Professional Documents
Culture Documents
Available atonline
www.sciencedirect.com
online
Available at www.sciencedirect.com
at www.sciencedirect.com
Available online at www.sciencedirect.com
ScienceDirect
ScienceDirect
Procedia CIRPCIRP
Procedia 00 CIRP
Procedia (2017) 000–000
106 (2022)
00 (2022) 168–173
000–000 www.elsevier.com/locate/procedia
Procedia CIRP 00 (2022) 000–000 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
gonomic score is subjective, as the performed activities are 2.1. Key-points detection
scored by the observation of only one person.
Given the trend towards flexible assembly systems [6], ini- Via the HigherHRNet CNN [12], pixel positions for the dif-
tiated by the demand for mass customized products, a static ferent key-points are retrieved. The key-points of interest are:
ergonomic risk analysis of a work cell no longer reflects the nose, eye (left/right), shoulder (left/right), elbow (left/right),
true ergonomic load of an operator due to the rapid change of wrist (left/right), hip (left/right), knee (left/right) and ankle
tasks. Next to the work environment, the ergonomic workload (left/right). This network follows the bottom up approach.
is mainly influenced by the operator’s own actions. Therefore,
it is of importance to encourage the operator to perform work 2.2. 3D inference
using the correct ergonomic techniques.
In literature, different approaches for continuous ergonomic To retrieve accurate projective points, all key-point positions
load monitoring are described [7, 8, 9]. In the scope of indus- are corrected via calibration data. Based on the camera posi-
trial applications, these principles are mainly used for physical tions, a ray (i.e. 3D line) is casted for each point through space.
human-robot applications [10]. Nevertheless, it is not known The intersections of these rays are used to calculate a 3D point
how this can be done over prolonged time periods without af- per key-point. As there is some error margin on the 2D points,
fecting or distracting the operator. Furthermore, no methods on it is possible that the rays from the different viewpoints do not
how ergonomic feedback effectively can be provided towards intersect exactly in space. For this reason, an approximate solu-
operators on the shopfloor is found. While digital work instruc- tion is computed by minimizing the square of the distance to the
tions (DWIs) are available to offer cognitive support towards rays. In this way, a three-dimensional skeleton can be inferred
operators within HVLV assembly environments [11], there is for an individual. When more than one individual is detected in
currently no integration to inform the operator on his ergonomic the images, a heuristic algorithm groups the views of the differ-
soundness. ent individuals by minimizing the projection error on the com-
This work provides a methodology for integrating er- puted key-point positions [13]. To be able to compute the po-
gonomics information into assembly information to provide sition of a joint, at least two camera views are required. When
feedback on ergonomic soundness of the task execution via occlusions occur, the maximum of the available rays are used.
DWIs. The remainder of the paper is organized as follows. If less than two rays are available, the joint status considered as
In Section 2, a video-based approach for monitoring the er- unknown and will be ignored for the rest of the processing.
