1 Introduction

The modern manufacturing industry has gradually shifted from make-to-stock production to make-to-order production (Wang and Jiang, 2017). Increasing individualization and shorter product life cycles require companies to be simultaneously flexible and cost-effective at a level (Khan and Turowski, 2016). To achieve the necessary degree of dynamic flexibility, simple machines transform into intelligent production systems in which equipment and systems are highly interconnected (Morlock et al., 2016). This evolution is shaped by the substitution of humans through automation and closer cooperation between humans and machines (Apt et al., 2018). The resulting socio-technical system is characterized by increased plant and process complexity. Hence, the interaction with machines and production systems becomes progressively unpredictable and less repetitive for employees (Thalmann et al., 2020; Ulber and Remmers, 2019). Under these aspects, shop-floor workers gradually transform into knowledge workers, for whom collaboration and technological assistance gain importance in production processes (Hartmann, 2015; Riedel et al., 2017).

Therefore, production describes the purposeful service creation of input factors (e.g., material goods and services), transformed into value-enhanced output factors (Kern et al., 1996; Neumann, 1996). In this context, production is defined as the ‘[…] directed use of goods and services, the so-called factors of production, to extract raw materials or to produce or manufacture goods and to generate services […]’ (Bloech et al., 2014). The quality and yield of the production process, and thus the product to be manufactured, is determined by the interaction of human labor, operating resources, and materials (Neuhaus, 2007; Voigt, 2008).

For production processes, there is a variety of literature-based categorizations. On a general level the distinction can be made between two production technologies: ‘discrete manufacturing’ and ‘process manufacturing’ (Blömer, 1999; Dangelmaier, 2009; Thomas et al., 2007). Discrete manufacturing processes are characterized by producing piece goods, such as those manufactured in mechanical engineering (Bakir et al., 2013; Bartneck et al., 2008; Blömer, 1999). Challenges of discrete production processes, especially during the elimination of production anomalies through shop-floor workers, are usually uncritical and can be remedied in a short time due to the possible interruptibility of the process (Müller and Oehm, 2019; Schult et al., 2015). On the other hand, process manufacturing is characterized by continuous material flows, where raw materials are transformed through chemical and physical processes using formulas and recipes (Dennis and Meredith, 2000; Fransoo and Rutten, 1994; Smith, 2008; Urbas, 2012). Characteristic of this process is the uninterrupted supply of input materials (e.g., plastic pellets) and the continuous completion of end products (e.g., printed edge band) by a so-called continuous production line consisting of a large number of coupled individual machines (e.g., extruder, cooling system, printing) (Blömer, 1999). In addition to the term process manufacturing, continuous production is often used synonymously (Blömer, 1999; Dangelmaier, 2009; Thomas et al., 2007; Volland, 2012). For an in-depth comparison of knowledge requirements for shop-floor operators in discrete and continuous manufacturing see Müller and Oehm (2019).

The application domain of this work represents such a continuous production scenario. During the production of furniture edge band, the individual machines coupled as a production line comprise a length of up to 100 m. Controlling up to 20 coupled machines comes with frequent parameter adjustments. Especially the coating and printing of the edge band requires a detailed adaption to specific customer requirements, which is characteristic for continuous production processes and often leads to anomalies (Hörner et al., 2020). These production anomalies occur abruptly (e.g., caused by a machine failure) or gradually (e.g., caused by dragging parameter deviation) and are defined as apparent occurrences that deviate from a defined process standard or what is expected (Saez et al., 2017). The effects of these production anomalies are significant and cause, among other things, schedule delays, machine failures, or inferior product quality, which are associated with high costs due to recalls or recourse claims (Eljack and Kazi, 2016; Venkatasubramanian et al., 2003). The associated costs can be minimized when an anomaly occurs, mainly through timely and situation-appropriate anomaly rectification. The elimination of these unforeseen anomalies in process manufacturing usually consists of three different stages: the detection of the present anomaly (e.g., the color of the edgeband is not OK), the diagnosis of the underlying cause (e.g., ink pump 3 is clogged), and the implementation of process-compliant action measures (e.g., flushing of the ink pump) to eliminate the anomaly (Lau et al., 2012; Patrick et al., 1999). While the detection of the present fault usually seems to be easily manageable (Müller and Oehm, 2019), the diagnosis of the cause and the derivation of suitable action measures are highly complex (Patrick et al., 1999; Perrow, 1984).

1.1 Related work

Despite current efforts to fully automate anomaly correction, it mostly remains the task of the operator to fix occurring anomalies (Webert et al., 2022). Digital interventions such as assistance systems can be used to support the operator in conducting this task. Those systems focus on facilitating human-technology interactions while increasing overall productivity and shortening new employees' training times (Apt et al., 2018). Assistance systems are defined as information technology systems that support people and other information technology systems in a goal-oriented activity (Minor, 2006). In production-related scenarios, a distinction can be made between physical, sensory, and cognitive assistance systems (Mark et al., 2021). In anomaly correction, cognitive assistance systems are the main interest. These systems support users in accomplishing mental tasks required to achieve a particular goal under the given conditions (Romero et al., 2016). Due to the changing demands of job requirements for a machine operator, those systems have gathered a substantial amount of research attention. Mark et al. (2021) provide a comprehensive state-of-the-art review of operator assistance systems. 80 of the 121 reviewed assistance systems are cognitive assistance systems. The analysis of the listed assistance systems shows that their manufacturing focus is on discrete manufacturing processes, while continuous production processes are only marginally considered (see Mark et al., 2021). Since discrete and continuous manufacturing processes differ significantly based on a wide variety of factors, including intervention complexity, and thus also have different support needs (Müller and Oehm, 2019), additional research is needed here.

In cognitive assistance system development and design one of the critical issues is integrating the complexity of operational problems as comprehensively as possible into the system and facilitating practical support on this basis (Niehaus, 2017). However, a wide range of operational issues can usually only be adequately described through employees' knowledge gained from experience (Niehaus, 2017). Since the therefore required knowledge is usually only available in tacit form, and there is a lack of concrete externalization methods in manufacturing-related application areas, the development of a standardized externalization approach is needed (Gavrilova and Andreeva, 2012; Gosling and Andrade, 2018; Gourlay, 2006; Niehaus, 2017; Okafor and Osuagwu, 2006; Shadbolt, 2005). Although many cognitive assistance systems do rely on employees knowledge, they usually fail to describe their externalization procedure in detail (see Mark et al., 2021).

