Insight Into Active Health Monitoring Methods Using Machine Learning

Current satellite validation tests involve numerous procedures to qualify the space vehicle for the vibrations expected during launch and for exposure to the space environment. Structural Health Monitoring methods are being considered in an effort to truncate the number of validation tests required for satellite checkout. The most promising of these monitoring techniques uses an active wave-based method in which an active piezoelectric transducer propagates a Lamb wave through the structure, where it is then received by a second sensor and evaluated over time to detect structural changes. Thus far, this method has proven effective in locating structural defects in a complex satellite panel; however, the attributes associated with the first wave arrival change significantly as the wave travels through ribs and interfaces. This study establishes a method to identify important features in the sensor signal that may otherwise be missed. In this work, an FE model of a plate is developed, and variability is introduced for each state observation. Different states are modeled as masses distributed at 12 locations. Signals are obtained through an explicit time-domain analysis. Matching Pursuit Decomposition is used to extract physically significant parameters from the signal, the features are reduced to a more manageable subset, and Support Vector Machines are utilized to identify important features in the sensor signal. The results of the classification indicate that for a plate specimen the extracted feature corresponding to anti-symmetric wave mode interaction with damage does the best job of differentiating damage states.