Towards a Cognition-Based Framework Describing Interdisciplinary Expert Team Processes for Cognitive Robotics in Industry 5.0 Technologies

Abstract

With the evolution of traditional production towards smart manufacturing, humans and machines interact dynamically to handle complex production systems in semi-automated environments when full automation is not possible. To avoid undesirable side effects, and to exploit the full performance potential of experts, it is crucial to consider the human perspective when developing new technologies. Specifically, human sub-tasks during machine operation must be described to gain insights into cognitive processes. This research proposes a cognition-based framework by integrating a number of known psychological concepts. The focus is on the description of cognitive (team) processes in the resolution of anomalies within a manufacturing process with interdisciplinary experts working together. An observational eye tracking study with retrospective think-aloud interviews (N = 3) provides empirical evidence for all cognitive processes proposed in the framework, such as regular process monitoring and—in case of a detected anomaly—diagnosis, problem solving, and resolution. Moreover, the role of situation awareness, individual expertise and (cognitive) team processes is analyzed and described. Further, implications regarding a human-centered development of future production systems are discussed. The present research provides a starting point for understanding and supporting cognitive (team) processes during intelligent manufacturing that will dominate the production landscape within Industry 5.0.

1. Introduction

Intelligent manufacturing in the era of Industry 5.0 is characterized by three key points: (1) sustainability and resource-efficiency, (2) agility and flexibility by intelligent production, and (3) strict human-centeredness (e.g., [1,2]). An important requirement for the successful transition from “traditional production” (i.e., large quantities, low product variations/flexibility, objective of high automation levels) to “human-centered, smart manufacturing” (i.e., flexible adjustment to current order/resource challenges, medium levels of automation with close and efficient human–machine collaborations) is the convergence of cyber-physical production systems (CPPSs; e.g., [3]) and humans (H) to human-cyber-physical production systems (HCPPSs; [4,5,6,7,8]). CPPSs integrate the cyber-linked computational level (C-system), which is increasingly based on artificial intelligence (AI), with physical-technical systems and processes (PP-system). That means that “embedded computers and networks monitor and control the physical processes, usually with feedback loops where physical processes affect computations and vice versa” [9] (p. 1). Thus, the physical component (e.g., machine) has a twin in the cyberspace in which data of the physical component (e.g., generated from sensors or human input) are stored and processed [3]. This reliable interconnection between the physical and cyber-world is a key challenge for flexible and agile production and needs to be established for concrete technical processes, such as forming, coating, or welding [6]. Technical developers that aim to address the human subpart in a suitable and harmonic way (i.e., human-centered) need to integrate the H-perspective right from the start. This can be achieved, for instance, by (1) analyzing tasks that might be cyber-automated or supported by intelligent robots in a systematic and psychologically plausible way (e.g., cognitive task analysis methods; procedures for the elicitation of human expert knowledge; [10,11]) and by (2) choosing AI algorithms that correspond to the given level of human consciousness [2,12] to align AI to operators’ cognitive processes [7,13]. This shift in technical development from a predominantly technical orientation to a joint human-technology orientation is necessary and crucial for Industry 5.0, as the human-CPPS relationship will significantly change human work tasks. While simple routine tasks involving rule-based activities have a high automation potential [14], most complex cognitive and manual tasks that require individual and situational problem solving are currently not fully substitutable by AI [15]. The ability of creative and flexible, context- and situation-specific problem solving, i.e., overcoming obstacles or closing gaps in the action plan to reach a certain goal (see, e.g., [16,17]), is a characteristic strength of humans as it requires a semantic association between observations, experience and knowledge [18]. The transferability of these complex cognitions to CPPSs is limited due to the strong context dependency [19]. Furthermore, in complex and dynamic manufacturing environments, the operator must process data that are often hidden or uncertain, vary over time and are not limited in number and interaction [20]. In future decentralized production systems, operators will interact dynamically to handle various (semi-)automated systems, while purely physical work will decrease [21]. Therefore, the cognitive tasks of operators, such as monitoring (including problem solving and decision making), will increase. Further, the monitoring of complex and dynamic production systems might require experts from different disciplines to detect errors at an early stage and ensure quality standards (e.g., [22]). For all manufacturing technologies and tasks that are not completely automatable, these challenges need to be met. Collaborative human-technology systems and intelligent robots have to develop in order to support humans’ cognition and manual work as well as the interaction in interdisciplinary expert teams. To avoid losing the human perspective and triggering undesirable side effects, such as cognitive overload, stress, or mistrust in automation, human aspects in technology development have to be involved consistently and appropriately. For this, it is necessary to describe tasks and sub-tasks in detail to gain insight into the cognitive processes of humans as well as interaction processes within the expert team.

The objective of the present paper is to use an exemplary manufacturing process—incremental robotic roll forming (“RoRoFo”)—to show how human centeredness can be implemented through transdisciplinary technology development (the HCPPS approach). We are concentrating on the human-centered aspect of Industry 5.0, as there is an urgent need here and comparatively little research has been carried out to date. To this end, we explicitly analyze higher cognitive capabilities of the operating/interdisciplinary expert team, such as monitoring, detection of anomalies, diagnosis, joint problem solving and team interactions. We address the research question of how cognitive team processes can be described during the supervision of a complex and dynamic manufacturing technology that requires experts from different fields. This is performed because describing is a precondition of explaining and predicting human behavior, which is, in turn, necessary in order to:

• Develop and design new productions systems, such as HCPPSs, by representing and integrating cognitive processes into such systems (in the long-term) (see, e.g., a similar approach for thermal spraying technology in [2]).
• Realize the full potential of expert teams by developing and designing (cognitive) support possibilities for the human team members (in the short-term) [21].

As the RoRoFo process is currently being developed by an interdisciplinary expert team, it is a good example for many other immature technical processes in the manufacturing area.

The paper is structured as follows: After a brief introductory presentation of central findings concerning human expertise and cognition, we propose a cognition-based framework for interdisciplinary expert teams developing technical processes (Section 2). This enables us to specify one facet of “human-centeredness” in the era of Industry 5.0. Thereafter, we show how this perspective can be used for the development of a specific manufacturing process. Based on a RoRoFo study that was accompanied by cognitive and human factor experts (Section 3), the cognition-based framework and the results of eye tracking and interview analyses are used to formalize the technical process concerning cognitive aspects (Section 4). From this, we derive the skills that an intelligent robot would have to learn in order to be able to offer human-oriented assistance services, discuss results, reflect on limitations, and summarize future prospects (Section 5).

