Unmanned Systems Technology 027 l Hummingbird XRP l Gimbals l UAVs insight l AUVSI report part 2 l O’Neill Power Systems NorEaster l Kratos Defense ATMA l Performance Monitoring l Kongsberg Maritime Sounder

84 Focus | Performance monitoring sample applications assembled into a single software package using software development tools. Another example of using a neural network for performance monitoring is an acoustic-based fault diagnosis technique for three-phase induction motors driving the wheels of a driverless car. Four states of the motor were analysed – a healthy motor, a motor with a broken rotor bar, a motor with two broken rotor bars, and a motor with the faulty squirrel-cage ring. A multi-layered neural network can analyse the output of a microphone and compare that to the different failure mechanisms. The neural network can then identify potential failures before they occur. Software monitoring Another way of monitoring the performance of a system is to monitor the data bus, which these days is often a Controller Area Network (CAN). Besides CAN, autonomous vehicles are usually equipped with Ethernet. The decision and control modules are connected to the CAN bus, while the sensors are connected to the Ethernet, although sensors such as IMUs now have CAN interfaces as well. CAN and Ethernet use different frame structures, so combining the data effectively to compress and send to the cloud is a major challenge. A data recorder for doing just that has been tested on an autonomous vehicle, and the tests show that the data can be reduced by a factor of seven. Data on the in-vehicle network (IVN) is compressed and transmitted to a remote centre for vehicle monitoring and data analysis. The input interface is either CAN bus or Ethernet. The data comes from sources including the perception module, decision module, control module, Lidar, radar, cameras, RTK GPS, IMU and the odometer. The data parser can tell if the message is for real-time monitoring or historical data recording. The real-time monitoring covers the vehicle’s location, speed, heading, mileage, diagnostic trouble codes and so on, and is sent to the remote centre every second. Meanwhile the historical data, including all the data on the IVN, is compressed into files every 10 minutes and sent when there is available bandwidth. Upon the arrival of a CAN message, the new data is mapped onto a 2D space associated with the specific CAN ID. The data recorder collects vehicle information and sends the data to a central server over the 4G cellular network. The real-time data receiver decodes some selected parameters for vehicle monitoring, while the historical data file receiver aggregates all the commands to the vehicle through the remote command transmitter. The modules on the test vehicle (an electric golf cart) generated more than 50 kinds of CAN messages with different periods. During each 10 minutes, 150,565 records were collected. Each CAN record consisted of arrival time, CAN ID, data length code and data bytes, producing over 2 Mbytes in the 10-minute sampling period. This was then compressed to 305 kbytes using techniques such as data grouping and differential sampling, easing the pressure on the comms link. Self-diagnosis Another approach is an integrated self- diagnosis system (ISS) using a multi- layered neural network. This collects information from the vehicle sensors, diagnoses itself and sends the results to a central server. One example uses an ISS with three modules. An in-vehicle gateway module collects the data from the in-vehicle sensors and transfers the data collected through various protocols in the vehicle. These include CAN, FlexRay and Media Oriented Systems Transport (MOST) protocols that are used to link to the OBD. The data collected from the in-vehicle sensors is transferred to the CAN or FlexRay protocol, and the media data is collected while driving is transferred to the MOST protocol. The various types of messages transferred are transformed into a destination protocol message type. A second module, the optimised deep- learning module, creates the dataset used for training a multi-layered neural network, taking the data collected from the in- vehicle sensors and assessing the risk of vehicle components failing, as well as the risk to other parts that might be affected by a defective component, thus August/September 2019 | Unmanned Systems Technology The structure of an integrated self-diagnosis system for an autonomous vehicle. It includes an in- vehicle gateway module that collects data to feed to an optimised deep-learning module, which is then fed to a data processing module that sends the data to the cloud (Courtesy of YiNa Jeong et al, Catholic Kwandong University)