Die Babybommer-Generation der 1960er Jahre wird in den kommenden Jahren verrentet. Dadurch verlieren Unternehmen langjährige erfahrene Beschäftigte und auch deren Wissen, wenn nicht effiziente Lösungen gefunden werden, dieses zu identifizieren, zu speichern und zu transferieren. Dieser Herausforderung widmet sich das Forschungsvorhaben „KI_eeper – Know-how to keep“. Im Projekt wird erforscht, inwiefern künstliche Intelligenz Möglichkeiten eröffnet, um das implizite Erfahrungswissen von Beschäftigten automatisiert im Arbeitskontext zu erfassen, zu verarbeiten und zu transferieren. Aktuelle Ansätze des Wissenstransfers sind aufwendig und häufig auch mit hohen Kosten verbunden. Dabei hat jedoch eben das implizite Wissen große Relevanz für Unternehmen, welche für kleine und mittlere Unternehmen (KMU) noch stärker ausgeprägt ist.
Am Ende des Projektes soll ein digitales Assistenzsystem entstehen, welches das gesammelte und ausgewertete implizite Wissen von Erfahrungsträgern allen Beschäftigen zugänglich macht und diese somit bei der Ausführung ihrer Tätigkeiten bedarfsgerecht unterstützt. Dabei soll zunächst der Fokus auf Tätigkeiten in der Produktionsfertigung gelegt werden. Aufgrund der großen Vielfallt an Fertigungsverfahren und damit verbundenen unterschiedlichen Daten, soll eine allgemeingültige Lösung für Verarbeitung der Daten mittels der künstlichen Intelligenz gewählt werden. In dieser Veröffentlichung sollen ausgewählte Algorithmen betrachtet werden, welche für eine potenzielle technische Lösung verwendet werden können.
1 Introduction: KI_eeper – Know-how to keep
The research project “KI_eeper – Know-how to keep”, founded by the German Federal Ministry of Education and Research (BMBF), consists of a consortium of research institutions, developers and two companies from the metal and electrical industry in Germany. These provide case studies on which the technical solution aimed in the project is to be developed and tested. The aim of this project is to identify, safe and transfer the tacit experiential knowledge of long-term employees by using artificial intelligence.
Knowledge can generally be distinguished as tacit and explicit. Explicit knowledge is understood as knowledge that can be clearly communicated. It can be expressed in words and numbers, as well as passed on to others as, for example, data or processes [1]. Tacit knowledge, on the other hand, is knowledge that cannot be fully or adequately communicated in language by the outgoing person. This knowledge is often unknowingly embodied in the minds of employees and is manifested in intuitive behavior [2]. Activities do not always consist of explicit or tacit knowledge alone, but of the mixture of both. The more tacit knowledge is required, the more difficult it is to record and transfer knowledge to another person. Tasks with high proportion of tacit knowledge include activities with complex procedures and non-standardized processes [3], which can often also be found in manufacturing companies. According to North [4], a basic prerequisite for the competitiveness of a company is that employees can apply their experiential knowledge by training-on-the-job. If a long-term experienced employee leaves a company without passing the tacit knowledge, this can result in a loss of competitiveness. On the other hand, successful handling of existing knowledge in a company can even lead to an increase in competitiveness [5].
Anzeige
Knowledge management is concerned with how to transfer knowledge to other employees within an organization to prevent the disadvantages described by North. Approaches such as coaching, mentoring or training offer the opportunity to pass on tacit specialist knowledge [6]. Due to the demographic change, such concepts can often no longer be pursued, as many specialist positions are not filled and thus the direct transfer of the necessary knowledge is no longer possible. An increasing number of companies are having problems recruiting skilled workers, because the shortage of skilled workers is being acute in manufacturing companies [7]. In addition, the transfer of knowledge is hindered by language barriers of employees with migration background. Therefore, a solution should be found how valuable knowledge can be absorbed and stored as efficiently as possible. This valuable implicit knowledge should subsequently be made available to unskilled workers by means of assistance systems.
