Decay Detection Of Marine Gas Turbine Based On One-class Classifier Algorithms

Marine gas turbine is an important main propulsion power plant.However,due to the operation in the harsh environment of salt fog erosion and changeable sea conditions,after a period of operation,it will often decay due to pollution,thermal corrosion,wear and other reasons after a period of operation.Therefore,timely prediction of the decay of marine gas turbine will effectively achieve the condition-based maintenance of this equipment,and then improve the operation reliability and the propulsion performance of the ship.Generally speaking,most of the samples collected by the ship automation system are normal operation data.At this time,the traditional supervised learning algorithm can not be directly used for the decay state evaluation of the equipment because it focuses on data sets with enough normal samples and abnormal samples and all tags are known.On the other hand,the normal operation data of ship equipment is likely to be contaminated by a small number of unmarked abnormal samples and noise samples.These pollution data mixed in the normal operation data,to a certain extent,weaken the decay detection effect of the model.However,the literature shows that the two one-class classifier algorithms,One Class Support Vector Machine and Isolation Forest,only need a single class of samples to realize anomaly detection,and the Isolation Forest algorithm also has a good ability to tolerate the pollution data.In view of this situation,this paper proposes two methods of ship equipment decay detection based on one-class classifier algorithms with different characteristics,that is,One Class Support Vector Machine decay detection method for pure normal data set and Isolation Forest decay detection method for pollution data set.In this paper,firstly,the decay data set of the equipment is preprocessed:the invariant and linear related attributes are removed and the data is normalized.At the same time,this paper also constructs 40 pollution data sets with pollution ratio from 0 to 100%.After the above processing,this paper uses pure normal data set and pollution data sets with different pollution ratios to construct multiple One Class Support Vector Machine and Isolation Forest decay detection models.At the same time,in order to fully compare the detection effect of One Class Support Vector Machine and Isolation Forest decay detection models,this paper uses the grid search/10 fold cross validation method to optimize the hyper-parameter n and kernel parameter g of One Class Support Vector Machine algorithm,the number of isolation trees T and subsampling size of Isolation Forest algorithm respectively.Among them,the hyper-parameter optimization process of Isolation Forest algorithm is introduced for the first time in this paper.Experiments show that under various stable conditions,the accuracy of One Class Support Vector Machine decay detection trained by normal operation data of equipment can reach more than 95%,and other evaluation indexes are also good.On the other hand,when the proportion of pollution data reaches 20%,the accuracy,recall,F1 score and AUC of the isolation forest decay detection model are still above 0.9.So we can see that the decay detection model based on One Class Support Vector Machine established in this paper,its novelty lies in:in the process of model training,it only needs the normal operation data of equipment,which greatly reduces the requirements of data marking.On the other hand,the decay detection method based on Isolation Forest proposed in this paper has the advantage that the model can still maintain high decay detection accuracy even if there are a large number of unmarked contaminated data in the data set.These two decay detection methods are complementary to each other,which can be used as decision support tools for ship equipment maintenance,and help to realize condition-based maintenance of ship equipment.