether a faulty insertion has occurred during each successive insertion. The trained machine learning model efficiently analyzes large amounts of data and recognizes patterns and correlations that would be difficult to identify using traditional methods. This enables well-founded decisions and new insights into complex processes. It allows for the optimization of processes, increased efficiency, and reduced costs by utilizing historical and real-time data. The method enables real-time responses, reduces uncertainties, and precisely identifies, for example,Errors, anomalies, or trends. In manufacturing technology, the trained machine learning model benefits from precise sensors and repeatable processes, which facilitate the detection and learning of incorrect insertions. The generated production-relevant data is available as extensive datasets. These extensive datasets can be systematically collected, processed, and utilized. This allows problems to be detected early and production errors and scrap to be minimized. The inventive method enables the reliable monitoring of machine and plant processes through the systematic collection, processing, and analysis of data.

[0020] Preferably, the method further comprises initiating a correction process to adjust the connection and / or an adjustment of process parameters for controlling the robot performing the connection, based on the verified connection. This is preferably done if the verification has revealed or indicated a faulty connection. This has the advantage that, based on the result of the verified connection, suitable countermeasures can be taken if a faulty connection is present, either to improve the connection of the electronic components or at least to improve future connection processes. Preferably, the correction process or the adjustment of the process parameters includes controlling and / or regulating the robot performing the assembly, depending on the verified connection or the result of the connection verification.

[0021] In a preferred embodiment of the invention, the trained machine learning model is based on supervised learning, wherein the analysis of the first and second process data includes assigning one of at least two predefined classes to each data point of the first and second process data, and wherein the plugging state is determined depending on the assigned class. In other words, the analysis of the first and second process data comprises applying the trained model to the first and second process data to assign one of at least two predefined classes to each data point of the first and second process data. The plugging state is then determined depending on the assigned class. In other words, the trained machine learning model was trained using supervised learning by employing appropriate algorithms.Preferably, the trained machine learning model was trained using labeled first training process data and labeled second training process data, wherein the training process data each comprise at least two predefined classes, or wherein the first and second training process data are each labeled. Preferably, one of the at least two predefined classes is a classification as an IO (OK) connection, which in other words denotes an error-free connection. The other of the at least two predefined classes is preferably a classification as a faulty connection, which in other words denotes an incorrect connection. Supervised learning offers particular advantages over other machine learning methods in connection with the detection of faulty connections.Supervised learning is particularly suitable because the abundance of labeled data and a clear definition of a goal allow for especially precise analyses, and it is easy to implement. The goal of error-free placement is clearly definable and verifiable, and the large number of necessary placements also provides a wealth of data.

[0022] Preferably, the trained machine learning model is based on reinforcement learning, and the process further includes receiving feedback by capturing feedback from the environment that represents a reward or success of the determined state of being stuck, and passing the received feedback to the trained machine learning model for adaptation. Preferably, in the case of reinforcement learning, the trained machine learning model was trained using reinforcement learning, whereby the model learned an optimal policy that maximizes rewards in a dynamic environment. The optimal policy is a central component of reinforcement learning and describes the best strategy that an agent can use to achieve the maximum cumulative reward in the long run in an environment.A policy is a function that specifies which action an agent should perform in a given state. The optimal policy maximizes the expected sum of all future rewards the agent can achieve through its decisions. The optimal policy is based on a value function that describes the expected total value of the reward if the agent starts in a given state and follows a specific policy. An optimal value function indicates the maximum reward value that can be achieved from a given state, while the optimal policy selects, for each state, the action that maximizes this value. This means that the optimal policy considers not only the immediate reward of an action but also its long-term impact on overall success. Various methods are used to calculate the optimal policy.This includes dynamic programming approaches such as policy iteration and value iteration, model-free methods like Q-learning, and modern approaches like deep reinforcement learning, which uses neural networks to learn optimal strategies in high-dimensional state spaces. The optimal policy is thus a strategy in reinforcement learning that enables the agent to make the best decisions in dynamic and complex environments and maximize long-term success. Adaptation, and therefore receiving and providing feedback, preferably occurs continuously or at regular intervals. Reinforcement learning offers an advantage for dynamic and interactive scenarios because it learns from feedback in real time and optimizes long-term processes. It does not require labeled data but is based on rewards or punishments, which makes it particularly valuable in complex environments such as robot control.The accuracy and reliability of connectors and their manufacturing are enhanced through continuous data learning, thereby minimizing human error and improving system performance. In other words, the method further includes adaptive connector state determination by adjusting the detection method for subsequent or future faulty connector detections based on received and transmitted feedback to account for dynamic changes in the environment, the robot, and / or the electronic components. This adaptation of the connector state determination is based on the adapted trained model. Dynamic changes preferably include varying production conditions such as temperature, humidity, and other process data parameters previously described.In other words, based on the feedback, the trained model is adapted and continuously trained, and the adapted model is used in further recurring iterations of the process, resulting in continuous optimization.