In this regard, the term ‘tacit knowledge’ is introduced – in a simplified manner – as non-codified knowledge gained through learning by doing (Roberts, 2000). The primary concern is with ‘knowing how’ to rectify an already detected anomaly. In particular, the practices applied by employees and how they are applied are of interest in the context described above. We do not aim to contribute to the controversial and longstanding discussion of defining knowledge itself. For a comprehensive and critical analysis of the emergence of tacit knowledge, see Schmidt (2012). For a thorough synthesis of the developments of knowledge and expertise sharing in the CSCW community, see Ackerman et al. (2013).

Since anomalies are especially challenging to fix in a continuous manufacturing environment, tacit knowledge's externalization and storage have high relevance in this manufacturing area (Mersch et al., 2011; Smith, 2014; Terhechte et al., 2019). On an organizational level, the need to capture and support employees with this knowledge is accelerated by a continuing shortage of skilled workers, demographic change, and increased employee turnover (Hartmann, 2015; Riedel et al., 2017). Making manufacturing-relevant knowledge of employees usable throughout the entire manufacturing process is a major challenge, which can be done through various procedures – one of them being the development and application of a method for the structured externalization, storage, and reuse of this knowledge (Hansen et al., 1999; Jasperneite and Niggemann, 2012; Nonaka, 1994). In particular, the initial step of knowledge externalization, i.e., the transformation process from tacit knowledge into explicit knowledge independent of persons, has been recognized as a significant challenge for manufacturing companies (Barley et al., 2018; Niehaus, 2017). The literature therefore repeatedly criticizes the lack of concrete externalization methods in manufacturing-related application areas (Gavrilova and Andreeva, 2012; Gosling and Andrade, 2018; Gourlay, 2006; Okafor and Osuagwu, 2006; Shadbolt, 2005).

Simply storing such knowledge in a static repository is not sufficient to foster organizational knowledge sharing (Ackerman et al., 2013; Nonaka and Takeuchi, 1995). To account for the socio-technical nature of modern shop-floor environments, an assistance system is developed as groupware concerning the context-specific support requirements of collaborative work (Schmidt and Bannon, 1992). Groupware is defined by Ellis et al. as: ‘computer-based systems that support groups of people engaged in a common task (or goal) and that provide an interface to a shared environment.‘ (Ellis et al., 1991, p. 40). While studying other domains extensively, the related research field of computer-supported collaborative work has given little attention to the domain of advanced manufacturing (e.g., 'flexible', 'order-driven') (Schmidt, 2013). Although groupware such as digital peer tutoring systems exist in the manufacturing domain (e.g. Clemmensen and Nørbjerg, 2019), little research focuses on the correction of anomalies.

In conclusion, literature revolving around assistance systems fails to ‘unveil’ their methodology for knowledge externalization. One contribution of this work is therefore the provision of a standardized procedure to externalize tacit manufacturing knowledge. By making the procedure explicit, we aim to foster the discussion on this issue. Furthermore, little work is to be found on the subject of supporting manual anomaly correction in manufacturing (Webert et al., 2022). The lack thereof is accentuated in the field of CSCW, since it has given little attention to advanced manufacturing in general. The second contribution is therefore, the design and implementation of a groupware assistance system. The major novelty of this work lies in merging the field work-oriented knowledge externalization procedure with the systems human centered design. We aim to thereby satisfy and unveil requirements of computer supported collaborative work in the context of anomaly correction in the continuous manufacturing industry.

1.2 Methodological approach

To address the listed issues, the research question proposed in this paper focuses on the method of externalization and collaborative reusability of tacit knowledge to eliminate manufacturing-related anomalies:

How can problem-solving knowledge for more efficient elimination of anomalies in continuous manufacturing processes be externalized and provided utilizing an assistance system?

The research design of this paper is based on the design-oriented research paradigm of Design Science Research (DSR) (Hevner et al., 2004). Here, novel artifacts are developed in an iterative process, where the execution of the process steps is not subject to strict process sequencing (Peffers et al., 2007). A dynamic approach and continuous feedback lead to a constant improvement of the partial artifacts (Peffers et al., 2007). The purpose of the artifact proposed in this paper focuses on tackling current problems in knowledge externalization, especially in continuous manufacturing processes. This is done by developing a standardized method for externalizing tacit knowledge to enable a structured knowledge base and collaborative knowledge sharing using an assistance system.

In detail, this paper aims to solve these problems by developing a mixed-method approach to externalize employee-bound knowledge and evaluating it in a manufacturing scenario. Building on the externalized knowledge, a development approach for a knowledge-based assistance system and the practical implementation is presented. Furthermore, an evaluation concept is introduced to test the assistance system in a manufacturing scenario. By developing the assistance system as groupware, research insights in computer-supported cooperative work are given in a real-world, continuous production scenario. The development approach takes the collaboration support requirements of the environment into account by designing and developing in close collaboration with the shop-floor workers. The focus of the presented insights is on the collaboration enabling technology.

2 Approach to externalize tacit knowledge

The literature describes various methods for knowledge externalization and elicitation (Gavrilova and Andreeva, 2012; Milton, 2003; Shadbolt and Smart, 2015). Although they all essentially pursue the same objectives, their suitability can vary depending on the given context. Potential influencing factors are for example the type of knowledge addressed or the knowledge carriers capability, motivation, and opportunity to externalize knowledge (Gavrilova and Andreeva, 2012). Related papers describe specific requirements for capturing employee-bound knowledge (Grandvallet et al., 2017; Guerra-Zubiaga, 2017; Johnson et al., 2019). These serve as an orientation to define requirements for the development of an externalization approach in industrial process manufacturing. The rationale behind the specific requirements (see Table 1) is explained in the following section.

Table 1 Properties of methods for knowledge externalization.

Since the type of knowledge to be externalized is employee-bound, the method should be capable to partially externalize ‘Tacit Knowledge’. Secondly, the requirement ‘Procedural Knowledge’ is derived from the necessity of capturing procedural expertise in the present continuous manufacturing environment. Procedural knowledge describes knowledge about action sequences and processes, which a manufacturing employee gains primarily through the repetitive execution of activities (Pawlowsky, 2019). Therefore, the method needs to capture the relationship between individual knowledge components and concrete process steps. Thirdly, the requirement ‘Analyst integration’ is included to counteract some of the practical limitations of externalization methods. As Gavrilova and Andreeva (2012) state, knowledge carriers in organizations might not be sufficiently motivated, capable, or presented with the opportunity to externalize their tacit knowledge. To counteract these tendencies, they suggest integrating an analyst into the knowledge externalization process so that the process is not solely dependent on the knowledge carrier. The last requirement, ‘1-h feasibility’ is included due to time limitations because shop-floor employees still must be able to accomplish their daily workload. The ‘Method-Class’ structure is mainly derived from Shadbolt and Smarts’ (2015) taxonomy of elicitation methods.

