Workforce fatigue prediction via wearable sensor data analytics

Work-related musculoskeletal disorders (WMSDs) remain a critical challenge in labour-intensive professions, often leading to disability, reduced productivity, and significant healthcare costs. Common ergonomic risk factors, such as repetitive tasks, heavy lifting, and awkward postures, contribute to localised muscle fatigue, which over time can cause chronic injury according to fatigue failure theory. Although observational risk assessments are widely used and influential in setting standards for WMSD prevention, they primarily offer static, reactive evaluations of ergonomic risk and fatigue. This thesis introduces a System for Activity-aware Fatigue Evaluation (SAFE) framework using torque-based biomechanical endurance modelling to predict the time to physical fatigue during static and dynamic tasks in an individualised and task generalisable manner. The framework addresses limitations of ergonomic assessments, force-based endurance models, and other wearable ergonomic systems. The feasibility of the framework was demonstrated using a custom MATLAB application supporting wearable data streaming, manual entry of anthropometric and task-related data, and output of endurance metrics. This solution shows potential for proactive fatigue forecasting in real-time and fills important gaps in the field. Preliminary testing with a torque-based endurance model showed large errors for one-handed static (28.62%) and two-handed dynamic (49.21%) fatiguing shoulder flexion tasks which led to exploring alternative models and inputs. Subsequently, six different models and four different methods of obtaining maximum torque (〖Torque〗_max), an input which indicates the maximum strength of the individual, were compared. This included regression models that predicted 〖Torque〗_max based on individual factors developed in the present work (R2≈.6, P<.05). However, a novel method to determine 〖Torque〗_max with the new improved consumed endurance (NICE) model produced the lowest errors across all model-〖Torque〗_max combinations for all tasks. The new method achieved RMSE=19.11 s compared to the errors of other models and methods which ranged from 41.08 s to 65.54 s. The novel method developed used data from a static task at a given load relative to the maximum voluntary contraction (MVC) force to solve for 〖Torque〗_max. The torque-based endurance modelling approach is an important contribution that replaces traditional force inputs with torque, improving prediction accuracy and allowing the model to account for multi-joint dynamics, and therefore different tasks, and inter-individual variability. The system developed to validate the SAFE framework comprises of inertial measurement units, pressure insoles, and a custom MATLAB application to continuously refine predictions based on upper extremity configuration, task type, and load, requiring minimal manual input. Prior to deployment of the framework, the influence of these sensor types on endurance predictions was assessed offline where predictions were calculated using non-wearable data (laboratory-based equipment and manual input) and wearables data. Pressure insoles were not found to be a reliable method for estimating the load (mean absolute errors of 29.8%), and it was decided to automate the input of task-related parameters, such as external load, component shape and geometry, using task prediction methodologies. Machine learning techniques were employed to identify task types using pressure insole data with high accuracy (83% for 5 s window), demonstrating the feasibility of automated load input to the endurance model while achieving explainability and enabling low data throughput. The SAFE framework was preliminarily tested during simulated industrial tasks with human subjects. Results confirmed the reliability of joint angle estimation compared to gold-standard technology (RMSD≤13.9°, MAE≤10.6°, .51≤CMC≤.95). Very strong correlations between the endurance metric describing cumulative fatigue, consumed endurance (CE), with subjective fatigue ratings (rmm≥.883, P < .001), and time (rmm≥.909, P < .001) demonstrated the strong potential of the framework as a task generalisable (adjusting predictions using task-specific coefficients) yet individualised (using the novel approach to calculate 〖Torque〗_max, consideration of anthropometrics and individuals’ perception of fatigue) tool for WMSD prevention. Endurance measures generated by the framework successfully differentiated between ergonomically high- and low-risk tasks (P<.050) while overcoming the shortcomings of self-report measures and ergonomic risk assessment tools. This is the first known implementation of an endurance model with task classification into a wearable ergonomic system, offering a biomechanical perspective grounded in real-world data rather than purely theoretical models, with applicability to various dynamic tasks involving the upper extremities. This work marks a step toward proactive real-time physical fatigue management and potentially WMSD prevention.

Grantor dc:publisher

University College Cork

Year dc:date.issued

2026

Author dc:creator

• O'Sullivan, Patricia

Advisors dc:contributor.advisor

• O'Flynn, Brendan
• Komaris, Dimitrios Sokratis
• Menolotto, Matteo

• Work-related musculoskeletal disorders (WMSDs)
• Wearable sensor systems
• Biomechanical modelling
• Fatigue modelling
• Machine learning
• Ergonomics
• Injury prevention
• Occupational health
• Explainable methods

Statement dc:rights

• © 2026, Patricia O'Sullivan.

Licence dc:rights.uri

• https://creativecommons.org/licenses/by/4.0/

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10468/18813

OAI identifier oai:identifier

oai:cora.ucc.ie:10468/18813