University of Southern Queensland FACULTY OF ENGINEERING AND SURVEYING DEVELOPMENT OF AN ARTIFICIAL NEURAL NETWORK (ANN) FOR PREDICTING TRIBOLOGICAL PROPERTIES OF KENAF FIBRE REINFORCED EPOXY COMPOSITES (KFRE). A dissertation submitted by Mr. Tyler John Griinke in fulfilment of the requirements of CourseENG8411 and 8412 Research Project towards the degree of Bachelor of Engineering (Mechanical) October 2013 i Abstract Study in the field of tribology has developed over time within the mechanical engineering discipline and is an important aspect of material selection for new component design. Most of these components experience failure due to this form of loading. It has been well established that there are several conditions or parameters that may influence the tribological performance of a material. Good correlations with experimental results are not clearly obtained or achieved from mathematical models. Artificial neural network (ANN) technology is recognised as an effective tool to accurately predict material tribological performance in relation to these influencing parameters. The benefit and importance is the ANN models capability to predict solutions by being trained with experimental data. They essentially catalogue the performance characteristics eliminating the need to refer to tables and the requirement for additional time consuming testing. This will aid in continuing research, development and implementation of fibre composites. The aim of the project was to investigate artificial neural network (ANN) modelling for the accurate prediction of friction coefficient and surface temperature of a kenaf fibre reinforced epoxy composite for specific tribological loading conditions. This study has verified the ability of an artificial neural network to make closely accurate generalised predictions within the given domain of the supplied training data. Improvements to the generalised predictability of the neural network was realised through the selection of an optimal network configuration and training method suited to the supplied training data set. Hence, the trained network model can be utilised to catalogue the friction coefficient and surface temperature variables in relation to the sliding distance, speed and load parameters. This is limited to the domain of the training data. This will ultimately save time and money otherwise used in conducting further testing. i ii Certification I certify that the ideas, designs and experimental work, results analysis and conclusions set out in this dissertation are entirely my own efforts, except where otherwise indicated and acknowledged. I further certify that the work is original and has not been previously submitted for assessment in any other course or institution, except where specifically stated. Tyler John Griinke Student Number: 0061004639 Signature Date iii Acknowledgements I extend thanks to my supervisor, Dr Belal Yousif for his guidance and support when it was needed throughout the project. I would like to thank my parents, Pieter and Trish and my grandfather, Vern for their support and patience. iv Table of Contents Abstract .............................................................................................................................. i Certification...................................................................................................................... iii Acknowledgements .......................................................................................................... iv List of Figure .................................................................................................................. viii Nomenclature .................................................................................................................... x 1 Introduction .................................................................................................................... 1 1.1 Project Topic ...................................................................................................... 1 1.2 Project Background ................................................................................................. 1 1.3 Research Aim and Objectives ................................................................................. 2 1.4 Justification ............................................................................................................. 3 1.5 Scope ....................................................................................................................... 4 1.6 Conclusion ............................................................................................................... 4 2 Literature Review ........................................................................................................... 5 2.1 Introduction ............................................................................................................. 5 2.2 Neural Networks (NNs) ........................................................................................... 5 2.2.1 Biological Neurons........................................................................................... 6 2.2.2 Artificial Neural Networks (ANN) .................................................................. 9 2.2.3 Node/Neuron Operational Structure .............................................................. 13 2.2.4 Layout ............................................................................................................ 17 2.2.4 Training and Training Functions .................................................................... 20 2.3 Tribology & ANN Applications ............................................................................ 26 2.3.1 Tribology ........................................................................................................ 26 2.3.2 Tribology Testing ........................................................................................... 26 2.3.3 Materials ......................................................................................................... 29 Fibres ....................................................................................................................... 30 2.3.4 Kenaf Fibre Reinforced Epoxy Composite (KFRE) ...................................... 31 2.4 Risk Management .................................................................................................. 34 2.4.1 Introduction .................................................................................................... 