Based on a systematic literature analysis according to Mayring (2012), Table 1 lists a selection of the identified externalization and elicitation methods. Additionally, it evaluates their methodological suitability for knowledge externalization in continuous manufacturing processes. A method is considered suitable if it fulfills all the requirements described above. Whether or not the requirements ‘Tacit Knowledge’ or ‘Procedural Knowledge’ are fulfilled is derived from the following literature: Miltons' (2003) categorization of elicitation methods, Shadbolt and Smarts (2015) remarks, and Burges' (1996) elicitation technique classification. The third requirement, ‘Analyst integration’ is derived from Gavrilova and Andreevas (2012) research, where they identify primary elicitation methods which integrate an analyst into the externalization process. The decision for ‘1-h feasibility’ is primarily based on Miltons (2003) analysis on whether an elicitation method is suitable to be explained and executed in a timeframe of 45 min to 75 min. The main concern is the invested time of the knowledge carrier. For example, external observations and protocol analysis only require a substantial time investment from the external analyst and are therefore considered to fulfill the requirement.

‘X’ in Table 1 represents that the externalization method fulfills a certain requirement. ‘/’ indicates uncertainty, which stems from substantial discrepancies between the categorization results among the authors listed above. Whenever the authors do not consider an externalization method, the method is rated based on its corresponding ‘Method-Class’ as depicted in Table 1 (e.g., ‘Fixed Probe Interviews’ is rated as ‘Interviews’). The literature research carried out in the first quarter of 2020 was developed with the help of the following academic search engines: Emerald Insight, De Gruyter, ScienceDirect, ResearchGate, IEEE Xplore, SAGE journals, SpringerLink, OPAC, and GoogleScholar. The search terms used include: 'Knowledge Externalization', 'Knowledge Elicitation', 'Knowledge Acquisition', and the specification with phrases such as 'Manufacturing', 'Shop-Floor', 'Methods' and 'Techniques'. Table 1 shows only an excerpt of elicitation and externalization methods available in the literature. For a more comprehensive summary and categorization see, for example, Burge (1996) or Shadbolt and Smart (2015).

It needs to be mentioned, that Table 1 only shows a degree of suitability to fulfill a certain requirement. One can not derive, for example, that ‘Reportory Grid’ is enterely incapable of eliciting procedural knowledge. Furthermore, Table 1 might suggest that one particular method could be used to elicit the targeted tacit knowledge with a satisfactory result. However, as Shadbolt and Smart (2015) argue, there is usually the need to apply a variety of methods in any knowledge elicitation program, ‘[e]ven when it appears that only one particular body of knowledge is being dealt with’ (Shadbolt and Smart, 2015, p. 190).

Therefore, the resulting methodology (see Figure 1) uses interview techniques and an observation technique to externalize the expert knowledge of shop-floor employees, which is only implicitly available. The degree of method structuring increases with the level of detail of the externalized knowledge. This ensures that the externalization is carried out entirely without predetermined restrictions, as would possibly be given by pre-defined structures (Kendal and Creen, 2007). In this way, the externalization is as comprehensive as possible while maintaining a high level of detail. Based on the information generated from observation and interviews, structured text and protocol analyses are carried out to obtain documentable results (Shadboltl and Smart, 2015). Figure 1 visualizes and describes the four-step process of knowledge externalization using a combination of (A) unstructured interviews, (B) external observation and semi-structured interviews, (C) structured textual analysis, and finally (D) protocol analysis.

Figure 1.
figure 1

Methodology to externalize production knowledge.

In the externalization of employee-bound knowledge, the additional cognitive burden on knowledge carriers' minimization plays a significant role. Therefore, the process shown in Figure 1 integrates a dedicated external analyst role (Kendal and Creen, 2007). In addition to the cognitive relief, this reduces the dependence on the knowledge carrier's intrinsic willingness and motivation to communicate his implicitly available expertise. The analyst's 'interposing' reduces the cognitive distortion of knowledge externalization and enables an intersubjectively comprehensible procedure (O'Hagan, 2019).

In step 'A', the analyst needs to gain a basic understanding of the process flow. Conducting unstructured interviews with process stakeholders enables the analyst to overview the manufacturing scenario and make an initial, rough categorization of the manufacturing process (Weiss and Kulikowski, 1984). The analyst thereby gains initial but crucial insights into the specific knowledge domain (Gavrilova and Andreeva, 2012).

The measures carried out in step 'B' describe the actual externalization of process-specific knowledge. For this purpose, the analyst observes shop-floor workers in the process of eliminating manufacturing errors and loosely records the observations (Bungard et al., 1996; Kleber, 1992). Subsequently, semi-structured interviews are carried out with the shop-floor employees, using the so-called teach-back method. The analyst describes the actions observed by the employee and asks him/her for additions and corrections (Johnson and Johnson, 1987). This enables efficient and formally correct knowledge externalization with a simultaneously low additional cognitive load on the shop-floor employee (Milton, 2012).

To structure and codify the loosely logged externalized knowledge, a qualitative and quantitative content analysis of the logs is carried out in step 'C', enabling deductive categorization of the identified anomalies and the corresponding solution strategies (Mayring, 2012).

Utilizing a protocol analysis with process stakeholders, which is oriented towards so-called ‘collegial verbalization’, a validation of the externalized knowledge and knowledge enrichment takes place in the final step 'D' (Erlandsson and Jansson, 2007). In addition to validating that the knowledge is process compliant, it becomes possible to eliminate duplicates and undesired action measures from the generated knowledge base. By externalizing the tacit knowledge of the process stakeholder, e.g., a process engineer, further enrichment of the knowledge base is possible.

2.1 Evaluation of knowledge externalization in a continuous manufacturing scenario

To evaluate the four-staged knowledge externalization approach developed in this paper (see Figure 1), the method is tested in the continuous manufacturing scenario already outlined in chapter 1. The evaluation aims to confirm the approach's functionality, check its practical relevance, and emphasize the application potentials of the externalized knowledge. As stated in the previous chapters, the externalization and provision of knowledge are particularly relevant in eliminating manufacturing anomalies. On the one hand, human errors are one of the leading causes of manufacturing downtime, but human productivity is also mainly needed to eliminate these anomalies (Angelopoulou et al., 2020; Habtoor, 2015; Barroso and Wilson, 2000; Zhang et al., 2020). This can be operationalized through digital assistance systems since they provide a suitable basis for reusing knowledge and collaboration (Dhuieb et al., 2015). One main challenge of digital assistance systems is integrating complex operational problems and the employee-bound knowledge required to solve these, the developed approach intends to counteract this challenge (Niehaus, 2017). Additionally, the particularities of a continuous manufacturing scenario lead to an increased need for research on the utilization of tacit knowledge in this field (Bakhrankova, 2009; Dennis and Meredith, 2000), which is not given to this extent in discrete manufacturing processes due to the less complex assignment of individual knowledge components to concrete process steps (Terhechte et al., 2019). The evaluation of the four-staged approach for knowledge externalization in the described manufacturing scenario is done over a period of 6 weeks. It follows the procedure shown in Figure 1 with the involvement of an external knowledge analyst and internal knowledge carriers.

