Issue 45 | Uncrewed Systems Technology Aug/Sept 2022 Tidewie USV Tupan | Performance monitoring | Bayonet 350 | UAVs insight | Xponential 2022 | ULPower UL350i and UL350iHPS | Elroy Air Chaparral | Gimbals | Clogworks Dark Matter

44 Focus | Performance monitoring aircraft is propagated using random control inputs. The controls and resulting states are recorded and fed to a neural network for training, along with the corresponding derivatives. The neural network gradually evolves the mapping between controls with their resulting states and the set of aerodynamic derivatives used. This stage is performed only once, since the resulting trained network has the ability to generalise the estimation of the aircraft’s dynamics. During the inference phase, the trained network is fed with real recorded flight data (controls and states) to infer the aerodynamic derivatives. For validation purposes, the recorded flight data can also be simulated. However, it is hard to demonstrate that the data is real, as some of the interactions between variables only happen in real life. Instead, the HIL system can be used as a verification tool for testing the framework once it has been developed. For example, the HIL simulator could include an engine failure, and see if the model is able to infer where the fault is happening and then how the autopilot responds. TinyML ML is an increasingly important element of performance monitoring inside uncrewed systems, and is leading to changes in the sensor architecture. Rather than feeding all the sensor data into an autopilot or a central ECU, monitoring can increasingly be performed locally using a specialist pre- trained ML framework running on low- power, low-cost microcontrollers. Called TinyML, this approach is optimised for real-time event detection in devices with constrained amounts of memory. It uses ML frameworks such as TensorFlow Lite or PyTorch that are optimised for specific applications, for example audio feeds from microphones listening to a motor. The frameworks can be trained on good data as well as examples of failures in the motor. This combination of a microphone and low-cost microcontroller, which can use an ML accelerator for low- power operation, can monitor the performance of UAV motors to detect early signs of problems. When particular audio patterns are detected, the microcontroller notifies the autopilot or ECU. This reduces the overall power consumption of the system, extending battery life and operation time by reducing the load on the more power- hungry central processor. There are a number of tools available to install these TinyML algorithms on a microcontroller, and they are increasingly becoming automated August/September 2022 | Uncrewed Systems Technology This autonomous aircraft is fitted with the latest Cyphal comms links for sensors (Courtesy of Katla.aero/Cyphal) Rather than feeding all the sensor data into an autopilot or central ECU, monitoring can increasingly be performed on an ML framework