Downloaded from https://www.everyspec.com ADS-79D-HDBK (indicating the need for maintenance action). Thresholds can be ―hard‖ where a single value is provided (example: bearing energy is normal below 1.25 ips) or ―variable‖ where a range of values is provided (example: marginal is between 3.2-3.3 ips). Estimation of RUL should provide a confidence interval identification of the incipient fault and the fault severity which is creating the degradation. If HI values are to be used to assess fault severity, sufficient data from fault validation testing and I/TDA should exist to fully understand the relationship of HI value to fault severity and the progression of fault severity with time.
HI values that are not well correlated to fault severity should not be used to estimate RUL. Prognosis, or the estimation of RUL, forms the basis for projecting the time at which maintenance action should be taken. Estimation of RUL through ―trend analysis‖ of HI values is only legitimate when: a. Data for the HIs is taken at frequent, regular intervals (application dependent based on the estimated time of failure growth). b. HI behavior with fault progression is not cyclical or highly non-linear. Prognosis through trend analysis should be biased to yield conservative estimates of RUL, with greater bias for cases where HI severity and failure progression data is incomplete or non-robust.
Estimation of RUL through model-based techniques is legitimate when: a. Baseline data for normal, non-faulted operation exists b. Baseline data for the specific serial number tracked item exists (taken within 10 hours of operation since installation). c. Fault data exists to sufficiently describe the behavior of the fault under the normal range of operational loading. The primary metric used to assess prognostic effectiveness is: Prognostic Accuracy 17 : A measure of how close a point estimate of failure time is to the actual failure time.
Assuming that, for the i th experiment, the actual and predicted failure times are ) ( i t af and ) ( i t pf , respectively, then the accuracy of the prognostic algorithm at a specific predicting time p t is defined as: i e N t ACCURACY , 1 ) ( 0 N D D p prognostic i 1 where ) ( ) ( i t i t D af pf i is the distance between the actual and predicted failure times, and 0 D is a normalizing factor, a constant whose value is based on the magnitude of the actual value in an application. N is the number of experiments. Note that the actual failure times for each 17 Vachtsevanos, G., F.L. Lewis, M. Roemer, A. Hess, and B. Wu. Intelligent Fault Diagnosis and Prognosis for Engineering Systems .
Wiley & Sons: New York, 2006. 75