2.1.1 A: Categorization of the manufacturing process

The categorization of the manufacturing process is based on an unstructured interview with process experts in the role of senior production engineers. Two engineers are asked about the structure of the manufacturing process, its target sequence, and process steps that have historically been significantly affected by anomalies. The result is a categorization of the manufacturing process into ten main categories.

2.1.2 B: Identification and externalization of tacit knowledge

The identification and externalization of tacit employee knowledge is carried out over four working days based on external observations on the shop-floor. In this process, the analyst observed 36 different shop-floor workers across shifts during the elimination of anomalies (e.g., activities to eliminate uneven product coating). It becomes apparent that operators usually discover a possible anomaly at the very end of the production line, either by pre-defined inspections or by chance. After the manual anomaly detection, the anomaly correction is also manually done by the operator. Therefore, the specific actions taken by the employees are recorded. When encountering a non-solvable anomaly incidence, a reoccurring pattern of knowledge inquiry is observed. An overarching anomaly category is described to a colleague, followed by a visual description and the request for anomaly solution measures. After observation, the shop-floor employees make additions and corrections to the recorded knowledge. Therefore, semi-structured interviews with the analyst in the style of the ‘teach-back’ method are used.

2.1.3 C: Structuring and explicating knowledge

The semi-structured and loosely recorded results from the previous step were then examined using structured text analysis. The analyst identified 228 rule sets for the elimination of manufacturing anomalies. A rule set consists of the unique combination of a two-staged anomaly knowledge and the corresponding solution knowledge. A rough classification of the present anomaly takes place with selecting the anomaly category. The visual appearance of the anomaly further refines this. The solution knowledge is finally derived from this distinct combination, whereby a quantity of solution knowledge can be assigned to the defined concatenation of anomaly category and appearance. An exemplary presentation can be found in Figure 3 (Chapter 3.3) of this paper.

2.1.4 D: Validation and enrichment of externalized knowledge

A protocol analysis is performed in collaboration with the senior production engineers from Step 'A' to ensure process conformity and further enrich the identified rule sets. The analysis clarifies the rule sets linguistically and eliminates duplicates and demonstrably nontarget elimination actions. Based on further process documentation, which is explicitly available, the control set is enriched with 33 additional action measures.

In total, the performed evaluation in the described manufacturing scenario results in 261 sets of anomaly resolution rules, consisting of 10 anomaly categories, 117 anomaly appearances, and 261 specific solutions to eliminate an anomaly. Examples of externalized anomaly categories are ‘paint and varnish’, ‘adhesion promoter’ or ‘spatial shape’. The categories are then further refined based on their visual appearances. Exemplary tuples of category and appearance are, e.g., ‘Paint and varnish—paint deposits on top and/or bottom side’, ‘Adhesion promoter—adhesion promoter carry-over on roller’ or ‘Spatial shape—wall thickness too small’. There is at least one process-compliant anomaly elimination solution for each individual combination of category and appearance. For the exemplary anomaly ‘Ink and varnish—ink deposits on top and/or bottom’, a solution measure for elimination is ‘Check all drying channels, printing units incl. primer printing unit for correct setting according to work instructions’.

The evaluation itself was conducted in Q1/2020 for about 50 h. Due to the four-shift system used in the application scenario, an extension of the analysis period was necessary to include as many different store floor employees as possible. Therefore, the analysis took place throughout four working days, spread over a period of six weeks. Table 2 shows additional data about the execution and participants of the evaluation. For each step, the number of persons involved and their age and job role in years are recorded and analyzed. In addition, the result of the respective step is listed.

Table 2 Results and additional data from the evaluation of knowledge externalization.

To evaluate the externalized knowledge, an interview is conducted with the production engineers consulted in step A and step D. Based on their knowledge and standardized process documentation, the possible solutions to eliminate anomalies are reviewed in the interview for their suitability and completeness. In the four-hour interview session, led by the knowledge analyst (cf. ‘step C’), all 261 sets of knowledge are evaluated in terms of their suitability for use in ongoing production operations, their process conformity, and their coverage of the production process. The evaluation is mainly based on the experience of the process engineers; in the case of uncertainties, an additional check is carried out based on existing process documentation (e.g., procedural instructions). Concerning process conformity and coverage of the entire production process, the experts confirm the suitability of the externalized knowledge. The experts thus attested that the knowledge for anomaly elimination completely covers the production process, from the plasticisation of the raw material to the packaging of the finished product. In order to adapt the knowledge base to changed framework conditions (e.g., changed machinery, product changes), it is possible to skip individual steps of the externalization method. For example, only the structuring step ‘C’ can be applied to newly integrated contents of process documentation in order to integrate changed framework conditions into the knowledge base. This enables continuous further development and adaptation of the knowledge base to changes or innovations without repeatedly running through the entire process. The effort of knowledge externalization to ensure the currency of the knowledge base is therefore correspondingly low compared to the initial execution of all steps A-D.

Table 3 shows the distribution of all 261 externalized rule sets referencing on the respective anomaly category. The allocation of the anomaly appearances to the ten created anomaly categories (cf. step A) shows a differentiated distribution. While category C-5 ‘Machine error / stop’ accounts for 19.7% of all anomaly appearances, the categories C-7 ‘Primer’ and C-10 ‘Tempering’ each represent only 2.6% of all anomaly appearances. It should be noted that the categories C-1 ‘Adhesion promoter’ and C-6 ‘Paint and varnish’ belong to a similar process, namely surface printing. In the categorization carried out with the production engineers (cf. step A), the surface printing process is split into the categories C-1 and C-6 due to the given relevance and need for fine granular differentiation. From this point of view, they would add up to around one-third of all appearances and thus represent by far the most critical category. While individual categories are affected by only a few potential anomalies (e.g., C-3 ‘Humidity / moisture’, or C-4 ‘Longitudinal distortion’), some categories contain a very extensive portfolio of possible anomalies. For example, anomalies in C-6 ‘Paint and varnish’ can present themselves in very different ways, increasing the number of potential resolution actions in this category. In order to keep the effort of identifying the appropriate appearance low, it is therefore necessary to pay attention to an appropriate quantity when defining the categories. At the same time, the number of categories must remain manageable with a more fine-grained segmentation of the appearances to not impair the clarity of the initial anomaly categories. In the interviews with process experts, a minimum number of 3 up to a maximum of 25 appearances per category is considered reasonable.

