Prediction Of Proximal Femur Loads Using Finite Element Analysis And Artifical Neural Network

• Main Author: Tan, Chuan Ser
• Format: Final Year Project Report
• Language: English
• Published: Universiti Malaysia Sarawak, (UNIMAS) 2020
• Subjects: TS Manufactures
• Online Access: http://ir.unimas.my/id/eprint/37086/2/TAN%20CHUAN%20SER.pdf; http://ir.unimas.my/id/eprint/37086/

There have been numerous studies upon the proximal femur to help urderstand the relationship between loading conditions acting on the bone and their corresponding bone properties. There are studies that employ artificial neural network (ANN) to predict the loading acting on the proximal femur using existing bone mechanical properties. Loading data predicted from mechanical properties of bone are important as they would give insight on prosthesis design of patient specific treatment, post-surgery monitoring, criminal science and others. Hence, this study aims to predicts loading data using ANN from equivalent elastic strain (EES) data via finite element analysis (FEA) with initial loading data input. Proximal femur finite element model from repository is recreated in SpaceClaim 2019 R3 prior to running FEA within ANSYS Workbench. Material definition of proximal femur as well as loading and boundary conditions are stated and defined to run FEA. Due to constrictions of ANSYS Workbench, 6 loading cases evolved into 12 loading cases due to axis specific force input. The DOE data for I 000 sample load cases are exported into csv file for input in ANN. ANN is done in PyCharm which uses Python Language. 200 sample sets are omitted until the last phase for testing of final ANN model. The selection of neural network architecture of ANN via hyperparameter tuning y by evaluating performance using trial-and-error method. After running training and testing to get the best results of neural network architecture, cross validation is ready to be applied to the model. 3-fold, 5-fold and 10-fold cross validations models are compared with respect to performance for model selection. After model selection, the final ANN is ready to be tested against 200 sample data sets which are regarded as "world data". Performance is evaluated to see whether ANN can predict loading accurately. The final model of ANN tested against the "world data" yields an accuracy of 92.44% which is satisfactory to be deployed in real life applications.