A feature extraction and visualization method for fault detection of marine diesel engines. (February 2018)

• Record Type: Journal Article
• Authors: Xi, Wenkui; Li, Zhixiong; Tian, Zhe; Duan, Zhihe
• Is Part Of: Measurement. Volume 116(2018)
• Page Start: 429
• Page End: 437
• Publication Date: 2018-02
• Subjects: Marine safety -- Diesel engines -- Condition monitoring -- Fault detection -- Feature extraction; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2017.11.035
• Languages: English
• ISSNs: 0263-2241

Abstract

Graphical abstract: A feature extraction and visualization method is developed for fault detection of marine diesel engines. It firstly constructs a time-frequency reference signal to supervise the vibration-source separation of kernel independent component analysis. Then perform the manifold learning algorithm (t-SNE) to visualize the fault features. Lastly, an extreme learning machine is applied to the fault pattern recognition. Highlights: Condition monitoring on marine diesel engines. An automatic vibration-source separation method using supervised kernel ICA. Feature extraction and visualization using manifold learning. Experimental data was used to evaluate the proposed method. Abstract: The reliability of marine diesel engines is extremely important for a safe navigation. For the condition monitoring of diesels engines, the independent component analysis (ICA) has been proven to be effective in separating useful vibration sources from the engine vibration signals. However, ICA still needs expert knowledge to identify the source of interest. To avoid human factors in ICA, an automatic vibration-source extraction and feature visualization method is proposed in this paper for fault detection of marine diesel engines. In this method, the Stockwell transform was used to construct a time-frequency reference signal to guide the separation process of the kernel ICA. Only the fault-related source was separated by this improved time-frequency supervised kernel ICA (TFSKICA). Then the t-distributed stochastic neighbor embedding (t-SNE) was employed to extract and visualize the fault features. Lastly, the extreme learning machine (ELM) based classifier was built to identify the engine faults in an intelligent manner. Experimental data acquired from a commercial diesel engine was used to evaluate the performance of the proposed method. The analysis results demonstrate that the TFSKICA is able to separate the vibration source of interest for distinct fault feature extraction by the t-SNE in a visualization manner. The fault recognition rate of the proposed method is also better than that of some existing approaches.