USE Network launch I UAV Works VALAQ l Cable harnesses l USVs insight l Xponential 2020 update l MARIN AUV l Suter Industries TOA 288 l Vitirover l AI systems l Vtrus ABI

87 AI systems | Focus That allows the hardware, software stack, simulation for design and validating the test tools to be developed at the same time for optimisation. This can deliver an efficiency of up to 95% on real workloads for the neural network, reducing the load on the CPUs. Again, using a simple state machine as the controller for the NNA makes it very deterministic, which is easier to certify to the ASIL-B safety standard and above. The compiler creates low-level commands for the state machine and uses an estimator to allow developers to see how a framework can be ported to the hardware. There are many challenges with moving from training to an embedded inference engine; it is an iterative process that requires a lot of engineering expertise. Using simulation to test different configurations, layers and the time interleaving between the layers allows developers to move around a network and explore the different capabilities. Battery monitoring Neural network frameworks can also be used to predict battery health with 10 times the accuracy of industry-standard approaches. Researchers in the UK for example have developed a new way of monitoring batteries by sending electrical pulses into them and measuring the response. The measurements are then processed by a machine learning algorithm to predict the battery’s health and useful lifespan. Their method is non-invasive and is a simple add-on to any existing battery system. Battery performance degrades as a result of a complex combination of subtle chemical processes. While each individual process has little effect on performance, collectively they can significantly shorten it as well as the battery’s lifespan. Rather than tracking the current and voltage during battery charging and discharging, the AI algorithm looks at the patterns of the many processes taking place inside the battery. The researchers performed more than 20,000 experimental measurements to train the model – the largest dataset of its kind – and tagged the important changes to separate them out from background noise. Recent advances in machine learning allow the whole dataset from the battery measurements to be used as an input into the model without handpicking features, and the model selects the most relevant variables. The model can also give hints about the physical mechanism of degradation. It can inform which electrical signals correlate most closely with ageing, which in turn allows developers to design specific experiments to probe why and how batteries degrade. Shipping A key autonomous shipping project is bringing neural networks together with path planning and expert systems. Over the past two years, the engineers developing the Mayflower autonomous trimaran have been training the ship’s AI models using more than a million nautical images collected from cameras in the Plymouth Sound on the UK’s south coast, as well as open source databases. This training of the vision system uses Nvidia V100 Tensor Core GPUs to build a framework that detects and classifies ships, buoys and other hazards such as land, breakwaters and debris. On board the vessel, the network runs on Nvidia’s Jetson AGX Xavier devices. The image classification is Unmanned Systems Technology | June/July 2020 AI is a key element in the ‘Mayflower’ autonomous ship project (Courtesy of IBM)