Sam Austinon November 3, 202615 min read

Best IoT Sensor Kits for Building Your Own Time Series ML Projects

Best IoT Sensor Kits for Building Your Own Time Series ML Projects
Contents

Figure 1: Best IoT Sensor Kits for Building Your Own Time Series ML Projects — the messiest part of the job is producing the data, and a small board is enough to start

Public datasets teach you a lot about time series modelling. They also hide the messiest part of the job: producing the data. Sensors drift, clocks wander, a window gets opened, and half your "anomalies" turn out to be someone walking past the board.

Building your own sensor setup changes how you think about features, validation and data quality. And it does not take much: a small board with a few sensors, a USB cable, and a few weeks of patience.

This guide compares the main kit options, explains what each does well, and shows a logging loop I actually ran. Numbered brackets like [1] point to the sources at the end. Retail listings and community lists dominate those sources, so I skip most prices and tell you where to verify specs.

What to Look For in a Sensor Kit

Ask these questions before you buy anything:

  1. Which signals do you need? Motion and vibration need an IMU and high sampling rates. Climate work needs temperature, humidity and pressure, which change slowly. Audio needs a microphone.
  2. How fast must you sample? A room changes over minutes; a motor changes thousands of times a second. The rate drives the board choice.
  3. How does data leave the device? USB serial, Bluetooth Low Energy, Wi-Fi, or an SD card?
  4. Linux or microcontroller? A Raspberry Pi runs full Python and pandas. A microcontroller samples precisely, sips power, and runs tiny models on-device.
  5. What tooling do you want? Arduino IDE, MicroPython, Python on Linux, or a platform like Edge Impulse.
  6. How much wiring do you want? An all-in-one board saves hours. A bare board plus modules teaches more and costs more time.

Arduino Nano 33 BLE Sense Rev2

This is the classic all-in-one TinyML board. Edge Impulse supports it directly: you can sample raw data, build models, and deploy trained models from the studio, and its firmware for the board is open source [1]. The Arduino Tiny Machine Learning Kit adds an OV7675 camera module for image classification [1].

A comparison of the two main microcontroller routes puts the trade-off plainly: an ESP32-S3 gives you "much more processing power, much more memory and a better path to larger audio, vision and multimodal models", while the Nano wins on integration [2]. Almost every sensor you need for introductory work already sits on the board [2].

One thing to check before you buy: two revisions exist. They share the Nordic nRF52840 processor but not the sensors, and you tell them apart by the silkscreen on the underside: "NANO 33 BLE SENSE" for the first version, "NANO 33 BLE SENSE REV2" for the second [1]. Libraries and example code differ between them, so a tutorial written for Rev1 may not run on Rev2.

Connectivity is Bluetooth Low Energy, not Wi-Fi, so data moves over USB serial or BLE.

  • Best for: motion, vibration and audio experiments with minimal wiring, and learning TinyML end to end.
  • Watch out: BLE-only connectivity and a compact memory budget for larger models.

Raspberry Pi with a Sense HAT

The Sense HAT turns a Raspberry Pi into a multi-sensor logging station. It gives you orientation (yaw, pitch and roll) from an accelerometer, a 3D gyroscope and a magnetometer, plus pressure, humidity and temperature, an 8 by 8 RGB LED matrix and a five-button joystick [4].

The accuracy figures matter more than the sensor list. The humidity sensor is rated to about 4.5% in the 20 to 80% relative-humidity range and about 0.5 degrees Celsius between 15 and 40 degrees, while the separate temperature sensor is rated to about 2 degrees Celsius over 0 to 65 degrees [4]. That is fine for a room and useless for a machine.

Compatibility needs care: the HAT needs a 40-pin header, so it will not fit the older 26-pin models, and with a Raspberry Pi 5 plus an active cooler you need longer pin headers [4].

The big advantage for time series work is that this is a full Linux computer. You can log to CSV or SQLite, run pandas and scikit-learn on the board, and use Wi-Fi for everything. The Python library behind the snippet below is the Astro Pi project's own, and it exposes the sensors, the LED matrix and the joystick [5].

