Investigation of Model-Based Diagnostic Methods, 09-R8104

Printer Friendly Version

Principal Investigator
Matthew B. Ballew
 

Inclusive Dates:  09/28/09 – 04/01/11

Background - This project explored various diagnostic techniques utilizing empirical data and physics-based models. Previous projects addressed anomaly detection technologies. This project is related to those previous efforts by addressing diagnostics, which is the next step of the detection, diagnostic, and prognostic process that enables a Condition Based Maintenance (CBM) system. In improving and automating the diagnostic process, project team members developed a configurable physics-based model to be fused with empirical sensor data acquired from in-flight data. The improved and automated methods developed for this project represent advancement beyond the state of the art for aerospace propulsion engine diagnostics.

Approach - The main objectives of this project were to create a configurable physics-based model and to develop analysis methodologies that fuse the physics-based information with sensor data to provide a high level of diagnostic automation and precision.

Accomplishments - With this approach, pattern/regime recognition and classification algorithms and particle swarm intelligence-based diagnostic algorithms were created and tested. The pattern recognition algorithms proved reliable for operational regime classification and sensor validation. The particle swarm diagnostic algorithms fuse the formerly disparate data sets of empirical data and physics-based models. These algorithms reduce diagnostic time by automatically directing maintenance personnel to the failed module, even if the failure is not predefined as a fault.

Future efforts for this project will address using the pattern recognition algorithms as a diagnostic tool for alerting users to previously defined failures, using the particle swarm optimization algorithms for selecting an optimal set of features to be used for diagnostics, and using the physics-based model to calculate and test the use of additional health parameters that are available once the model and data are fused.

Figure 1. Cost information associated with one particle during a diagnostic analysis. By communicating with other nearby particles, the particle iteratively changes its position to achieve the lowest possible total cost. This information is then mapped to diagnostic metrics identifying the failed component.


Figure 1. Cost information associated with one particle during a diagnostic analysis. By communicating with other nearby particles, the particle iteratively changes its position to achieve the lowest possible total cost. This information is then mapped to diagnostic metrics identifying the failed component.

 
 
 

2010 Program Home