gonomic surroundings of an operator is proposed. Next, the
integration of ergonomics information within assembly infor- 2.3. Ergonomics assessment
mation is described in Section 3. First validation tests for the
proposed methodology are conducted and described in Section Based on the 3D key-points, obtained in the previous step,
4. Finally, the conclusions and an outlook is to be found in Sec- all the body joint angles are computed. First, the sternum and
tion 5. pelvis locations are estimated based on the centre of the rays
from the shoulders and hips, respectively. The back flexion
is represented by the angle between the vertical axis and the
spine. Arm and leg torsions can only be computed if there are
enough flexions in the elbows and knees. Therefore, wrist and
ankle angles cannot be inferred reliably within this system and
2. Ergonomics monitor will therefore be ignored. To calculate the angles of the upper
body, a standard frame is attached to the sternum. This frame is
To warrant objective and continuous ergonomic load mon- aligned with the spine axis and the shoulders. The head frame
itoring, a vision-based ergonomics monitor is developed. The is registered from the nose, eyes, and ears positions. The neck
set-up consists of low-cost, low intrusive sensing solutions that flexion, rotation and lateral flexion are derived from the chained
allow for measurements over prolonged time periods without rotations between the sternum and the head frame. The arm
affecting or distracting the operator. Based on multiple video frames are created via the wrist, elbow, and shoulder joint po-
streams, recorded at an assembly work cell, the operator’s pos- sitions. Via these frames, arm rotation, flexion, and adduction
ture, and finally a real-time ergonomic score, is calculated. Fig- angles are calculated. In the same way, hip flexion, adduction
ure 1 gives an overview of all analysis steps within the er- and rotations are computed using the ankle, knee and hip joint
gonomics monitor. First, a convolutional neural network (CNN) positions. Elbow and knee flexion are computed from the angle
is used to detect body key-points in a frame image. Next, key between the upper and lower arms and legs respectively. Simi-
points coming from the different video streams are intersected lar to the sternum frame, a hip frame is attached to the pelvis.
in space to determine the three-dimensional (3D) position of The back torsion is calculated by the chain rotation between
the body joints. Finally, based on a standard ergonomic norm the sternum frame and the hip frame. Based on the calculated
(REBA), an ergonomic score is computed via the joint angles. joint angles, a partial REBA analysis is performed. Two groups
The algorithms described below are written in Python and make are scored: group A (trunk, neck, and legs) and group B (up-
use of the Pytorch library for CNN processing and OpenCV for per arms, lower arms, and wrists). As indicated above, the wrist
image processing. joint is ignored in the proposed ergonomics monitor. For each
2
170 Arno Claeys et al. / Procedia CIRP 106 (2022) 168–173
A. Claeys et al. / Procedia CIRP 00 (2022) 000–000 3
Thus, primary pieces of context for one or more entities can or correcting instruction components. This process can be par-
be used to derive secondary context for that same entity. This tially automated by suggesting instruction components based
principle is used to derive assembly related context form both on context. Based on the ergonomics parameters, obtained from
historical data (e.g. manufacturing data) and sensor data. Possi- the ergonomics monitor, suggestions are made for adding rele-
ble types of secondary assembly context categories include ex- vant instruction components (e.g. an awareness icon for lifting
perience, condition, requirements and, within the scope of this operations) that inform the operator to use the most optimal
work: ergonomics. way for executing the task ergonomically. In some cases, more
Based on the ergonomic scores obtained from the er- detailed information in the form of a guideline is required to
gonomics monitor, the ergonomics module performs several ac- inform the operator about the correct ergonomic method. Here,
tions as illustrated in Figure 3. Below, a detailed description of a procedure is updated by adding a guideline instruction to the
these different actions is given. workflow.
An ergonomic risk level report (Figure 3, a) is an overview By making ergonomics information context-aware, content
of all monitored ergonomic threats for a certain operation. As filtering can be applied. Based on historical performance data,
stated before, a report is generated for each step within a task. stored in the reports, decisions can be made to show or hide
The output from the ergonomics monitor is normalized and a certain instruction. For example, when an operator starts
logged per second and is added to the report. In the scope of falling back into old (bad) habits, an ergonomic guideline can
this work, the REBA scores per joint (e.g. legs, shoulders, etc.) be shown at the next execution of the task. Decision rules
and per group (A, B and C) are stored. Per step, an ergonomic may be implemented in a static (e.g. periodic, ergonomic er-
score is determined by taking the maximum moving average rors, etc.) or dynamic way (e.g. recommender system based
score. Hence, short impulses are filtered and idle recordings are on operator feedback). Allowing the operators to give feedback
eliminated. Based on these reports, flags can be placed on tasks on the instructions is highly necessary when applying context-
when the ergonomic score exceeds a given threshold, or when awareness, as this may be used to refine the context filter rules.