Table 3 Distribution of externalized knowledge based on anomaly categories.

Due to the direct relationship between the appearance and the resolution measure, the relative numbers of anomaly elimination solutions behave similarly to the appearance numbers above. On average, the number of assigned elimination solutions on the associated anomaly appearances is 2.2 solutions per appearance. Generally speaking, an operator can choose from more than two possible elimination solutions for an identified anomaly. Given the externalization methodology, which focuses mainly on the experience knowledge of operators, it can be assumed that the anomaly appearances and elimination solutions with the highest occurrences are of above-average importance for anomaly remediation steps. Although the knowledge is now externalized, it needs to be integrated into the appropriate technological tool to foster knowledge and expertise sharing. The ‘ethnographic’ insights gathered during the externalization procedure are therefore used to inform the development of the assistance system (Chapter 3).

3 Design and development of a knowledge-based assistance system

The developed assistance system therefore aims to assist shop-floor workers in their ongoing work, while being framed as a knowledge-based assistance system. However, merely storing knowledge in a repository without considering the social context is insufficient (Ackerman et al., 2013; Nonaka and Takeucchi, 1995). Past research has shown that one of the main challenges is the creation of an environment where employees use and expand the available knowledge (Hislop et al., 2018). To assist in creating such an environment, the design and development phase conceptualizes the assistance system as groupware concerning the common information space of the shop-floor (Marca and Bock, 1992). According to Johansen's (1988) CSCW-Matrix, the assistance system enables asynchronous (time-independent) but collocated (same-place) collaboration. The already externalized knowledge (see Chapter 2) is digitized in the form of a knowledge repository and creates the foundation for time-independent knowledge exchange. In the development process of the assistance system, the aim of the previously conducted externalization procedure is to provide structure and initial usefuleness of the system. Building on this foundation, the assistance system integrates further functionalities to foster asynchronous knowledge and expertise sharing on the shop-floor. Therefore, knowledge exchange is possible across shifts, independent of currently present knowledge carriers but limited to the same location. However, introducing such a digital assistance system on the shop-floor also interferes with established workflows and organizational structures. Challenges in the design approach thus not only exist in the selection of suitable technologies but also regarding social and ergonomic factors. There are specific prerequisites to creating practical value in human–machine interaction: the active involvement of the employee in the work process and familiarity, acceptance, and usage of the technical environment (Fischer et al., 2019). Therefore, designing a successful human–machine interaction requires the users' involvement (shop-floor workers). By integrating users, their expertise is utilized while simultaneously establishing familiarity, acceptance, and understanding of the new work processes (Schenk et al., 2016).

ISO 9241:210 (2010) presents a human-centered design approach for developing interactive systems. This development approach aims to make systems usable and useful by involving the users and concentrating on their demands and requirements. ISO 9241:210 claims to increase the resulting system's effectiveness, efficiency, accessibility, and sustainability. Additionally, possible negative impacts on human health, safety, and performance that might result from using the system are countered. Multiple authors (Nelles et al., 2016; Quandt et al., 2020) consider the approach as best practice when developing digital assistance systems in a manufacturing environment. According to the Design Science Research paradigm, the core of the human-centered design approach stands an iteration cycle of four stages, which are usually conducted in the following sequence (see Figure 2): understand and specify the context of use (1), specify the user requirements (2), produce design solutions (3) evaluate the design against requirements (4). Before entering the iteration cycle, a non-recurring stage of planning the design process must be conducted. If the evaluation (4) concludes that the design solution meets the requirements, the development is completed (Apt et al., 2018).

Figure 2.
figure 2

Development and methods used for the design of an assistance system in continuous manufacturing scenario based on ISO 9241:210 (2010).

While ISO 9241:210 describes the human-centered design process's fundamental structure, potential methods to carry out the given steps are not specified thoroughly. To identify possible methods for each design phase, a meta-analysis of five literature reviews according to Mayring (2012) is carried out (Apt et al., 2018; Benyon, 2014; Jokela et al., 2003; Nelles et al., 2016; Schild et al., 2019). The analysis shows that there are 32 methods applicable to the specified stages, which are derived from various disciplines. Among others, the range from workshops and user-requirement personas to interactive prototyping design. The design's evaluation against the requirements can be conducted using heuristic evaluation, experimental or expert evaluation, cooperative evaluation, or questionnaires. For this study, an analysis of the physical and organizational environment, the application context, and the end-users' technical and task-specific requirements is performed. The resulting requirements are implemented in an initial mock-up and evaluated with a user test. This represents one completed iteration cycle. The evaluation is used to query the priority of user requirements and, if necessary, to adapt them accordingly or add new derived ones. With the user requirements revised, a second functional prototype is created, which is evaluated with a user test. This iteration step is repeated with a third improved prototype. Figure 2 illustrates the four-tier iteration cycle and the corresponding methods.

3.1 Understand and specify the context of use

An initial workshop is conducted to understand and specify the context of use, including one management member, three process experts, the development team, and two experienced shop-floor workers (users). This aims to get an overview of the systemic context of use, recorded in a document during the workshop. A brainstorming session is held with the process experts, management, and the development team to further refine the use context. Based on the created document of the context of use, containing a first set of ideas regarding the system's functionality and the company's goals, semi-structured interviews are conducted with the shop-floor workers. User stories and personas are created based on the consolidated information of the previous activities'. These include characteristics about the person (e.g., age, education, job qualification, job experience, IT skills), the task (e.g., typical work task and subtasks, issues at work and improvement suggestions, the overall goal of using the system), the environment (e.g., lightning, noise, security and health concerns, social environment) as well as the current working equipment (hardware and software usability, improvement suggestions) (ISO/IEC 25063, 2014; Maguire, 2013; Nelles et al., 2016). Subsequently, the documented context of use is evaluated by shop-floor workers in a focus group. Non-confirming and irrelevant information is adjusted, and new information is integrated into a final context of use document. Whenever further relevant information is gathered in the following iterations, the context of use is adjusted accordingly.

