Applying Machine Learning to IoT Telemetry

Working with IoT telemetry is different from standard tabular datasets. The data is often noisy, out-of-order, and arrives at irregular intervals. When applying Machine Learning to these systems — such as predicting battery failure or optimizing charging slots — you need a specialized pipeline.

The Noise Problem

Sensors are imperfect. A voltage sensor might report 48V, 48.1V, and then a sudden spike to 99V due to an electrical transient. Standard ML models will overfit to this noise unless you apply proper filtering.

Common techniques include:

• Moving Averages: Smoothing out jitter.
• Median Filtering: Removing extreme outliers without blurring edges.
• Kalman Filters: Estimating the true state of a system when your measurements are noisy.

Feature Engineering for IoT

Time is your most important feature. Instead of just using raw values, consider:

• Rolling deltas: How much has the SOC changed in the last 10 minutes?
• Frequency domain features: Using FFT (Fast Fourier Transform) to find vibration patterns.
• State transitions: How many times has the relay flipped in the last hour?

In my work at Mooving, we use these techniques to monitor thousands of batteries in real-time, ensuring safety and uptime for the entire swapping network.