Many situations arise in engineering where it is desired to model a system of complicated input and output variables. However, analytical difficulties arise when these systems exhibit nonlinear behavior. Neural networks have proven useful for such applications because they are able to model complicated nonlinear systems through exposure to a database including input parameters and the desired outputs. One such complicated system consists of the unknown relationships between flight variables and structural loads on helicopters. The development of an accurate neural network based model would allow indirect monitoring of these loads so that fatigue-damaged components could be replaced according to load history. In this thesis, an extensive database of real-time flight records has been effectively used to teach a multilayer feedforward artificial neural network nonlinear relationships between common flight variables and the resulting component loads. The trained network predicts time-varying mean and oscillatory load records corresponding to flight variable histories. Component loads in both the fixed and rotating systems of a military helicopter have been resolved over a variety of standard maneuvers. Predictions under the present conditions are on the order of 90 to 100% accurate. Although the range of maneuvers presently considered is rather limited in comparison to the total helicopter flight spectrum, the present results justify further pursuit of this neural network application. / Master of Science

Helicopter load histories applied to notched metal samples are taken as examples, and their fatigue lives are predicted by using a simplified version of the local strain approach. This simplified method requires an input load history in the form of the rain-flow matrix and places bounds on the fatigue life. A peak-valley reconstructed history is generated based on the standard spectrum Helix. A second history studied is a more irregular one based on actual flight data. lt is used to generate three reconstructed histories based on three principles: peak-valley, to-from, and rain-flow. Emphasis is given to the rain-flow reconstruction method, and different reconstruction methods based on rain-flow cycle counting are presented. Life predictions are presented for all of the above cases, and the comparison with test data and other considerations suggest that the most promising reconstruction approach is one based on rain-flow cycle counting. Finally, a method is presented which reconstructs a history with the same rain-flow cycles and also the same distribution of relative time increments between adjacent peaks and valleys. This reconstructed history gives the same fatigue life as the original history. / Ph. D.