34 2.4.2 Identification of Risks .................................................................................... 34 2.4.3 Evaluation of Risks ........................................................................................ 35 2.4.4 Risk Control ................................................................................................... 35 3 Research Design and Methodology ............................................................................. 37 3.1 Introduction ........................................................................................................... 37 v 3.1.1 ANN Development Process ........................................................................... 37 3.1.2 Implementing MATLAB ............................................................................... 38 3.2 Collect & Process Data .......................................................................................... 41 3.3 Generate Optimal Model ....................................................................................... 43 3.3.1 Select Transfer Function ................................................................................ 44 3.3.2 Select Training Function ................................................................................ 45 3.3.3 Select Layer Configuration ............................................................................ 46 3.4 Improved Generalisation ....................................................................................... 46 3.4.1 Generalisation Technique............................................................................... 47 3.4.2 Validation ....................................................................................................... 49 3.4.3 Train and Test Generalised Model ................................................................. 49 3.5 Simulate and Compare ANN Results .................................................................... 50 3.6 Resource Analysis ................................................................................................. 50 4 Results and Discussion ................................................................................................. 52 4.1 Generate Optimal Model ....................................................................................... 52 4.1.1 Select Transfer Function ................................................................................ 52 4.1.2 Select Training Function ................................................................................ 57 4.1.3 Select Layer Configuration ............................................................................ 60 4.1.4 Training and Testing without Generalization ................................................ 63 4.2 Generalizing .......................................................................................................... 66 4.2.1 Training with Generalization ......................................................................... 66 4.2.2 Generalising Technique ................................................................................. 66 4.2.3 Train and Test Generalised Model ................................................................. 70 4.3 Simulate and Compare ANN Results .................................................................... 72 4.3.1 Predictability Outside Trained Domain .............................................................. 74 5 Conclusions .................................................................................................................. 77 5.1 Introduction ........................................................................................................... 77 5.2 Derived Model Configuration and Training .......................................................... 77 5.3 Network Testing, Simulation and Comparison ..................................................... 78 5.2 Conclusion ............................................................................................................. 78 6 Recommendations ........................................................................................................ 79 6.1 Introduction ........................................................................................................... 79 6.2 Limitations and Challenges ................................................................................... 79 6.3 Recommendations for future work ........................................................................ 79 List of References ........................................................................................................... 80 vi Appendix A: Project Specification ................................................................................. 84 Appendix B: DATA ........................................................................................................ 85 Appendix C: MATLAB Code Example .......................................................................... 86 Three_Hidden_Layer.m (Script File) .......................................................................... 86 Appendix D: Training, Testing and Simulation Results ................................................. 89 vii List of Figure Figure 1- Biological Neuron (www.neuralpower.com) ................................................................................ 7 Figure 2– Connection and impulse transfer of two biological neurons (www.optimaltrader.net). .............. 8 Figure 3– ANN structure representing interconnected organised parallel operating nature of numerous individual neurons (www.optimaltrader.net). ............................................................................................ 10 Figure 4 – ANN summarised as black-box that computes outputs from various input parameters. (www.optimaltrader.net). ........................................................................................................................... 10 Figure 5 - The predicted vs. experimental values for experimental motor performance parameters (Yusaf et al. 2009). ................................................................................................................................................. 12 Figure 6 - The predicted vs. experimental values for experimental motor performance parameters (Yusaf et al. 2009). ................................................................................................................................................ 12 Figure 7 - Elementary Neuron Model (Demuth and Beale 2013) .............................................................. 