[0023] Gathering feedback from the environment preferably involves a manual and / or electronic check of the connection. In other words, the feedback comprises the result of this check. This result can be a confirmation of the determined connection state or a rejection / negation of the determined connection state. A reward corresponds to the confirmation. In other words, reinforcement learning is based on a manual and / or electronic check of the connection and on passing a reward, based on the manual and / or electronic check and being positive or negative, to the trained machine learning model.

[0024] The trained machine learning model was particularly favored for supervised learning and subsequently for reinforcement learning. In this way, the machine learning model initially learns through observation (as in supervised learning) and later improves its strategy through experimentation (as in reinforcement learning).

[0025] In a further preferred embodiment of the invention, the trained machine learning model comprises a neural network based on supervised and / or reinforcement learning. In the case of supervised learning, the neural network is trained with labeled first and labeled second training process data to distinguish at least two predefined classes, and / or in the case of reinforcement learning, it is trained to maximize rewards. In other words, in the case of reinforcement learning, the trained model is trained to learn an optimal policy that maximizes rewards. Neural networks based on supervised and / or reinforcement learning approaches offer numerous advantages that make them particularly suitable for the detection of errors.Supervised learning enables precise predictions and classifications from large, labeled datasets, while reinforcement learning operates in dynamic environments and optimizes long-term strategies through continuous feedback. This combination creates diverse application possibilities in technical fields characterized by large datasets, complex decision-making processes, or dynamic requirements. Such approaches are particularly well-suited for detecting malfunctions in automotive engineering, especially in robotics (for precise control and navigation). The neural network's ability to recognize precise patterns and adapt flexibly to new conditions enables reliable detection.

[0026] Preferably, analyzing the first and second process data includes feature extraction and processing by analyzing and transforming the first and second process data to pass them to the neural network as a feature vector. The analysis of the first and second process data preferably includes applying the neural network by processing the first and second process data to generate a classification into a predefined class in the case of supervised learning, and / or to determine the state of the process based on the dynamic environment in the case of reinforcement learning.

[0027] According to a further embodiment of the invention, the trained machine learning model is based on anomaly detection, wherein the analysis of the first and second process data includes assigning a category to each data point of the first and second process data, and wherein the plugging state is determined as a function of the assigned category. In other words, the analysis of the first and second process data includes applying the trained model to the first and second process data to assign a category to each data point of the first and second process data. In other words, the trained machine learning model was trained using anomaly detection algorithms. One category of anomaly detection is preferably a misplugging. In other words, a misplugging is considered an anomaly.A data point is therefore preferably assigned a category if it fulfills a specific characteristic that the other data points do not exhibit. This is preferably a characteristic that the data points of the process data exhibit under error-free or, alternatively, error-prone conditions, thus enabling differentiation and detection. In other words, analyzing the first and second process data involves applying the trained model to each data point of the first and second process data to calculate an anomaly score for the data point, which expresses the probability that the data point deviates from normal patterns. Preferably, each data point is classified as "normal" or "anomalous" based on a predefined threshold for the anomaly score. The anomaly score is a numerical value that indicates how much a data point deviates from expected patterns or a normal data distribution.It is used in anomaly detection to distinguish normal data from anomalies or outliers. The score is calculated by the machine learning-based model and serves as a measure of a data point's deviation. Points with high anomaly scores are flagged as potential anomalies, while points with low scores are considered normal. A predefined threshold enables this differentiation. The state of the data point is thus determined based on its categorization. Anomaly detection essentially means identifying irregularities. In other words, anomalies differ not only from the general majority of the data but also from the normal state. These deviations are identified through anomaly detection. The primary goal of detecting incorrect data points is to identify potential errors, damage, and operational failures at an early stage.The process data is usually continuously recorded by sensors and must be evaluated promptly. The continuous stream of data and the potential for mutual interference between machines make it difficult for humans to quickly and reliably recognize complex relationships. The method according to the invention avoids this disadvantage.

