Cruise & Ferry Review - Spring/Summer 2021

1 0 5 disrupted by system or component failure. In such cases, the engineer will often manually download and inspect measurements and data from the relevant systems or subsystems to pinpoint the root cause and propose an appropriate solution. As part of this process, they classify the incident or fault. The data set concerning a fault may consist of categorical data in the form of alarm lists, measurements from shipboard sensors and text-based information in the form of service reports or communication between the support team and crew. Unlike unsupervised machine learning, in which little or nothing is known about the data, supervised machine learning relies on the data being accurately labelled. The labelled input data is frequently divided into training/ validation data sets and test data sets, which are used to prepare and verify the machine learning algorithm. In the case of fault diagnostics in complex marine systems, a classification algorithm is employed. The underlying engine used to train the classifier varies depending on the data set, but the training and testing process, as well as the deployment of the trained model into a production system, can all be integrated into the ABB Ability ecosystem. Unsupervised algorithms cannot learn structure from labelled data, and instead must identify that structure independently. In cooperation with researchers from Norway’s University of Oslo and as part of the BigInsight research project it is funding and contributing to, ABB has developed an unsupervised system that takes output from marine systems and finds structure or clusters in the unlabelled data set. The algorithm is based on advanced Bayesian statistical methods, accounting for the fact that data sets collected from marine applications do not constitute big data on the scale of other industries, such as the consumer applications market. Each of the structures corresponds to a mode of operation of the equipment, including fault or failure modes. However, for these results to be useful, domain knowledge is needed to accurately recognise and name the various clusters or operational modes. The developed model allows engineers to manually edit the cluster. For example, for a given fault class, an engineer may decide that a certain alarm or message delivered by the system is irrelevant and should not be considered as an indication by the algorithm. The methodology will be applied to automated diagnostics and fault detection for complex marine systems. It will be integrated into the digital service offering and support on- site engineers in their maintenance work. ABB is augmenting its support and troubleshooting services not by replacing human staff with AI, but by developing systems that combine the benefits of data and machine learning with the skills and knowledge that only experienced engineers can offer. The more input the experts provide, the more intelligent the systems will become. This will ultimately improve the level of service provided to the customer, increase vessel efficiency, and make maintenance work easier and safer for the engineer. CFR Morten Stakkeland is a data scientist at ABB Marine & Ports