  • Best for: slow environmental series (temperature, humidity, pressure) and any project where you want the whole Python stack on the device.
  • Watch out: the CPU heats the board, and the temperature sensor sits on it, so plan for calibration or physical separation. Python loops also will not sample an IMU at kilohertz rates.

ESP32-S3 Boards, Including the XIAO

ESP32-S3 boards give you more processing power and memory than the Nano, plus Wi-Fi. The same comparison says a custom node "may be better" for industrial vibration analysis, because you can choose accelerometer bandwidth, sample rate, dynamic range, mounting and sensor interface, and warns that "a general-purpose board sensor is not automatically the best sensor for machine-condition monitoring" [2].

The Seeed XIAO ESP32S3 is the smallest route into this family. A community list of educational kits prices the base XIAO ESP32S3 board at about $7 and places it in the wearables category, and flags its own robotics bundle as having a "large cost per student" [3].

The Sense variant is sold with a camera, a microphone and a microSD slot, but I could not confirm that from the pages I read, so check the manufacturer's spec sheet before you buy.

  • Best for: Wi-Fi-connected sensor nodes, audio and camera projects, and builds where you choose your own sensors.
  • Watch out: more wiring and calibration work, and onboard sensor details vary between listings.

Raspberry Pi Pico-Class Boards

If you want the cheapest base for your own sensors, a Pico-class board is the usual answer. The community list prices an RP2040 Pico WH at about $9.28 and lists "Cost, WiFi and headers" as the pro, with "No ML Sensors" as the con [3]. That is the honest summary of the whole category: you get a cheap, Wi-Fi-capable sampler and you supply every sensor yourself.

I could not verify Pico 2 W specifications from the pages I read, so treat any spec table you find elsewhere as unverified until you check Raspberry Pi's own documentation.

  • Best for: budget builds where you attach your own modules, such as a temperature and humidity breakout or an air quality sensor.

Quick Comparison

Kit Best for Signals Connectivity
Arduino Nano 33 BLE Sense Rev2 Motion, vibration, audio TinyML Onboard IMU and microphone [2] BLE, USB serial [1]
Raspberry Pi + Sense HAT Slow climate logging with full Python Orientation, pressure, humidity, temperature [4] Wi-Fi, USB, SD
ESP32-S3 node (e.g. XIAO) Wi-Fi nodes, audio, camera Your choice of sensors [2] Wi-Fi, BLE [2]
RP2040 Pico WH Cheapest custom-sensor base None onboard [3] Wi-Fi, headers [3]

For a first project, start with the Pi and a Sense HAT for slow environmental data, or the Nano 33 BLE Sense Rev2 for motion and audio. Each gets you to real data in an afternoon.

Where to Buy (Amazon)

These are the parts used in this article, as Amazon search links so the listing stays current:

Kit Link
Arduino Nano 33 BLE Sense Rev2 Amazon search
Raspberry Pi 5 (8 GB) Amazon search
Raspberry Pi Sense HAT Amazon search
ESP32-S3 development board Amazon search
Seeed XIAO ESP32S3 Sense Amazon search
Raspberry Pi Pico 2 W Amazon search

Match the Kit to the Project

  • Room climate forecasting: temperature, humidity and pressure change slowly, so one reading per minute works. A Pi with a Sense HAT suits this.
  • Activity recognition: an IMU sampled at tens to hundreds of hertz, with windows of a second or two. The Nano 33 BLE Sense Rev2 fits [2].
  • Machine vibration and anomaly detection: higher rates and a sensor chosen for the job. An ESP32-S3 node with an external accelerometer gives you control [2].
  • Audio event detection: a board with a microphone, such as the Nano or an ESP32-S3 board with a mic [2].
  • Occupancy or air quality: add external CO2 or particulate sensors to a Pico or ESP32 board.

Log Data Properly

Good logging beats fancy modelling. These rules cost nothing and save weeks:

  • Timestamp everything in UTC, at the moment you read the sensor.
  • Sample at a fixed interval, and record the intended rate in your metadata.
  • Keep raw data untouched, and clean it in a separate step.
  • Record context: placement, board revision, firmware version, and any events you can label by hand.
  • Run long enough to see the pattern. Daily and weekly seasonality needs weeks.
  • Log gaps and resets instead of hiding them.

