A residual estimation based approach for isolating faulty parameters

Abstract

This paper presents a new residual estimation based diagnostic approach that includes detection and fault isolation using the Mahalanobis distance (MD). The faulty performance parameter isolation approach is based on the analysis of residual MD values. The residual value is calculated by taking the difference between MD values estimated in two different scenarios: first, when a performance parameter is present, and second, when that performance parameter is absent. The residual of the MD values for each parameter is obtained by using training data from several experiments as part of the training data analysis planned by the design-of-experiment concept to analyze the impact of each parameter. The distribution of residual MD values for each parameter is analyzed and a 95% probabilistic range is established. This range represents the expected contribution by parameters toward a healthy system's MDs, and it is used to identify the parameters that are responsible for the anomalous behavior of a system. Parameters that fall below the lower bound of the 95% probabilistic range are considered candidates for the anomalous behavior, and the parameter that has the lowest residual value is isolated as the faulty parameter. A case study on computers is presented to demonstrate and test the suggested new approach's ability to isolate faulty parameters. 1, 2

Sensor Systems for Prognostics and Health Management

“…Fault isolation can also be done by using mathematical models, such as principle component analysis (PCA) [ 1 , 28 ] and residuals estimation [ 29 ], to analyze the data from “general” sensor systems. However, if the considered sensor system has the ability to detect and isolate the faults or failure mechanisms, it will improve the efficiency of PHM and provide more direct information.…”

Automatic data mining for telemetry database of computer systems

“…In traditional statistics analysis, principal components analysis is used to reduce the dimensions of parameters. Kumar et al [6,7] used Mahalanobis distance (MD) to isolate the faulty parameters from the observed data and determined the health condition of a computer system. They showed that MD is a good health index for measuring multi-variate monitored parameters.…”