1.1. Background

Transitioning toward a circular economy is important for sustainable manufacturing and resource efficiency ^1,2,3. In this context, the disassembly of end-of-life products and efficient assembly processes play essential roles, particularly in remanufacturing systems ⁴. However, many of these operations still rely on manual labor because automation remains challenging amid variations in products and operating conditions ⁵. This reliance on manual labor, coupled with labor shortages due to birthrate shortages, necessitates high operational efficiency. In such settings, a significant disparity in work efficiency often exists between skilled and novice workers, which directly affects overall productivity ^6,7. Therefore, a clearer understanding of the relationship between work motion and learning effects is needed to enhance the efficiency of manual tasks.

1.2. Literature Review

Several studies have investigated worker proficiency and motion characteristics by using data-driven and motion-based approaches. Saeki et al. ⁸ quantified the intuition and skills of experienced workers in welding operations for die manufacturing by using motion capture to support skill transfer and human resource development. Sato et al. ⁹ evaluated aircraft wing polishing tasks using motion capture and multiple regression analysis to support the training of craftsmen and reduce the unevenness in finished products across different workshops. Kawane et al. ¹⁰ proposed an assembly motion analysis method that combined motion capture and machine learning to classify motion differences between experienced and inexperienced workers. In addition, Hayashi et al. ¹¹ classified proficiency before and after learning in a nut-loosening task based on working time using machine learning techniques such as random forest and gradient boosting decision trees.

Furthermore, our previous study, Hayashi et al. ¹², used optical motion capture to analyze proficiency differences in a disassembly task and showed that task repetition was associated with reduced variance in joint angles related to rapid upper limb assessment (RULA) ¹³ and shorter task completion times, although the RULA grand score itself showed no significant differences between skilled and novice workers. While ergonomic assessment systems, such as RULA, are widely recognized assessment methods used in ergonomic investigations to evaluate the physical workload of the neck, trunk, and upper limbs ^{13,14,15,16,17}, previous applications have often been limited to assessing musculoskeletal disorder risks in specific occupational settings or validating muscle activity.

However, previous studies, including ours ¹², have primarily examined assembly and disassembly as separate tasks. Although these tasks often involve similar physical motions, their similarities and differences in learning effects and motion characteristics remain insufficiently explored. Therefore, this study follows the analytical framework established in Hayashi et al. ¹² and extends research area to compare disassembly and assembly tasks directly. By applying the same motion-based evaluation criteria to both processes, this study aims to identify the common and distinct characteristics of proficiency acquisition in these two fundamental manufacturing operations, nut-loosening and nut-tightening. This study is based on the hypothesis that as workers gain proficiency in a repetitive task, their posture becomes more stable, as reflected in reduced variance in joint angles. Therefore, this paper uses the variance of each joint angle as an indicator to evaluate task proficiency. In addition, although reductions in task completion time are a primary measure of learning during repetitive work, this study does not use this measure because the prolonged measurements required are difficult to conduct in an experimental setting. Furthermore, this study aims to break down the process by which joint angles influence learning effects.

1.3. Research Questions

Nut-loosening and nut-tightening tasks were selected as the primary research targets in this study. In mechanical assembly lines, fastening operations using screws and bolts account for approximately 38% of the total amount of work, and represent one of the most frequent operations in mechanical assembly processes ¹⁸. Accordingly, nut tightening constitutes a representative assembly task, and nut loosening is a fundamental operation in disassembly. Analyzing both tasks within a unified framework enables a direct comparison of learning-related changes between the assembly and disassembly tasks. Building upon our previous study on disassembly ¹², this study aims to identify both common and distinct learning patterns across these symmetric operations.

This study aimed to clarify the learning-related changes in workload, posture, and movement coordination during repeated nut-loosening and nut-tightening tasks. Optical motion capture was employed to obtain joint angle data during task execution, and RULA was used to evaluate the workload-related posture characteristics. Furthermore, to capture learning-related changes that may not be sufficiently reflected in discrete RULA grand scores, principal component analysis (PCA) was applied to the joint angles that constitute the RULA assessment. PCA enables the extraction and visualization of the dominant movement coordination patterns and their changes with task repetition. The following research questions (RQs) were formulated:

1. RQ1:

   Does RULA grand score vary between nut-loosening and nut-tightening tasks, and are these scores affected by task repetition?

2. RQ2:

   What differences in joint angles derived from the RULA can be observed between the first and last ten trials, and which of these changes are common to both nut-loosening and nut-tightening tasks?

3. RQ3:

   What differences are suggested by comparing the first and last ten trials in both the nut-loosening and nut-tightening tasks using PCA?

