Full-life dynamic identification of wear state based on on-line wear debris image features
Introduction
Wear is one of the most important indicators of engine's performance. It affects not only the wear behavior but also the dynamic and thermodynamic performance. State monitoring is effective method to acquire the changing performance of a running machine. With the development of condition-based maintenances, real-time monitoring becomes the focus of engine's state monitoring [1]. However, wear state monitoring and identification is becoming a bottleneck in real-time monitoring due to the complex characteristics of tribological systems, such as system-dependent, time-dependent and physical coupling [2]. Therefore, the study of real-time wear state monitoring and analysis is critical to engine's full-life performance maintenance.
Although wear mechanisms have been comprehensively studied with various specific tribo-pairs, researchers still encountered many difficulties in understanding the wear state of a machine in service [3]. In the following sections, we will discuss the progress of wear monitoring technology focusing on two main aspects: monitoring methods and analysis methodologies. For monitoring methods, people have tried to obtain the real-time wear state by means of many indirect measurements such as vibration, temperature and performance parameter monitoring. Eventually, only the information of failure was obtained, while the initiation and propagation of the fault were absent [4]. In essence, wear is a process characterized with systematic effects and multi-field couples, thus indirect monitoring seems difficult to provide profound information about the mechanisms. Oil analysis, a direct analysis means of wear mechanisms, has been adopted in many engineering applications for wear and lubricant monitoring [5]. When applied to the real-time monitoring of an engine's performance, traditional oil analysis encounters two problems: one is the lack of on-line monitoring means, which retards obtaining the real-time information, and the other is the shortage of knowledge about full-life variations, which prevents reliable fault identification and maintenance decision making. In analysis methodologies, as analytical models are generally difficult to build for most engineering problems, most of the literatures focused on identification models based on monitoring data, namely data models [1]. Some were confined to condition data, others included event data. Event data can be used to assess current condition indicators and their performance, but needs the involvement of human. Some time-sequence data models, like the ARMA model by Jihong Yan [6] and the discrete Markov process by D. Banjevic [7], focused on trend predication and abnormality warning [8]. However, for data models, the accuracy of predication is the main problem for monitoring. Self-learning data models, like the artificial neural network model [9], have higher identification accuracy with self-adjusting property. However, the accuracy of an artificial neural network model is determined by not only suitable variables and initiations but also large number of both normal and abnormal samples [10]. Generally, data models can reflect the regularities among the mass of data, but not physical mechanisms, e.g. wear mechanisms.
Accordingly, comprehensive understanding of wear state monitoring and identification should include the following aspects:
- (1)
Wear states are fully determined by micro-scale wear mechanisms and macro-scale wear quantity. Therefore, data models, without physical mechanism involvements, have congenital deficiencies for characterizing wear state.
- (2)
Wear state is highly dependent on machines with different tribo-components and working conditions, thus is incomparable for different machines. Therefore, each machine needs a particular prediction model.
- (3)
Wear is affected by many factors including material properties, lubricants and even conditions, thus wear properties are random over a short period. On the other hand, wear is also a process of structure damage and material loss, thus wear properties are regular over a long term. Therefore, both dynamic and statistical methods should be taken into consideration for modeling.
In this paper, a new wear state characterization modeling method was investigated with on-line wear debris images. The characteristics of the wear debris images were adopted for characterizing wear rate and wear mechanism. For full-life monitoring, an automatic identification model was investigated with two categories: normal and abnormal. Finally, the method was examined with real-time image data sampled from an engine bench test.
Section snippets
Wear state characterization based on features extracted from on-line wear debris image
Wear state characterization is the premise of wear analysis. Furthermore, wear monitoring is the base of wear state characterization. As described above, direct monitoring is necessary to understand wear state more in depth. Wear debris is the by-product of wear process, thus the images of wear debris contain profound information of not only wear quantity but also wear mechanisms. Besides, on-line monitoring is necessary for real-time analysis. Correspondingly, an on-line visual ferrograph
Dynamic matching model of wear state based on Support Vector Data Description
Wear is different from break failures as it is a gradual and accumulative process with distinguished stages. Therefore, wear performance remains similar in the same stages because of similar wear mechanisms and characteristics. However, there is no definite boundary between every two stages. Accordingly, dynamic matching based on wear characteristics, other than turning point, is suitable for stage identification. In this section, a dynamic matching model for wear characterization was studied
Wear analysis of engine bench test with on-line ferrograph data
Two stages were included in the engine bench test: the running-in and the durability stages. The test was carried out in the Institute of Automotive Engineering of First Automobile Works, commonly known as FAW. The running stage lasted for about 20 h continuously and oil was changed at the end of the stage. The entire test lasted over 200 h.
The on-line images were sampled at intervals of 2 h. Nine representative images in different stages are shown in Fig. 8. The two indicators were calculated and
Conclusions
Aiming at real-time monitoring of wear states, quantitative wear state characterization was studied. On-line wear debris images were used to extract two quantitative indicators for characterization. Furthermore, the full-life features of the wear state evolution were analyzed for dynamic identification. The corresponding modeling method and the effects in an engine's application were studied. The main conclusions drawn are as follows:
- (1)
Wear state was characterized by wear mechanism and quantity.
Acknowledgments
The financial support of the present research was provided by The National Science Foundation of China (Grant no. 51275381), the China Scholarship Council financial support (Grant No. 201206285002), and the China Postdoctoral Science Foundation funded project (Grant no. 201003672). The author would also like to acknowledge The Institute of Automotive Engineering of First Automobile Works, commonly known as FAW. The author is also most grateful to the anonymous referees and the Editor for their
References (17)
- et al.
Development of an optimized condition-based maintenance system by data fusion and reliability-centered maintenance
Reliability Engineering and System Safety
(2010) - et al.
Towards the development of an automated wear particle classification system
Tribology International
(2006) Wear particle analysis—utilization of quantitative computer image analysis: a review
Tribology International
(2005)On-line and indirect tool wear monitoring in turning with artificial neural networks: a review of more than a decade of research
Mechanical Systems and Signal Processing
(2002)- et al.
Support vector domain description
Pattern Recognition Letters
(1999) - et al.
A boundary method for outlier detection based on support vector domain description
Pattern Recognition
(2009) Ferrography-then and now
Tribology International
(2005)- et al.
Application of ferrography in condition based maintenance
Strojarstvo
(2010)