Skip to content
Multigrid

Sensor and IoT data

Sampling rates, drift, fusion and edge inference — modelling telemetry from devices that are unreliable by nature.

A sensor does not report the world. It reports a voltage that has been digitised, scaled by a calibration constant somebody wrote down once, and stamped with a time from a clock that is drifting. Every one of those steps introduces an error with its own shape, and modelling telemetry well is mostly a matter of knowing which error you are looking at: quantisation noise is white and harmless, aliasing is irreversible, and calibration drift is a slow bias that no amount of averaging removes.

These pages work outward from that. They start at the sampling decision, because it constrains everything downstream and cannot be undone in software, and then cover what to do with the stream once it exists: fusing several noisy channels into one estimate, catching a sensor that is failing before its output looks wrong, aligning clocks that disagree, and deciding what gets computed on the device rather than shipped. Where a number matters, it is derived on the page from inputs you can replace with your own.

Sensor Fusion Algorithms Explained

How two noisy measurements of the same quantity combine into one estimate that is better than either, worked through from inverse-variance weighting to the Kalman recursion.

10 min read

Anomaly Detection Across Correlated Sensor Streams

Why a fault often shows up as a broken relationship between channels while every individual channel stays inside its limits, and how to detect that.

10 min read

Edge Inference for Wearable Sensor Data

What running a model on accelerometer and PPG streams on the device itself actually involves: windowing, feature extraction, memory arithmetic and the duty cycle that decides battery life.

10 min read

Compressing Sensor Data Streams for Storage

What delta-of-delta and XOR encoding do to a telemetry stream, with the storage saving derived from a stated sampling rate and channel count.

10 min read

Choosing a Sampling Rate for IoT Data Collection

The Nyquist limit sets the floor, energy and bandwidth set the ceiling, and this derives data volume and current draw at three rates on one device.

10 min read

Detecting Sensor Calibration Drift Over Time

How to separate a slow bias in a sensor from a slow change in the thing it measures, using periodic references, peer comparison and a bias state.

10 min read

How Much Battery an On-Device Model Actually Uses

Energy per inference derived from a stated power draw and duration, turned into daily drain and device lifetime at three inference rates.

9 min read

Streaming Sensor Data Into a Model in Real Time

Build a windowed inference consumer that reads sensor messages from a queue, buffers them per device, and scores complete windows without blocking the reader.

10 min read

Building a Real-Time Digital Twin From Live Sensor Telemetry

How a continuously updated model of a physical asset stays in sync with telemetry, and what the residual between model and measurement is actually for.

10 min read

What Changes in Cost Moving Inference From the Cloud to the Edge

Derives the fleet size and inference volume at which amortised edge hardware becomes cheaper than per-inference cloud pricing, from labelled assumptions.

10 min read

Handling Schema Drift in a Sensor Data Pipeline

How a firmware update silently changes what a field means, why type checks do not catch it, and the contract a pipeline needs so that it does.

10 min read

Smart Building Occupancy Sensing With AI

Estimating headcount in a room by combining CO2 concentration with motion, using the mass balance that makes CO2 a quantitative occupancy signal.

10 min read

Fleet Telematics Anomaly Detection

Detecting abnormal vehicle behaviour from CAN bus and GPS streams, including why per-vehicle baselines matter more than fleet-wide thresholds.

10 min read

Labeling Challenges in Industrial IoT Sensor Data

Why ground truth for continuous machine telemetry is expensive and often unobtainable, with a worked labelling-cost example and the weak-supervision alternatives.

10 min read

What It Costs to Ingest a Million Sensor Readings a Day

Derives message, storage and query cost for a million readings a day from stated payload sizes and unit prices, and shows why batching dominates the answer.

10 min read

Detecting a Failing Sensor Before It Reports Bad Data

The variance and rate-of-change signatures that precede total sensor failure, and how to monitor a sensor's health separately from the quantity it measures.

10 min read

Time-Aligning Data From Sensors With Different Clocks

Correlated events that do not line up across sensors, diagnosed as a clock offset and fixed by estimating and correcting the offset per device.

10 min read

Choosing Between Edge and Cloud Preprocessing for Sensor Data

Deciding which processing stages run on the device by weighing bandwidth reduction against the compute the device has and the information the reduction discards.

10 min read

Other topics