Application of neural networks to integrated ergonomics

Leistungen

Realistic simulation of human behaviour plays a primary role in modern industrial ergonomics. The approaches commonly used in movement modelling, are based on traditional bio-inspired strategies that do not involve the intrinsic variability, due to subjective perception of the contextual quantities. For these purposes, a data driven approach based on a collection of physical experiments, capable of grasping the natural essence of human movements, has been developed and it is herewith presented. Motion analysers, artificial neural networks and open CAD environments have been involved to produce a toolset to perform virtual experiment through the simulation of human motion, fully integrated with a methodology. Results carried out investigating specific industrial scenarios are also presented. The innovative aspect was the use of NN in order to learn the movement features through experimental evidence that is to perform physical experiments, without any assumption about the kinematics and dynamics of the human being. The next sections report the results achieved with the completion of this project. The developed system can be considered composed of two main parts: a theoretical one, that is a methodology to be followed in order to define and perform physical experiments aimed to collect data - and virtual ones for data usage. A software suite, based on a neural network (NN) engine, was integrated into a CAD representation of the analysed environments, allowing the visualisation of experiments (both physically performed and simulated ones), and completed by a human movement database. The software suite was furnished with a user methodology guide and a step-by-step procedure for the software suite usage. The natural data flow implicitly involves task analysis, artificial scenario implementation, motion acquisition, anthropometrical and kinematics identification, neural training and simulation. The possibility to evaluate results of w.r.t simulation is possible through the database, which completes the systems' features. Conclusions: The experimental evidence clearly showed that the intrinsic variability, intended as discrepancy among repetitions of the very same task, could not be foreseen in any model, and thus proved to be a great source of information to compute any ergonomic index. Previous paragraphs have shown the reliability of the ANN data driven approach to simulate human motion in application-oriented contexts, learning features emerged in data collection campaigns. Good generalisation capabilities emerged, even if the trials were concentrated in three well-defined scenarios. New involved aspects about the simulation of the postural behaviour became visible. The open point is the classification and the comparison of different behavioural strategies on the base of the statistic distribution of task, subject, and object and environment information. The related possibility of employ NN in learning cinematic time sequences of statistically clustered individual appears very interesting.

Leistungen

Diese Seite teilen

Herunterladen