Chair: Dr. Jinyang JIAO (Xi'an Jiaotong University)
Reliable Faults Diagnosis of Roller Bearings using Hybrid Feature Models and Improved Multiclass Support Vector Machines with Classifiers Discriminant Analysis Mr. Manjurul ISLAM and Jongmyon KIM (University of Ulsan)
This paper presents a reliable multi-fault diagnosis scheme using hybrid feature models and an improved one-against-all multi-class support vector machines (OAA-MCSVM) classifier. Distinct feature extraction methods are concurrently applied on an acoustic emission (AE) signal to detect each fault condition about bearing defects. These fault features are composed of time- and frequency- domain statistical parameters and complex envelope spectrum analysis. While a dimensional features vector is further utilized with standard OAA-MCSVM classifier for diagnosis and identification, such classification method ignores individual classifier competence when results from multiple classes are agglomerated for final decision and therefore yields undecided and overlapped feature spaces where classification accuracy is severely degraded. To solve this unreliability problem, this study introduces a dynamic reliability measure (DReM) technique for individual SVM in the one-against-all framework. This DReM accounts for the spatial variation of the classifier’s performance by finding the local neighbor region of a test sample in the training sample space and defining a new decision function for the OAA-MCSVM. The efficacy of the proposed OAA-MCSVM with DReM classifier is verified with a fault diagnosis application for accurately identifying single and multiple-combined faults in low-speed rolling element bearings under various operating conditions. Experiment results demonstrate that the proposed classifier method is superior to the three state-of-the-art algorithms, yielding 6.19% to 16.59% improvement in average classification accuracy.
A Reliable Technique for Remaining Useful Life (RUL) Estimation of Rolling Element Bearings using Dynamic Regression Models Mr. Wasim AHMAD, Sheraz KHAN, Manjurul ISLAM and Jongmyon KIM (University of Ulsan)
Induction motors most often fail due to faults in the rolling element bearings. Sudden failures in a system result in long unscheduled downtimes, which cause huge economic losses. Prediction of imminent failures and estimation of the remaining useful life (RUL) of a bearing is essential for scheduling prior maintenance and avoiding abrupt shutdowns of critical systems. This paper presents a prognostics technique for rolling element bearings that utilizes dynamic regression models, which are updated recursively, to estimate the evolving trend in a bearing’s health. These models are then used to predict the future value of the bearing health indicator and estimate the RUL of the bearing. The proposed algorithm is tested on the bearing prognostics data from the Center for Intelligent Maintenance Systems (IMS). Experimental results demonstrate excellent prognostic performance and bearing’s RUL estimates within the specified tolerance bounds by effectively determining the time to start prediction (TSP) and dynamically calibrating the models to adopt to the evolving behavior of the bearing health indicator.
Application of Hidden Markov Model to Fault Diagnostics of Glass De-chuck System Mr. Hwanoh CHOI, Kyoungrae NOH, Sangmun YUN and Daewhan KIM (LG Electronics)
One of the methods to convey stuff in vacuum state, a dechuck operation which attaches object to glues and detaches object from glues has been used in engineered field. The dechuck operation installed in production line of glasses or wafers sometimes causes damages on product in detaching process. Since most of failures in dechuck system results from abnormal states of glues such as aged states or strong adhesive states, diagnostics of abnormal glues and a timely replacement of glues are the issue of the dechuck system. In this study, a testbed simulated dechuk system is devised and a hidden Markov model (HMM) is applied for diagnostics of the problematic glues in dechuck system of testbed. In order to apply the HMM, states, observations, and a sequence should be defined. The states are the elements of cause events set, the observations are the elements of result events set, and the sequence is the order of arbitrarily extracted observations. The HMM is usually utilized to estimate the posterior probabilities (also called Bayesian probabilities or conditional probabilities) that the observation at the specific time point is resulted from certain state given a sequence of observations. So as to adjust the dechuck system to the HMM, the states of glues are defined as the states, the features in load variation of glass and trace of load center variation of glass are chosen to the observations, and the chronological order of the features is employed to the sequence. As a result, detection rate of abnormal glues in dechuck system is improved by using posterior probability of HMM. We expect that the diagnostic method applying HMM is extended to the real dechuck system of glass production line so that this method will contribute to cost savings and reduction of defect rate.
Validation of Remaining Useful Life Prediction of Li-Ion Battery based on the Voltage Drop Ms. Yuri YUN, Seokgoo KIM, Kyusung JUNG and Joo-ho CHOI (University, Korea Aerospace University)
Batteries, which are used for the energy storage and power distribution, tend to degrade, and their capacity declines with repeated charging and discharging cycles. The battery is considered to fail when it reaches 80% of its initial capacity. In general, the battery state under operation can be characterized by identifying the state of health (SOH) and state of life (SOL), which refer the capacity degradation and remaining useful life respectively. Recently, authors have found that the SOH can be indirectly estimated based on the observation that the slope of voltage curve under charging is proportional to the capacity degradation. In the study, only the full charge and discharge cycles under room temperature were conducted with Li-ion battery, which is not the case in reality. In this study, more research is conducted to find out more reliable and robust measurement of the capacity and voltage drop that may be independent of the degradation conditions. Several tests are made under various C-rates, charging stabilization time and surrounding temperatures. Once succeeded, the regression model is established between the capacity and voltage drop, that is used in the estimation of the SOH. Adaptive Particle filtering (APF) framework is then applied during the battery usage to estimate the SOH and predict the RUL in the form of a probability distribution. In the APF, the recursive state transition and measurement functions are given by the empirical degradation model and the regression model, respectively. The APF performs the two functions at the same time which are the anomaly detection and prognostics. Experiments are conducted for a Li-ion battery by repeating full charge discharge cycles, in which a fault is imbedded to change the degradation pattern at a certain moment of the cycle to illustrate the technique.