Time Series at Edge

A temperature sensor, accelerometer, microphone, or power meter does not produce a single answer-ready value—it produces a stream of measurements over time. Time series at edge means analyzing that stream directly on the device that collects it, so the device can recognize patterns, changes, or future risks without constantly sending raw data to the cloud.

Edge devices break an incoming signal into short, overlapping windows: for example, one second of vibration data from a motor or two seconds of accelerometer readings from a wristband. A model then examines the values and their order. It might use raw samples, or compact features such as average, variance, frequency peaks, or signal energy. Unlike a photo classifier, which sees a complete image at once, a time-series model must keep up with a continuing stream and preserve enough recent history to make sense of the present.

• A battery-powered sensor detects an unusual vibration pattern that signals bearing wear in a pump.
• A smartwatch recognizes walking, running, or a fall from motion data.
• A wake-word detector continuously listens for a short audio pattern while keeping the microphone data local.
• A smart electricity meter spots abnormal load behavior before equipment fails.

At the edge, sampling rate, window length, model size, and inference frequency are engineering decisions, not afterthoughts. A Cortex-M microcontroller might have only a few hundred kilobytes of RAM, so it cannot retain long histories or run a large recurrent network. Developers commonly use small convolutional models, feature-based classifiers, or compact anomaly detectors in runtimes such as TensorFlow Lite for Microcontrollers. The model must also meet a power budget: running every millisecond can drain a coin-cell battery even when the model itself is tiny. Careful windowing and low-power triggering let the device react locally, remain useful offline, and avoid transmitting sensitive raw sensor streams.