[0028] The trained machine learning model is further preferably trained to use one or more predetermined categories to perform the analysis of the first and second process data for categorization or classification. One of the several categories is preferably a normal connection, i.e., an error-free connection, and another of the several categories is preferably an anomalous connection, i.e., a faulty connection. In other words, determining the connection state further includes categorizing or classifying the first and second process data using one or more predetermined categories based on the analysis of the first and second process data.

[0029] Preferably, an isolation forest algorithm is used for anomaly detection. In other words, each data point of the first and second process data is analyzed by the isolation forest model to calculate the anomaly score, with data points that can be easily isolated being classified as potentially anomalous and data points that are more difficult to isolate being classified as normal.

[0030] In a further preferred embodiment of the invention, the trained machine learning model is furthermore based on explainable artificial intelligence (AI) and is trained with first and second training process data, which may be labeled or unlabeled, and includes explainability mechanisms to provide at least a basis for determining the plug state. In other words, the trained machine learning model was trained with labeled or unlabeled first and labeled or unlabeled second training process data and includes explainability mechanisms to make the decision-making basis of the trained model comprehensible to humans.In other words, determining the plug state further includes generating explanatory information by analyzing the internal decision-making processes of the trained machine learning model for the determined plug state using an explainable AI method, wherein the explanatory information provides a basis for determination. The method preferably also includes providing the generated explanatory information by outputting it.

[0031] Preferably, the method further comprises determining the cause of the faulty connection between the electronic components, depending on the basis for the investigation. In other words, the explainable artificial intelligence includes at least a root cause analysis. Preferably, determining the cause of the faulty connection includes identifying the first and second process data and / or a type of first and / or second process data responsible for the faulty connection. This has the advantage that the basis for the investigation or the generated explanatory information is highly likely to represent the cause of the problem. Once the cause is known, it can be used in the context of a continuous improvement process to manually, automatically, or semi-automatically adjust a robot control system to prevent future faulty connections.At the same time, error feedback extending into production planning and product development enables sustainable error correction.

[0032] A second aspect of the invention relates to a method for training a machine learning model for the inventive method for detecting faulty connections between electronic components during robot-guided assembly, wherein the machine learning model is trained to recognize the connection state of the electronic components. The method comprises providing a training dataset containing first training process data of the robot performing the connection and second training process data of the electronic components, as well as the respective connection states of the electronic components, and receiving the training dataset. Furthermore, optionally, data preprocessing is performed, preferably comprising at least normalizing, transforming, and / or cleaning the training dataset to optimize it for use in machine learning.The method comprises defining a machine learning model, preferably by selecting or constructing the machine learning model with a predefined architecture, and initializing model parameters by setting initial values ​​for the model parameters of the machine learning model. Furthermore, the method comprises training the machine learning model by adjusting the model parameters, preferably through an iterative optimization process based on updating the model parameters using an optimization algorithm depending on a deviation between predictions of the machine learning model and target values, and validating the machine learning model, preferably by testing the machine learning model on a validation dataset.Furthermore, the training is terminated, preferably when a predetermined termination condition, preferably a minimum error rate or a maximum number of iterations, is met, and the machine learning model is defined or output as a trained machine learning model. The procedure also includes storing the adapted model parameters of the trained machine learning model in a computer-readable medium for later use in the procedure.

[0033] The training process data is similar in nature to the process data, so the preceding statements apply accordingly to the training process data.

[0034] The inventive method enables the generation and provision of a trained machine learning model for the inventive method for detecting faulty connections between electronic components during robot-guided assembly, thus allowing for precise control and sustainable optimization of the corresponding assembly processes. In particular, robot-guided assembly can be improved by reliably detecting and correcting assembly errors. By using the trained machine learning model, it is possible to evaluate whether a faulty connection has occurred during each successive connection. By training the model using the training process data and associated connection states, a control unit or system is able to assess whether a faulty connection has occurred during each successive connection.