The remainder of this paper is organized as follows: Section 2 describes the experimental procedures, data acquisition using optical motion capture, and RULA and PCA methodologies. Section 3 presents the results of RULA and PCA. Section 4 discusses the learning-related similarities and differences between nut-loosening and nut-tightening tasks based on the extracted principal components. Section 5 presents the conclusions and outlines the directions for future research.

2.1. Overview

In this study, the motion data of the nut-loosening and nut-tightening tasks were analyzed based on the methodology established in our previous study ¹². Whereas the previous study focused exclusively on disassembly (nut-loosening), the present study extends the scope to include both disassembly and assembly (nut-tightening) to investigate similarities and differences in proficiency development across these tasks.

The analytical procedure comprised four steps. In Step 1, 3D time-series coordinated data were acquired via optical motion capture for both tasks. To ensure reliable motion capture and consistent coordinate system definitions, the same marker set and experimental conditions used in the previous study ¹² were employed. In Step 2, the joint angles were calculated for each frame, and the RULA grand score was obtained. Notably, while the previous study analyzed four joint angles, this study incorporated two additional angles—the left forearm and left upper arm—to capture the coordination of the non-dominant hand. In Step 3, PCA was performed to examine how these six RULA-related angles differed among workers and across tasks. Finally, Step 4 discusses the changes in workload and the specific characteristics of the learning effect by comparing skilled and novice workers across both assembly and disassembly tasks.

2.2. Experimental Procedure and Motion Data Acquisition

The experimental setup for this study followed the protocol established in our previous work ¹². Data were collected for both nut-loosening (disassembly) and nut-tightening (assembly) tasks to enable direct comparison of the two tasks. As illustrated in Table 1, five workers participated in the experiment, including one skilled worker with extensive field experience. Workers A, B, C, and D performed the nut-tightening task, and workers A, B, C, and E performed the nut-loosening task. Workers D and E did not participate in the loosening and tightening tasks, respectively, because the experiments were conducted over two days; therefore, workers D and E participated in only one of the two tasks. While the hardware and environment, such as the table height (81 cm) and marker configurations, were identical to those in the previous study (see ¹² for details), the scope was expanded to include assembly trials. The participant’s posture and the experimental environment are shown in Fig. 1. The participants performed 30 repetitions of each task as quickly as possible using a wrench (LOBSTER UM30) to align the target markers on the M16 nuts. As illustrated in Fig. 2, 12 markers were used to capture the motion: ten on the participant’s clothing and two on the wrench. Trajectories were recorded at 60 fps using an optical motion capture system ¹⁹. By maintaining the same conditions as those in ¹², we focused on identifying specific differences in the motion patterns and workload between loosening and tightening.

Table 1. Attributes of research participants.

2.3. Motion Data Processing

Missing values in the trajectory data were manually interpolated where possible, and trials with extensive data loss were excluded to ensure analytical reliability. In contrast to our previous study ¹², which utilized manual stroke division, this study implemented a more rigorous segmentation procedure to distinguish between functional motion phases.

The motion data from both nut-loosening and nut-tightening tasks were segmented into individual strokes. A single stroke was defined as a complete cycle consisting of wrench rotation followed by the return motion. Unlike the uniform division used in the previous study ¹², each stroke in this study was specifically categorized into two functional phases: the “operation phase” (active tightening or loosening) and the “returning phase.” Preparing for target tasks, these phases are defined as follows:

1. ・

   Tightening: The operation phase of the assembly task, involving clockwise rotation.

2. ・

   Loosening: The operation phase of the disassembly task, involving counterclockwise rotation.

3. ・

   Returning: The interval during which the wrench returns to the starting position.

These phases were extracted using a signal processing-based algorithm involving peak and valley detection of wrist marker trajectories. To maintain the focus of this study on the comparative analysis of proficiency, the technical details of the segmentation algorithm, including parameter selection and validation, are provided in Appendix A.

Each operation phase was further divided into three equal-duration segments. This subdivision was necessary because the dynamic characteristics of nut rotation vary substantially throughout the motion, making an overall average insufficient for distinguishing proficiency levels.

2.4. Workload Evaluation Using RULA

To evaluate the workload during assembly and disassembly tasks, RULA ²⁰ was employed. The calculation procedure for the RULA grand score (GS) in this study followed the methodology established in our previous work ¹². Worker posture was calculated using trigonometric functions based on 3D time-series coordinated data obtained via motion capture. For example, forearm angle is defined as “the angle formed by the longitudinal axes of the upper arm and the forearm (representing the degree of elbow flexion), with the arm extended straight down defined as 0 degrees” in this study.

As detailed in the previous study, certain joint angles that could not be obtained directly from the marker trajectories were approximated. Specifically, based on video observations, wrist flexion/extension was assigned an intermediate score of 2, whereas wrist twist and leg balance were fixed at the lowest risk scores. These approximations were consistently applied to both the nut-loosening and nut-tightening tasks to ensure a fair comparison. Other ergonomic factors, such as muscle usage and force/load, were defined as zero because the tool weight was less than 2 kg and the task duration for each stroke was short.

