This project predicts the next values of motor health/operation signals from time-series sensor data using a Long Short-Term Memory (LSTM) network.

Data (source & signals)

• Data comes from an ABB Smart Sensor attached to an electric motor (fan). The platform provides ~90 days of measurements.
• Signals include health (overall vibration, bearing condition, skin temperature) and operational (vibration axial/tangential/radial, motor supply frequency, output power, peak-to-peak for three axes, etc.).
• The workflow downloads two Excel files (health + operational), then merges them into one dataset for modeling.

Method

1. Cleaning & smoothing: remove irrelevant entries, handle missing data, and filter outliers; visualize each signal.
2. Correlation & dimensionality reduction: keep variables with Pearson correlation > 60% to overall vibration.
3. Model: LSTM RNN (bidirectional, many-to-many) with ReLU activations and Adam optimizer; trained with 2 LSTM layers and ~400 epochs.
4. Sequence setup: use a 10-step input window to predict subsequent values.
5. Train/test split: data is split into training and testing sets and evaluated with plots.

Results (summary)

• The notebook shows target vs. predicted overlays for key signals (e.g., overall vibration and peak-to-peak axes).
• An accuracy check (after rounding continuous values) is included; results are noted as limited by small data size.

See Report for full details.