A neural network model for predicting postures during non-repetitive manual materials handling tasks

Posture prediction can be useful in facilitating the design and evaluation processes for manual materials handling tasks. This study evaluates the ability of artificial neural network models to predict initial and final lifting postures in 2-D and 3-D scenarios. Descriptors for the participant and condition of interest were input to the models; outputs consisted of posture-defining joint angles. Models were trained with subsets of an existing posture database before predictions were generated. Trained models predictions were then evaluated using the remaining data, which included conditions not presented during training. Prediction errors were consistent across these data subsets, suggesting the models generalised well to novel conditions. The models generally predicted whole-body postures with per-joint errors in the 5 degrees -20 degrees range, though some errors were larger, particularly for 3-D conditions. These models provided reasonably accurate predictions, even outperforming some computational approaches previously proposed for similar purposes. Suggestions for future refinement of such models are presented. The models in this investigation provide a means to predict initial and final postures in commonly occurring manual materials handling tasks. In addition, the model structures provide information about potential lifting strategies that may be used by individuals with particular anthropometry or strength characteristics.