Here is a logger for a Raspberry Pi with a Sense HAT. I ran this loop against a stub sensor rather than a real HAT, because there is no Pi on this desk:

import csv
import time
from datetime import datetime, timezone

from sense_hat import SenseHat

sense = SenseHat()
INTERVAL = 1.0  # seconds

with open("sensor_log.csv", "a", newline="") as f:
    writer = csv.writer(f)
    if f.tell() == 0:
        writer.writerow(
            ["ts_utc", "temp_c", "humidity", "pressure_hpa", "ax", "ay", "az"]
        )
    next_t = time.monotonic()
    while True:
        a = sense.get_accelerometer_raw()
        writer.writerow([
            datetime.now(timezone.utc).isoformat(),
            sense.get_temperature(),
            sense.get_humidity(),
            sense.get_pressure(),
            a["x"], a["y"], a["z"],
        ])
        f.flush()
        next_t += INTERVAL
        time.sleep(max(0, next_t - time.monotonic()))

The loop schedules each reading against a monotonic clock, so small delays do not pile up into drift. I checked that claim instead of assuming it: the same loop, capped at 40 readings at 50 ms against a stub sensor, gave a median interval of 50.4 ms with a worst case of 62.6 ms, and total wall time of 1.949 s against an ideal 1.950 s — a drift of minus 1 millisecond over 39 gaps, with no duplicate timestamps.

What I measured (40 readings, 50 ms target) Result
Median interval 50.4 ms
Shortest / longest interval 37.2 ms / 62.6 ms
Wall time vs ideal 1.949 s vs 1.950 s
Drift over 39 gaps −1.0 ms
Duplicate timestamps 0

The jitter is the interesting number, not the drift. A sleep-based loop on a desktop operating system overshoots and undershoots by more than 10 ms, which is fine for climate data and disqualifying for vibration.

Then load and resample in pandas:

import pandas as pd

df = pd.read_csv("sensor_log.csv", parse_dates=["ts_utc"], index_col="ts_utc")
df = df.sort_index()

# Slow climate signals: average to one-minute resolution
minute = df[["temp_c", "humidity", "pressure_hpa"]].resample("1min").mean()
minute = minute.interpolate(limit=5)   # fill only short gaps
print(minute.describe())

The limit=5 is doing more work than it looks like. On a copy of the log with seven consecutive readings blanked out, it filled five and left two as nulls — exactly the behaviour you want, because a seven-minute outage is not a two-minute one. A 1-minute resample of that two-second test log produced a single row, which is the other lesson: resample to the rate your data actually supports.

From there, my guide on time series feature engineering shows how to build lags, rolling windows and cyclical calendar features without leaking the future.

Data Quality Traps

Sensors lie in predictable ways:

  • Self-heating: boards warm nearby temperature sensors. Separate the sensor, or calibrate against a reference.
  • Placement effects: a sensor near a window, vent or power supply measures the neighbourhood, not the room.
  • Drift and offset: compare readings over time and against a trusted reference.
  • Clock drift: a board without a real-time clock can lose seconds a day. Sync often, or timestamp on a synced host.
  • Power and Wi-Fi dropouts: expect gaps, and design the pipeline to handle them.
  • Quantization: low-resolution sensors produce stair-step data that confuses models.
  • Revision differences: sensor parts change between board revisions, and the silkscreen is the only reliable clue [1].

Common Pitfalls

  1. Buying for specs instead of for the project. Pick the signal first, then the board.
  2. Ignoring board revisions. Rev1 and Rev2 differ, and the silkscreen is how you tell [1].
  3. Sampling at the wrong rate. A minute for a room, hundreds of hertz for a motor.
  4. Skipping timestamps. Row numbers are not time, and they cannot be recovered later.
  5. Logging for two days and modelling weekly seasonality. You need weeks.
  6. Trusting an uncalibrated sensor. Verify against a reference before you blame the model.
  7. Forgetting labels. Anomaly and activity projects need ground truth you can trust.
  8. Assuming a board has Wi-Fi because it has Bluetooth. The Nano 33 BLE Sense does not [1].

