Structural health monitoring is a critical part of aircraft maintenance. To avoid potential disasters, aircraft components must be repaired or replaced before they fail. However, to reduce unnecessary cost, components should not be repaired or replaced before action is needed. To address this balance, predictive maintenance of components is increasingly used. Predictive maintenance is the process of using data from observations, sensors, records, and other sources to estimate when maintenance needs to be performed. The goal of predictive maintenance is to provide cost savings over preventative maintenance, the more costly and conservative approach of replacing or repairing components well before maintenance is needed to avoid a catastrophic failure.

Traditionally, aircraft predictive maintenance is performed by observing and monitoring a specific component over time to identify damage in its early stages and thereby avoid structural failure. This can be done by taking and inspecting images of a structural stress point using the visible spectrum, infrared, and x-rays. However, the process of collecting and inspecting such images is burdensome and can be costly. More recently, researchers have turned to data fusion from onboard aircraft sensors for accurate predictive maintenance. These sensors typically already collect data and therefore avoid the need to collect and analyze images separately.

Extreme vibrations, acceleration, and other forces present on HH-60G helicopters can sometimes produce periods of severe data corruption or dropout. When data dropout occurs, damage prediction from sensors often defaults back to worst-case modeling, diminishing cost advantage. Therefore, reconstruction of sensor records during periods of dropout is of great interest. In this manuscript, we investigate the use of machine learning approaches to sensor data reconstruction for predictive maintenance on HH-60G helicopters. In particular, the sensor of interest is an accelerometer which yields three outputs for vertical, longitudinal, and lateral acceleration. Dropout present in one or more accelerometer readings will be reconstructed using data from more than 30 other sensors on board the aircraft.

We investigate two main machine-learning approaches. The first approach is a deep neural network that treats each time sample from the data records as an independent input. Though well vetted in previous studies, this approach ignores correlation between subsequent time samples. Therefore, as a comparison our second approach uses long short-term memory (LSTM) networks to exploit correlation in the time series data.  For each approach, we analyze the reconstruction effectiveness for purposes of predictive maintenance for the HH-60G.


Sensor-Data Reconstruction for Helicopter Structural-Health Monitoring Using Deep Learning

Division / Sponsors: 
Aerospace Division

Paper Proceedings: https://asmedigitalcollection.asme.org/SMASIS/SMASIS2020/volume/84027

SMASIS2020 - Virtual Conference

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.