2. Human-Centeredness in Intelligent Production

For the planning and implementation of hybrid human-technology development paths, it is necessary to specify the assignment of functions based on the complementary strengths of humans and cyber-physical parts [18]. For example, due to their capabilities, intelligent robotic systems can perform certain tasks that humans cannot (easily) perform or cannot perform without stress or danger. The functions to be assigned can relate either to physical capabilities (e.g., lifting heavy loads) or to cognitive capabilities (e.g., path planning). The basis for the implementation of specific abilities in CPPSs is to analyze and describe the starting point, the target state, and the possible courses of actions or operations that are necessary for task execution. For physical tasks, the technical solution ultimately implemented may differ (significantly) from the human solution. However, for cognitive tasks, the technical solution must still be understandable and explainable in terms of the functions to be supported or automated, as cyber-intelligence is not as easily “visible” as physical capabilities, such as lifting loads. Even if the actual (AI-based) algorithmic solution does not imitate the human solution path exactly, cognitive abilities (specifically human information processing) need to be used as a guide to create transparent benefits for the human-technology interaction and to ensure human-centeredness. Furthermore, human-centeredness in production means that the real-world tasks (and not only virtual ones) are addressed, as a significant part of value creation in manufacturing is linked to real technical processes and produced goods.

2.1. Human Information Processing and Cognition

Human situation awareness and behavior rely on human information processing and cognition. Models of cognitive information processing describe cognitive stages in varying degrees of fineness and can be used to analyze single (cognitive) processes when interacting with dynamic systems [23]. As key components the (1) perceptual-, (2) cognitive-, and (3) action-related levels can be distinguished [24] (see Figure 1).

The perceptual stage includes the detection, discrimination, and identification of stimuli, which is in principle comparable to technical recording of sensory information and parts of information analysis (see Figure 1, left). The cognitive stage comprises more in-depth processing. Based on the perceived input, prior knowledge is retrieved from memory and integrated into a mental representation of the current situation in working memory, which is the basis for problem solving and decision-making processes. Finally, within the action stage, responses are selected, prepared, and activated [23,24,25].

Figure 1. Integrated information processing models from cognitive psychology and human factors (adapted from [24,26]) including different processing stages and specifications of cognitive subprocesses for “anomaly resolution” in human–machine interaction processes.

Figure 1. Integrated information processing models from cognitive psychology and human factors (adapted from [24,26]) including different processing stages and specifications of cognitive subprocesses for “anomaly resolution” in human–machine interaction processes.

All three phases include some challenging processes in the work tasks of intelligent manufacturing that should be considered in the technical development in order to avoid unintended side effects and to promote good assistance and automation solutions. Specifically, during the cognitive stage, requirements for humans’ cognitive systems are high, which often leads to constraints in human performance [24]. For instance, if the interpretation of gathered information shows that a certain goal cannot be reached under the current circumstances, problem solving and decision-making are necessary. To solve a problem, its cause(s) must usually be identified and eliminated, which is a challenge in complex real-world applications due to the information required often not being immediately available and needing to be collected over time [27]. A number of prescriptive and descriptive models are discussed in the literature for describing and understanding complex cognitive processes, such as problem solving (e.g., [28,29]). In technology design and development processes, problem solving is a very common sub-task allocated to humans. It is driven by the following cognitive schemata (see Figure 1, middle): Once a dysfunction (i.e., anomaly) is detected during regular process monitoring, the discrepancy between the current state and goal has to be determined. Then, the possible causes for this discrepancy are diagnosed in order to be eliminated; therefore, a set of action alternatives is generated and evaluated before, finally, the best action alternative is chosen [29]. Using this cognitive schema is one key factor of successful problem solving [30] and, therefore, anomaly resolution.

The quality of solutions and action alternatives is determined by domain-specific schemas. This is in line with research on expertise suggesting that domain-specific knowledge furthers the understanding of new information (i.e., problem identification and diagnosis) but also affects information search (perceptual-stage) and action execution (action-stage) (e.g., [20,30]). These rather holistic approaches of human information processing and problem solving when interacting with dynamic systems can be used to understand the successful process control of complex and dynamic technical processes of human experts.

The competence level of the humans in interdisciplinary expert teams is crucial and must be considered in two ways when realizing human-centered Industry 5.0: (1) in the technology development phase, the cognitive information strategies of experts must be analyzed to ensure that the knowledge to be represented in the intelligent assistance system is correct and as complete as possible; (2) the competence levels of future operators must be anticipated to ensure the maximum suitability or adaptability of the system.

2.1.1. Characteristics of Expert Performance

Experts are humans with exceptional knowledge and skills in a certain domain (area of knowledge) that enable excellent performance [31]. They have deep and wide domain-specific knowledge [32,33], which results in a deep-level problem representation [20,34] and understanding of causal relations [20,35]. Domain-specific features are structured in long-term memory [36] and organized in hierarchies of chunks, i.e., a larger information unit (see, e.g., [37]) as well as single information elements, all within a broad knowledge network. Due to long-term experience, the knowledge networks’ size, stored chunks, and interrelations increase [33]. Only a “pointer” to the chunk remains in working memory [38], relieving the time-and-space-limited working memory and enabling efficient problem solving. In sum, experts are able to make fine discriminations, understand how things work, can perceive large patterns of information, have routines and tactics, know more facts, are able to run mental simulations, and are also able to spot anomalies and detect problems by spending more time on situational analysis [39].

Empirical studies in the context of engineering (e.g., on robot programming) highlight the importance of human expert knowledge by analyzing and formalizing expert strategies compared to novice approaches [40]. Expertise is related to a more efficient information search behavior, indicating that the expert is better able to process domain-specific information, resulting in faster task solving and fewer errors. Moreover, ref. [20] reviewed the role of expertise during supervision tasks in dynamic environments and found that expertise is crucial for monitoring, diagnosis, and decision-making processes, as experts are more able to produce inferences, anticipate situations, and look at situations more globally and functionally than novices (i.e., they have better situation awareness).

2.1.2. The Role of Situation Awareness

Expertise is strongly associated with situation awareness (SA), which can be defined as “the perception of the elements in the environment within a volume of time and space, the comprehension of their meaning and the projection of their status in the near future” [41] (p. 97). There are many cognitive mechanisms that are important for developing and maintaining SA which are influenced by expertise, including attention (i.e., attention must be directed to meaningful environmental cues to enable goal-directed information search), working memory processes (i.e., perceived information must be integrated and compared with other information and projected into the near future), and long-term memory processes (i.e., knowledge in the long-term memory must be activated to be available for retrieval and processing in the working memory) [42]. It has been found that SA is important for supervision tasks in different application fields [43], including, for example, human–robot collaboration [44], as well as process or air traffic control [45,46].

