Development of innovative prognostic methods for EMAs

In recent years, the adoption of Electro-Mechanical Actuators (EMA) in the aerospace sector, mainly as secondary actuation devices, is strongly increasing, in particular in the more electric and all electric design philosophies. These approaches aim at creating a single form of secondary power, in order to drive all users (systems); the only form of energy so versatile to be capable of accomplishing such task is electric energy. While using electricity, EMAs provide the natural electro-mechanical interface needed to convert secondary power, electricity, to useful mechanical work. At this moment, the use of EMAs in large scale commercial aircrafts is relegated to secondary flight control actuation (flaps, slats, airbrakes), while the use as main flight control surfaces actuation is still limited to small UAVs. The manufacturers’ choice has multiple reason: firstly, in large power applications, electro-hydraulic actuators are still lighter and more compact; in second place, EMAs are still recent technology, so there is not a complete literature on their failure modes nor an established prognostic methodology. One approach that could led to fault detection and subsequent isolation is the use of a properly trained Neural Network, evaluating the response of the EMAs and/or the motor to a given signal and outputting an estimate of convenient values measuring a relative level of performance degradation. This is the aim of this work, to test and validate innovative methods for Electro-Mechanical Actuators (EMAs) prognosis, based on the use of Neural Networks. In this work, the motors analyzed and modeled are 3-phases Brushless DC trapezoidal motors (BLDCs), widely used in the aerospace sector. In order to collect the substantial amount of data needed to properly train the Network, a very detailed Simulink model has been used, modeling in detail both electrical and mechanical components; the model derives directly form Eng. M. Dalla Vedova PhD dissertation [10] . This approach has multiple benefits compared to physical testing: low cost, fast deployment, good accuracy. There are five variables modeling all the possible faults condition analyzed: three representing a partial shortage of each phase, one modeling the static eccentricity value and the last representing the phase of such imbalance. The Simulink model has been run thousands of times, imposing each time a different set of faults conditions; current, voltage and position have then been logged for each iteration. Using the aforementioned physical data, a MATLAB algorithm has been used in order to properly reconstruct the Counter Electro-Motive Force (CEMF); the CEMF signals have then been suitably sampled and used as training set for various neural networks. Performance are evaluated on a test set not previously used in training, highlighting the difference between different architectures and sampling strategies.