Predictive Anomaly Detection for Vertical Transport
AI Lead · 2025
Predictive maintenance for elevators, built from sensor data nobody understood: I defined the events I wanted, hand-annotated them with a millisecond chronometer, and read each one's signature straight off the charts. No machine learning required.
Setup
IoElevator was a non-invasive IoT device installed on or above an elevator or inclinator cabin, equipped with accelerometer, pressure, humidity, temperature, and magnetometer sensors. The goal was to move elevator maintenance from scheduled and reactive to monitored and predictive: continuous insight into how each lift was behaving, with anomaly detection on top.
The first challenge was simply how to use the data at all. Nobody knew what each sensor could actually tell us, and I had never worked with anything like it. It was a completely new field with no map.
Why it was hard
There was no documentation and no prior experience to lean on. The more I studied the streams, the more I realized how unusual the situation was: raw signals from five sensors, and no idea which of them encoded anything useful. When there is nothing to rely on, you have to create the thing you wish existed.
Impact
What began as five unlabeled sensor streams became a working maintenance tool: one that catches problems before they become failures, from misleveling to vibration anomalies traced back to the part that caused them. Predictive maintenance for vertical transport, built entirely on geometry and arithmetic, with not a model in sight.
Takeaways
Not everything needs a state-of-the-art model to create real value. Often the basics, intuition, and plain problem-solving get you further.
When there is no documentation, write your own. Defining the events precisely was the thing that made them findable.
Sometimes the move is to forget you are a senior with years of experience. Be a child, have fun, and solve the puzzle.