2.1.3. Characteristics of Cognitive Team Processes

Working tasks are often not executed by a single human operator; rather, they are executed by operator teams with different professions, areas of expertise, and roles throughout the task. Therefore, the understanding of knowledge distribution among team members, roles, and team interaction processes are relevant and must be described and understood in order to fully account for the human perspective in future production systems. According to [47], expert teamwork is characterized by an interdependent interaction of the team members to reach shared goals. These are known to be crucial criteria for human–machine teaming, which is an advanced form of human-centered assistance and automation [4,8,48,49]. Thereby, collaboration strategies are adapted using coordination, cooperation, and communication. Moreover, the team must have a collective understanding of the task. Due to the heterogeneous composition of teams, including experts from different fields, the team has a broad pool of knowledge consisting of individual knowledge and knowledge overlapping among team members [50]. Individual knowledge or internalized team knowledge includes specific, individual expertise as well as individual knowledge about the team structure and the expertise of other team members, whereas externalized team knowledge is the common knowledge basis shared and accepted by all team members (also containing an aggregation of knowledge [51]). Therefore, shared mental models play a significant role within teams [52]. A mental model can be defined as a knowledge structure that contains a representation of system functions, including their influencing factors (see, e.g., [53]). That is, team members must predict what actions the others will undertake and what they need for a successful performance [52]. Two types of shared mental models exist: task work mental models, which is the knowledge of the team members regarding the task, relevant equipment/systems and their applications, and teamwork mental models, which is the knowledge of the team regarding the others’ tasks, their knowledge, skills, and action routines [52]. A high fit of these models between the team members ensures an efficient knowledge exchange [52,54]. Based on the shared mental model concept, SA can be developed by a team [55]. Team SA is described as the degree to which each team member has an accurate knowledge or SA of shared items [55]. Thereby, team SA is the current state of the shared mental models [42].

2.2. A Cognition-Based Framework Describing Interdisciplinary Expert Team Processes

So far, there is no holistic framework including different cognitive concepts, such as expertise, SA, and cognitive team processes, influencing the supervision process in complex and dynamic production environments. Therefore, a cognition-based framework was developed based on the summarized findings (see Figure 2). With reference to general models of human information processing (see Figure 1), our framework specifies the cognitive processing phase for interdisciplinary teamwork and focuses on the cognitive subprocess “anomaly resolution”. This is a frequent and relevant sub-task in technical design and development processes, as well as in daily machine operation.

More precisely, the framework systematizes the interplay of the cognitive processes (a) process monitoring, (b) anomaly detection and diagnosis, and (c) problem solving and resolution in their temporal development (see Figure 2, middle) within the context of a team interaction. It considers influencing factors, such as context, team processes, as well as individual expertise and SA (see Figure 2, circles at left and right side). Monitoring is a central task during supervision processes, including regular system control and the detection of signs for dysfunctions in the form of anomalies [20]. Thus, within (a) regular process monitoring (see Figure 2, middle), information is detected and possible current and/or predicted discrepancies between current and goal states are determined [24,29]. If no anomaly is detected, the human operator remains in the monitoring process. However, if an anomaly occurs, a specific diagnostic process is required (see (b) anomaly detection and diagnosis in Figure 2). This means that the possible causes of the deviation between the actual and target states need to be determined on the basis of individual expertise [20,29]. The results are mental problem models of each team member [20,23,51], which leads to a set of individual action alternatives [29] that might subsequently be exchanged and evaluated within the team (see (c) problem solving and resolution in Figure 2). Finally, the best action alternative is chosen jointly [51]. Due to collective learning within the problem solving process, the acquired (team) knowledge can be used for subsequent problem solving processes [50].

There are various cognitively inspired models and experiments that investigate cognitive subprocesses in more detail, under controlled laboratory conditions, and with specifically constructed material (e.g., [27,56,57]). Even though many results are interesting and potentially transferable, the aim of basic research approaches is not primarily geared towards real-world applications. Therefore, with regard to system engineering in the digital age [58], we have chosen an industry-oriented application environment to evaluate our cognition-based framework.

3. Method

In order to describe the cognitive processes during anomaly resolution in a complex and dynamic production area based on the developed cognition-based framework, an observational eye tracking study with retrospective think-aloud interviews is conducted. The study has an explorative character.

3.1. Incremental Robotic Roll Forming

For the present study, incremental RoRoFo was chosen as example manufacturing process (see, e.g., [59,60]). It represents a forming process to produce angled profiles (in L, C or Z shape) from plain metal sheets. Therefore, a pair of rollers, which are guided by a robot along the planned bending edge of a metal sheet, induces bending stresses into the sheet, thus forming it incrementally (see Figure 3, right, from [60]) (meaning that the sheet is formed incrementally). Compared to common forming processes, RoRoFo is especially suitable for fast, agile, and economical production of small batch sizes (e.g., in the aircraft industry) and, hence, is related implicitly to sustainability and resilience in production. Because the process is quite complex, immature, and hence, not yet completely controllable, RoRoFo development is currently largely based on human expertise concerning the different sub-tasks as well as the overall task integration to be solved (see Figure 3, left). At present, the human developer team takes all decisions with almost no assistance from a cyber-component. The whole process can be distinguished into five sub-processes (or sub-tasks) that interact with each other:

• Manufacturing order (i.e., based on a component drawing a suitable forming method and tool configuration is selected);
• Analytical definition/control data (i.e., based on the target geometry of the work piece a suitable forming strategy, i.e., analytical curves procurement, is chosen);
• Teach-In/ robot control (i.e., the robot is programmed/taught to realize the planned curve);
• Forming process (i.e., the prepared work piece is angled gradually by the robot);
• Quality control (i.e., comparison of current and target-value of the component).

Figure 3. Set-up of the forming process with the operators/experts wearing eye tracking glasses (left) and the robot forming a prepared work piece (right, from [60]).

Figure 3. Set-up of the forming process with the operators/experts wearing eye tracking glasses (left) and the robot forming a prepared work piece (right, from [60]).

Within the present study, sub-task 4, i.e., the forming process, was selected for a more detailed investigation. Currently, several experts with different types of domain knowledge interact with each other to monitor the forming process, detect anomalies in due time, and solve them in order to achieve components of sufficient quality.

3.2. Participants

A well-established team of three qualified operators (one female, two males, aged between 32 and 42 years) took part in the study. The operators are employees of the Fraunhofer Institute for Machine Tools and Forming Technology (Chemnitz, Germany), hold a university degree in different areas (i.e., mathematics, engineering), and have professional experience of between 4 and 16 years.