the score is remarkably higher compared to similar tasks. In
this way, non-ergonomic situations can be detected in an earlier 4. Validation
stage (i.e. before injuries occur). Flagged tasks can be evaluated
by an ergonomist and ergonomics parameters (e.g. ergonomic The methodology proposed in this work is implemented in a
risk, REBA threshold) can be updated for that task. Based on proof of concept based on an industrial use case. The tasks used
these parameters, decisions can be made to optimize the work- during the validation process are originated from the assembly
station, and the ergonomic performance of an operator can be of a pneumatic air compressor. The ergonomic risks identified
monitored more accurately. Moreover, based on individualized within the tasks are: lifting heavy objects (> 8kg), rotation and
operator reports (Figure 3, b), operators can be reallocated in bad posture. The use case assembly work cell is replicated in
order to balance the ergonomic load [17]. a controlled lab environment equipped with a set of sensors,
including four industrial cameras (Allied Vision Manta G201C,
3.2. Formulate and update work instructions 1/1.8” 4.4 11mm Varifocal Lens) connected to a computer with
Cuda capabilities (Nvidia RTX 3070 GPU).
After ergonomics analysis, DWIs may require updates (Fig- Figure 4 gives a schematic overview of the proof of concept
ure 3, c). Instruction content can be edited by adding, removing system architecture. The central part is the Assembly Knowl-
4
172 Arno Claeys et al. / Procedia CIRP 106 (2022) 168–173
A. Claeys et al. / Procedia CIRP 00 (2022) 000–000 5
edge Store (AKS), containing the ontology schemas described the target is the ergonomic risk that corresponds to the guide-
in Section 3. Assembly information, DWIs and ergonomic line, and the context status is the ergonomic error flag. Thus,
guidelines are managed in a cloud based DWI platform1 . Dur- if an ergonomic error is made for the ergonomic risk that cor-
ing task execution, timestamps and explicit operator feedback responds to the guideline, then the guideline is considered as
are logged and stored in a report that corresponds to a product relevant. The DWI platform is configured to communicate with
order. Via an application programming interface (API), all rel- the AKS via WebSocket. Based on the context filter rule, a con-
evant manufacturing data (task, scheduling, and performance text request is send to the AKS. Based on the relation between
data) are exported. Via the assembly information configurator the person (active user of the DWI platform) and the ergonomic
(Figure 2), the DWI platform data model is transformed to ISA- risk, a relevance score is returned that determines if a guideline
95 models and transmitted as B2MML messages. In this way, should be presented.
all manufacturing data coming from the DWI platform are in- The proposed proof of concept is validated via an experi-
tegrated in the AKS. When new information is available, the mental approach. Fifteen participants (engineering students and
DWI platform pushes it to the AKS. academic personnel) performed ten assemblies each. The par-
Via an ergonomics configurator, both assembly and DWI ticipants were divided into three groups. The first group (sce-
data are enriched with ergonomics information. For the proof nario A) only received assembly related information and was
of concept, ergonomic risks (e.g. lifting), ergonomics param- thus unaware of ergonomic soundness. The second group (sce-
eters and a REBA threshold are assigned manually to a task nario B) received static ergonomic guidelines as a part of the
step. The threshold scores are determined by comparing er- assembly instructions. Here, the operator received the same in-
gonomic scores coming from the ergonomics monitor to eval- formation for each iteration. For the last group (scenario C),
uations done by an ergonomist. Based on this information, the ergonomics information was contextualized meaning that in-
system determines if an operator’s execution could be ergonom- structions are filtered based on the participant’s ergonomic per-
ically improved. formance.
