Methodology for developing a neural network leaf spring model

This paper describes the development of a neural network that is able to emulate the vertical force-displacement behaviour of a leaf spring. Special emphasis is placed on aspects that affect the predictive capability of a neural network such as type, structure, inputs and ability to generalise. These aspects are investigated in order to enable the effective use of it to model leaf spring behaviour. The results show that with the correct selection of inputs and network architecture, the neural network's ability to generalise can be improved and also reduce the required training data. The resulting 2-15-1 feed-forward neural network is shown to generalise well and requires minimal data to be trained. Experimental data was used to train and validate the network. The methodology followed is not limited to the application of leaf springs only but should apply to various other applications especially ones with similar nonlinear characteristics.