3.3. Procedure

After a short welcome and calibration of the eye tracker, the examiner gave instructions regarding the aim and procedure of the study and the concrete task of the operator team. The task was to form a (within sub-tasks 1 to 3 already prepared) work piece using incremental RoRoFo. This included loading the control data, testing the fit between the planned curve of the machine tool and geometry of the work piece, conducting empty run(s) for collision control, fixating the work piece into the clamping device, forming the work piece robot based (including partially releasing clamping points to prevent collision with the rolls and fixing them after passing again), and evaluating the forming result subjectively. This should be performed within a maximum of two hours.

Once the task was executed, video-based retrospective think-aloud interviews (see, e.g., [61,62]) were conducted by two interviewers with each operator separately. The interviews lasted about 90 min. Initially, the operators were asked about their own task, role, and expertise within the forming process, as well as about the task, role, and expertise of the other operators. Afterwards, the operators watched their own scene video captured by the eye tracker (operator 1 and 2) or the camera (operator 3). The operators were instructed to explain all sub-tasks and their respective sub-goal, as well as all action steps and their underlying intentions. The interviewers asked specific questions when the operator reported an anomaly (i.e., critical event) that could have an impact on the quality of the forming process or result. The specific questions were:

• What information do you receive?
• Where does the information come from (i.e., how do you or (one of) the other operators recognize the anomaly)?
• What are the potential causes for this anomaly?
• What are the potential subsequent consequences of this anomaly?
• How could the problem be solved? What are the pros and cons?
• Which action alternatives were chosen? And why?

3.4. Material

The entire forming process was recorded with a Garmin VIRB ULTRA 30 GoPro that was positioned above the entrance of the robot cell, keeping all three operators (as well as the technical set-up) in the picture. The camera recorded with a resolution of 1920 px × 1080 px and a sampling rate of 25 pictures/second. In addition, two of the three operators that were mainly involved with task execution (i.e., robot control and controlling the forming process) were equipped with mobile eye-tracking devices. Gaze data of operator 1 were sampled with binocular Tobii Pro Glasses 3 with a rate of 20 Hz, using an integrated full HD scene camera that covers a field of view of 106° (H: 95°, V: 63°). Gaze behavior of operator 2 was recorded with an eye tracker from SensoMotoric Instruments GmbH with a sampling rate of 60 Hz. A scene camera recorded surroundings with a resolution of 1289 px × 960 px. Both eye-tracking devices also recorded audio with integrated microphones. Excel 2016, iMotions 9.3, as well as Python (package Matplotlib; [63]) and IBM SPSS statistics V.28, were used to analyze gaze data such as fixation count, fixation duration, fixation proportions, and fixation sequences. Moreover, retrospective video-based think-aloud interviews were conducted based on a short interview guide (see Section 3.3) and recorded by a dictation device.

3.5. Analysis

Gaze data of both eye-tracking devices were synchronized and imported into iMotions. Areas of Interest (AOIs) were defined as those areas/objects that the three operators reported as sources of relevant information in order to maintain SA throughout the process. Operators mostly shared relevant AOIs, and only small differences arose from the different tasks and responsibilities (see Figure 4). For instance, operator 1 was in charge of operating the robot, and therefore held its control panel. In contrast, operator 2 was in charge of documenting the process, resulting in corresponding looks. Fixations were then mapped to these AOIs over the entire process. In order to compare gaze behavior of the operators, we used fixation proportions. As the eye-tracking devices delivered unequal data quality, proportions were calculated as summed fixation durations toward specific AOIs divided by the sum of all relevant fixations of this specific operator during a defined period of time. The recorded interview data were transcribed, summarized, and categorized, meaning that statements regarding (a) the distribution of roles and tasks within the forming process, (b) the sub-tasks within the forming process, and (c) the observed anomalies and subsequent problem solving steps were summarized into predefined categories, which were derived from the developed cognition-based framework (see Figure 2).

4. Results

4.1. Systematization and Formal Description of the Forming Process

As an information search is goal-directed (see Section 2.1.2), it is necessary to structure the forming process into sub-goals or sub-tasks; otherwise, it is not possible to determine the causes of operators’ information search behavior. Video and interview data revealed seven different sub-tasks with different sub-goals (see also bottom of Figure 4):

• Definition of the starting point (sub-goal: determining an optimal starting point of the robot, right in front of the first touching point between the set of rolls and—later—the sheet metal);
• Kinematic configuration (sub-goal: controlling the fit between the implemented curve progression of the machine tool guided by the robot and the work piece without work piece contact);
• Setting the starting point/teach-in (sub-goal: transferring the specific starting point coordinates to the robot);
• Empty run(s) (sub-goal: excluding collisions of the robot with the work table and clamping device during the forming process without a clamped work piece);
• Resetting the clamping device (sub-goal: optimizing the clamping device after checking the actual curve progression of the machine tool guided by the robot);
• Fixating the work piece (sub-goal: fixating the work piece correctly into the clamping device);
• Forming the work piece (sub-goal: forming the work piece gradually by the robot according to the predefined target geometry).

These sub-tasks were mentioned by all operators during the interviews consistently and were characterized by clear breaks/pauses between the sub-tasks. The sub-tasks lasted approximately between 2.50 and 13.45 min—specifically: 1. defining the starting point: 2.50 min; 2. kinematic configuration: 4.00 min; 3. setting the starting point: 6.45 min; 4. empty run 1: 4.45 min, empty run 2: 5.00 min; 5. resetting the clamping device: 3.50 min; 6. fixating the work piece: 6.00 min; 7. forming the work piece: 13.45 min (note that eye tracking data recording stopped after 7.00 min for operator 2). The the sub-task 4, “empty-run”, was conducted twice as, halfway through the sub-task, operators detected a high collision potential with the clamping device. Therefore, the clamping device was reset before the empty run was executed a second time to test the adjusted set-up.

In the following, the roles and responsibilities of the operators during the whole forming process are reported. Then, the monitoring process is described according to the above-mentioned sub-tasks, as monitoring is goal-driven and defined by measurable sub-goals. The sub-tasks 2 “kinematic configuration”, 4 “empty run(s)” and 7 “forming the work piece” were chosen exemplary to validate the cognition-based framework, as, within these sub-tasks, anomalies were detected by the operators, which indicates that they are particularly error-prone and, therefore, of special interest.

4.2. Operators’ Roles and Responsibilities during the Forming Process

During the interviews, operator 1 was reported to be responsible for programming the robot and controlling the speed and direction of the robot, which also involves the supervision of safety regulations and the access to the emergency shut-off of the robot. Operator 2 was reported to be responsible for the definition of the control data, as well as the physical forming of the work piece and its documentation for further process development. In addition, both operators shared the task of checking the trajectory of the machine tool with the overall goal of reaching the desired target geometry of the work piece. Operator 3 was reported to act as a “jumper”, meaning that operator 3 was able to take over the tasks of the other two operators when required. Operator 3 acted together with operator 2 when fixating the work piece into the clamping device and resetting the clamping device. Moreover, operator 3 was responsible for the overall evaluation of the forming process and their subsequent consequences regarding the research project. An overview of the roles and responsibilities of the three operators is shown in

Table 1. Overview of roles and responsibilities of the operators based on interview data.