[0035] Preferably, the machine learning model includes at least one or more neural network layers, decision trees, or other model-based algorithms.

[0036] A third aspect of the invention relates to a system for training a machine learning model for the inventive method for detecting faulty connections between electronic components, comprising a processor and a memory, the memory comprising instructions which, when executed by the processor, cause the control unit to execute the inventive method for training a machine learning model. The advantages achieved with the method can be achieved analogously with the system. The disclosed combinations of features of the method are analogously transferable to the system. Therefore, a repetitive description of the features and advantages is omitted.

[0037] A fourth aspect of the invention relates to a system for detecting faulty connections between electronic components during robot-guided assembly. This system comprises the machine learning model trained according to the invention for the inventive method of detecting faulty connections between electronic components, and includes a processor and a memory containing instructions which, when executed by the processor, cause the control unit to execute the inventive method for detecting faulty connections between electronic components. The advantages achieved with the method can be achieved analogously with the system. The disclosed combinations of features of the method are analogously transferable to the system. Therefore, a repetitive description of the features and advantages is omitted.

[0038] Further preferred embodiments of the invention result from the other features mentioned in the dependent claims.

[0039] Unless otherwise stated in individual cases, the various embodiments of the invention mentioned in this application can be advantageously combined with one another.

[0040] The invention is explained below using exemplary embodiments with reference to the accompanying drawings. These show: Fig. 1 a schematic representation of a method according to the invention for detecting a faulty connection between electronic components during robot-guided assembly according to a first embodiment, Fig. 2 a schematic representation of a method according to the invention for detecting a faulty connection between electronic components during robot-guided assembly according to a second embodiment, Fig. 3 a schematic representation of a method according to the invention for detecting a faulty connection between electronic components during robot-guided assembly according to a third embodiment, Fig. 4 a schematic representation of a method according to the invention for training a machine learning model for the method according to the invention for detecting the faulty connection between the electronic components during robot-guided assembly.

[0041] Fig. Figure 1 shows a schematic representation of a method according to the invention for detecting a faulty connection between electronic components during robot-guided assembly according to a first embodiment.

[0042] In a first process step S100, initial process data 101 from a robot performing the mating operation and second process data 102 from the electronic components are received. The reception of the initial and second process data 101, 102 preferably includes direct reading of the process data 101, 102, for example from appropriately designed sensors, or receiving the process data 101, 102 from a control unit or a server. This includes, among other things, sensors for continuous data acquisition from the robot performing the mating operation, the electronic components, and the processes occurring during this process.The first process data 101 includes, for example, the position of a robot performing the assembly in space, its travel speed, its acceleration, but also physical state variables including motor currents and torques of the robot, temperatures such as a working environment or the robot's operating temperature, humidity of the working environment, and / or information on wear and required maintenance of the robot. The second process data 102 preferably includes information on the electronic components relevant for the connection, such as their position and orientation in space, their relative position to each other and to the robot, their installation location, their complexity, for example, their respective geometric dimensions and shapes and interaction with each other, but also information on the connector itself.The electronic components in question are those required and used during the production of vehicles, such as cars. These electronic components include, for example, wiring harnesses and control units that are connected to each other via plug connectors.

[0043] In process step S200, the first and second process data 101, 102 are provided to a trained machine learning model 201. The trained machine learning model 201 is trained to recognize the plug-in state of the electronic components.

[0044] In process step S300, the mating state of the electronic components is determined by analyzing the first and second process data 101, 102 using the trained machine learning model 201. A faulty mating includes matings such as partial and incorrect matings, which are faulty because they are not fully connected mechanically and / or electrically. Accordingly, a mating is characterized by the completeness of the connection of the components involved. A faulty mating therefore includes a mechanical and electrical connection that deviates from a complete mating. The mating state is derived from the first and second process data 101, 102 through analysis using the trained machine learning model 201.

[0045] Furthermore, in process step S700, the connection is checked based on the determined connection state. The determined connection state contains information about the connection between the electrical components. The trained machine learning model 201 thus provides this information based on the determined connection state, making it possible to determine whether a faulty connection exists or not.

