Investigation of Tribotechnical Characteristics of Epoxy Composites Containing Al 2 O 3 , ZnO , and PTFE Using Neural Networks

ABSTRACT The paper investigates the dependence of wear intensity and friction coefficient of epoxy composites on the content of fillers—aluminum oxide (Al 2 O 3 ), polytetrafluoroethylene (PTFE), and zinc oxide (ZnO). To process the results of tribotechnical characteristics, models were constructed using neural networks combined with Akima and spline interpolation methods, which made it possible to increase the accuracy of experimental data processing. To generate a set of experimental data and for the subsequent use of neural networks, tribological studies were conducted under conditions of reciprocating motion of a cylindrical counterbody with a diameter of 24.5 mm. The tests were conducted at a sliding speed of 0.125 m/s and an indenter load of 43 N. The temperature in the friction zone during the experiments did not exceed 22°C–24°C, which minimized the impact of thermal degradation of the polymer matrix on the research results. It was found that ZnO and PTFE provide the lowest wear intensity, I = 3 × 10 −9 , and Al 2 O 3 , I = 15 × 10 −9 . The introduction of PTFE and ZnO into the material significantly reduces the friction coefficient. The analysis of residual values confirmed the high accuracy of the models. The residual values are concentrated near zero. The effectiveness of processing using the Akima method for detailed analysis of local deviations and complex nonlinear dependencies was revealed. Splines are better suited for larger deviation values when smoothing data. The deviation of residual values for the friction coefficient according to the Akima method is between −0.002 and 0.002, and when using the spline method, between −0.001 and 0.001. This means that when studying the tribotechnical parameters of epoxy composites and using neural networks, we were able to significantly increase the significance of the data, reducing residual values and ensuring a high level of correlation between experimental and predicted data. The results of the study confirm the reliability of the developed models for processing experimental results, which will ensure the creation of new materials with improved properties. The proposed approach can be applied to the analysis of other classes of composites.