2 Use Cases
In the current context of demographic change, the aim of the research project is to record tacit knowledge and convert it into explicit knowledge which can then be presented to all employees. In particular, new inexperienced employees should be able to be trained more quickly in this way. This is to be elaborated based on two very different case studies. The use cases in the companies have been selected in such a way that the input and output for the intended assistance system differ greatly. In addition, the outputs of the assistance system should be selected in such a way that they are presented as simply as possible, so that people with language barriers or semi-skilled employees can also use the system and manage the working process professionally.
2.1 First Use Case: Straightening Flat Steel
For the first use case, an activity was chosen that requires a high degree of tacit knowledge on the part of an individual person. This involves the manual straightening of flat steel, which can be available in different dimensions and material properties. The raw material can be up to three meters long, hardened or non-hardened and weigh up to 25 kg [8]. For the straightening itself, an oscillating hydraulic joining press is used, in which only the ram adjustment is infinitely settable. The change in the ram adjustment results in a change in the force acting on the flat steel. According to their own statements, the experienced employees at this workstation cannot describe how to straighten the flat steel. They work intuitively, which is derived from their long-term experience at this station. They cannot convey why they use a certain force to straighten at certain points. They feel the rhythm of the press. They intuitively recognize which of the two edges can be straightened best. They adjust the pressure without having to look at the display. They can see with their eyes whether the flat bar still has a slight deflection. That is why it takes at least one years and more to learn this job.
The actual and target condition of the flat steel are used as the first input parameter. The target condition can be retrieved digitally from technical drawings. The actual condition of each flat steel is to be recorded with the use of a scanning sensor. For the training of the AI, the setting of the path limitation is also required. This can be retrieved digitally from the hydraulic press. Furthermore, the exact positioning of the flat steel in relation to the pressure point on the press is required so that the tacit knowledge can be processed by the AI. At present, a combination of distance sensors and the application and reading of a vernier is required for this. Thus, positioning is chosen as the third input variable for the flat steel. From these three input variables, the artificial intelligence is to calculate the correlation between the acting force and the pressure position on the bending of the flat steel. This correlation is to be made available to the inexperienced worker in the form of position data – again with the distance sensors and the vernier – as well as a setting suggestion for the path limitation.
Anzeige
Later, during the standard operation of the assistance system, only the actual and target states are to be used as input parameters. However, position data and path limitation settings are still to be defined as output parameters.
2.2 Second Use Case: Sub-Process of a Surface Technology System
The surface technology system of the other applications company was selected as the other use case. This is a plant process in which sheet metal parts are coated by means of powder coating and wet coating. In this process, the components are first clamped in a frame, which is then hung on a carriage of a rail system. This rail system transports the components to the individual stations of the system: cleaning, coating, curing, and hanging off with quality control. The complete system can be divided into several subsystems. The first station at the beginning of the process, where the components are hung up and the machine is controlled/operated, has the most influence on the efficiency and workload of the other subsystems. At this point, not only the frame is selected to match the components, but also the number of items, the positioning, the orientation of the components and the clamping elements used as all as the masking positions are determined. The masking serves to ensure that desired areas are not coated, e.g., threats or threaded holes.
The number of different coating products is higher than 2600 with a total of 250 various colors. Actual there is not much information/instruction on how the components are to be positioned on the frames, which auxiliary materials are to be used or which masking measures are to be taken. This information can be partially derived from existing technical drawings – including the masking measures. This results in the target state of the technical drawings as the first input variable for the training of the artificial intelligence. Furthermore, the AI must be fed with the experiential knowledge of the employees. For this purpose, an input is required where the employees enter the type of suspension, number of components, etc. In addition, a direct link between the AI and the technical drawings should be established at this point. In this case, a correlation to the recorded data as well as the target state should also be defined. As a third input variable, an image of the suspended components is to be recorded after the suspension process. This should be further associated with the other input variables. During the recording, not only the individual carriages are to be considered, but also their sequence. Since different sized components require various times during powder coating as well as other baking times, a certain sequence of diverse items must be adhered to maintain the desired cycle time. Currently, this planning of the sequence is only possible through the experience of longtime employees.
