PSAM 16 Conference Paper Overview
Welcome to the PSAM 16 Conference paper and speaker overview page.
Lead Author: Tamer Tevetoglu Co-author(s): Bernd Bertsche bernd.bertsche@ima.uni-stuttgart.de
| A Machine Learning Approach to Enhance the Information on Suspensions in Life Data Analysis |
|---|
| Increasing digitalization and implementation of sensors in systems result in high data availability, which enables and benefits data-driven approaches. Commonly, these approaches revolve around predictive maintenance, anomaly detection, or clustering. In this paper, we analyze the practicality and performance of life data analyses based on neural networks. To this end, the Weibull analysis is extended with a machine learning approach and compared with conventional approaches in a laboratory test setup. Reliability engineers usually have budget and time constraints regarding testing strategies. These constraints manifest as an inability to accurately verify a system’s reliability with a pre-defined confidence due to small sample sizes, insufficient number of failures from testing, or inadequate choice of life data analysis methods. Conventional approaches in life data analysis counteract these constraints by taking suspensions into account or allowing to correct the bias when computing parameter estimates and confidence bounds. Hence, engineers only have limited number of tools in order to deal with constraints in reliability testing. Previous studies have shown that these counteracting measures may not be effective under certain circumstances, i.e. despite taking suspensions into account or using bias-corrections, parameter estimates may differ substantially from the ground truth. This may lead to a false sense of security regarding the operational life of a product. As data-driven approaches become steadily more important in other reliability engineering areas, e.g. Prognostics and Health Management (PHM), the focus of this paper lies on the analysis whether some shortcomings in life data analysis can be mitigated by using data-driven approaches in addition to or instead of conventional approaches. We develop a data-driven model that uses a neural network to recognize patterns in sequences of data, e.g. numerical times series data emanating from sensors. A trained model is able to output the remaining useful lifetime (RUL) of a system based on sequential sensor data like temperature, vibration, etc. In life data analysis, failures and their respective sensor data can be used to train data-driven model. This trained model is then being used to predict the RUL of the suspensions. If the predicted failure times are close to the unknown real failure times of the suspensions, one may use the predicted failures in addition to the actual failures, and thus may obtain more accurate parameter estimates and confidence bounds. In order to verify this proposition, we conduct a study to determine how using neural networks to increase the number of failures by predicting the RULs of suspensions actually performs against conventional approaches like bias-corrections. For this purpose, we use a turbofan engine data set from NASA and compare the performances of three Weibull analysis approaches to each other: 1. Maximum likelihood estimation (MLE) with bias-correction 2. MLE with machine learning 3. MLE with machine learning and bias-correction For each approach, parameter estimates and confidence bounds are evaluated for a censored subset of the data set with varying sample sizes and censoring shares. The first approach is based on the MLE in combination with the Hirose and Ross bias-correction. This bias-correction method performed best in a previous study. The second approach requires training a machine learning model with actual failures and subsequent prediction of the suspensions’ RULs. Then a conventional Weibull analysis is conducted with the actual and predicted failures. The third approach includes a subsequent bias-correction after using the machine learning model. Based on this simulation study, this paper’s main objective is to conclude on whether the use of neural networks can mitigate above mentioned shortcomings, and if so, what the precise prerequisites are. These prerequisites include the sample size, number of failures, number of suspensions, censoring share, and choice of methods. Our special attention lies on the minimum number of actual failures during testing that are needed in order to train an adequate data-driven model. |
Paper TA24 Preview
Author and Presentation Info
Lead Author Name: Tamer Tevetoglu (tamer.tevetoglu@ima.uni-stuttgart.de)
Bio: Tamer Tevetoglu studied Mechanical Engineering at the University of Stuttgart in Germany and received his academic degree Master of Science in 2017. Since 2017, he is a researcher in the Reliability Engineering Department at the Institute of Machine Components at the University of Stuttgart and pursues his PhD studies.
Country: Germany
Company: University of Stuttgart
Job Title: PhD Student
Bio: Tamer Tevetoglu studied Mechanical Engineering at the University of Stuttgart in Germany and received his academic degree Master of Science in 2017. Since 2017, he is a researcher in the Reliability Engineering Department at the Institute of Machine Components at the University of Stuttgart and pursues his PhD studies.
Country: Germany
Company: University of Stuttgart
Job Title: PhD Student