[0046] Optionally, the trained machine learning model 201 can also be based on supervised learning, and the analysis of the first and second process data in step S300 in this case comprises applying the trained model to the first and second process data 101, 102 such that each data point of the first and second process data 101, 102 is assigned one of at least two predefined classes. The mating state is then determined depending on the assigned class. One of the at least two predefined classes is, for example, a classification as an IO mating (OK mating), which in other words denotes a mating without errors. Another of the at least two predefined classes is preferably a classification as a faulty mating, which in other words denotes a faulty mating.Accordingly, the data consists of first process data with assigned predefined class 101' and second process data with assigned predefined class 102'.

[0047] Fig. Figure 2 shows a schematic representation of a method according to the invention for detecting a faulty connection between electronic components during robot-guided assembly according to a second embodiment. With the Fig. 1. Described components; corresponding elements are not explained again, and the corresponding description refers to these. Fig. 1 to apply.

[0048] In this variant, after determining the mating state of the electronic components by analyzing the first and second process data 101, 102 using the trained machine learning model 201 in step S300, feedback 401 is received in a process step S400 by capturing feedback from the environment, which represents a reward for the determined mating state. Capturing feedback from the environment includes, for example, a manual and / or electronic inspection of the mating. In other words, independently of the trained model 201, the mating is additionally inspected to determine whether it is faulty or fault-free. Feedback 401 then comprises, in other words, the result of this inspection. This result can then represent feedback 401 as confirmation of the determined mating state or as a rejection / negation of the determined mating state.In other words, it is analyzed whether the trained machine learning model 201 made a correct statement.

[0049] In a further process step S500, the received feedback 401 is then passed to the trained machine learning model 201 for adaptation of the trained model 201. If the analysis by the trained model 201 is correct, a reward is given.

[0050] Fig. Figure 3 shows a schematic representation of a method according to the invention for detecting a faulty connection between electronic components during robot-guided assembly according to a third embodiment. With the Fig. 1 or the Fig. The two components described, corresponding elements, are not explained again, and the corresponding description refers to them. Fig. 1 to apply.

[0051] In this variant, after determining the connection state of the electronic components by analyzing the first and second process data 101, 102 using the trained machine learning model 201 in step S300, a cause 601 of the faulty connection between the electronic components is determined in a process step S600. Accordingly, the trained machine learning model 201 is based on explainable artificial intelligence and was trained with first and second training process data 101, 102, which are labeled or unlabeled, and includes explainability mechanisms to provide at least a basis for determining the connection state.Accordingly, a root cause analysis is carried out here by identifying the cause 601 of the faulty connection by identifying the first, second process data 101, 102 and / or a type of first and / or second process data 101, 102 that are responsible for determining the faulty connection.

[0052] Fig. Figure 4 shows a schematic representation of a method according to the invention for training a machine learning model 201 for the method according to the invention for detecting the faulty connection between the electronic components during robot-guided assembly. In other words, an embodiment of a training method is shown here, in the execution of which a machine learning model 201 is then used for one of the components described in the Fig. 1, Fig. 2 to Fig. The 3 methods shown will be obtained.

[0053] In process step S110, a training data set 110 is provided, containing first training process data 111 of a robot performing the insertion and second training process data 112 of the electronic components, as well as the corresponding insertion states 113 of the electronic components. The training process data 111 and 112 are identical in nature to the process data 101 and 102, so the preceding statements apply accordingly to the training process data. However, here, the corresponding insertion states 113 are also present for the training process data; in other words, the result that should be obtained when a trained model is applied to new process data. This provides the method or model with a result during training, which it can use to learn how to analyze the process data. Furthermore, in process step S210, the training data set 110 is received.Furthermore, in process step S310, a machine learning model is defined. Then, in process step S410, model parameters are initialized by setting initial values ​​for the model parameters of the machine learning model. In process step S510, the machine learning model is trained by adjusting the model parameters. Subsequently, in process step S610, the machine learning model is validated. Furthermore, in process step S710, the training is completed and the machine learning model is defined as a trained model for machine learning 201. In process step S810, the adjusted model parameters of the trained model for machine learning 201 are stored in a computer-readable medium for later use in the process. Reference symbol list S100 Receiving first process data and receiving second process data S200 Providing the first and second process data for a trained model S300 Determining the plug-in status of electronic components S400 Receiving Feedback S500 Passing the received feedback to the trained model S600 Determining the cause of the faulty connection S700 Checking the connector S110 Providing a training dataset S210 S210 Receiving the training data set S310 Defining a machine learning model S410 Initializing model parameters S510 Training the machine learning model S610 Validating the machine learning model S710 Ending training and defining the machine learning model S810 Saving the adjusted model parameters 101 first process data 101' First process data with assigned predefined class 102 second process data 102' second process data with assigned predefined class 110 training data set 111 first training process data 112 second training process data 113 associated plug states 201 trained machine learning models 401 Feedback 601 Cause of the faulty connection