After the data training, the cognitive assistance system should provide all employees with suggestions regarding the hanging process when retrieving certain articles as an output. Furthermore, the sequence of articles to be processed should be determined and communicated by the artificial intelligence.
3 Methodological Approach
In order to find suitable AI models for the use-cases described, some criterion must first be selected so that a more detailed selection can be made.
At the time of this publication, the technical solutions only exist as concepts, which is why it is not yet possible to make a more precise statement regarding the sensor technology to be selected. However, IoT sensors will be used in both applications. Therefore, the selected model must be able to handle sensor data in general.
It was also determined that the model must be as transparent as possible or present data in such a way that an employee can cognitively grasp the learned properties of the model. Therefore, the internal processes of the model should have as little complexity as possible so that every employee can process the results of the AI. This results in complexity as a further criterion for selection. For the same reason, it must also be possible to process the interactions between the input variables when capturing the tacit knowledge. Considering these two prerequisites, it becomes clear that an explainable AI is being searched for, since the actual calculation processes running in the background are not available as a so-called “black box”, but as a “white box”. These processes are therefore available to the user and can be understood. This results in another positive aspect: from a socio-technical point of view, prejudices of employees could be reduced if it is recognizable what the actual model calculates or works with the data.
The scalability of the AI model also plays a major role. A generalist approach can result in both large and small amounts of data for different applications. Thus, the targeted AI must also be able to cope with large databases. In addition, accuracy is used as a criterion for the choice of model to obtain the highest possible resolution of the implicit knowledge.
To obtain a suitable selection of AI models, reference databases such as IEEE Xplore and ScienceDirect were accessed. These databases were first roughly searched for machine learning approaches which were rated using evaluation criteria. The generated data set was compared with the previously described criteria, e.g. that the model is explainable. Some of the resulting algorithms will be briefly presented in the following chapter.
4 Presentation of Different AI Models
Due to the variability of the two use cases, the biggest challenge in this project is to develop a universal AI-based solution, to capture the tacit knowledge of the experienced employees. Although in both cases there is digital data available for training that can be directly defined as input into the AI. The type of input of the experiential knowledge is very different. In the first case, the sensor system captures the knowledge completely passively. Whereas in the second use case, although certain data is passively recorded by taking photos and by an active input. The employees enter information into the technical system at the beginning of the technology introduction, e.g., using selection menus and images in one of the use cases. Over time, the system has collected so much data that employees have to actively enter less and less information. The system should then be able to make suggestions for improved task execution based on the collected data. In the following, some viable models that were identified due to research will be briefly presented.
4.1 Knowledge-Base Building Procedure
Huang et al. [10] addressed the involvement of the human expert knowledge in digital image processing using a machine learning method to build a knowledge-based model. They assume that each object can be represented by its attribute value class vector, such as [attribute a_1,…, attribute a_n, class – i]. Therefore, they chose a subset of data, S, as a representation of data, so that this subset of data includes all possible classes. Then, a decision algorithm is applied to the subset of data to partition S into subsets of S_1, S_2, …, S_n based on some production rules. These production rules express that the object can be assigned to a certain class if the rules are followed. For instance: if attribute 1 > 80 and attribute 2 < 10 then object will be belong to class 2.
4.2 GAMI-Net
The fundamental idea of the GAMI-Net described by Yang et al. [9] is to capture the main effects and pairwise interactions of the model inputs. From this point of view, the structure of the GAMI-Net consists of two disjoint networks “Main Effects Module” and “Pairwise Interactions Module”, each performing its own tasks. The Main Effects Module tries to capture the nonlinear main effects of the model inputs using so-called subnetworks. Each of the subnetworks gets one input and propagates it through intermediate layers to calculate one output; while the pairwise interactions get two inputs to capture the partwise interactions of the model inputs. Hence, what distinguishes these two modules from each other is the number of input neurons in their input layers. At last, a linear combination of the modules outputs forms the final output. At last, a linear combination of the module outputs forms the final output. The following Fig. 1 depicts the structure of the GAMI-Net.
