Estimating the Life of an Aircraft Engine

For NASA or any other company within the aerospace industry, the ability to monitor the degradation of an aircraft engine is pivotal. In this blog, we will investigate how to increase a Machine Learning model’s ability to help predict when engine breakdowns will occur and how to estimate the remaining useful life on an aircraft engine. The dataset in this model is simulated for commercial aircraft and, therefore, if an unexpected breakdown occurs, it will likely do so while in the air with customers aboard, causing a significant problem. As such, being able to monitor the condition of an aircraft engine is invaluable. Furthermore, aircraft engines are far from inexpensive, meaning there is an added economic benefit to understanding its useful service life.

For the purposes of this test, we will look at a standard turbofan engine. The study is simulated by C-MAPPS, the most accurate method of simulating a turbofan engine, and only uses pre-installed sensors on the turbofan engines such as temperature and pressure sensors. The data was simulated with a low sample frequency of 1Hz. Therefore, the longest life-cycle for this dataset is 362 cycles. For this test, the dataset was applied to 7, 30, and 100 breakdowns. Let’s take a look at each one.

When testing a aircraft model with 7 breakdowns, the first 4 breakdowns were used for training, 1 for validation, and the final two for testing. This model showed train and validation trending in the right direction, but failed to predict the initial breakdown. However, when testing a learning model, it is more important to be able to understand the trend than the initial breakdown. In a model with 30 breakdowns the first 20 are used for training, the next 4 are for validation, and the final 6 are for testing. In this model, the training, validation, and test losses are lower than the first model. However, because this model has more breakdowns, it is not fair to compare them directly. Finally, in a model with 100 breakdowns, the first 75 are for training, the next 10 are for validation, and the final 15 are for testing. Once again, the errors for training, validation, and testing are impaired when compared to the first model. However, the model improves by having 100 breakdowns rather than 7. This means the model successfully predicts the trend as well as the breakdowns in the validation data.

In conclusion, improval of the model for predictive maintenance could help NASA in saving lives as well as money. As stated, going from 7 to 100 breakdowns showed that the model does a better job at predicting the engine life, as well estimating breakdowns. Ultimately, the testing determined that if you have the right data in the correct sample frequency and sample interval, you will be able to obtain an accurate Machine Learning model for estimating engine’s remaining useful life by having 7 breakdowns available.

For all types of aircraft engine parts and much more, look no further than Purchasing Union. Owned and operated by ASAP Semiconductor, we can offer customers access to an inventory of more than two billion new and obsolete parts. Within this inventory you will be able to find all types of parts for the aerospace, civil aviation, defense, electronics, and IT hardware markets. Dedicated account managers are always available and ready to help you find all the parts and equipment you need, 24/7-365. For a quick and competitive quote, email us at sales@purchasingunion.com or call us at 1-714-705-4780. Let us show you why we consider ourselves the future of purchasing.