A multi-layer back propagation artificial neural network (BPNN) was used to predict the friction coefficient and the specific wear rate of short fiber reinforced polyetheretherketone (PEEK) composites. The best network structure for property forecast is decided as 5-[29] 1 -2. The train function is trainlm, and the transfer functions from input to hidden layers and from hidden to output layers are logsig and purelin, respectively. For the network of ingredient forecast, the best network structure is 2-[300] 1 -[150] 2 -4. The train function is trainscg, and logsig, tansig are the transfer functions from input to hidden layers and within the hidden layers, respectively. Purelin is transfer function between hidden and output layers. The results show that ANN techniques can effectively be used to predict the tribology behavior and the components of composites. (Abstract) Keywords-back propagation artificial neural network (BPNN); artificial neural networks (ANN); PEEK composite; tribology behavior; component (key words)
Ⅰ. INTRODUCTIONFriction materials normally consist of more than 10 ingredients. The type and the amount of each ingredient in the friction material have been determined mostly on the basis of empirical observations. The lack of good prediction method is a barrier to a deep mechanical and chemical understanding of complex materials. Recently, artificial neural networks (ANN) are used as an interdisciplinary tool in many applications. Another pioneering works in this research field has been performed on the ANN-modeling and ANN-prediction of the wear properties of short fiber composites [1,2] . The major advantage of neural networks is their possibility to predict dependencies between many parameters, and to apply them to any given situation in a "black box" fashion.Therefore, the present work is focused on the application of ANN on prediction of formulation and properties of stainless steel/carbon fiber reinforced PEEK composites developed in our early research [3] . A back propagation neural network (BPNN) was developed, i.e. a multiple-layer feed-forward network with non-linear differentiable transfer functions. A database, containing material compositions and wear characteristics of PEEK composites, was used to train and test the neural network. Afterwards, the well-trained neural network was employed to predict the wear properties according to new input data. The quality of the prediction was also evaluated. In this paper, the modeling and forecasting functions of neural network were realized by software Matlab6.5.
Ⅱ. FORECAST OF TRIBOLOGY PROPERTIES FROM MATERIAL
COMPOSITIONAn ANN is conventionally constructed with input, output and hidden layers. Hidden layers can contain one or several layers for its particular application. Each layer has different numbers of neural elements. As in nature, the network function is largely determined by the connections between these elements. The process of creating an ANN for materials research can be summarized as: database collection; training of the neural ne...