Inputs and outputs of neural networks in identification of wear debris

Artificial feedforward neural networks were used for identification of wear debris generated under various loads and at various sliding distances in pin-on-disk steel sliding experiments. Wear debris characteristics were described using four parameters, namely representative diameter, elongation, roundness and reflectivity, and the averages of these parameters were used as inputs to the networks. It was found that the percentage of correct identification depends on the size of the sample used for the averaging. Computer simulation of network learning was conducted using normally distributed random data for study of the effects of sample size. The results showed that the distance between the averages normalized by the standard deviation should be larger than 1.2 for successful identification, which corresponds to a sample size of 800 debris in the present experiments. For identification of continuous variables such as load and sliding distance, use of analog outputs from the networks was proposed, and was shown to work well if wear particle parameters changed monotonously with these variables.