The bow-tie method takes its name from the shape of the diagram used to identify causes and effects. It gives a visual summary of all plausible accident scenarios that could exist around a certain hazard and by identifying control measures, displays what can be done to control those scenarios. “So, by using AI-enabled systems we can make this approach even more advanced,” says Professor Yanwei Fu. “As an example of what we are working on with DNV, we focus on computer vision and IoT devices for use onboard autonomous ships. Deep learning is increasingly being used and there are also multiple traditional sensors which can be used as data sources. And cameras are also being added and are covered in our work to further advance AI assurance on vision capabilities.”
Playing tricks on the eyes of AI
One of the major issues with autonomous ships is how they will react to other vessels, fixed and floating objects in a safe and environmentally responsible manner. The ship will be receiving data from systems such as radar, AIS and GPS along with optical and perhaps Infrared images.
In a perfect world this would seem sufficient but at sea, faulty sensors, bad weather, poor quality data and even the images used in training the AI to identify what a ship looks like will affect the AI’s decision-making ability. This will be compounded by the fact that both the autonomous vessel and other ships will be constantly moving potentially allowing images to be misinterpreted.
There is a large body of published papers on this apparent vulnerability of AI in several applications and it will be something that needs to be overcome before the abilities of AI systems can be assured. The partners will therefore focus on using generative adversarial networks to test AI algorithms for testing detection and classification of ships.
“Our idea is to train a GAN on a set of customer-provided images, and then use it to generate new images that are realistic but slightly different from the training set. These new images would be used to check the robustness of a customer-created deep learning algorithm for detection and classification of ships. Eventually, such an approach will help us to build a toolbox for AI algorithm certification, complementing the approach described in DNVGL-RP-0510, “Framework for assurance of data-driven algorithms and models” says Michael Chen.
“Another example of our collaboration is the work we are doing on scenario modelling for risk-based assurance. Different standard levels will have different quantitative requirements. As we know, many hazard cases are due to human factors. One of the things we are aiming to do is to provide a standard library which manages knowledge and scenarios. The data-driven scene graphs will enable, for example, AIS analysis & modelling, AIS prediction & generation, scene classification and game engine simulations,” adds Professor Yanwei Fu.
According to Professor Yanwei Fu, the project is not yet at a stage where any findings or results can be published but he says this could happen soon and he is confident the work with DNV will produce beneficial results.