The ergonomics monitor calculates a real-time REBA score
based on the postural body angles, and the ergonomics parame- Table 1. Average REBA scores per ergonomic risk per scenario
ters (force/load, coupling, and activity score) obtained from the
Ergonomic risk Scenario Av. REBA
AKS for the active task step. The ergonomic measurements are
returned to the AKS and are linked to performance data orig- A 4,39
inated from the DWI platform. In this way, all measurements Lifting B 4,35
C 4,17
are linked to an identifiable task step. After each task iteration,
A 3,23
the ergonomic scores are evaluated per ergonomic risk, and Posture B 2,67
stored as a relation between the person and the corresponding C 3,04
ergonomic risk. When the evaluated score exceeds the thresh- A 3,23
old, the evaluation is flagged. Rotation B 3,35
C 3,10
Ergonomic guidelines are equipped with context filter rules,
that check if a context status for a defined target is met. Here,
scenario. Here, a positive evolution is visible when an opera- of the academia-industry collective project Ergo-EyeHand, Er-
tor receives ergonomic guidelines (scenario B and C). Never- gonomic Monitoring and Improvement (HBC.2018.0250).
theless, when guidelines are delivered too frequently (scenario
B), the operator ignores the guidelines. Consequently, the op-
References
erator falls back into old habits which has a negative impact
on the ergonomic score. Based on explicit operator feedback, [1] S. Bevan, “Economic impact of musculoskeletal disorders (MSDs) on work
a negative rating was visible for scenario B while for scenario in Europe,” Best Practice and Research: Clinical Rheumatology, vol. 29,
C, operators were more satisfied with the received information. no. 3, pp. 356–373, 2015.
Although this proof of concept is validated over a short period [2] J. De Kok, P. Vroonhof, J. Snijders, G. Roullis, M. Clarke, K. Peereboom,
with a limited number of participants, the positive impact of P. van. Dorst, and I. Isusi, Work-related musculoskeletal disorders : preva-
lence, costs and demographics in the EU. 2019.
context-aware ergonomic guidelines is visible. [3] J. Grobelny and R. Michalski, “Preventing work-related musculoskele-
tal disorders in manufacturing by digital human modeling,” International
Journal of Environmental Research and Public Health, vol. 17, no. 22,
5. Conclusions and further research pp. 1–19, 2020.
[4] A.-C. Falck, R. Örtengren, and D. Högberg, “The impact of poor assem-
bly ergonomics on product quality: A cost–benefit analysis in car manu-
This work proposes a methodology for integrating er- facturing,” Human Factors and Ergonomics in Manufacturing & Service
gonomics information in contextualized digital work instruc- Industries, vol. 20, no. 1, pp. 24–41, 2010.
tions. To monitor the operator without interfering the task [5] L. McAtamney and S. Hignett, “Rapid Entire Body Assessment,” Hand-
at hand, a vision-based ergonomics monitoring system is de- book of Human Factors and Ergonomics Methods, vol. 31, pp. 8–1–8–11,
veloped. Via vision techniques, ergonomic features are esti- 2004.
[6] L. V. D. Ginste, A. D. Cock, A. V. Alboom, Y. Singh, E.-H. Aghezzaf, and
mated and are used to calculate an objective ergonomic score J. Cottyn, “A formal skill model facilitating the design and operation of
based on the REBA evaluation technique. Based on first tests, flexible assembly workstations,” in IFIP International Conference on Ad-
the obtained scores are found accurate. Next, a methodol- vances in Production Management Systems, pp. 108–116, Springer, 2021.
ogy to integrate ergonomics information within context-aware [7] U. Jayaram, S. Jayaram, I. Shaikh, Y. Kim, and C. Palmer, “Introducing
operator support systems is proposed. The methodology en- quantitative analysis methods into virtual environments for real-time and
continuous ergonomic evaluations,” Computers in industry, vol. 57, no. 3,
ables the automatic generation of ergonomic risk level reports, pp. 283–296, 2006.
through which shopfloor operators can be informed with their [8] A. Yazdani, R. S. Novin, A. Merryweather, and T. Hermans, “Ergonomi-
ergonomic soundness and DWIs can be updated or reformu- cally intelligent physical human-robot interaction: Postural estimation, as-
lated based on ergonomic performances. To validate the pro- sessment, and optimization,” arXiv preprint arXiv:2108.05971, 2021.
