Smart Energy Monitoring System

Finding the fault before buying the hardware

Factory energy monitors are normally tested once they are bolted to a machine. This one was tested first, in a simulated factory where every fault was known in advance — including the faults it turned out to miss.

Compressor 0130 kW · runs, then idles between jobs
Drawing0.0 kW
Doing
Anomaly score0.00
Monitor saysLearning baseline
  • Power drawn
  • Fault actually happening
  • Monitor raised an alarm

Try both buttons. The same fault, at two moments in the same machine's cycle, does not get the same answer.

Running in this page, in your browser. No server, no account, nothing sent anywhere.

The problem

A month of electricity arrives as one number

Energy is a fifth to two fifths of what it costs to make things in energy-intensive manufacturing. Large plants meter every machine. Smaller manufacturers usually get one figure a month for the whole site.

Two very different problems hide inside that figure, and the bill cannot tell them apart. A machine idling all afternoon waiting for work burns money for nothing. A machine with a failing bearing draws more than it should and is on its way to stopping the line. The first wants a scheduling change, the second wants an engineer, and neither is visible until something breaks.

Why it does not get bought

Four reasons smaller manufacturers say no

The research literature on adoption keeps returning the same four objections. Each one shaped a decision in this design.

It costs too much to start

The sensing node is commodity parts — a microcontroller and clip-on current sensors — costed against named commercial meters. That comparison is a priced estimate, not a purchase that has been made.

It might disturb the line

Clip-on sensors close around a cable. No wiring is cut and no machine is taken out of service to fit them.

Nobody here can run it

The interface is a web page. It opens on a phone on the shop floor, with nothing to install.

We cannot tell whether it will pay back

This is the one the project attacks hardest. You can watch the system find faults, and miss them, before committing money.

The method

Build the factory first, then test the monitor on it

Detection accuracy is usually measured against recordings from a real plant. Those recordings have a flaw: nobody wrote down when the faults really happened. A fault the monitor misses leaves no trace, so misses are invisible and the score always flatters the product.

  1. Simulate the machines

    Five of them, on steady, cyclic and idle-active duty, with realistic noise, drift and electrical behaviour, all driven from a fixed random seed.

  2. Inject faults deliberately

    The simulation chooses when each fault starts, how large it is and how long it lasts, and writes that down. The monitor never sees the record — the same arrangement as the two buttons above.

  3. Mark the monitor's answers

    Every alarm is checked against the record, and every miss counts against it. Because the seed is fixed, anyone can rerun the same factory and get the same result to the last digit.

What it found

Trustworthy in most conditions, blind in one

Across 143 deliberately injected faults:

8 in 10Faults caught before the episode ended
97%Of alarms were real faults rather than false alarms
25,000Readings reproduced exactly on a rerun

The failure worth more than the successes

A fault that began while a machine was idling between jobs was almost never noticed. The same fault during production was caught nearly every time.

98.1% caught — fault begins while the machine is working
1.2% caught — fault begins while the machine is idling

An idling machine draws so little that a fault hides inside the ordinary variation. This blind spot cost an afternoon to find. Found after installation, it would have cost a great deal more — and it tells the production system exactly what it has to fix: it needs to know which phase a machine is in, not only how much power it is drawing.

These figures come from the simulation, not from a factory installation. They measure the design and the detection method, not a finished product.

The full instrument

Run the whole factory floor

The compressor above is one machine from a five-machine testbed. The full version adds the fleet view, the cost tile, the event feed, editable machine parameters, saved profiles and CSV export of every reading with its ground-truth fault flag.