Data-driven ship models for rapid virtual prototyping

Many of us have raced in a narrow streets or landed a jumbo plane in a small airport, at least in simulator. Virtual reality is very convenient for learning new competences and system rules. Furthermore, the consequences of a crash are minimal: you can start over after a game over. It is sadly not the case in reality. April 2007 capsized Bourbon Dolphin, a one-year-old anker handling vessel, during operation. The crew had not got enough training on the emergency procedures. 8 people died or are still missing. Training in a simulator is a reasonable alternative to training in real premises. For the training to be effective, one needs to train with real equipment, perfect visualization, and realistic physics models. The project is about improving the physic models of ships, cranes and other offshore and maritime equipment by feeding the physics models with real data in order to improve their realism. Virtual Prototyping is the process of virtually testing hardware, software, and/or procedures for products and operations before, during, and after their creation in the real and tangible world. It allows shorter time to market design-production cycles, by improved quality assurance, early identification of logical mistakes user testing and so on. The accurate modelling behaviour the ships is of crucial importance not only during their design phase but also during simulation of virtual operations. A typical example of virtual prototyping in the offshore industry involves the lifting of several hundred tons of heavy subsea objects from the deck of an offshore subsea vessel onto a sea bed laying hundreds up to thousands of meters below sea level. A typical question virtual prototyping has to answer is then: what are the environment conditions (Beaufort scale, wind, current) under which the whole system (crane + ship + humans) can perform the operation. Another typical use of simulation is to crew train on new or virtual products or operations. When humans are involved in the simulation or when the simulation is used to make predictions in real-time, one needs to have fast algorithms. In the past, this has come at the cost of accuracy. In the advent of cost-efficient and faster processing computing power, artificial intelligence algorithms have reached or outreached human capabilities in many domains: chess, go game, air conditioning system optimization, pattern recognition, handwriting reading, reading comprehension, car or truck driving etc. The project is to create faster and more accurate predictive simulations based on data gathered offshore and in smart cities.

The phd project has open a door on digital twins for smart cities.

Because interfacing with real world contraptions is both expensive and time-consuming, the first phase focuses on interfacing with the digital twins in simulated environment. This allows cost-effective and efficient interfacing with emulated hardware and software output, and the possibility to run simulations faster than perceived human real time and gives the possibility to train various controlled weather and wave conditions. The benchmark will be the simulated ship itself, hereby providing a 100% guarantee that the algorithm is fit for purpose. Model environment and emulate sensors: wave, wind and current based on sensor data. Create or improve realistic sensor data from simulator: wind sensor, Radar, Lidar, Wave Radar, MRUs, wave buoys, Acoustic Doppler Current Profiler, thruster command and actual output, crane cargo specific sensor, Ballast information, model various cargo and ballast loads and ballast loads, as well as mooring forces in a simulator. Focus here is to model exact current, wind, and wave environment, which is key to precisely predict ship motions, possible applications are on-board real-time augmented tools like energy optimization tools, and collision prevention and other prediction systems. Parallel to sensing and modelling environment and ships, a standard method for testing digital ship models in virtual sea trials has to be agreed upon, both theoretical and programmatically. Once the environment and benchmarking are in place, data-driven machine learning algorithms can be developed and tested in a fast and effective manner (third step). For this purpose, develop an API to interface with simulation framework for algorithms to learn, analyse and predict the simulated ship motions, hereby making a data-driven ship model. When algorithms are stable enough and trustable against the benchmark, the next step is to switch from the realistic simulated data to the offline and online real data (step 4).