PredictEdge: Predictive Maintenance on a Raspberry Pi

For Tata Technologies InnoVent 2026 I built PredictEdge: a prototype that predicts when industrial machinery is going to fail — running entirely on a Raspberry Pi, no cloud required.

The problem: Remaining Useful Life

Predictive maintenance comes down to one number: Remaining Useful Life (RUL) — how many cycles a machine has left before it degrades past a threshold. Get it right and you service equipment just in time instead of too early (wasteful) or too late (catastrophic).

Data: NASA CMAPSS

I trained on the CMAPSS turbofan-degradation dataset — multivariate sensor time series with run-to-failure trajectories. The model learns to map a window of sensor readings to an RUL estimate.

Getting it onto the edge

A model that needs a GPU is useless on a factory floor. So the pipeline was:

1. Train a compact regression model.
2. Quantize to INT8 with TensorFlow Lite.
3. Deploy the .tflite model to a Raspberry Pi and run inference locally.

converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_types = [tf.int8]
tflite_model = converter.convert()

INT8 quantization cut the model size dramatically with only a small accuracy hit — exactly the trade you want at the edge.

Why edge, not cloud

• Latency: decisions happen where the machine is.
• Resilience: works even when connectivity drops.
• Privacy & cost: raw sensor streams never have to leave the site.

PredictEdge is still a prototype, but it proved the core idea: a credible RUL model can fit on a $35 computer.