Telemetry data analysis using multivariate sequential probability ratio test

摘要

One embodiment provides a system that analyzes telemetry data from a monitored system. During operation, the system periodically obtains the telemetry data as a set of telemetry variables from the monitored system and updates a multidimensional real-time distribution of the telemetry data using the obtained telemetry variables. Next, the system analyzes a statistical deviation of the multidimensional real-time distribution from a multidimensional reference distribution for the monitored system using a multivariate sequential probability ratio test (SPRT) and assesses the integrity of the monitored system based on the statistical deviation of the multidimensional real-time distribution. If the assessed integrity falls below a threshold, the system determines a fault in the monitored system corresponding to a source of the statistical deviation.

主权项

1. A computer-implemented method for analyzing telemetry data from a monitored system, comprising, in at least one computer, performing one or more operations for:
periodically obtaining the telemetry data as a set of telemetry variables from the monitored system; using the obtained telemetry variables to update a multidimensional real-time distribution of the telemetry data that represents a multidimensional distribution of pointwise differences in values of the telemetry variables observed at the monitored system and values of the telemetry variables observed at a known-good system, wherein a first dimension for the multidimensional real-time distribution corresponds to a first type of the telemetry variables, and wherein a second dimension for the multidimensional real-time distribution corresponds to a second type of the telemetry variables that is different from the first type; using a multivariate sequential probability ratio test (SPRT) to analyze a statistical deviation of a parameter of the multidimensional real-time distribution from a parameter of a multidimensional reference distribution for the monitored system that represents a multidimensional distribution of pointwise differences in values of the telemetry variables observed at a normally operating system and values of the telemetry variables observed at the known-good system, and wherein using the SPRT comprises determining a drift in values of a matrix for the parameter of the multidimensional real-time distribution from values of a matrix for the parameter of the multidimensional real-time distribution; assessing the integrity of the monitored system based on the drift, wherein assessing the integrity comprises, when the drift falls below a threshold, determining a fault in the monitored system corresponding to a source of the statistical deviation; determining a degrading field replaceable unit (FRU) of the monitored system that caused the statistical deviation by using at least some data in the multidimensional real-time distribution; and replacing the degrading FRU.