The context of use analysis identifies the shop-floor workers as the primary users of the intended system. Due to increased product variety, their main issue is adjusting the machinery when encountering an anomaly in the production line. The adjustments tasks range from purely mechanical jobs, like tightening a screw, to information technology tasks, for example, the supervision of manufacturing execution computers. The capabilities to handle the anomalies vary greatly depending on the experience and qualification of the employee. Therefore, the company aims to make knowledge of more experienced workers available to inexperienced employees. It is emphasized – again – that the goal is not to create a knowledge repository but a system that acts as a collaboration space where the users can integrate feedback and update and expand upon the currently available knowledge.

3.2 Specify user requirements

The gathered material is structured and synthesized to derive user requirements from the previously conducted context of use analysis. As ISO 9241:210 (2010) states, the user requirements should be put in relation to the business objectives and the intended context of use. To do so, the aforementioned business objectives are structured by the three categories of usability measurement provided by ISO 9241:11 (see Table 4). From this perspective, effectivity mainly focuses on improving the anomaly handling success rate, resulting in a more precise, complete, and appropriate troubleshooting process. Efficiency is concerned with the shop-floor resources (e.g., time, human effort, material resources) allocated to troubleshoot an occurring anomaly. The goal is to reduce labor time, and human effort spent to fix an anomaly, thereby increasing the ratio of error-free production time. The primary purpose of the satisfaction category is to ensure long–term use from the business perspective. These objectives set the frame for the specification on user requirements and are of significant concern in the evaluation section.

Table 4 ISO 9241–11 concerning the primary business objectives.

The user requirements depicted in Table 5 are based on the sequential user tasks necessary to achieve the goal of rectifying an occurring anomaly. The order of user tasks represents the knowledge inquiry process among shop-floor workers. More precisely, it mimics the inquiry process of an employee who does not know how to solve an anomaly and is seeking in-person support from a more experienced employee. During the execution of the externalization procedure, observations showed that the process usually starts with categorizing the anomaly. This is an initial vague description of the encountered anomaly category (e.g., Paint and varnish anomaly, c.f. Table 3) by the knowledge-seeking employee. To further specify the anomaly category, a refinement takes place by defining the visual anomaly appearances through the shop-floor worker in need. Therefore, the derived requirement for the system is to integrate this structure in the anomaly selection process. In the next step, the experienced employee needs to develop possible solution measures that could potentially rectify the described anomaly — followed by a more in-depth description of the most appropriate solution measure. The knowledge-seeking employee is then in the position to execute the description and to evaluate whether the measure resulted in solving the problem or not. This can then be followed by other possible measures provided by the more experienced employee. If none of the described measures fix the occurring anomaly, the employees collaborate to solve the occurring anomaly. This is considered to be the task ‘come up with solutions to new anomalies’, where no currently present employee is knowledgeable enough to come up with an appropriate measure without ‘trying out’ a new solution measure procedure. Therefore, the system needs to provide a feature for integrating this newly created practice/knowledge.

Table 5 User tasks and derived requirements.

3.3 Produce design solutions

The design solutions are created based on the material from the previous steps as well as a literature research and competitor analysis. The personas integrate the sequence and timing of tasks and aid the information architecture (Nelles et al., 2016). The user dialogue and the graphical user interface design are derived from the specified requirements. Based on an initial conceptional interaction design, a first prototype is created. The first prototype is built as a system mock-up as described in Camburn et al. (2017). This method allows rapid assessment of essential system characteristics while possibly stimulating team discussion. The mock-up is represented in Microsoft PowerPoint and visualizes the system's necessary interaction behavior and knowledge structure. The embodied knowledge is based on the externalized tacit knowledge presented in Sect. 2.1. Figure 3 represents the knowledge structure divided into anomaly knowledge and the appropriate anomaly solution knowledge. Anomaly knowledge is further categorized as anomaly category and visual anomaly appearance. Solution knowledge is specified to concrete solution instructions. The mock-up integrates parts of the previously externalized knowledge to demonstrate the intended information structure.

Figure 3.
figure 3

Exemplary knowledge structure of the assistance system.

Considering the evaluation of the first prototype mock-up and the adjusted requirements, a second prototype is developed. The second prototype is realized as an HTML-based 'click-dummy', a prototype that looks like the goal system but does not provide any functionality (Sadabadi, 2013). The prototype's basic workflow consists of three main features and is based on the knowledge structure shown in Figure 3. The user interaction starts by specifying the anomaly in question by selecting the anomaly category. When a shop-floor worker notices a manufacturing anomaly, he/she categorizes them into a parent category to make a rough pre-selection about the anomaly at hand. The second step is the refinement of the detected anomaly by choosing the anomaly appearance, which is based on a visual representation and further short textual descriptions. The anomaly elimination process begins in a final step, and solution knowledge is provided. A ranking of possible anomaly solutions is made available to the user based on user feedback. The feedback relates to whether the solution knowledge to eliminate the anomaly was found to be helpful in the past or not. A short and easy-to-understand textual instruction is provided when selecting a possible solution to an anomaly. Practical tips and visual material enrich this instruction, like pictures or videos. To further assist in locating the anomaly occurrence, the affected production machine is highlighted in a pictorial representation of the production line.

The third prototype is a fully functional prototype integrated into the production line control device and is accessible via multiple displays across the production line. Clicking on an 'Assistant’-button takes the user to the prototype's main dashboard (see Figure 4). Section 'A' gives access to simplified process documentation based on machine groups in the depicted production line. This allows for a more profound troubleshooting process if the provided solution knowledge does not lead to a satisfactory anomaly resolution. The feature shown in section ‘B’ describes a dynamic search function to access solution suggestions quickly. This is especially helpful when assigning a detected anomaly to the presented anomaly category is not possible for the shop-floor employee. Section ‘C’ represents the main feature of the prototype, the anomaly solution process. As described before, the user first specifies the detected anomaly category, selects the visual anomaly appearance, and finally gets anomaly solution knowledge provided (see Figure 3).

Figure 4.
figure 4

Main dashboard of the functional prototype.

After carrying out the possible solutions for an anomaly, the shop-floor worker is encouraged to provide feedback regarding the suggested solutions. Therefore, users can integrate binary feedback, which describes the success probability (solution helpful yes/no) of the shown solution to influence the displayed knowledge's ranking order. Additionally, a user can provide textual feedback, which can be used to expand the system's knowledge base. The textual feedback includes supplementing existing solution proposals and suggesting new solution proposals by a shop-floor employee. To ensure process conformity, a quality keeper must review feedback and, ideally, add the new solutions to the knowledge base. The feedback features make the system dynamic by continuously updating the knowledge structure (binary feedback) and content (textual feedback).

