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

42 AI for performance monitoring From a UAV operation perspective, performance monitoring is mostly about alarms or maintenance. Normally the UAV communicates through the base station, where all the alarms are colour-coded. An operator doesn’t have time to take numeric values on the ground station’s display into account, so it is about alarms and how to make the data easily available. For example if the IMU is not operating correctly then safety actions have to be taken, such as opening the parachute to ensure that the UAV is recovered safely. All the data from a mission – navigation drift, speed, control commands and so on – is also available at the end of a mission via a log file for analysis. One of the most common questions after a flight is regarding the flare attitude, pitching up the UAV to get altitude for landing. If this is performed at too high an altitude, it can cause damage; too low and it can be ineffective. The autopilot checks the altitude, the real distance from the UAV to the ground and the indicated airspeed, as well as the maximum pitch, and generates the automatic actions to correct the craft’s attitude if these are outside a preset range. These alarms are the standard way for the operator to check the correct operation of the UAV, as they will notice if a value falls outside the normal range. However, it is important to avoid overwhelming an operator with too many potential alarms, and this is where ML is helping. One UAV developer has written a tool that gives access to all the parameters and the most common alarms so that operators can start adding more parameters to the screen via a configuration file. For each flight, the tool opens a different file with all the data. This is done semi-automatically, and even during the mission the operator can access the graphics, which is useful for the engineers to check that everything is working as expected. For a UAV to monitor itself in operation, however, the autopilot is the critical component. It takes data from sensors around the craft, from navigation and the IMU to the servo controller and the engine sensors, but the key is knowing whether that data is in the safe operating range, so it needs labelling. This is where ML and all the collected data come in. Using techniques such as reinforcement learning can provide a framework for making decisions about the performance data during the mission. To do that, data collected from UAV development flights needs to be labelled. With 150 to 200 variables for each flight – and data being collected every second – that is a considerable task. After each flight the log files are put on a server with information from the alarms. Labelling this data can be time- consuming, so there are tools to automate the process, taking the data from the log files and formatting it for the ML system. This includes labelling whether a sensor is outside its normal operating range, which ties into the alarm data in the config files. The sizes of the log files depend on the flying time – 10 to 11 hours is a typical mission, and that can generate August/September 2022 | Uncrewed Systems Technology The training and inference process for performance monitoring in a UAV (Courtesy of UAV Navigation) The autopilot is critical for a UAV to monitor itself, as it takes sensor data from around the craft, but the key is knowing if the data is in the safe range