Sensor data are multidimensional streams of observations from various sensor systems. In this IO we work mainly on sensor systems in the maritime and industrial sector. In addition, we consider the research activity with Statistics Norway as ´sensoring´ society.

For maritime safety surveillance we develop new approaches based on the availability of large arrays of sensors, which monitor condition and performance of vessels, machinery, and power systems. Sensor data are becoming increasingly available on global ship fleets, with efficient broadband connectivity to shore. Our methodology is however very often of generic value. We suggest new approaches to condition and/or performance monitoring, which is the process of identifying changes in sensor data that are indicative of a developing anomaly or fault. In addition to using previous failure data and pattern recognition techniques to detect anomalies, we test model-based approaches that exploit knowledge on the sensors and the conditions they assess. We also rely on other data sources such as AIS data and meteorological data.

Scalable change and anomaly detection

The PhD funded through the dScience center at UiO, has been working on a new method for fast, sparsity-adaptive change point detection. The new method allows detection and localization of an unknown number of changes in the mean-vector sequence of high-dimensional Gaussian vectors. What is very promising, is that we can prove mathematically that the new method successfully localizes all change-points with a near-optimal error rate and under minimal conditions in all sparsity regimes. The method searches for multiple change-points through Seeded Binary Segmentation, and is therefore highly computationally efficient, obtaining a log linear computational complexity. This project is hugely relevant for the partners of this IO and is jointly supervised from NR and UiO. A paper and an R-package are almost finalized.

This PhD student, Per August Moen, and co-supervisor Martin Tveten from NR also won a hackaton on sensor-data from the energy sector arranged by Statkraft – Krafthack 2022.

In 2022 the concept of detection of overheating in electrical propulsion motors, which we have worked together with ABB on earlier, was revisited, in order to perform training of final models to be implemented on board the ships. During this work, it was discovered that minor changes in the configuration of the motor had fundamentally changed the parameters of the model. The system had experienced “concept drift” - changes in the underlying system that affects the machine learning model. This has initiated a new project to investigate if and how concept drift can be handled in data based models in a safety critical application.

Combining AI and expert knowledge for more efficient monitoring

In 2022, the PhD project using both event log files interpreted as multivariate point processes, and continuous sensor data from ABB systems, for failure detection/classification and prediction for these systems, finalized the first paper. The work extended gradient boosting methods to temporal event data, where intensity functions for each event type are learned as flexible functions depending on both “standard” covariates (e.g. sensor data) and the history of events. The main application of this framework is for predicting future events. A second project focusing on the sensor data and in particular identification of which subsets of sensors that trigger specific events, has also been a topic this year, relying more on parametric models and domain expertise in conjunction with data driven methods. This is still work in progress.

A “nærings-PhD” project with ABB is also progressing. It is dedicated to early detection of incipient bearing faults in rolling element bearings. A first paper is published, based on a adaptive division of the vibration signal into a number of frequency bands, time-domain segmentation algorithm and high-resolution maximum likelihood frequency estimation to discover small vibration pulses excited by the defect in the bearing. In 2022, we have been working on a second paper focusing on improved bearing fault detection by identification of shock pulses originated from the incipient bearing fault and a third paper on identification of different bearing degradation states using prior information from previous measurement points.

A project with data from ABB where we integrate topology drawings of ships’ power systems with sensor data and textual log messages in order to automatically detect and classify power blackouts was completed successfully in 2021. The more challenging prediction problem was investigated in 2022, and has continuated internally within ABB. A significant effort on implementing and updating data models in order to implement automated detection and later prediction has been laid down, but the work is not yet completed.

Inferring the effect of marine bio fouling on loss of performance

In 2022, we finalized a scientific paper on the work we have done with DNV on modelling of loss of performance due to bio growth (fouling) on hull and propellers, using Bayesian GAM and INLA. Loss of performance means for example increased fuel consumption. Data come from various ships, with timepoints for hull and propeller washing and a large amount of time series of relevant operational measurements. A detailed report and code has been transferred to DNV at an earlier stage.

Towards zero emission vessels – li-ion battery health diagnostics and prognostics

Our battery sensor data project with DNV and their collaborator Corvus Energy, which is a major producer of maritime batteries operating in several ferries in Norwegian fjords, has made considerable progress in 2022. We develop methods for data driven monitoring of battery health, based on historical data from operating vessels provided by the battery producer. We started with a study of SOH (State of Health) degradation modelling using publicly available data, because there are extremely few measurements of state of health in the operational data. A paper on this, using multivariable fractional polynomials, was published in 2022.

In order to be able to do the same type of SOH degradation modelling on the operational ferry data, we have developed and published a method for extracting pseudo capacities that can be used instead. This is a semi-supervised learning method, that exploits the relationship between the capacity of the battery and the discharge capacity of similar labelled cycles This method has been of great interest to the battery system producer and has caught a lot of attention.

Furthermore, two master theses on battery data analysis have been successfully defended in June 2022. One was conducting a comparative study of the whole spectre of machine learning methods for the SOH degradation with the publicly available data. The other master project was a study of error-in-variable models for total capacity estimation, based on the operational data from Corvus.

Uncertainty quantification for ship emission models

A new project with DNV on quantifying uncertainty in an emission model, was initiated in 2022. Maritime transport is responsible for a large amount of global Greenhouse Gas (GHG) emissions every year. Stricter rules and regulations on limiting GHG emission put pressure on shipping to reduce its emissions, and the industry is seeking to improve the ship efficiency. Modelling of ship fuel consumption and emission is a fundamental input to evaluate the impacts of shipping on environment and climate, and to evaluate the new measures for reducing GHG emissions. DNV has developed the VERDE model, that estimates ship fuel consumption and emission, based on ship hydrodynamical models, information from the Automatic Identification System (AIS), ship parameters and met-ocean data. The VERDE model produces predictions of fuel consumption, that give satisfactory results on aggregated levels like the total fleet. However, the uncertainty in predictions on ship level is large, and we work on giving a measure of uncertainty in these predictions. This is still ongoing work.

Combining data sources with misclassifications, maintaining privacy (SSB)

With partner Statistics Norway (SSB) we work on estimating the proportion of different, non-overlapping, classes in the population (currently employed/unemployed due to data availability, but could be all types of classes) using data from two parties, where both parties have data with misclassifications. The goal is to develop methods that make Statistics Norway better equipped to utilise data from external companies while safeguarding privacy as much as possible. A paper on adjusting misclassification using a second classifier with an external validation sample, was very well published in 2022.