Operator	Subjective Statement about Own Role and Responsibilities	Supplementary Information Given by the Other Team Members
1	Programming the robot Controlling speed and trajectory of the robot Avoiding collision of the robot with the work table and clamping device Supervising and complying the safety regulations	Checking the trajectory of the machine tool/robot for correctness (with focus on the quality of the forming result)
2	Defining analytical definition/control data Supervising the physical forming of the work piece Fixating the work piece into the clamping device; resetting the clamping device	Checking the trajectory of the machine tool/robot for correctness (with focus on the quality of the forming result and possible consequences for analytical definition/forming strategy)
3	Evaluating the whole forming process Acting as a “jumper” (i.e., is able to assume tasks of operator 1 and 2) Fixating the work piece into the clamping device; resetting the clamping device	Taking overall responsibility for the forming process/research project

.

Gaze data revealed differences and commonalities regarding the information search behavior of operator 1 and 2 over the entire forming process (see Figure 4). For instance, both operators directed their focus mostly towards the clamping device (31.0% and 22.9% of all relevant fixations). However, operator 1 looked more to the machine tool (27.4% of all relevant fixations) than to the work piece (15.1% of all relevant fixations), whereas operator 2 attended more to the work piece (20.4% of all relevant fixations) than to the machine tool (19.4% of all relevant fixations). However, it has to be noted that differences between the machine tool and the work piece were not as pronounced with operator 2 (difference of about 1%) as they were with operator 1 (difference of about 12%). This has to be attributed to the fact that the forming zone during the forming operation was located between the rolls and adjacent sheet areas. Operator 1 also looked more to the robot (3.7% of all relevant fixations) compared to operator 2 (2.1% of all relevant fixations). Further, only operator 1 attended to the control panel (6.6% of all relevant fixations). In contrast, only operator 2 focused on devices needed for process documentation (1.2% of all relevant fixations).

The gaze data were in line with the reported roles and responsibilities based on the interviews (see Table 1). Hence, as operator 1 was responsible for controlling the speed and trajectory of the robot, which also included the supervision of safety regulations, it was necessary to understand and predict the path of the robot and the attached machine tool. As the machine tool worked along the work piece, operator 1 mainly looked to the work piece, anticipating the course of the machine tool/robot (representing individual situation awareness). Thus, operator 1 integrated all relevant information from the control panel and machine behavior, resulting in glances to the machine tool and control panel (representing individual expertise).

As operator 2 was responsible for the physical forming of the work piece and documentation purposes, they especially paid attention to the areas where forces were induced from the machine tool into the work piece in order to reach its target geometry. Therefore, operator 2 anticipated the physical forming of the work piece (representing individual situation awareness) and the integrated relevant information of the material, machine tool, and applied forces (representing individual expertise). Moreover, as operators 1 and 2 shared the task of reaching a target geometry of the work piece in high quality, they looked to the work piece periodically, representing the (cognitive) team processes of operators 1 and 2.

4.3. Anomaly Resolution during the Forming Process

4.3.1. Regular Process Monitoring

During sub-task 2, “kinematic configuration”, the operators’ goal is to control the fit between the implemented curve progression of the machine tool and the work piece; therefore, the machine tool attached to the robot takes a path that is supposed to be parallel to the work piece without actual work piece contact. The primary information gathered is regarding whether the size of the gap between the work piece and machine tool remains the same throughout the process (see

Table 2. Gaze proportions to relevant AOIs in percent for operator 1 and 2 dependent on analyzed sub-tasks.

Sub-Task	AOI	Gaze Proportions in %
		Operator 1	Operator 2
Kinematic configuration	Work piece	12.2	30.8
	Machine tool	66.3	35.0
	Clamping device	15.2	9.6
	Operator 1	-	19.8
	Operator 2	2.7	-
	Operator 3	3.0	1.1
	Robot	0.6	3.7
	Work table	-	-
	Control panel	-	-
	Documentation	-	-
Empty run (1st time)	Work piece	-	-
	Machine tool	28.6	6.1
	Clamping device	34.1	41.4
	Operator 1	-	4.0
	Operator 2	1.5	-
	Operator 3	0.5	0.9
	Robot	0.8	0.6
	Work table	30.9	40.9
	Control panel	3.6	-
	Documentation	-	1.2
Empty run (2nd time)	Work piece	-	-
	Machine tool	30.0	12.5
	Clamping device	24.6	23.5
	Operator 1	8.3	32.6
	Operator 2	-	-
	Operator 3	-	-
	Robot	3.0	3.2
	Work table	33.6	28.2
	Control panel	0.5	-
	Documentation	-	-
Forming the work piece	Work piece	8.8	25.2
	Machine tool	35.0	21.3
	Clamping device	40.7	30.8
	Operator 1	-	10.5
	Operator 2	9.6	-
	Operator 3	0.2	-
	Robot	5.7	4.1
	Work table	-	0.9
	Control panel	-	-
	Documentation	-	5.5

). If there are deviations, the path of the machine tool is not congruent to the work pieces’ geometry and may need to be adjusted. In order to evaluate the consistency of the gap, both operators focused primarily on the work piece and the machine tool. Operator 1 clearly monitored the path of the machine tool (see Figure 5), which is evident by two thirds of the relevant fixations that were directed towards it.

Operator 2 divided their attention between the work piece and machine tool while also showing a significant proportion of fixations towards operator 1, who was responsible for controlling the robot (see Table 2).

The goal of the sub-task 4, “empty run(s)”, is to prevent collisions of the robot with the work table or the clamping device during the forming process; therefore, the path of the machine tool attached to the robot is tested with its actual starting coordinates but without the clamped work piece. During the first empty run, both operators focused on the machine tool, the clamping device, and the work table, with a strong emphasis on the elements of the set-up that might be in the path of the machine tool (i.e., clamping device and work table; see Table 2). As problems were detected, the first empty run was interrupted in order to subsequently adjust the clamping device. A second empty run was then conducted to examine the improvement of the clamping device set-up. In addition to the focus on the elements with collision potential, operator 2 showed a large fixation proportion on operator 1. Video data showed that this is due to an ongoing discussion of the problem (i.e., collision potential with the clamping device). Operator 1 engaged in the discussion without losing sight of the set-up, whereas operator 2 did not expect further problems, allowing for continuative discussions regarding future projects.