The Bayesian Rule List by Letham et al. [11] is about the generation of a list of rules for classifying data. The model consists of two main steps:
1.
Obtaining frequent patterns within the dataset using the FP-tree (frequent-pattern tree) algorithm.
2.
Learning of a decision list by selection of the previously obtained patterns.
Thus, the first step is to learn frequently occurring patterns. The authors use the FP-tree algorithm as an example, but according to them it also works with all other algorithms that extract patterns in data sets. Accordingly in the further process a general representation of the finding patterns, which also happens with the FP-tree algorithm. Due to the more detailed explanation of FP-tree will be omitted. Patterns are here both the characteristics of a property within a data set as well as the combination of characteristics. The modeler determines when a data pattern occurs frequently.
4.4 Deep Learning for Case-Based Reasoning through Prototypes
Li et al. [12] attempted to present a neural network architecture (Deep Learning for Case-Based Reasoning through Prototypes – DLCBRP) through which without using a second model an explanation for each prediction could be possible. To do this, they proposed an encoder and a decoder. During training, the data set is processed by the encoder in such a way that the dimension of the input is reduced and useful properties for a prediction are retained. This reduced training data set is then used to learn prototype vectors. In this context, a prototype means that it is as close or identical to an observation of the training data set. The learned prototype vectors, whose validation might be visualized by an encoder network, reside in the prototype layer, which is then connected to the next fully connected layer by weights. Finally, by normalizing the weighted sum, a softmax layer output is given, which represents a probability distribution of all classes (Fig. 2).
In the context of machine learning, not only the choice of an adequate approach should be addressed, but also the data that is fed into a model is most important. In the described cases, there is tacit knowledge that someone has acquired over years of working in a company. Collecting this tacit knowledge is one of the most challenging tasks, as any kind of sensitive knowledge must be taken into account in terms of the overall data quality. Hence, the model will learn from its training data more effectively if more correct training data can be provided. It should be considered that this kind of knowledge may not always can be expressed in form of data.
Another aspect that should be addressed is the explainability of the chosen models. After all, in the context of preparing or teaching someone who has not been familiar with a system, there is this expectation that after some hours or days they should be able to perform their work independently without the help of supervisors or others. Once a suitable model has been trained, it should be able to interpret which components or how much of each component is involved in a final product. This paper has therefore first tried to select some of the approaches that belong to the explainable models.
The AI models described before have different properties. However, none of the discussed algorithms provides a universally valid solution for the inclusion of tacit knowledge. With the paper, an intermediate goal has been achieved in which algorithms for the use case have been investigated. It is evident that further work is required to achieve the overall project goal.
Acknowledgements
The presented results are part of the research project „KI_eeper – Know how to keep“ (FKZ: 02L20C500- 02L20C505). We would like to thank the German Federal Ministry of Education and Research (BMBF) for funding this project. Furthermore, we thank the Project Management Agency Karlsruhe (PTKA) for supporting the project. The authors are responsible for the content of this publication.
Open Access Dieses Kapitel wird unter der Creative Commons Namensnennung 4.0 International Lizenz (http://creativecommons.org/licenses/by/4.0/deed.de) veröffentlicht, welche die Nutzung, Vervielfältigung, Bearbeitung, Verbreitung und Wiedergabe in jeglichem Medium und Format erlaubt, sofern Sie den/die ursprünglichen Autor(en) und die Quelle ordnungsgemäß nennen, einen Link zur Creative Commons Lizenz beifügen und angeben, ob Änderungen vorgenommen wurden.
Die in diesem Kapitel enthaltenen Bilder und sonstiges Drittmaterial unterliegen ebenfalls der genannten Creative Commons Lizenz, sofern sich aus der Abbildungslegende nichts anderes ergibt. Sofern das betreffende Material nicht unter der genannten Creative Commons Lizenz steht und die betreffende Handlung nicht nach gesetzlichen Vorschriften erlaubt ist, ist für die oben aufgeführten Weiterverwendungen des Materials die Einwilligung des jeweiligen Rechteinhabers einzuholen.
