A Low Frequency Uni-variate Model for the Effective Diagnosis and Prognosis of Bearing Signals Based Upon High Frequency Data

Jamie L. Godwin and Peter Matthews
Submission Type: 
Full Paper
phmc_14_004.pdf595.26 KBSeptember 11, 2014 - 3:42am

Prognosis of rotating machinery is of vital importance to ensure ever increasing demands of availability, reduced maintenance expenditure and increased useful life are met. However, the prognosis of bearings typically employs techniques in the frequency or time-frequency domain due to the high frequency nature of the data involved (typically >20 KHz). This data quickly becomes unmanageable in practice and often has inferior prognostic horizons in comparison to those techniques which are based upon low frequency data analysis.
This paper presents a novel methodology based upon the computation of the deviation from the empirically derived cumulative density function (CDF) of bearing data. For this purpose, the non-parametric, two sample, uni-variate Kolmogorov-Smirnov test is employed for the analysis. In particular, this paper focuses on mitigating the requirement of a-priori knowledge for bearing prognosis.
Initially, assumptions regarding the underlying structure of high frequency bearing data are explored on publically available data, and found to deviate from what would be expected.
Exploiting this, we use the non-parametric two-sample uni-variate Kolmogorov-Smirnov test to define normal operational behaviour, whilst mitigating the requirement for a-priori knowledge. This reduces the computational complexity of the system whilst having the prospect to reduce the inherent noise within the high frequency bearing signal.
Strong trends of degradation which can be used to derive prognostic maintenance conditions are observed, with sound statistical analysis performed. In particular, statistically significant degradation is found to occur 75 hours before failure occurred (representing identification at 54.2% of bearing life). Both the Kolmogorov-Smirnov D statistic and p-value are employed as health metrics to which degradation can be inferred from. A series of 4 experiments is presented, showing the versatility of the described technique and cases where the technique cannot be employed.
The technique is validated on a failed bearing and then verified on an independent, healthy bearing, and is shown to correctly identify the bearing of question in each case, enabling the prioritisation of maintenance actions which can be used to assist in reducing overall maintenance expenditure.

Publication Year: 
Publication Volume: 
Publication Control Number: 
Page Count: 
Submission Keywords: 
Applied statistics
High frequency
Submission Topic Areas: 
Data-driven methods for fault detection, diagnosis, and prognosis
Model-based methods for fault detection, diagnostics, and prognosis
Physics of failure
Submitted by: 

follow us

PHM Society on Facebook Follow PHM Society on Twitter PHM Society on LinkedIn PHM Society RSS News Feed