This proposal seeks funding for the Center for Intelligent Maintenance Systems studies conducted by the University of Cincinnati site (lead), the Missouri University of Science and Technology site and the University of Michigan site. Funding Requests for Fundamental Research are authorized by an NSF approved solicitation, NSF 10-507. The solicitation invites I/UCRCs to submit proposals for support of industry-defined fundamental research.
The proposed research focuses on methods for controlling and evaluating the quality of data used in prognostic applications of system health and indicating when maintenance is needed. The proposal is well conceived, well organized, and the goals and objectives of the research are presented well. The tasks to be accomplished over the two years are clearly outlined, as well as which of the cooperating institutions will carry out the work.
The proposed research answers a significant research question raised by the industrial partners. For many companies assuring the quality of datasets before actually performing prognostics can avoid unnecessary investment in redundant prognostics analysis due to poor quality datasets. Assured data quality will improve prognostics results, which leads to better maintenance decisions and significant cost saving. The IMS Center has actively involved minority and female graduate students and has provided a number of Research Experience for Teachers and Undergraduates (RET and REU) projects.
In the past, significant resources and effort was spent on developing data processing and health monitoring models for condition based monitoring applications, often to only discover at the end, that the initial data collected was flawed and not suitable. Through this research grant, a methodology and set of tools were developed to evaluate the data quality using several metrics, in which one would know whether the data was suitable for developing reliable health monitoring models and methods. A novel spectral clustering method was used to detect outliers in the data and also used to evaluate whether the data was suitable for diagnosing different failure modes. A set of signal validation metrics were also evaluated for determining whether the vibration signal is correct and should be used for later analysis. Lastly, a set of metrics for evaluating whether the data is suitable for prognostics (failure prediction) were developed, including monotonic and trendability metrics. The developed methodology and tools were evaluated on two separate applications, with the first example dealing with a machine tool feed axis system used in manufacturing. The data quality evaluation tools were applied to this feed axis condition monitoring data, and effectively removed outliers and determine which data sets could be used to develop models to diagnosis the different failure modes. The second application dealt with renewable energy, and in particular a wind turbine condition monitoring system. The vibration signal validation metrics removed erroneous files. The monotonic metrics evaluated the health monitoring methods and determined that the component health trends were suitable for prognostics (life prediction). The data quality methodology, tools, and case studies were disseminated in two Industrial Advisory Board Meetings (IAB 21 and IAB 23), a book chapter, a journal paper, and two dissertations. The methodology and tools developed for data quality significantly impact the attended intellectual merits of this research grant. By proposing and then evaluating a systematic methodology for data quality for condition monitoring and prognostic purposes, scientists and engineers now have a procedure to follow when developing these predictive monitoring systems. No longer should significant resources be used to develop health monitoring models, to only realize at the end that the data quality was not suitable. The data quality methodology and tools were successfully demonstrated with two case studies, in which erroneous data was removed and health models could be developed in a more rapid and reliable manner. It is believed that the scientific community and industry will further improve and adopt this methodology for researching and developing condition based monitoring and prognostic systems. There is significant broader impacts that could be impacted by this data quality methodology and tools. Beyond the current manufacturing and renewable energy condition monitoring applications, there are vast other condition monitoring applications that would need this data quality evaluation (railway, solar panel degradation, electric vehicles). These broader applications would need a customized version of the data quality tools to ensure that the data is correct for further analysis and health modeling. In addition, as the "big data" aspect in manufacturing and other sectors become more prevalent, it is going to be even more imperative to have a suitable set of data quality tools to evaluate all this data and determine which is useful for further analysis. It is also quite understandable, that data quality metrics would be useful for a variety of other fields, including health care, social media data, financial data, surveillance/security, and so on.
