Prototype of a vessel collision detection system

Biti lastnik plovila ali flote plovil je stresna naloga, še posebej, če s plovilom upravlja nekdo drug ali pa je plovilo nenadzorovano privezano v marini. Zgodi se lahko marsikaj. Plovilo lahko trči v drugo plovilo, obalo ali drug objekt. Za oddaljen nadzor plovil obstaja mnogo naprav, ki so zmožne lastniku posredovati marsikatero informacijo. Vendar taka, ki bi omogočala zaznavo trka, še ni ponujena na tržišču. V ta namen se v zadnjem obdobju v naprave vgrajujejo pospeškometri. Zgolj podatki, ki jih ti senzorji beležijo, niso prav koristni, dokler se jih ne osmisli. Zato je cilj diplomske naloge osmisliti pridobljene podatke. Ker trki na morju niso prav pogosti, je večina podatkov, ki se lahko kvalificirajo kot trk, lažno pozitivnih. Za boljše razumevanje smo podatke iz senzorjev povezali s podatki o vremenu, poiskali korelacije med meritvami pospeškometra in ostalimi atributi ter klasificirali vsako večje odstopanje vrednosti pospeškometra. Uspeli smo razložiti približno tri četrtine takih točk in tako pripraviti podatke na nadaljnjo obdelavo, najverjetneje z metodami strojnega učenja.Owning a vessel or a fleet can be stressful, especially if someone else is steering it or it is left unsupervised in the marina. Many things can go wrong and a collision with another vessel, the shore or other objects is possible. Of course there are many devices, created especially to monitor one\u27s vessel that are capable of transmitting numerous different informations to the owner, but none of those are capable to detect collisions. Lately, accelerometers are being implemented into these devices, but the data they provide is not useful until it is processed properly. The goal of this thesis is to process vessel sensor data and classify it. Because collisions at sea are fairly rare, most of the data recognised as a collision is false positive. To better understand what exactly is going on with the vessel at any given moment, the accelerometer and all other vessel data is merged with weather data. After that, correlation between accelerometer data and any other given attribute is calculated. Outliers are then classified, according to the previous analysis. Roughly three quarters of all outliers were classified correctly and the data is prepared for further processing, most probably by applying machine learning algorithms