Research On Community Detection Method In Complex Networks And Its Application To Fault Diagnosis In Marine Engineering

Complex networks is a new emerging discipline. It combines the knowledge from math, physics, computer science and sociology, and attracts many scientists from different fields in 21st century. A complex network is composed of many vertices and many edges. In most real networks, there exist some groups of vertices, the edges in which are denser. These groups are called communities. Community detection is an important topic of complex networks, which can help people understand complex systems and different phenomenons. Newman fast algorithm and label propagation algorithm are two classic methods for community detection. They have near linear time-complexity and need not preset the number of communities. The community detection methods are usually used to real network clustering, while data clustering, as the other branch of clustering problem, gets less attention, which is important for fault diagnosis for marine diesel engine. The marine diesel engine is the heart of a ship, and the fast fault diagnosis for marine diesel engine using Newman fast algorithm and label propagation algorithm is meaningful to ensure the safety of navigation. In this dissertation, the label propagation algorithm is generalized and is improved, and the fault diagnosis methods for marine engine system based on community detection theory are proposed. The research is developed from four aspects.1. Due to the automatic number of classes using Newman fast algorithm in data clustering, a fault diagnosis method for marine diesel engine based on the algorithm is proposed. Consider a sample as a vertex, and the similarity between two samples as edge weight, and a weighted network is constructed. The criterion function used in Newman fast algorithm is as the criterion function in bottom-up hierarchical clustering, and a cluster model is constructed. After clustering the samples, use cluster results to recognize the pattern of samples in fault diagnosis. The experimental results show this method is a fast, accurate and stable method. Moreover, the method can recognize the fault pattern, which is not in the history data.2. Label propagation algorithm is generalized to make it applicable to weighted networks and be used in clustering fault data from marine diesel engnine. The three important contents of the algorithm are the initial assignment, the rule and the termination condition of label propagation. By calculating the probability that two neighbor vertices are in the same community according to the principle of multiple edges, the rules of label propagation and the termination condition are weighted, and then the algorithm is generalized into the weighted networks. According to the community detection results on networks, the generalized label propagation algorithm is applicable to community detection for weighted networks; according to the clustering results on classic test dataset and fault dataset from marine diesel engnie, the generalized label propagation algorithm is applicable to data clustering.3. To delay the occurrence of trivial solution in label propagation process, a label propagation algorithm with prediction of explosive transition is proposed. Due to the randomness in label propagation, the trivial solution is obtained frequently, which affects the speed and the accuracy of the algorithm. By transforming the label propagation process into the construction of a network, the explosive transition phenomenon in the construction process of random network is connected with the occurrence of trivial solution, and then the prediction process of explosive transition is added to delay the occurrence of trivial solution. The concept of neighbor purity is generalized into weighted networks, and the incomplete update condition with vertex degree is given. The experimental results of community detection on networks show the improved algorithm is stable, less time-consuming, and sensitive to small communities. The experimental result of clustering on fault dataset from marine diesel engine, the improved algorithm won’t miss the small-scale classes, which make it applicable to the cases the scales of samples from different faults are quite different in fault diagnosis.4. Due to information loss in fault diagnosis using single clustering method and preset parameters or methods needed in fault diagnosis using multiple clustering method, a fault diagnosis method for marine diesel engine based multi-time label propagations is proposed. The result of label propagation algorithm may be various. The improved algorithm can be used to multi-time cluster fault data of marine diesel engine. The multiple results are aggregated or a unique result is confirmed and then the final cluster result is obtained. The clustering center is used to recognize the pattern of new fault sample. The experimental results show the method is a fast, accurate and stable method.