A decision-support / predictive-maintenance study: given streams of sensor readings from a fleet of aircraft turbofan engines, estimate each engine's Remaining Useful Life (RUL) — how many operational cycles are left before failure. Accurate RUL estimates let operators schedule maintenance just in time, avoiding both unexpected failures and wasteful early part replacement.

The data: NASA C-MAPSS

The project uses NASA's C-MAPSS turbofan degradation dataset, made of four subsets of increasing difficulty:

Subset Operating conditions Fault modes
FD001 1 (sea level) 1 (HPC degradation)
FD002 6 1
FD003 1 2 (HPC + fan)
FD004 6 2

Each engine is a multivariate time series of 3 operational settings and 21 noisy sensor measurements. Engines start healthy and degrade until failure; the goal is to predict the remaining cycles on the test set, where the series is truncated before failure.

From raw cycles to training samples

For training, the RUL label of each cycle is simply how many cycles remain until that engine's last recorded cycle:

def add_rul(df):
    grouped = df.groupby(by="engine_nr")
    max_cycle = grouped["time_cycles"].max()
    result = df.merge(max_cycle.to_frame(name="max_cycles"),
                      left_on="engine_nr", right_index=True)
    result["RUL"] = result["max_cycles"] - result["time_cycles"]
    return result

Each engine's signals are then cut into sliding windows of 10 cycles, so the models see a short slice of recent history rather than a single snapshot:

for i in range(0, len(engine_data) - window_size):
    X_train.append(xs[i:i + window_size, :])     # 10 cycles × 21 sensors
    y_train.append(ys[i + window_size - 1])      # RUL at the window's end

Sensors are standardized with StandardScaler, and an optional Butterworth low-pass filter can be applied to denoise the raw signals before windowing.

Models compared

The same pipeline feeds four regressors, from simple to sequential:

  • Linear Regression and Gradient Boosting — on the flattened 10 × 21 window (a quick, strong classical baseline). SVR was also explored.
  • LSTM and GRU — recurrent networks that consume the window as an actual sequence:
model = keras.Sequential()
model.add(layers.LSTM(input_dim=len(sensor_names), units=5))
model.add(layers.ReLU())
model.add(layers.Dropout(0.3))
model.add(layers.Dense(1))
model.compile(optimizer="adam", loss="mse", metrics=[RootMeanSquaredError()])

Training uses Adam with MSE loss and early stopping (patience 4 on validation loss). Every model is evaluated on all four datasets with MSE, RMSE and R², so it is easy to see how each approach copes as the number of operating conditions and fault modes grows.

Takeaways

RUL prediction is a textbook prognostics problem and a great test bed for comparing paradigms: lightweight classical regressors versus recurrent sequence models, on time-series data that is noisy, multivariate and uneven in length. The harder subsets (FD002 and FD004, with six operating conditions and two fault modes) are markedly tougher than the single-condition FD001 — a reminder that, in real predictive maintenance, operating context matters as much as the model.