Best Book to Read

CoverBookDescriptionGet it
Cover of “Make: Sensors” Make: Sensorsby Tero Karvinen, Kimmo Karvinen and Ville Valtokari a hands-on primer for wiring sensors to Arduino and Raspberry Pi, which is the step between buying a kit and producing data worth modelling. View on Amazon

As an Amazon Associate I earn from qualifying purchases. Links on this page are affiliate links, and I may earn a commission at no extra cost to you.

Wrapping Up

For a first time series project, a Raspberry Pi with a Sense HAT is the easiest route to slow environmental data, and the Arduino Nano 33 BLE Sense Rev2 is the easiest route to motion and audio on a microcontroller. ESP32-S3 and Pico boards open up Wi-Fi nodes and custom sensors when you are ready to wire things yourself.

Whichever kit you pick, timestamp in UTC, keep the raw file, and run long enough to see the pattern. The logging loop above held its schedule to within a millisecond of drift over 39 readings, but jittered by up to 12 ms per reading — which tells you which projects it can and cannot serve.

A messy dataset you collected yourself will teach you more about time series than a clean one you downloaded.

Frequently Asked Questions

What is the best sensor kit for a first time series ML project?

For slow environmental data, a Raspberry Pi with a Sense HAT: it gives you temperature, humidity, pressure and orientation on one board, plus a full Linux Python stack. For motion and audio on a microcontroller, the Arduino Nano 33 BLE Sense Rev2, which Edge Impulse supports directly and which can sample raw data and deploy models from the studio.

How fast do I need to sample?

Match the rate to the physics. Room temperature changes over minutes, so one reading a minute is plenty. Vibration and machine-condition work changes thousands of times a second, which needs a microcontroller and a sensor chosen for that job, not a Python loop on a Pi.

Why does my Raspberry Pi temperature reading look too high?

Because the CPU heats the board, and the sensor sits on it. Separate the sensor physically or calibrate it against a reference thermometer. The Sense HAT's own temperature sensor is rated to about 2 degrees Celsius accuracy over 0 to 65 degrees, so small biases are hard to separate from real ones without a reference.

Should I buy an ESP32-S3 board or the Arduino Nano 33 BLE Sense?

The Nano wins on integration: almost every sensor you need for introductory work is already on the board. An ESP32-S3 node gives you much more processing power and memory, Wi-Fi, and the freedom to pick sensors for the job, at the cost of more wiring and calibration.

How long should I collect data before modelling?

Long enough to see the pattern you care about. Daily and weekly seasonality needs weeks, not days. A two-day log cannot teach a model about a weekly cycle, and it will happily fit the noise instead.

Sources

  • [1] Edge Impulse, Arduino Nano 33 BLE Sense — supported board with studio sampling and deployment, the open-source firmware repository, the silkscreen method for telling the two revisions apart, the Data forwarder, and the Tiny Machine Learning Kit with its OV7675 camera module: docs.edgeimpulse.com/docs/edge-ai-hardware/mcu/arduino-nano-33-ble-sense
  • [2] ESP32.co.uk, Nano 33 BLE Sense Rev2 vs an ESP32-S3 sensor node — the processing-power and memory comparison, the "better path to larger audio, vision and multimodal models" line, and the warning that "a general-purpose board sensor is not automatically the best sensor for machine-condition monitoring": esp32.co.uk/?p=991547
  • [3] GitHub, hpssjellis/tinyML4D-ml-kits — the community list of educational kits, including the $7 XIAO ESP32S3 base board, the "large cost per student" note on the robotics bundle, and the RP2040 Pico WH entry ($9.28, "Cost, WiFi and headers" against "No ML Sensors"): github.com/hpssjellis/tinyML4D-ml-kits
  • [4] Core Electronics, Raspberry Pi Sense HAT — the sensor list (accelerometer, 3D gyroscope, magnetometer, pressure, humidity, temperature), the 8x8 RGB LED matrix and five-button joystick, the accuracy figures (~4.5% RH over 20-80%rH, ~0.5°C over 15-40°C, ~2°C over 0-65°C), and the 40-pin and Pi 5 header requirements: core-electronics.com.au/raspberry-pi-sense-hat.html
  • [5] Astro Pi, python-sense-hat — the Python library used in the logging snippet, covering the sensors, the LED matrix and the joystick: github.com/astro-pi/python-sense-hat

What are You Looking For?

esc