Figure 5 shows the interactions between the assistance system and the described free-text feedback module. The application database integrates the initial externalized knowledge through semantic rule sets, which users can access with the functionalities described above. To minimize the burden on the quality keeper to ensure the free text comments' process conformity, a natural language processing (NLP) component is added to the system. Thus, pre-processing is performed on the free-text comments by assigning them to semantic clusters. This ensures that comments with the same or similar suggestions are already combined, and the quality keeper only has to check these clusters of comments. The entire NLP process structure is based on the clustering process of Halkidi et al. (2001), whereby the k-means algorithm is used to form the semantic clusters (Kanungo et al., 2002; Lee, 2019).

Figure 5.
figure 5

Overview of dynamic feedback integration function.

As shown in Figure 5, user input as free-text comments about new solution knowledge is forwarded to the NLP component. It includes the addressed anomaly category and its description, a timestamp, and the provided feedback. This allows for context assessment of the provided feedback in later stages. The NLP component then filters the input, clusters the comments into semantically similar groups and temporarily stores them in a shared database. The accumulated results can then be displayed through the system's backend by an authorized process administrator. The process administrator, who acts as a quality keeper, validates and verifies the clusters before entering them into the application database. This validation is of particular importance in manufacturing since non-process-compliant knowledge could lead to severe errors and even injuries. The presented semi-automated feedback functionality ensures knowledge quality, continuous learning in the system and is the primary facilitator of an ongoing knowledge externalization process. Table 6 summarizes how the produced design solution targets the previously established user requirements (c.f. Table 5).

Table 6 User requirements and corresponding assistance system features to support knowledge sharing on the shop-floor.

3.4 Evaluate design solutions

The first prototype is evaluated by users applying heuristic evaluation (Nielsen and Molich, 1990), using the dialog principles of ISO 9241:110 as usability heuristics. It reveals a need for a visual representation of the workspace (production line) to assist navigation. Furthermore, the repeated expression of the users desires to integrate the assistance system into already used monitoring interfaces. The second prototype evaluation uses cognitive walkthrough with experts (John and Packer, 1995) and cooperative evaluation with shop-floor workers (Monk et al., 1993). Here, requests for enabling semi-automated feedback are proposed, which can potentially provide a steadily growing knowledge base while keeping the operator workload low.

However, the current state of the assistance system is the third prototype where a more comprehensive evaluation is aimed for. As data for a possible usability evaluation is not available yet, the paper focuses on a productivity impact evaluation. As Hold et al. (2017) observe: ‘The primary objectives of [digital assistance systems] are the increase of productivity’ (Hold et al., 2017). Keller et al. (2019) provide a set of key performance indicators (KPIs) to evaluate the productivity impact of a digital assistance system in assembly work stations. However, the application domain of the developed assistance system is a continuous manufacturing scenario. Therefore, certain KPIs are not feasible (e.g., a KPI based on ‘number of parts produced’). Such KPIs are adjusted accordingly (e.g., ‘number of parts produced’ is replaced by ‘total length produced’).

Table 7 presents the resulting adjusted KPIs to measure the productivity impact of an assistance system in a continuous manufacturing scenario. On the workplace-level the ‘Production Gradient’ (PG) describes the Production Time (article being produced) in relation to the Operating Time (production line being active). The Article Failure Time in ‘Inverse Article Failure Time Gradient’ (IAFG) is the time of faulty articles being produced (e.g., pattern of produced edge-band not OK). Machine Failure Time in ‘Inverse Machine Failure Time Gradient’ (IMFG) is related to machine malfunction durations (e.g., an error occurs in the material infeed). It is worth mentioning that production may continue during ‘Machine Failure Time’. The sole purpose behind taking the inverse in IMFG and IAFG is to allow for interpretation consistency. This way, an increase in IMFG and IAFG indicates a productivity improvement, which is in line with the other presented KPIs. The ‘Availability Rate’ (AR) is the actual operating time concerning the scheduled operating time on the process-level. This KPI aims to identify potential downtime losses due to break downtimes or setup and adjustment procedures (Nayak et al., 2013). ‘Performance Rate’ (PR) describes the real production rate concerning the theoretical or ideal run rate. By increasing the PR, primarily speed losses due to more minor stops and reduced speed are countered (Nayak et al., 2013). ‘Quality Rate’ (QR) is simply the length of products produced which pass the quality control concerning the total length of produced units. This KPI measures quality losses due to startup rejects or production rejects (Nayak et al., 2013). Finally, the ‘Overall Equipment Effectiveness (OEE) is the multiplication of the three KPIs from the Process-level (AR, PR, QR). The OEE can be considered as a primary guiding metric for decreasing and eliminating the most common causes of economic efficiency losses in the manufacturing process (Nayak et al., 2013; Keller et al., 2019). The OEE is therefore an appropriate measure for a holistic benefit evaluation of a digital assistance system in manufacturing (Keller et al., 2019).

Table 7 KPIs for the evaluation of an assistance system in continuous manufacturing.

In the first step of the evaluation procedure, a pre-/post-comparison is conducted. Measurements of KPIs are collected for a single production line. For 3 months (from 11.20 – 01.21), production line 48 was not equipped with the developed assistance system. This represents the pre-stage. The resulting average KPI performance during the pre-stage is depicted in ‘Line 48 No AS’ (see Table 8). Afterwards, also for a 3-month period (from 02.21 – 04.21), production line 48 was equipped with the assistance system. This represents the post-stage. The average KPI results of this time interval are presented as ‘Line 48 With AS’ in Table 8. The corresponding ‘Delta’, is the difference between ‘Line 48 with AS’ and ‘Line 48 No AS’. Therefore, it is the observed change in KPI performance when comparing the pre- and post-stage.

Table 8 Single production line pre-/post-comparison, without and with assistance system.

As showcased in Table 8, with the introduction of the assistance system, a performance increase for all measured KPIs occurred. However, there is some uncertainty whether this improvement might not be traced back to changes to other general conditions (e.g., seasonal differences, employee training, process improvements). To obtain insight into the assistance system's influence, an additional benchmark with a reference production line pool is conducted. This pool of three production lines was not equipped with the assistance system during the same two time periods as before (see Table 9). The production lines included in the pool are located on the same shop-floor, are equipped with similar machinery, produce the same article type, and have a similar production lot size. Table 9 shows the results of this benchmark. ‘Line 48’ is the assistance system test line, already portrayed in Table 8. ‘Pool’ indicates the average measurements of KPIs for the three production lines included in the reference pool. ‘Delta Total’ is the difference between the ‘Line 48 Delta’ and the ‘Pool Delta’. It describes the difference in performance change between those two.

Table 9 KPI change comparison.

