Modeling machine degradation with Hidden Markov Models
Access full-text files
Date
2018-05
Authors
Whittemore, Anthony Ryan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Description
The objective of this project is to evaluate and justify the use of a threshold condition based maintenance policy. HMM based methods will be developed and explored for the diagnostics and prognostics of stochastically degrading machines over a fixed time horizon. HMMs are a powerful tool with a formal probabilistic basis. By applying Bayesian probability methods, the parameters that describe an HMM can be adjusted in response to a set of observations emitted from the model. One such famous use of HMMs is in the field of automatic speech recognition (Rabiner, 1989). They are a simplified version of dynamic Bayesian networks, which is a probabilistic graphical model used to represent a set of random variables and their conditional dependencies over adjacent time steps. Through the popularity of automated learning techniques, HMMs have seen use in areas of manufacturing (analyzing bearing vibration data) (D. Tobon-Mejia, 2011), and even energy production (accident classification in nuclear power plants) (KC Kwon, 1999). Many systems found in the manufacturing realm are complex and dynamic in nature. The complexity aspect arises from a manufacturing facility’s random machine downtimes and repair times. The dynamic aspect arises naturally because machines accumulate a cost if they go unrepaired, producing faulty products. In addition, HMMs can be used as classification models and are impartial to the types of data being passed into them. This provides flexibility in the model structure needed to adjust the model, efficiently providing robust results. Los Alamos National Lab has initiated new manufacturing facilities with large, expensive, and dangerous equipment. They have ramped up production for the creation of thermal power supply (TPS) units and intend to continue to meet increasing demand. After each assembly, the TPS units undergo shock, thermal, electrical and vibration tests. These tests are used to both verify that they will provide power under extreme conditions, and that the manufacturing process is within specifications. During the tests, the only observable quantity is the voltage profile recorded during each test. If the voltage falls below a threshold it is labeled as “fail” and sent to another facility to be decommissioned. The overall vision of this report is to investigate optimal threshold maintenance policies that will provide base line estimates of proper time to repair machines at LANL. Currently, there is no maintenance policy set in place. This begs the question of where to begin in designing quality control standards and procedures. With no quality control policies in place, cost will dramatically increase due to the release of faulty products, devastating damages to customers, or possible in-house catastrophic failures. HMM methods will be utilized in estimating the optimal time to schedule machine maintenance. The model proposed in this paper will provide a general framework to build upon to create standardized quality control practices across the lab.