Research On Health Condition Assessment Of Marine Gas Turbinepropulsion System Based On Data-Driven Method

The marine gas turbine propulsion system is prone to performance decay due to its working characteristics and marine environment,which leads to various failures.Among them,the compressor blades and turbine blades are prone to fouling and corrosion under the combined action of air salt spray and the adverse working environment of high temperature,high pressure,and high speed inside the gas turbine.The propeller and hull are prone to marine fouling when they are immersed in seawater for a long time.Decay of the marine gas turbine propulsion system will lead to the reduction of ship economy,and may eventually lead to serious failures.For marine engineers,it is essential to learn the health state of the marine gas turbine propulsion system timely and accurately,and make a reasonable maintenance plan accordingly,which is the key to ensure the safe,economical and efficient operation of the ship.The thesis mainly studies several key technical problems in using data-driven methods to evaluate the health state of the marine gas turbine propulsion system,including qualitative decay detection,estimation of decay degree,qualitative detection of main decay direction,estimation of decay imbalance,and simultaneous decay detection.It is difficult to obtain a large number of fault samples from the real ship because the marine gas turbine propulsion system is often not allowed to run with faults,which seriously restricts the application of data-driven methods in the health assessment of the marine gas turbine propulsion system.In order to reduce the dependence on fault data and promote the application of data-driven intelligent diagnosis technology in the field of the shipping industry,the following research work has been carried out.A gas-path health assessment method for the marine gas turbine propulsion system based on one-class support vector machine is proposed.The main functions of the method include qualitative decay detection,estimation of decay degree,qualitative detection of main decay direction,and estimation of decay imbalance.Different from the traditional supervised learning algorithms,one-class support vector machine only needs one-class data to train the model.For decay detection,only normal samples are necessary to train the model.The model detects whether the new samples are within the decision boundary to judge whether the system has performance decay.Because only normal samples are used to train the model,the feasibility of the algorithm in real ship deployment is greatly enhanced.On the basis of qualitative decay detection,the decay degree of the system is estimated by using the decision values of one-class support vector machine.Comparing the estimated results with the decay degree determined by the decay coefficients,it is found that they are basically consistent in the trend.The estimation of decay degree enables marine engineers to have a quantitative understanding of the overall health state of the system and make a rough prediction of the remaining service life.The marine gas turbine propulsion system is very complex,with many components,and the decay degree of different components is likely to be different.That is to say,there is a problem of imbalanced decay.In order to locate and analyze the decay state of the system,the main decay direction and decay imbalance are studied.The purpose of main decay direction detection is to estimate which component is more seriously decayed for better maintenance.Under the assumption that the decay of each component develops along with different directions in high-dimensional space,the main decay direction of the system is determined by comparing the degree of the sample tending to each decay direction.Especially,when the whole system is in a state of moderate decay,there may be a serious decay of one component and only slight decay of other components.In this case,only the severely decayed component needs to be maintained.Main decay direction detection helps to reduce maintenance workload and avoid blindness of maintenance work.In order to describe the imbalanced decay of the system quantitatively,a method to calculate the imbalance degree of the system decay is proposed,which can estimate the imbalance degree of two-components decay quantitatively.It is found that the main decay direction and decay imbalance estimated by the proposed method is in good agreement with the actual state determined by the decay coefficients.The performance of six one-class classification algorithms,including one-class support vector machine,support vector data description,isolation forest,global k-nearest neighbors,local outlier factor,angle-based outlier detection,in health assessment of the marine gas turbine propulsion system is compared,which provides a reference for related applications.It is found that when the training set is pure normal samples,the one-class classification algorithms can easily solve the problem of qualitative detection of single component decay.However,it is difficult to extract completely normal training samples in reality.The normal training set is likely to be mixed with some abnormal samples,resulting in data contamination and a decline in algorithm performance.In order to test the tolerance of one-class classification algorithms to contaminated data,a series of comparative studies were carried out on the artificially constructed contaminated dataset.The results show that the performance of one-class classification algorithms decreases with the increase of the contamination ratio of the training set.Most one-class classification algorithms have little tolerance to contaminated data,and slight data contamination will lead to algorithm failure.The exception is that the isolated forest algorithm has a strong tolerance to contaminated data.In order to enhance the tolerance of the one-class classification algorithms to contaminated data,the algorithms are improved from the aspect of usage in the two cases of contamination ratio known and contamination ratio unknown.The performance of the improved algorithms on contaminated data sets has been significantly improved.Finally,the one-class classification algorithms such as one-class support vector machine,support vector data description,global k-nearest neighbors,local outlier factor are used to estimate the decay degree of a single component.Comparing the estimated results with the actual decay degree determined by the decay coefficients,it is found that they are in good agreement.The application of multi-label learning in simultaneous decay detection of the marine gas turbine propulsion system is studied,and a multi-label support vector machine algorithm based on binary relevance is proposed for simultaneous decay detection of the marine gas turbine propulsion system.Multi-label learning algorithms can assign multiple decay labels to a single sample and can deal with the problem of missing labels.It is suitable for the problem of simultaneous decay detection of the marine gas turbine propulsion system.Multi-label learning is a kind of supervised learning algorithm,which needs a certain amount of labeled samples to train the model.When all decay combinations are considered,the types of decay modes increase exponentially with the number of decay labels.In practice,it is difficult to collect so many labeled simultaneous decayed samples.Referring to the characteristics of multi-label learning which can deal with the problem of missing labels,only normal samples and single decayed samples which are easy to obtain are used to train the multi-label classification model.It is found that the proposed method has an excellent performance in detecting unknown simultaneous decay,and it outperforms binary relevance support vector machine,classifier chains support vector machine,multi-label k-nearest neighbors,binary relevance k-nearest neighbors for simultaneous decay detection of the marine gas turbine propulsion system.