As Table 9 shows, the Delta of Line 48 and Pool is positive among all KPIs. This indicates that a certain degree of performance improvement is indeed due to other general conditions. However, ‘Delta Total’ also highlights that the introduction of the assistance system increases production performance above average among all KPIs (compared to the reference pool). The lowest Delta increase is observed in IAFG, which can be explained due to its measurement. Error codes such as Article Failure Time and Machine Failure Time don’t co-occur in the production control system. Therefore, if not handled quickly and correctly, Article Failure Times might lead to Machine Failure Times. This is in line with the observation that the highest ‘Delta Total’ is achieved in the IMFG (+ 2.0). Since the IMFG indirectly measures how long machine failures persist, it is assumed that the assistance system supports employees in fixing anomalies that might cause or are caused by machine failures. Therefore, this assistance system support manifests in an accelerated and more successful anomaly handling process. The OEE, as a holistic KPI and the second-highest Delta Total achieved, indicates that the overall efficiency of the production line could be increased due to the introduction of the developed assistance system. Therefore, it is concluded that the introduction of the assistance system prototype does have a positive effect on productivity.

4 Discussion and outlook

4.1 Implications

The implications of this study are relevant to both research and practice. Considering the progressive change from shop-floor workers to knowledge workers in increasingly complex manufacturing environments, it is apparent that systemic knowledge-support for shop-floor employees in their daily tasks is needed (Hartmann, 2015; Riedel et al., 2017). The major contribution of this paper lies in conceptualizing and applying a knowledge externalization approach and combining it with the human-centered design of an assistance system.

Through the standardized approach for tacit knowledge externalization, employee-bound knowledge can be codified and shared time-independently. We outlined the feasibility of our approach through its practical implementation in a real manufacturing scenario, whereby 261 structured anomaly correction measures have been codified. Although various methods for knowledge externalization exist, a combination thereof is missing in the context of shop-floor practices. The lack of literature in this domain makes it hard to contrast our work with others. By presenting a methodological approach to externalizing knowledge, we hope to encourage other researchers to contrast their work with ours. To do so we point to several literature resources on the topic of knowledge externalization methods. Other researchers can thereby build upon, diverge from, and analyze our and previous methods from differing epistemological traditions.

By combining the knowledge externalization insights with the assistance system design, the systems interaction structure is mimicking the knowledge inquiry process among shop-floor workers. We hypothesize that this, in combination with the development and design of the system in close collaboration with shop-floor workers, warrants knowledge representation and knowledge retrieval tailored to the shop-floor requirements. Furthermore, by integrating the externalized knowledge into the system, we claim to provide initial usefulness. Those claims are supported by the positive effects of the system on production performance. In this regard we cannot pinpoint those effects to certain system functionalities or undermine them with statistical significance. A positive effect was hover observed in a 3-month pre-/post comparison and a simultaneous benchmark with similar production lines. Since little research has been done on cognitive assistance systems in the continuous production domain, we invite other researchers to further analyze the support requirements and opportunities for collaborative assistance in the continuous production domain. It also needs to be mentioned that other digital interventions can positively effect knowledge work on the shop-floor (see e.g., Hannola et al. (2020)).

Moreover, the system integrates employee feedback and knowledge sharing with a subsequent semi-automated feedback processing to enable ongoing collaboration among shop-floor workers and a continuously updated knowledge base with minimized maintenance effort. The aim is to thereby encourage system use and acceptance, potentially fostering collaborative knowledge exchange. However, we cannot yet make claims about the long-term collaborative effects of our system. In this regard it needs to be mentioned that peer-tutoring systems could for example foster collaborative knowledge exchange equally well. According to Johansen's (1988) CSCW-Matrix, our system supports collocated, but time-independent collaboration. Therefore, it should be emphasized that investigations on the topic of same time and different place groupware could also lead to equally beneficial results.

There are several factors which support the validity of this design science research approach (on the validity of design science, see for example Hevner et al. (2004)). Since shop-floor workers require additional knowledge support, utility is provided through a system which allows the workers to access and share their anomaly correction knowledge. Through the systems design they gain a common information space to time-independently collaborate on solving anomalies, which is especially crucial for less experienced employees. The novelty of the created artefacts (externalization method, assistance system) lies in the combination of already existing methods, tailored to the context of continuous manufacturing and the practical implementation thereof. The claims of utility are supported by the outcomes of the externalization method (261 structured anomaly correction measures) and the created system (positive effects on productivity). As mentioned before, we did not finish a usability evaluation for the system. Claims of ease of use can therefore not be supported through data. We believe however, that the human-centered design in collaboration with shop-floor workers promotes usability of the system. In summary, the possible advantages of the externalization and provision of knowledge through the developed assistance system are diverse. They include, among others, a faster training period for new employees, increased process stability due to faster anomaly correction, and the depreciation of scrap production and thus costs.

4.2 Limitations

This paper's results must be considered under certain limitations that need to be addressed in future studies about the results' generalizability. First, the developed method for knowledge externalization has only been carried out in a single real manufacturing scenario. Prospective studies must be made in different manufacturing scenarios to provide further generalizability. Second, as the knowledge database is continuously updated and evolves with the users' knowledge through feedback integration, it is crucial to implement incentive structures for users to share their knowledge. Gamification is a suitable approach for knowledge-sharing incentive design (Friedrich et al., 2020). Integrating gamification features, such as rewardability, visibility, and competition incentives, will be part of future research (Ayoung and Christian, 2017). Third, the knowledge-based assistance system's evaluation mainly aims to affect the overall productivity performance positively. Once sufficient usage data and experience are available for the third iteration of the prototype, it is considered necessary to carry out a usability evaluation. Finally, the presented assistance system is conceptualized as groupware and focuses on technological research. Consequently, the social, psychological, and organizational implications are not the focus of this paper and have therefore not been analyzed comprehensively (Rama and Bishop, 2006).

4.3 Conclusion

The systemic knowledge support of shop-floor employees in their daily tasks constitutes a significant concern in operations research and practice. This paper presents a novel approach to leverage anomaly elimination knowledge by developing a standardized knowledge externalization methodology. The subsequent design and prototypical development of a knowledge-based assistance system creates an operationalization and collaborative reuse of externalized knowledge. Integrating a semi-automated user feedback function ensures that the knowledge base is always up to date. It is applied in a continuous, real manufacturing scenario to determine its relevance and practical suitability for the externalization method. The evaluation results confirm the developed methodology's suitability by the quantity and quality of externalized, tacit knowledge. The productivity impact evaluation of the assistance system confirms an increase in production efficiency. However, the long-term practical applicability needs to be discussed in further studies.