Proud, Jasmine K ORCID: 0000-0002-6969-8377 (2022) Know your limits. Predicting lift capacity using time series spine kinematics for a military manual handling task. PhD thesis, Victoria University.

Abstract

In Australia, 41% of industry injury claims are due to manual handling tasks, costing $14.58 billion annually. In the Australian Army, 78% of physically demanding tasks are considered manual handling, which increases the risk of musculoskeletal injury. Of injuries in active service personnel for the Australian Army, 22% occurred in the trunk [3, 4] with manual handling recognised as the cause for 5% of all injuries [3, 4]. This has led to the need for an exoskeleton system that can support, move and adapt to repetitive, fatiguing tasks. The predecessor to this exoskeleton system is the development of an assist-as-needed control algorithm that will predict when personnel are lifting above their maximum acceptable weight of lift (MAWL), which is indicative of an increased injury risk. This algorithm could also be deployed on a simpler stand-alone wearable device that could assist personnel in reducing risk factors associated with injury due to manual handling tasks, through providing visual or auditory feedback. Laboratory experiments using biomechanical task analysis based on military manual handling protocols were performed with a sample size of 32 participants. Inertial measurement units (IMUs) were used in a six-segment spine model for data collection. The normalised (for time) kinematic output of the IMUs for participants during lift-to-platform tasks were analysed for the relationship between changes in spine kinematics and increasing external load. Statistical parametric mapping was performed to determine significance in the IMU variables. Additionally, polynomial correlation of discrete features were analysed for use as predictive factors of external loading above a participant’s capability which resulted in poor correlation. Machine learning was performed due to its ability to find trends and features in data that may not be apparent via statistical inference. Supervised machine learning algorithms capable of classifying multivariate time series data were compared. The Random Convolutional Kernels (ROCKET) algorithm had the highest accuracy for its ability to classify a high risk (at or above MAWL) or low risk (below MAWL) lift, with a 10-fold cross validation mean accuracy of 91.2 ± 2.7%. A moderate f1-score was maintained through dimensionality reduction of the spine segments and data frames per feature. Reducing the spine segments to one (middle lower thoracic) and data frames to half (50) resulted in a f1-score of 86%. This research contributes an accurate novel predictive model that uses machine learning to classify spine kinematics from IMUs into high and low risk lifts, based on MAWL. In future work, the novel predictive model developed in this thesis will contribute to the development of a stand-alone device providing user-feedback. The model will also be part of an assist-as-needed control system for the development of an active exoskeleton that could provide augmentation to Defence personnel during manual handling. These devices aim to reduce injuries caused by lifting above an individual’s capacity.

Search Google Scholar

Repository staff login