While the original RULA provides a snapshot of postural risk, this study aggregated these scores for each defined motion phase. The RULA-based scores were averaged over one-third of the duration of the operation phase, consistent with the segmentation described in Section 2.3. This approach allows for a longitudinal analysis of how physical load evolves as workers repeat the assembly and disassembly tasks.

2.5. PCA for Task Comparison

PCA was used to visualize and compare the transition of movement patterns across repeated trials. To ensure consistency with our previous evaluation ¹², the analysis focused on the first and last ten trials for each worker, representing the initial and stabilized stages of task execution.

While the previous study analyzed only four joint angles during the disassembly task, this study expanded the feature set to six angles by incorporating the left forearm and left upper arm. This modification is based on the suggestion by Hayashi et al. ¹¹ that non-dominant hand coordination may influence proficiency acquisition. Furthermore, PCA was applied separately to the “operation” and “returning” phases defined in Section 2.3. This enabled a more granular investigation of how learning effects differ between functional and auxiliary movements.

To capture each phase of postural transition, the data within each phase were divided into three temporal segments: the first, second, and third segments. By applying PCA to a combined dataset comprising nut-tightening and nut-loosening tasks, this study aims to clarify the physical interpretation of each principal component. Furthermore, this study aims to examine how the distributions of principal component scores shift from the first ten trials to the last ten trials, identifying shared and task-specific characteristics of proficiency in assembly and disassembly.

3.1. RULA GS for Nut-Loosening and Nut-Tightening Tasks

RULA GS were calculated for each segment of the loosening, tightening, and returning phases to evaluate the physical workload. Figs. 3 and 4 show the transitions of GS values during the nut-loosening and nut-tightening phases, respectively.

Across both tasks, a common trend was observed: the GS values for each worker remained almost constant, regardless of task repetition. In the nut-loosening task, workers A and C exhibited a constant GS of 5, whereas workers B and E showed a consistent GS of 3. Similarly, in the nut-tightening task, although skilled worker A showed slight fluctuations between 3 and 5 in the early trials, the scores eventually stabilized at 5. Workers B and D consistently exhibited a GS of 3. These results, which are representative of all analyzed phases, including the returning phase, indicate that the RULA GS does not improve with task repetition, which is consistent with the findings of our previous study ¹² regarding the disassembly task.

This absence of change suggests that the RULA GS, a discrete measure with coarse scoring increments, has limited sensitivity to subtle improvements in movement patterns or postural stability resulting from repeated task execution. A major contributor to the persistently high GS values observed for certain workers (e.g., A and C) was the trunk angle, which frequently exceeded the 60° threshold according to our calculations. Although visual inspection of the task videos suggested that such extreme trunk flexion was unlikely to have occurred during the practice, implying that the calculated values may reflect neck or upper-body inclination rather than true trunk motion, the dominance of this high trunk score effectively obscured changes in other joint angles. Furthermore, no substantial differences in GS were observed between the operation phases (loosening/tightening) and the returning phase because the workers maintained a similar global posture throughout the entire motion cycle.

3.2. Comparative Analysis of Joint Angle Obtained via RULA

Because the RULA GSs exhibited limited sensitivity, the mean and variance of the individual joint angles were analyzed to capture subtle changes associated with task repetition. Individual analysis of the mean and variance of the angles of specific body parts is essential for observing task progression and postural adaptation. Table 2 presents the differences in mean angles (\(\Delta\)) for the nut-loosening task, and Table 3 presents the differences in variance for the same task, respectively. Tables 4 and 5 present the differences in the mean and variance of the nut-tightening task, respectively.

3.3. PCA for Nut-Loosening and Nut-Tightening Tasks

To extract comprehensive motion patterns among the workers, PCA was performed using six joint angle features: forearm, upper arm, jaw, trunk, left forearm, and left upper arm. In this analysis, PCA was applied to the combined dataset, including all phases and segments of both the nut-loosening and nut-tightening tasks. This unified approach was adopted to construct a common low-dimensional feature space, enabling direct comparison of motion characteristics across tasks and phases under the same principal component axis.

The original six-dimensional joint angle data were compressed by PCA. Fig. 5 shows both the cumulative explained variance ratio and the corresponding eigenvalues for each principal component. The left vertical axis represents the cumulative explained variance ratio, right vertical axis indicates the eigenvalues, and horizontal axis represents the principal components from PC1 to PC6. The cumulative explained variance ratio reached approximately 64% with the first two principal components (PC1 and PC2) and increased to 94% with PC4, indicating that most of the variance in the joint angle data was preserved within the first four principal components.

