Statistical methods for high dimensional sensor based monitoring of ship systems

Hovedmålet med denne oppgaven er å utvikle datadrevne og sensorbaserte modeller, og å studere hvordan disse kan implementeres og brukes i sikkerhetskritiske domener. Når konsekvensen av feil prediksjoner er lav, kan veien fra algoritme og modellutvikling til fullskala implementering være relativt kort. For slike applikasjoner er økt nøyaktighet ofte nok til å rettferdiggjøre implementering. For sikkerhetskritiske applikasjoner er imidlertid tillit og robusthet viktig, og tolkninger og forklarbarhet kan noen ganger være en forutsetning. Oppgaven er utført i nært samarbeid med DNV GL, et globalt selskap for kvalitetssikring og risikostyring. Den tilbyr klassifisering og uavhengige rådgivningstjenester til forskjellige bransjer. En sentral utfordring for DNV GL er å sikre og verifisere systemer og bruk av systemer som er basert på datadrevne modeller. I denne avhandlingen utvikler vi datadrevne prediksjonsmodeller og anomali-deteksjonsmodeller for applikasjoner i den maritime industrien. Ulike modifikasjoner foreslås for å forbedre eksisterende påvisningsteknikker for anomali, og for å forbedre effektiviteten. Vi undersøker forskjellige evalueringsteknikker, for eksempel kryssvalidering, for å sikre robust ytelse på usett fremtidige datasett. Siden ytelsen til datadrevne modeller er avhengig av opplevelsen i treningsdataene, foreslås grundig datasettanalyse. Det foreslås en ny treningsdatasentrisk tilnærming for å forklare og tolke datadrevne modeller.

The doctoral rpoject concerns statistical sensor-based regression models and anomaly detection methods, evaluation methods, performance measures, and explainability and interpretation of black box models. The main contributions of the thesis include: - Development of a novel method to enhance the explanability and interpretations of opaque machine learning and statistical models. - Modifications of the semi-supervised anomaly detection framework based on reconstruction with AAKR and residuals analysis using SPRT to: 1) reduce computation time; 2) make it possible to distinguish between explanatory and response variables; and 3) use a regional credibility estimation in the residuals analysis to improve accuracy. - Demonstrations of the usefulness of data-driven methods for anomaly detection in maritime applications - Demonstrations of the importance of appropriate cross validation techniques for performance evaluation, especially for sequential sensor data.

Maintenance and inspections of ships performed by DNV GL have traditionally been based on a preventive scheme where components have been overhauled or maintained according to a time schedule. This philosophy is based on the assumption that a component has a defined lifetime, after which its failure rate increases. However, estimates of lifetime have large uncertainties and a large percentage of failures are not age-related, and are therefore not adequately addressed by preventive scheduled maintenance. DNV GL aim to develop radically new statistical approaches based on the recent availability of large arrays of sensors, which monitor condition and behaviour of machinery and power systems. Data are ubiquitous: almost every activity in which DNV GL engages produces and requires data. More and more ships get sensors installed, collecting more and more data. Data are the critical inputs into almost all decisions. Statistical inference is needed to turn data into knowledge, to understand unexplained mechanisms, discover hidden patterns, and predict future behaviours. As the accessibility, volume and complexity of data increases, new model based statistical and machine learning methods must be developed. Sensor data from ship generate very high dimensional time series, which relate to each other either in a known way (network of sensors) or in terms of stochastic dependence (correlation, coherence). These will be analysed for motif discovery, anomaly detection and classification. The purpose is to automatically alert about shifts in trends, variability or extremes, about changing patterns of behaviour or about any potential deviations from the norm. We will in particular consider state space models and other Bayesian hierarchical models. Both classification of faults, prediction of condition and faults, and optimization of performance are possible key aspects of this PhD.