Research On Data-driven Decay Detection Of Marine Gas Turbine Propulsion System

Marine gas turbine propulsion system is prone to performance decay due to its own working characteristics and marine environment,which leads to the reduction of ship economy and safety.Therefore,it is very necessary to timely and accurately assess the health status of the propulsion system and formulate reasonable maintenance plan accordingly.It is difficult to obtain a large number of marked decay data from the real ship because the ship propulsion system is often not allowed to operate with faults,the current situation that decay data is difficult to obtain seriously restricts the application of data-driven methods in marine gas turbine propulsion system decay detection.In order to solve this problem,from the point of view of minimizing the demand of algorithm model for decay data,this thesis carried out the following research work:The performance of six one-class algorithms in the single component decay detection of marine gas turbine Propulsion system is compared and studied,which provides a reference for the application of this method in the real ship.When the performance of single component of marine gas turbine propulsion system decay,considering the difficulty of data collection in practice,this thesis uses one-class algorithm to detect decay.Unlike the traditional supervised learning algorithm,one-class algorithm only needs a kind of normal data to train the model,and the requirements of training data are low.In this thesis,a series of comparative experiments is carried out from the influence of the size of training set,the influence of training sample combination,the distribution of decay coefficient of misclassified samples and the tolerance for contaminated data.Experiment results show that these six algorithms can achieve certain decay detection effect,but the performance is different.When the training set is a pure normal sample,the comprehensive performance of the ABOD(Angle-based Outlier Detection)algorithm is relatively good;When the training set is slightly polluted,the detection accuracy of the i Forest(Isolated Forest)algorithm is the highest;When the training set is severe polluted,the detection accuracy of LOF(Local Outlier Factor)algorithm is relatively good,but only 70%.A multi-label simultaneous decay detection model for marine gas turbine Propulsion system based on SVM is designed.As the propulsion system usually includes multiple components,not only the decay of a single component should be paid attention to,but also the simultaneous decay of multiple components.Learning from the characteristics that multi-label learning algorithm can deal with the problem of missing labels,a multi-label model based on SVM is designed,the model is trained with normal data and single component decay data.Experiments are carried out on a public data set verified by real ship data.experimental results show that the multi-label decay detection model trained with only normal data and single component decay data can accurately detect most of the simultaneous decay.Compared with the single label multi classification algorithm,the proposed simultaneous decay detection model requires fewer kinds of training data,the model is easy to expand,and has the ability to detect unknown simultaneous decay,the practical feasibility is high.