Υλοποίηση και πραγματοποίηση αλγορίθμων προληπτικής συντήρησης
Design and implementation of predictive maintenance algorithms
View/ Open
Keywords
Διορθωτική συντήρηση (Reactive maintenance) ; LSTM ; SVM ; Matrix profile ; Τεχνικές ανίχνευσης ακραίων τιμών ; Συντήρηση ; Μηχανική μάθηση ; One class support vector machine ; Προγνωστική συντήρηση (Proactive maintenance) ; Προβλεπτική συντήρηση (Predictive maintenance) ; Προληπτική συντήρηση (Preventive maintenance ή PMS) ; Gradient boosting ; Anomaly detection ; Είδη ανωμαλιών ; Ανίχνευση ανωμαλιών στη ναυτιλία ; Level anomaly detector ; Gradient boosting classification approach ; Weighted permutation entropy ; Second-level anomaly detector ; Single channel LSTMAbstract
As maritime vessels send increasing amounts of sensor data to their data centers, improved data analysis tools are needed to help operations engineers, who monitor this data, to reduce operational risk. Given that, Predictive Maintenance is valuable part of reducing the overall costs and operational risk. In this essay, we demonstrate the effectiveness of some Anomaly (or Novelty) Detection approaches on time-series data, based on Machine Learning and Statistics, such as, Long Short-Term Memory (LSTMs) networks, One Class SVM, Matrix Profile, Cointegration check etc. in predicting the condition of certain parts of vessel main engine. The data on which we studied the different approaches came from data streams of sensor installed in 10 vessels of a Major Greek Maritime Company. Finally we propose a complementary model consisted of a combination of the approaches studied promising betters results. The proposed models is comprised of two steps, the first step makes a prediction raising an alarm for high potential of an anomaly in the incoming data, whereas the second one gives a probability of this anomaly to indicated the studied defect about Crosshead Bearing.