Turbo Engine RUL Prediction: A Machine Learning Approach

Summary

The video demonstrates how to predict a turbofan engine’s remaining useful life (RUL) using NASA’s CMAPSS engine dataset. As engines accumulate operating cycles, their condition deteriorates, but the degradation pattern varies between engines. A fixed maintenance schedule can therefore replace some engines too early while allowing others to approach failure.

The presenter reframes the prediction target from RUL to a life ratio: the current cycle divided by the engine’s end-of-life cycle. This produces a value from 0 at the start of an engine’s life to 1 at failure, which the presenter argues is easier for a machine-learning model to predict. If the end-of-life cycle is known, the ratio can be converted back to RUL.

Methodology

Load and prepare the dataset

• Download the NASA turbofan engine data, provided in text files.
• Read the files into pandas and rename the columns for clarity.
• Remove columns 26 and 27, which the presenter says contain only missing values.
• Remove sensor features that do not vary, since constant readings provide no useful predictive information.

Create a target variable

Treat each engine’s operating history as a sequence of cycles, ending at its failure cycle in the training data. Calculate the life ratio for each observation:

current cycle ÷ end-of-life cycle

The target is 0 at the beginning of an engine’s life and 1 at the end. The presenter notes that RUL can span very different ranges across engines, while the life ratio stays between 0 and 1.

Explore and select features

The presenter plots sensor readings and examines their correlations with the life ratio. Some sensors correlate with the target, while settings 1 and 2 have weak correlations and are omitted from the model. Engine ID is also described as weakly related to the target.

Train and evaluate a random forest

The model uses sensor features to predict the life ratio.

• FD001 training set: The presenter first trains a random forest with 100 trees, using a 20% test split. The reported mean squared error (MSE) is about 0.0036 on that split. Performance is worse on the separate test data, which the presenter identifies as overfitting.
• All four training subsets: The presenter then expands training to include FD001–FD004. The reported MSE is about 0.0040 on training data and 0.0049 on test data.
• Predictions by engine: Plots of predicted versus true values show how closely the model tracks test engines’ life ratios. Accuracy varies among engines.

Compare with neural-network approaches

A basic neural network with three hidden layers of 256 neurons each, Leaky ReLU activations, and one output prediction achieves a reported test MSE of about 0.0053, similar to the random forest.

The presenter also discusses CNN and LSTM approaches that use sliding windows of sensor readings as model inputs. In the demonstrated run, these perform worse, with an MSE of about 0.031.

The presenter’s practical conclusion is that the random forest is the preferred approach in this demonstration: it trains quickly and gives the best reported result among the tested models.

Speakers and Sources

• Speaker: The presenter, whose name is not stated in the subtitles.
• Dataset source: NASA’s turbofan engine (CMAPSS) dataset, described as downloadable through Kaggle.
• Code source: The presenter says the code is available on GitHub.
• Additional source mentioned: An earlier introductory video by the presenter about fault prognosis.