Anomaly Detection Using Transient Engine Performance Data, 09-R9776

Principal Investigator
Matthew B. Ballew

Inclusive Dates:  01/01/08 – 12/31/08

Background - Jet engine performance data are analyzed for adverse trends or step changes based on an engine's operating baseline. An undetected engine component failure can lead to unscheduled maintenance actions, reduced asset availability, and catastrophic engine failures soon after the initial component failure. Currently, only steady-state data, such as cruise data, are being used for Engine Trending and Diagnostics (ET&D). High-stress transient conditions (e.g., throttle advance during takeoff) are currently being ignored. This project will investigate the feasibility of analyzing currently ignored transient data because it is thought that impending engine failures could be detected earlier in high stress transient conditions than in steady-state conditions.

Approach - The objective of this project was to determine if anomaly detection and diagnostic methodologies could be developed using transient engine performance data to better assess the health and condition of expensive assets. The following tasks were performed to reach this objective:

• Task 1: Gather Data. This task required gathering engine performance data files and maintenance data to support the following tasks.
   
• Task 2: Develop Reader. An existing software program (originally designed to ignore transient conditions) was altered to extract all performance data from the files.
   
• Task 3: Develop Method of Retrieving Conditional Data from Database. During this task, project team members explored various ways to define a transient condition and then refined the algorithms that extract the desired transient data.
   
• Task 4: Develop Method to Analyze Transient Data for Anomalies. The objective of this task was to develop a method to analyze the filtered data for anomalies. This included developing a method to present data in a repeatable format and developing multivariate statistical algorithms to automate the analysis of those data.
   
• Task 5: Compare Anomalies to Engine Maintenance Records. During this task, performance anomalies were detected and then correlated to historic maintenance data to test the accuracy and precision of the algorithms.

Accomplishments - All tasks were successfully completed. Methods of conditioning the nonlinear transient data have provided linear baselines from which to detect adverse engine conditions. The multivariate statistical algorithms proved to be an effective method of identifying anomalies and charting trends. The selected decomposition method has proved to be effective in identifying which parameters caused observations to signal. Findings from this project resulted in a demonstration tool that can rapidly be applied to data sets for proof-of-concept purposes.

Figure 1. This chart shows analysis results from an engine that had sustained compressor damage. This damage eventually caused a mission abort. SwRI-developed algorithms detected this problem one month before the mission abort.

2009 Program Home