Skip to content
CO
← All work
03 / 09

Predictive Maintenance for Turbofan Engines

A comparative evaluation of eight AI approaches for predicting the remaining useful life of degrading aircraft turbofan engines using NASA C-MAPSS data.

Type
Artificial Intelligence · Predictive Maintenance · Research
Role
AI Engineer / Researcher
Contribution
Data preparation · Modelling · Evaluation · Research
Year
2026

The problem

Aircraft turbofan engines degrade gradually. Maintenance scheduled too early wastes serviceable life; maintenance scheduled too late risks failure. Remaining useful life (RUL) prediction estimates how many operating cycles an engine has left from its sensor history. Many approaches have been proposed, from classical regression to deep and neuromorphic networks, but they are often reported under different protocols, which makes them hard to compare fairly. This project evaluated eight of them side by side on the same data, splits and test conditions.

What I built

A single experimental pipeline on NASA’s C-MAPSS FD001 dataset, run identically for eight approaches: Ridge Regression, Extra Trees, Histogram Gradient Boosting, a 1D CNN, an LSTM, a Transformer, an Echo State Network and a spiking neural network built as a Liquid State Machine in the NEST simulator.

Sensor readings are prepared once, split by engine and turned into fixed-length windows, so every model sees the same inputs. Each approach is tuned under a controlled evaluation, then scored once on the 100 unseen engines of the official test set, alongside error by degradation stage and computational cost. The spiking model encodes 17 retained sensor features as 34 ON/OFF spike channels, drives a reservoir of 300 leaky integrate-and-fire neurons, and reads out RUL through a ridge regression layer with 301 trainable parameters.

Approach

One protocol, eight models

  1. 01Data preparation
  2. 02Engine-level split
  3. 03Window generation
  4. 04Model training
  5. 05Controlled evaluation
  6. 06Official test

Key decisions

  • Split engines, not windows

    Windows from the same engine are highly correlated. Splitting by engine rather than by random window keeps an engine’s history out of both sides of the split and prevents leakage.

  • One protocol for all eight

    Every approach receives the same preprocessing, splits and metrics, so differences in the results reflect the models rather than the setup.

  • Keep the official test set sealed

    The 100 NASA test engines are used once, for final evaluation only, so no tuning decision is informed by the numbers being reported.

  • Read accuracy alongside cost

    Models are compared on computational cost as well as RMSE. A slightly less accurate model that is far cheaper to train and run can be the better engineering choice.

Architecture of the NEST spiking neural network: 17 sensor features encoded as 34 ON/OFF spike channels feed a recurrent reservoir of 300 leaky integrate-and-fire neurons, read out by ridge regression to predict remaining useful life.
The spiking model: event-driven reservoir with a ridge regression readout.

Inside the spiking model

Input
17 sensor features → 34 ON/OFF spike channels
Reservoir
300 LIF neurons · 240 excitatory, 60 inhibitory
Readout
Ridge regression · 301 trainable parameters
Window
30 cycles · 1 ms simulation step

Outcome

Best official-test RMSE (cycles)
12.57
AI approaches evaluated
8
Unseen test engines
100

Histogram Gradient Boosting achieved the lowest official-test RMSE at 12.57 cycles, narrowly ahead of the LSTM (12.85) and Extra Trees (12.89). The spiking Liquid State Machine placed last on accuracy at 20.17, but trains only a 301-parameter readout, which is the trade-off the study was designed to expose: accuracy has to be read alongside computational cost.

Tools
Pythonscikit-learnNEST SimulatorpandasNumPyMatplotlib
Capabilities
Machine LearningDeep LearningTime-Series ModellingPredictive MaintenanceSpiking Neural NetworksModel Evaluation

Working on something similar?

Start a conversation