The sub-task 7, i.e., the actual forming of the work piece, is led by the goal to achieve the predefined target geometry of the work piece. Therefore, operator 2 focused on the physical forces that are applied to the work piece by the machine tool (see Table 2). However, as the trajectory of the machine tool was already implemented and tested, operator 1 also attended to previously identified potential problems, as collisions decrease the quality of the forming result and may even harm the clamping device. Therefore, operator 1 executed a large proportion of fixations to the clamping device, which allowed operator 1 to adjust the speed of the robot/machine tool according to anticipated collision potential.

4.3.2. Anomaly Detection, Diagnosis, Problem Solving and Resolution

During the regular process monitoring, the operator team detected five anomalies. Based on the interview data, the problem solving process was analyzed for all identified anomalies according to the developed cognition-based framework. Hence, the responses of the three operators to the interview questions (see Section 3.3) were categorized by information detection and state-goal-comparison, which is part of the regular process monitoring; identification of causes and expertise alignment, which is part of the anomaly detection and diagnosis; and individual generation of action alternatives, joint team evaluation, and choice of action, which is part of the problem solving and resolution process (see

Table 3. Overview of detected anomalies and the subsequent anomaly resolution process.

Regular Process Monitoring	Anomaly Detection and Diagnosis	Problem Solving and Resolution
(1) Anomaly regarding trajectory of the machine tool/robot (sub-task: kinematic configuration)
Information Detection Supervision and control whether the size of the gap between the work piece and machine tool remains the same throughout the process	Identification of Causes Operator 1: Deviation due to incorrect cutting of the work piece or due to incorrect analytical definition/control data Operators 2 and 3: Deviation due to a false fixating of the work piece into the clamping device (as a result of a measurement error)	Individual Generation of Action Alternatives Operator 1: Analysis and adjustment of the analytical definition/control data Operator 2: Continuation of the process despite possible quality losses
State-Goal-Comparison Operator 1 detects a deviation from the ideal trajectory; other operators are notified	Expertise Alignment Operator 2 Operator 3	Joint Team Evaluation [There was no evaluation of action alternatives.]
		Choice of Action Continuation of the forming process
(2) Anomaly regarding the distance between machine tool and work piece (sub-task: empty-run)
Information Detection Supervision and control whether the distance between machine tool and work table/ clamping device is large enough to avoid collisions	Identification of Causes [There was no identification of causes because of missing experience.]	Individual Generation of Action Alternatives Operator 1: Adjustment of the analytical definition/control data Operator 2: Adjustment (i.e., relocation) of the clamping device instead of releasing and fixating the clamping device more frequently
State-Goal-Comparison Operator 1 detects a distance between machine tool and work table which is too small and might lead to collision(s); other operators are notified	Expertise Alignment Operator 2	Joint Team Evaluation Both action alternatives increase the distance between machine tool and work table; however, the adjustment of the clamping device decreases quality losses due to uncontrolled residual stress and enables a continuation of the process (and thus, learning experience)
		Choice of Action Adjustment (i.e., relocation) of the clamping device
(3) Anomaly regarding work piece quality (sub-task: forming the work piece)
Information Detection Supervision and control of the actual forming of the work piece in comparison to the target geometry	Identification of Causes All operators: Expectation that too many parts of the clamping device were released at the same time, which led to uncontrolled residual stress of the work piece	Individual Generation of Action Alternatives All operators: There is no possibility to solve the quality defect; therefore: discontinuation or continuation of the process
State-Goal-Comparison Operator 2 detects a quality defect on the work piece; other operators are notified	Expertise Alignment Operator 2	Joint Team Evaluation Both action alternatives have a negative impact of the forming result; however, a continuation of the process increases learning experience
		Choice of Action Continuation of the forming process
(4) Detection of a slight collision (sub-task: forming the work piece)
Information Detection Supervision and control of the forming process to avoid collisions with the clamping device (and thus, to minimize the frequency of releasing and fixating the clamping device)	Identification of Causes All operators: Expectation that the clamping device was not sufficiently long released	Individual Generation of Action Alternatives Operators 1 and 3: Exclusion of a discontinuation of the process; instead: returning the robot (and releasing the clamping device) or continuing the ongoing process (i.e., pushing the machine tool through the wedging)
State-Goal-Comparison All operators detect a slight wedging between the machine tool and clamping device	Expertise Alignment Operator 1 Operator 3	Joint Team Evaluation The risk for damages on the robot or machine tool when continuing the ongoing process are expected to be small
		Choice of Action Continuation of the forming process
(5) Detection of a serious collision (sub-task: forming the work piece)
Information Detection Supervision and control of the forming process to avoid collisions with the clamping device (and thus, to minimize the frequency of releasing and fixating the clamping device)	Identification of Causes It is a novel situation for all operators (i.e., the collision occurred on a point that has previously been classified as non-critical); therefore, no causes are identified	Individual Generation of Action Alternatives Operator 1: There is no solution regarding robot control to avoid damages Operator 3: Continuation of the ongoing process (i.e., pushing the machine tool through the wedging)
State-Goal-Comparison All operators detect a serious wedging between the machine tool and clamping device	Expertise Alignment Operator 1 Operator 3	Joint Team Evaluation Both a discontinuation as well as continuation of the ongoing process might lead to damages due to high forces on the machine tool and clamping device Damages on the machine tool as a result of a discontinuation are expected to be higher and of more importance (because of the associated costs) than damages on the clamping device as a result of a continuation of the process
		Choice of Action Continuation of the forming process

). There were no inconsistent statements between the operators regarding the anomalies and categories. However, in some cases, information was only reported by one or two operators; therefore, only aggregated data were reported.

The first anomaly was detected during the sub-task 2 “kinematic configuration”. Operator 1 detected a deviation from the ideal trajectory and notified the other operators, with the operators then gathering possible reasons for the deviation. According to operator 1, the deviation could be due to an incorrect cutting of the work piece or incorrect definition of control data. According to operators 2 and 3, the deviation could be due to a false fixation of the work piece into the clamping device as a result of a measurement error. The operators reported that the expertise for solving the problem in the long-term lies within the areas of operators 2 and 3. Possible action alternatives are an analysis and adjustment of the analytical definition/control data (proposed by operator 1) or a continuation of the process despite possible quality losses (proposed by operator 2). The operators agreed with operator 2 and decided to continue the process.

