Industrial Energy Intelligence Platform

A real industrial challenge set by Kilani Group: monitor and cut a factory’s energy consumption. The obvious architecture streams every sensor reading to a server and decides there — which means the factory floor stops being able to detect its own faults the moment the network does.

So the detection was pushed onto the hardware. A quantised neural network runs on the ESP32 node itself, inspecting readings every ten seconds and flagging abnormal ones before they are ever published to the MQTT broker. No server contact is required for a fault to be caught.

• A sliding-window next-step predictor: the model takes the last five readings across all five sensor channels and predicts what each channel should read next. Input(25) → Dense(16, ReLU) → Dense(8, ReLU) → Dense(5, Sigmoid), roughly 597 parameters.
• Detection is fully unsupervised — the network is trained only on normal operation and never sees a labelled anomaly. A fault is caught because the model cannot predict it: when the worst channel’s normalised error exceeds 0.15, the reading is flagged and that channel is named as the culprit.
• One multi-output model rather than five separate detectors. It catches compound faults in a single inference call, exploits cross-sensor correlation — the two temperature probes normally track each other, so a divergence is caught even when neither channel looks bad alone — and takes about a fifth of the on-device footprint.
• Training data was recorded live from the hardware over the hackathon LAN: 128 rows of normal operation at a ten-second interval. Thirty labelled fault rows were generated separately and held out entirely for validation.

Sensors on factory equipment feed a system that shows where electricity is actually going, so a plant can cut the waste it could not previously see.