Abstract
This work presents a scenario in which machine learning (ML) adds value to the usability of an SMA actuator. The considered actuator is a locking device which is actuated by two antagonistically arranged SMA wires. The wires are activated using joule heating. The actuator is operated in aircraft interiors at ambient temperatures between −20°C and 70°C. Preliminary work has shown that the locking device can only be reliably operated in a temperature range from approx. 4°C to 40°C without adjusting the activation parameters. Below these temperatures, the wires must receive more heating energy to actuate the device. Above 40°C, the heating energy must be decreased. Otherwise, the wires could be severely damaged. Currently, a temperature sensor and thus an additional component is required for temperature detection. It is known from literature review and from our preliminary work that the characteristic course of electrical resistance during activation of SMA wires depends, among other things, on the ambient temperature. Therefore, it is possible to eliminate the temperature sensor and determine the ambient temperature by monitoring the electrical resistance during activation of the actuator wires. However, the resistance is additionally influenced by the state of wear which in turn is influenced by the actuator-specific load case and the activation frequency. Thus, temperature detection using monitoring the electrical resistance during activation is difficult to generalize beyond a specific load case. In this paper, the authors examined whether an ambient temperature between −20°C and 75°C can be correctly matched to a 5°C interval using a neural network trained with data from the course of the resistance and taking into account the state of wear for a specific actuator. To generate the necessary data, the actuator is operated in a climatic chamber until one of the wires breaks. The ambient temperature is varied between the two end temperatures. This test was carried out twice in total. A neural network was trained to test whether the ambient temperature of the wires can be determined. This procedure worked within the experiments. In a second step, the network was trained with data from experiment 1 to determine the ambient temperatures of experiment 2 and vice versa. This did not lead to a satisfactory result. Two different persons installed the wires in the actuator for the two different experiments. Therefore it can be concluded, that the installation of the actuator wires has a considerable influence on the applicability of machine learning in this scenario.