To examine whether combining different task types reduced the explanatory efficiency of dimensionality reduction, PCA was performed separately for the nut-loosening and nut-tightening datasets (Fig. 6). In both tasks, the cumulative contribution ratio exceeded 90% for PC4. Although the cumulative contribution ratio of the unified PCA was lower than that of the task-specific PCAs for earlier principal components, it was comparable to that of PC4. This result suggests that the unified PCA framework can adequately represent both tasks when the first four principal components are considered while allowing the two tasks to be compared within a common principal component space.

Figures 7 and 8 present the distributions of the first and second principal components (PC1 and PC2), and those of the third and fourth principal components (PC3 and PC4), respectively, for each worker during both the nut-loosening and nut-tightening tasks. In these figures, the first sessions (1st–10th trials) are shown in blue, whereas the last sessions (21st–30th trials) are shown in red. Different marker types denote the compressed task phases, and crosses indicate the centroids associated with each session. From top to bottom, the subplots depict the tightening and returning phases of the nut-tightening task, followed by loosening and returning phases of the nut-loosening task. The subplots are arranged from left to right for workers A, B, C, and E (and D in the nut-tightening task), resulting in a 4 \(\times\) 4 subplot configuration based on the task and workers.

As shown in Fig. 7, although the loosening and tightening tasks involved different movements, as reflected in the differences in the centroid shift, the PC1–PC2 distributions were similar for skilled worker A, indicating comparable joint angle patterns across the six joint angles. In contrast, worker B exhibited a wider distribution of principal component scores, particularly along PC2, in both phases of the nut-loosening task. Compared to novice worker B, skilled worker A exhibited smaller variations in PC1 and PC2 values, indicating more stable joint angle patterns across trials, regardless of whether the phase was operation or returning. This tendency was particularly evident in the nut-loosening task. Thus, these findings imply that for novice workers, stabilizing motion patterns might be more challenging in the nut-loosening task than in the nut-tightening task.

As shown in Fig. 8, in the nut-tightening task, no significant differences were observed between the first and last sessions or among workers, indicating limited changes in PC3 and PC4. However, in both phases of the nut-loosening task, the centroids in the last session tended to shift in the negative direction of PC3. Regarding PC4, the direction of the centroid shift differed among workers. A shift in the negative direction was observed for skilled worker A, whereas shifts in the positive direction were observed for novice workers B and C.

To evaluate the changes associated with repeated task execution, the first (1st–10th trials) and last sessions (21st–30th trials) were compared. Tables 6 and 7 present the changes in principal component variance between the last and first sessions (δ = Last − First) for nut-loosening and nut-tightening tasks, with operation and returning phases shown side by side. The columns show the variances of the individual principal components (PC1–PC4), whereas the rows represent workers, task segments, worker-wise averages, and the overall average.

As shown in Tables 6 and 7, the rows are organized by worker and motion segments, whereas the columns are organized by task and principal components (PC1–PC4). During the return motion of the nut-loosening task, the changes in the PCA variance were negative for all four principal components, indicating a consistent reduction in variance. During the return motion of the nut-tightening task, the PCA variance increased slightly for PC1 and PC2, but decreased for PC3 and PC4, suggesting a weaker tendency toward reduced PCA variance.

To facilitate the interpretation of each principal component, the principal component loadings were examined to identify the joint angles that contributed the most strongly to each component, as shown in Fig. 9. As shown in Fig. 9, PC1 exhibits negative loadings for the trunk, upper arm, forearm, and jaw angles. Therefore, a shift of PC1 scores toward the negative direction corresponds to an overall increase in these joint angles. As shown in Fig. 7, for skilled worker A, the PC1 scores gradually shifted toward the negative direction with repetition in both the nut-loosening and nut-tightening tasks. This suggests that repeated tasks could lead to a posture characterized by greater trunk flexion and movement of the upper limb.

PC2 showed positive loadings for both forearm angles and negative loadings for the upper arm and jaw angles. As shown in Fig. 7, workers A, B, and D exhibited a shift of PC2 scores toward the positive direction owing to task repetition in the nut-tightening task. This result suggests that workers bend their elbows to move their forearms more freely with repeated nut-tightening tasks.

Furthermore, PC3 showed positive loadings for the upper arm and left forearm angles and negative loadings for the jaw and right forearm angles. As shown in Fig. 8, for skilled worker A and novice workers B and C, the PC3 scores shifted toward the negative direction with task repetition in the nut-loosening task. Thus, this result suggests that, while reducing the upper arm and left forearm angles, workers tend to lean forward and move the dominant hands holding the wrench more extensively. Finally, PC4 is interpreted as a task-specific fine component that becomes fixed during the nut-tightening task but varies across individuals during the nut-loosening task.