Maintenance and inspections of ships are traditionally based on a preventive scheme where components have been maintained according to a time schedule. This approach 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 in this way. We develop new approaches based on the recent availability of large arrays of sensors, which monitor condition and operation of machinery and power systems.
Sensor data are becoming available for the first time on global ship fleets, with efficient broadband connectivity to shore. We suggest new approaches to condition 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. For the design of sensor monitoring systems, a key challenge is to determine the level of resolution in time and sensor density needed to have a precise dynamic picture of the actual health of the system. Borrowing strength across sensors and ships in a fleet is an important aspect, leading to increased safety of a whole fleet.
Overview of data driven statistical methods for condition monitoring systems
With focus on data-driven methods, we have reviewed statistical methods specifically relevant to condition monitoring of ship machinery systems. A number of statistical methods and approaches which are relevant for diagnostics and prognostics of ship machinery systems based on sensor data have been identified and studied. This study has resulted in big insight and a paper published in the International Journal of Condition Monitoring in 2018.
Time-efficient on-line anomaly detection methodology for the maritime industry
Because the method based on AAKR and sequential testing for maritime sensor data that we published in 2016, is highly computer intensive, we have developed and tested several modifications of the methodology in order to increase time-efficiency and still produce robust results. Especially, we have modified the way training data is stored and used in the reconstruction step, by clustering the training data in AAKR, which leads to significant speed-up. The improved one-line anomaly detection schemes have been tested on several multidimensional maritime sensor data sets. A conference paper is published in the Proceedings of the IEEE International Conference on Sensing, Diagnostics, Prognostics, and Control. A journal paper is in revision for publication.
Cluster based anomaly detection
We have developed unsupervised anomaly detection methods based on various clustering techniques. The idea is to identify clusters in sensor data in normal operating conditions, and assess whether new data belong to any of these clusters. Applied to sensor data from a marine diesel engine, this class of methods turns out to provide an efficient, completely unsupervised initial screening of data streams for anomalies.
Anomaly detection using dynamical linear models
As an alternative to AAKR for signal reconstruction under normal operations, we have used a number of multivariate dynamic linear models for the reconstruction, and combined with sequential testing of residuals. A challenge to this type of modeling is the heterogeneous correlation structure in time of our maritime data series. A conference paper is published in the Proceedings of the Annual Conference of the Prognostics and Health Management Society. A journal paper is accepted for publication in International Journal of Condition Monitoring and Diagnostic Engineering Management.
Anomaly/fault prediction from ship operation log files
One of the condition monitoring systems in a ship sends messages regarding the operational mode of the ship at irregular time points. This log file works on a finite alphabet of possible events, and the purpose is to analyze this log file to detect sequences of events which appear to be abnormal or correspond to failures, preferably giving a warning early enough to be able to take action. We have attacked this problem in different ways, building on feature extraction, penalized variable selection, hidden Markov models and machine learning methods from speech recognition. The results are promising and two papers are in preparation.
Sequential detection of changes using dimension reduction techniques
We have developed methods for sequential change detection in high dimensional streaming data, with the aim of detecting changes in the distribution of the data as soon as possible, keeping false positives as a minimal level. We have extended sequential detection methods for changes in mean level to yield also changes in variance and covariance between streams. Such changes are impossible to detect with univariate monitoring methods, of course, but might represent important failure modes in a ship monitoring sensor system. Furthermore, we have studied properties of dimension reduction techniques such as sketching and PCA in connection to sequential change detection, assuming sparsity, that is, that changes happen only in a few of the monitored streams. A master thesis on this topic was finalized in May 2017, and a paper is in preparation.
Change point detection in maritime sensor data using PELT
We are investigating the use of the PELT-methods (Pruned Exact Linear Time, developed by our collaborators in Lancaster) for multiple change point detection in maritime sensor data. A master thesis will be finalized in 2018 and we will then understand if the approach is useful for early change point detection in the marine context.
Hull condition monitoring
We have been working with the construction of an emulator for the fatigue rate of a ship, given its physical characteristics and operating history, including environmental and weather data. The emulator is compared to calculations from a complex physical model run by DNV-GL. The aim is to be able to use the emulator as a fast and precise supplement to the full model. First analyses comparing results from the physical model and the emulator show that the correspondence is adequate in most conditions. Also in connection to hull condition monitoring, we have collected sensor data on bending moments from several vessels together with their history. The aim is to evaluate the precision of an existing physical model for these quantities. First analyses comparing sensed bending and calculated bending for the same ships show that the correspondence is satisfactory in certain conditions, but not always.
Surprise detection for a motor cooling system
We study situations where interest lies in a multivariate response variable to be monitored in time and prone to anomalies which can be related to a set of covariate time series. Regression methods allow studying the residuals and investigating surprises appearing in these residuals. This approach has been applied to the case of a motor cooling system with good results. We are transferring the method to the partner.