The second anomaly occurred during the sub-task 4, “empty run(s)”. Operator 1 detected a distance between machine tool and work table, which was too small and, therefore, might lead to collision(s). There was no identification of possible causes for this anomaly by the operators. However, all operators reported that finding a lasting solution for this anomaly falls into the area of responsibility of operator 2. Two action alternatives were reported to solve the anomaly: first, an adjustment of the analytical definition/control data (reported by operator 1) and, second, the relocation of the clamping device (reported by operator 2). The team decided to relocate the clamping device. Thus, the distance between machine tool and work table could be increased and uncontrolled residual stress on the work piece due to releasing and fixating the clamping device more frequently could be avoided. Moreover, this enabled a continuation of the process and, therefore, learning experience.

The next three anomalies were detected during the sub-task 7, “forming the work piece”. First, operator 2 detected a quality defect on the work piece, which belonged to their area of expertise. All operators expected that too many parts of the clamping device were released at the same time, which led to uncontrolled release of residual stresses on the work piece and, thus, to an unfavorable forming result. Because all operators believed that there was no possibility of solving the quality defect, it had to be decided between a discontinuation or continuation of the process. The latter increased the learning experience; therefore, the process was continued. Second, the operators detected a slight wedging between the machine tool and clamping device, probably because the clamping device was not sufficiently long released. Operators 1 and 3, which were the experts in solving this anomaly, excluded a discontinuation of the process. Instead, it was proposed to either return the robot and release the clamping device or to continue the ongoing process, i.e., to push the machine tool through the wedging. All operators agreed to continue the process, as the risk for damages on the robot or machine tool were expected to be small. Finally, the operators detected a serious wedging between the machine tool and clamping device. This was a novel situation for all operators, i.e., the collision occurred on a point that had previously been classified as non-critical. The operators identified no causes for this anomaly. According to operator 1, there was no solution regarding robot control to avoid damages; therefore, operator 3 proposed to continue the ongoing process. Because damages on the machine tool as a result of a discontinuation were expected to be higher and of more importance (because of the associated costs) than damages on the clamping device as a result of a continuation of the process, the process was continued in agreement with all operators.

5. Discussion

The present research aimed to validate a theoretical cognition-based framework that was developed by integrating a number of known psychological theories/aspects within a specific use case, i.e., industrial application. Thereby, we focused on the description of cognitive (team) processes during the anomaly resolution within a complex and dynamic industrial process with interdisciplinary experts.

5.1. Summary of Results

5.1.1. The Anomaly Resolution Process

Gaze data indicated that monitoring during the different sub-tasks of the forming process is guided by the goals of the corresponding sub-tasks and, therefore, differs between the sub-tasks. Within the present study, regular process monitoring led to the detection of five anomalies. According to the proposed framework, subsequent anomaly resolution processes were initiated, and interview data documented cause identification, information integration into individual and team expert knowledge, and the (joint) generation and evaluation of action alternatives (see Figure 2). More specifically, during the sub-task of kinematic configuration, the fit between the implemented curve progression of the machine tool and the work piece was tested. Therefore, the work piece and machine tool received the most attention from operators 1 and 2. However, as the operators had various areas of expertise and resulting points of view, they emphasized different information sources. While operator 1 focused on the path of the machine tool (which she/he implemented in the robot), operator 2 attended to the relation between the work piece and machine tool, as it is evident in almost equal gaze proportions. All operators reported to have a mental representation of the goal state of the forming result; thus, they continuously evaluated the current state with regard to its deviation from the goal state. This resulted in an anomaly detection, i.e., there was a deviation from the current trajectory to the ideal one, which led to subsequent diagnosis, problem solving, and resolution (see Figure 2). However, the operator team did not explicitly evaluate the action alternatives jointly during these processes. Instead, there was an implicit agreement between the team members. Because the deviation from the ideal trajectory was a well-known problem for the operator team, which occurred due to either an incorrect cutting of the work piece, an incorrect analytical definition of the control data, or an inaccurate or insufficient fixating of the work piece into the clamping device, there was already a repertoire of action alternatives used in the past. One of the operators proposed to continue the process despite possible quality losses. All of the other operators agreed with this proposal nonverbally without actually having evaluated this in detail. The investigated team had been working for a long time together, and thus, was well functioning. In such cases, there are implicit team processes that cannot be made explicit, and so this advantage of an experienced team is not transferable to a technical system. However, a HCPPS could provide action plans and action alternatives regarding new and inexperienced team members to compensate for possible and unexpected absences of single experts.

The goal of the empty run(s) was preventing collisions of the robot with the work table and clamping device. During the first empty run, the operator team detected a deviation that was not tolerable for achieving the goal of the sub-task (i.e., preventing collisions) and, consequently, the goal of the entire process (i.e., sufficient quality). By monitoring the clamping device, machine tool, and work table, with a strong emphasis on the elements of the set-up that might be in the path of the machine tool, operators concluded that the distance between machine tool and work table was too small in regard to preventing collisions. This was not compatible with the quality goal; hence, an anomaly was detected, and subsequent diagnosis, problem solving, and resolution steps were initiated. During the second empty run, the detected anomaly was further discussed with regard to its implications for future tasks. Simultaneously, operators monitored whether the implemented anomaly resolution was successful, which indicates that post-processing occurred.

The last sub-task, the actual forming process, also showed that operators constantly monitored deviations from the goal state (i.e., achieving a sufficient target geometry of the work piece). Thereby, operators attended to those information sources that have proven to be the most error-prone throughout the entire process (i.e., clamping device) and those that show the development of the final forming result (i.e., the machine tool actually forming the work piece). Three anomalies were detected within this sub-task that led to a specific diagnosis, problem solving, and resolution within the team. However, for some of the detected anomalies, the causes were not identified during the diagnosis process. One of the reasons for this could be due to a lack of experience. For example, in the present study, the operators detected a serious wedging between the machine tool and clamping device during the forming of the work piece. This was a novel situation for all operators, meaning that the collision occurred on a point that had previously been classified as non-critical. Because the causes could not be eliminated immediately, it was more important for the team to address the “symptoms”, i.e., solve the wedging, in order to finish the forming task and avoid lasting damaged. Hence, in such (novel) situations, an identification of causes is probably postponed to the future. However, the debriefing of the forming process was not part of the present investigation.

In sum, the present investigation shows how regular process monitoring and, if necessary, subsequent diagnosis, problem solving, and resolution processes work during the supervision of complex and dynamic production systems (as proposed by [20,29]). However, as the supervision of complex and dynamic production systems is based on the composition of the respective team, some components of the analyzed process are concealed to the applied research methods.