posed methodology, a proof of concept, based on an actual in- [9] M. Peruzzini, F. Grandi, and M. Pellicciari, “Exploring the potential of op-
erator 4.0 interface and monitoring,” Computers & Industrial Engineering,
dustrial case is built in a controlled lab environment. A first vol. 139, p. 105600, 2020.
experimental test shows that context-aware ergonomics infor- [10] L. Gualtieri, I. Palomba, E. J. Wehrle, and R. Vidoni, “The opportunities
mation and guidelines have a positive impact on the operator’s and challenges of sme manufacturing automation: safety and ergonomics
motivation to perform actions in an ergonomically correct way. in human–robot collaboration,” in Industry 4.0 for SMEs, pp. 105–144, Pal-
While these first results are promising, there are some limita- grave Macmillan, Cham, 2020.
[11] S. Makris, G. Pintzos, L. Rentzos, and G. Chryssolouris, “Assembly sup-
tions that should be marked. The scores returned by the er- port using ar technology based on automatic sequence generation,” CIRP
gonomics monitor were only calculated via the REBA evalu- Annals, vol. 62, no. 1, pp. 9–12, 2013.
ation method, which is not the ideal assessment tool for all er- [12] B. Cheng, B. Xiao, J. Wang, H. Shi, T. S. Huang, and L. Zhang, “Higherhr-
gonomic situations (e.g. lifting). During further work, more er- Net: Scale-aware representation learning for bottom-up human pose esti-
gonomics evaluation methods will be added to the ergonomics mation,” Proceedings of the IEEE Computer Society Conference on Com-
puter Vision and Pattern Recognition, pp. 5385–5394, 2020.
monitor. Also, it is hard to interpret the ergonomic scores with- [13] S. Dehaeck, C. Domken, T. Bey-Temsamani, and G. Abedrabbo, “A strong
out manual intervention. This could be resolved by applying geometric baseline for cross-view matching of multi-person 3d pose esti-
machine learning algorithms that can derive additional context mation from multi-view images (in press),” ICIAP 2021 Conference Pro-
and add it to the ergonomic reports (e.g. the intention of an op- ceedings, 2021, 2021.
erator). The tests were performed over a short period of time [14] A. Claeys, S. Hoedt, H. Van Landeghem, and J. Cottyn, “Generic Model for
Managing Context-Aware Assembly Instructions,” IFAC-PapersOnLine,
and with a limited amount of participants in a controlled lab en- vol. 49, no. 12, pp. 1181–1186, 2016.
vironment. As a next step, the proposed methodology should be [15] A. Claeys, S. Hoedt, M. Schamp, H. V. Landeghem, and J. Cottyn, “Onto-
tested for a longer period in a real industrial setting. Based on logical Model for Managing Context-aware Assembly Instructions,” IFAC-
these results, steps could be taken to configure a recommender PapersOnLine, vol. 51, no. 11, pp. 176–181, 2018.
system based on the ergonomic performance of an operator. [16] A. Dey, G. Abowd, and D. Salber, “A Conceptual Framework and a Toolkit
for Supporting the Rapid Prototyping of Context-Aware Applications,”
Human-Computer Interaction, vol. 16, no. 2, pp. 97–166, 2001.
[17] G. Michalos, S. Makris, L. Rentzos, and G. Chryssolouris, “Dynamic job
Acknowledgements rotation for workload balancing in human based assembly systems,” CIRP
Journal of Manufacturing Science and Technology, vol. 2, no. 3, pp. 153–
This research is supported by VLAIO (Flanders Innovation 160, 2010.
& Entrepreneurship) and Flanders Make, the strategic centre for
the manufacturing industry in Flanders, within the framework
6