Claims

[1] Method for detecting a faulty connection between electronic components during robot-guided assembly, comprising the following steps: Receiving (S100) first process data (101) of a robot performing the insertion and second process data (102) of the electronic components; Providing (S200) the first and second process data (101, 102) to a trained machine learning model (201), wherein the trained machine learning model (201) is trained to detect a plug-in state of the electronic components; Determine (S300) the plug-in state of the electronic components by analyzing the first and second process data (101, 102) using the trained machine learning model (201); Check (S700) the connection depending on the determined connection status. [2] Method according to claim 1, wherein the trained machine learning model (201) is based on supervised learning, wherein the analysis of the first and second process data (102) comprises assigning one of at least two predefined classes to each data point of the first and second process data (102), and wherein the plugging state is determined depending on the assigned class. [3] Method according to claim 1 or 2, wherein the trained machine learning model (201) is based on reinforcement learning and wherein the method further comprises: Receiving (S400) feedback (401) by capturing feedback from the environment that represents a reward for the determined plugging state, and Passing (S500) the received feedback (401) to the trained machine learning model (201) to adapt the trained machine learning model (201). [4] Method according to any of the preceding claims, wherein the trained machine learning model (201) comprises a neural network based on supervised and / or reinforcement learning, wherein in the case of supervised learning the neural network was trained with labeled first and labeled second training process data to distinguish at least two predefined classes, and / or in the case of reinforcement learning was trained to maximize rewards of a determined stuck state. [5] Method according to one of the preceding claims, wherein the trained machine learning model (201) is based on anomaly detection, wherein the analysis of the first and second process data (101, 102) comprises assigning a category to each data point of the first and second process data (101, 102), and wherein the plugging state is determined depending on the assigned category. [6] Method according to any of the preceding claims, wherein the trained machine learning model (201) is further based on explainable artificial intelligence and has been trained with first and second training process data, which are labeled or unlabeled respectively, and includes explainability mechanisms to provide at least one basis for determining (S300) the plug state. [7] The method of claim 6, wherein the method further comprises: Determine (S600) a cause of the faulty connection between the electronic components depending on the basis of investigation. [8] Method for training a machine learning model for the method for detecting the faulty connection between the electronic components during robot-guided assembly according to any one of claims 1 to 7, wherein the machine learning model is trained to recognize the connection state of the electronic components, comprising the following steps: Providing (S110) a training data set (110) containing first training process data (111) of a robot performing the plugging and second training process data (112) of the electronic components as well as the respective plugging states (113) of the electronic components; Receiving (S210) the training data set (110); Defining (S310) a machine learning model; Initializing (S410) model parameters by setting initial values ​​for the model parameters of the machine learning model; Training (S510) the machine learning model by adjusting the model parameters; Validating (S610) the machine learning model; Finish (S710) the training and define (S710) the machine learning model as a trained machine learning model (201); and Storing (S810) the adapted model parameters of the trained machine learning model (201) in a computer-readable medium for later use in the procedure. [9] System for training a machine learning model for the method for detecting the faulty connection between the electronic components according to any one of claims 1 to 7, comprising a processor and a memory, the memory comprising instructions which, when executed by the processor, cause the control unit to execute the method for training a machine learning model according to claim 8. [10] System for detecting a faulty connection between electronic components during robot-guided assembly comprising the machine learning model (201) trained according to claim 8, comprising a processor and a memory, the memory comprising instructions which, when executed by the processor, cause the control unit to execute the method according to any one of claims 1 to 8.

Images