5.1.2. The Impact of Different Psychological Concepts on Anomaly Resolution and Resulting Cognitive Support Possibilities

The analysis of gaze behavior, as well as interview data, allowed for a clear role allocation between the three operators that took part in our study. Gaze and interview data discriminated between individual tasks and tasks that were shared between two or all three operators. This is in line with the proposed framework, as it highlights individual and team processes equally. More specifically, all operators were reported during the interview to have domain-specific knowledge (see, e.g., [32,33]), which is a central characteristic of experts and leads to associated responsibilities within the team and operating process. This means that operator 1 is mainly responsible for controlling the speed and trajectory of the robot, as well as avoiding collisions of the robot with the work table and clamping device; operator 2 focuses on the physical forming of the work piece and the clamping device settings; and operator 3 takes the overall responsibility for the forming process. This individual expertise enables the operators to run mental simulations and spot anomalies (see, e.g., [39]) as well as to anticipate situations and evaluate situations globally (see, e.g., [20]), which has also been found in the present study.

For example, operator 1 anticipated the trajectory of the machine tool guided by the robot. This was evident by gaze behavior that did not focus on the machine tool itself but preceded its position at the work piece (see Figure 5). Hence, the work piece received much more attention than the machine tool by operator 1. Due to the expert knowledge of operator 1, they can successfully perceive and integrate information from the work piece, control panel, and robot behavior to project or anticipate their status in the near future. Operator 2 anticipated the physical formation of the work piece using their experience and knowledge by integrating relevant information of the work pieces’ material, the machine tool, and the applied forces. Hence, the work piece and machine tool seemed to be almost equally important for operator 2. In combination with the interview data, this might indicate that the operators are well aware of the most relevant information sources/elements that have to be focused on to comprehend their meaning and anticipate their status in the near future [41,42]. Hence, as information searching is highly dependent on responsibilities, task goals, and individual expertise, human-centered systems need to account for these different roles by providing the appropriate information for each responsibility. To identify the needed information, it is important to formalize production processes not only based on their technical means (e.g., goals, resource efficiency), but also based on how the process, its sub-tasks, and its possible difficulties are perceived by the respective operator. Often, the operator is not aware of all relevant information sources needed, as increasing expertise results in more implicit knowledge. Therefore, it is crucial to use behavior measures, such as eye tracking and video-based retrospective think-aloud interviews, when investigating and formalizing production processes. As a result, information searching can be supported by enhanced sensory data or system-provided information.

In addition to the individual tasks, all operators evaluated the trajectory of the machine tool with respect to the quality of the formed work piece. Over the entire process, the clamping device received most attention from operators 1 and 2. Further, interview data showed that the clamping device was also of high importance for operator 3. As the clamping device is not considered during planning and programming the trajectory of the machine tool, it is the opaque variable throughout the process. Therefore, the emphasis on the clamping device for the entire team can be interpreted as a shared task that is based on externalized team knowledge [51], as all three operators engaged into a superordinate evaluation of the entire process and have a common knowledge basis. Following the framework (see Figure 2), the clamping devise was identified as a meaningful process variable for all three operators and their tasks. Once identified, knowledge about the common variable(s) could be integrated into technical systems, thus providing the team members with relevant shared knowledge and reducing their workload.

Moreover, interview data showed a high fit of task and teamwork mental model [52] between the operators, which enabled an efficient (and sometimes even implicit) knowledge exchange. This was especially evident in the diagnosis, problem solving and resolution processes. Here, operators came to similar approaches to solve a problem (due to already well-known problems) and often did not need to evaluate them. As teams are not always be as experienced as in the present study, inexperienced teams can be supported by assigning clear roles and providing a shared teamwork mental model, as this seems to be a key factor for a successful collaboration of team members.

5.2. Critical Reflection of the Present Research

The goal of the present research was to evaluate a holistic framework on cognitive (team) processes for complex and dynamic production processes using well-established psychological research methods that were combined and applied to a special industrial application. This means that it was aimed to evaluate the cognition-based framework with subjective interviews as well as objective behavioral data (i.e., eye tracking data) on a holistic level. In the next step, each element of the framework should be investigated with more fine-grained measures, allowing for a deeper description and understanding of single cognitive (team) processes, such as (team) situation awareness, enabling the development of human-centered assistance systems. The industrial application used for the present study was incremental RoRoFo, a process that is not completely controllable, is currently under development, and where the operators take all decisions and actions (i.e., there is no assistance from a cyber-component). In addition, the investigated interdisciplinary expert team was well-functioning due to long-term collaboration. It is therefore not clear if the team is representative of a “real” production context. In future studies, the framework should be applied to other industrial applications with (a) varying automation degrees and (b) varying team compositions. In sum, the methodological approach used in this study provided empirical evidence for all cognitive processes proposed in the framework, such as regular process monitoring and, if necessary (i.e., in the case of a detected anomaly), diagnosis, problem solving, and resolution. Moreover, the roles of situation awareness, individual expertise (and hence, individual roles/responsibilities) and (cognitive) team processes for anomaly resolution were elaborated on (see Figure 2), and our data were analyzed and interpreted on a cognitive level. In the next step, research needs to test whether the developed framework, as well as the used methods, are transferable to similar use cases within the production context. So far, this research was rather explorative, limiting its generalization.

6. Conclusions, Implications and Outlook

With growing flexibility and agility in production, the humans’ tasks will be changed; therefore, operators will interact dynamically to handle various complex systems while physical work will decrease [21]. This leads to an increasing involvement in cognitive tasks, such as supervision and anomaly resolution. As the interaction with decentralized and dynamic systems often requires experts from different disciplines (to detect errors, solve problems and ensure quality standards) the description of cognitive team processes is necessary to realize the full potential of such expert teams, e.g., by developing and designing (cognitive) support possibilities for the human team members. The proposed cognition-based framework for interdisciplinary expert teams offers opportunities to describe and, consequently, understand (cognitive) team processes during the supervision of complex and dynamic production processes. As a result, human-centered assistance proposals can be deduced. The applied methods (i.e., observational eye tracking study with retrospective think-aloud interviews) proved to be appropriate to examine the functioning of interdisciplinary operator teams based on the developed framework. For example, for the case investigated in the present study (i.e., incremental RoRoFo), it could be shown that the human still takes all decisions and actions, i.e., there is no assistance from a cyber-system. Hence, according to [26], the degree of automation is low (or non-existent). In the future, a cyber-system could support the operator team during diagnosis, problem solving and resolution after an anomaly is detected, as this process is very time consuming, experience dependent, and therefore, error prone. For example, the cyber-systems could offer a (narrowed) set of decisions/ actions alternatives based on the respective cause(s) of the detected anomaly. Thus, information searching and selection would be supported, relieving the human, and improving productiveness and product quality.

In sum, the present research provides a starting point on our way to understanding cognitive (team) processes during complex and dynamic production tasks that will dominate the production landscape within Industry 5.0. Future research must prove whether the developed holistic framework specifying the cognitive processing phase of anomaly resolution during interdisciplinary teamwork is transferable to similar use cases within the production context.