14 Figure 8 – Log-Sigmoid Transfer Function (Demuth and Beale 2013) ..................................................... 16 Figure 9 – Tan-Sigmoid Transfer Function (Demuth and Beale 2013) ...................................................... 17 Figure 10 – Linear Transfer Function (Demuth and Beale 2013) .............................................................. 17 Figure 11 – General Feed forward network (Demuth and Beale 2013) ..................................................... 19 Figure 12 - Two-layer tan-sigmoid/pure-linear network (Demuth and Beale 2013) .................................. 20 Figure 13 - Schematic drawing showing the most common configurations of tribological machine for adhesive and abrasive testing. (a) block on disc (BOD), (b) block on ring (BOR) and (c) dry sand rubber wheel (DSRW) (Yousif 2012). ................................................................................................................... 28 Figure 14 - A three dimensional drawing of the new tribo-test machine. 1-Counterface, 2-BOR load lever, 3-BOD load lever-, 4-third body hopper, 5-BOD-Specimens, 6-BOR-Speceimen, 7-Lubricant Container, 8- Dead weights (Yousif 2012) ................................................................................................................... 28 Figure 15 – Plant fibre structure (www.ccrc.uga.edu) ................................................................................ 31 Figure 16– Pin-on-Disc machine (Chin & Yousif 2009) ............................................................................ 33 Figure 17– Orientation of fibres with respect to sliding direction (Chin & Yousif 2009). ......................... 33 Figure 18 - Flowchart illustrating steps in developing the ANN model (Nirmal 2010). ............................ 37 Figure 19 – Example MATLAB training window ..................................................................................... 39 Figure 20 – Example of MATLAB performance plot ............................................................................... 40 Figure 21 – Example of MATLAB regression plot .................................................................................... 41 Figure 22 – MSE plot for all data subsets illustrating early stopping for a 3-[25-15-10]-2 network trained with the trainlm algorithm.......................................................................................................................... 47 Figure 23 – Comparison of transfer function performance of single hidden layer networks ..................... 53 Figure 24 – Comparison of transfer function performance of single hidden layer networks ..................... 54 Figure 25 – Performance comparison of transfer function combinations in the 3-[25-10]-2 network. ...... 54 Figure 26 – Performance comparison of transfer function combinations in double hidden layer networks. .................................................................................................................................................................... 55 Figure 27 – Performance comparison of transfer function combinations in triple hidden layer networks. 56 Figure 28 – Performance comparison of training functions in a 3-[25]-2 network .................................... 58 Figure 29 – Performance comparison of training functions in single hidden layer networks .................... 58 Figure 30 – Performance comparison of training functions in double hidden layer networks .................. 59 Figure 31 – Performance comparison of training functions in triple hidden layer networks ..................... 60 Figure 32 – Performance comparison hidden layer configurations ............................................................ 61 Figure 33 –Performance of various node volumes ..................................................................................... 62 Figure 34 - Performance of various node volumes .................................................................................... 63 Figure 35 – Selected ANN model training with trainlm over 2001 epochs ............................................... 64 Figure 36 – Friction coefficient results from ANN predictions and Experimental training data at 2.8m/s with 50N force ........................................................................................................................................... 65 Figure 37 – Friction coefficient results from ANN predictions and Experimental training data at 1.1m/s with 50N force ........................................................................................................................................... 66 Figure 38 - Average achieved training MSE values for variant hidden layer networks implementing early stopping generalisation, trained with trainlm. ............................................................................................ 67 viii Figure 39 - Average achieved R values for variant hidden layer networks implementing early stopping generalisation, trained with trainlm............................................................................................................ 68 Figure 40 - Average achieved training MSE values for variant hidden layer networks and volumes trained with trainbr. ............................................................................................................................................... 70 Figure 41 - Average achieved R values for variant hidden layer networks and volumes trained with trainbr. ....................................................................................................................................................... 70 Figure 42 – MSE Performance plot for double hidden layer network 3-[25-10]-2 trained with automatically generalising trainbr training function. ................................................................................. 71 Figure 43 – Final trained optimal model regression plots. ......................................................................... 72 Figure 44-ANN predictions and experimental data for surface temperature for various sliding distances 73 Figure 45- ANN predictions and experimental data for friction coefficient for various sliding distances . 74 Figure 46 - ANN predictions and experimental data for friction coefficient for various load force ......... 75 Figure 47 - ANN predictions and experimental data for friction coefficient for various load force. ......... 75 Figure 48 - ANN predictions and experimental data for friction coefficient for various load force .......... 76 ix