Founding Industrial Partners cohort is open — one year of Luviner free for selected manufacturers Apply by May 31 →

From Sensor to CSV

A practical guide to collecting data from your sensors and preparing it for AI training.

Where your data comes from

Sensor data can come from many sources. Here are the most common ones and how to export them.

Source How to export Typical format
PLC / SCADA Built-in data export or historian software (TIA Portal, FactoryTalk, etc.) CSV, XLSX
Arduino / ESP32 Serial.println() to Serial Monitor, or write to SD card with the SD library Serial text, SD card CSV
Data logger USB Connect the logger, download the file via manufacturer software CSV, TXT
Excel / Google Sheets Spreadsheet with manual readings — export as CSV (File > Save As > CSV) XLSX → CSV
OPC-UA / MQTT Use a client (e.g. Node-RED, Python opcua) to subscribe and log to CSV JSON → CSV

Step 1: From raw data to clean CSV

Raw sensor exports often include timestamps, text status columns, or metadata that Luviner doesn't need. Here's a before/after example:

# Raw export from SCADA (has timestamp and text status) 2026-03-25 10:00:01,23.4,0.12,1013.2,OK 2026-03-25 10:00:02,23.5,0.11,1013.1,OK 2026-03-25 10:00:03,45.8,3.45,1012.8,ALARM # Clean CSV ready for Luviner (numeric only, label as last column) temperature,vibration,pressure,label 23.4,0.12,1013.2,0 23.5,0.11,1013.1,0 45.8,3.45,1012.8,1

What to do:

  • Remove timestamp columns — Luviner handles temporal patterns automatically from row order
  • Convert text values to numbers — replace status strings (OK/ALARM) with integers (0/1)
  • Move the label (class) to the last column
  • Save as UTF-8 CSV with comma delimiter

Step 2: Clean common issues

Quick fixes for the most common problems in sensor data exports.

Problem Solution
Empty cells or missing values Delete rows with missing values, or fill with the previous reading (forward fill)
Text in feature columns (e.g. "N/A", "ERR") Replace with a numeric value or remove the row
Mixed units (e.g. some rows in °C, others in °F) Convert all values to the same unit before export
Decimal comma instead of dot (1.234 vs 1,234) Replace commas with dots in numeric columns, or use semicolon as CSV delimiter
Duplicate rows or irregular sampling intervals Remove duplicates; for irregular intervals, resample to a fixed rate if possible

Step 3: Labeling your data

Each row needs a label — an integer that tells Luviner which class that reading belongs to. There are several ways to assign labels depending on your situation.

Method When to use Example
Known fault types You already know what each condition looks like (from maintenance logs or domain knowledge) normal=0, bearing_fault=1, misalignment=2
From SCADA/PLC alarms Your system already flags abnormal conditions — map alarm codes to integers OK=0, ALARM=1
By time window You know when a fault occurred — label all rows before that moment as 0, and rows during/after as 1 "Everything before March 15 = normal (0), after = fault (1)"
Start simple. Even a basic normal/anomaly split (0 vs 1) is enough to get your first model. You can refine the labels later.

Examples by sector

Predictive maintenance (vibration)

Accelerometer data from an electric motor. Labels: 0 = normal, 1 = bearing fault, 2 = misalignment.

accel_x,accel_y,accel_z,rms,peak_to_peak,temperature,label 0.12,0.98,-0.05,0.34,1.2,45.2,0 2.34,1.87,0.95,4.56,8.9,67.8,1 3.12,2.45,1.23,6.78,12.3,72.1,2

HVAC monitoring

Air handling unit readings. Labels: 0 = normal, 1 = low airflow, 2 = compressor overload.

temp_supply,temp_return,humidity,pressure_diff,fan_speed,power_w,label 22.1,18.3,45.0,120,1450,340,0 22.0,22.1,78.0,15,1440,345,1 28.5,19.0,42.0,250,1800,520,2

Food storage & cold chain

Cold room monitoring for perishable goods. Labels: 0 = safe conditions, 1 = spoilage risk.

temp_chamber,humidity_rel,co2_ppm,weight_kg,time_hours,label 4.1,85.0,410,12.5,0,0 4.2,85.5,415,12.4,24,0 8.5,92.0,680,11.8,72,1

Energy & power quality

Electrical panel readings. Labels: 0 = normal, 1 = harmonic distortion, 2 = undervoltage.

voltage,current,power_factor,thd,frequency,label 230.1,12.5,0.98,2.1,50.00,0 228.5,18.2,0.72,15.3,49.95,1 185.0,25.0,0.65,22.8,49.80,2

Pre-upload checklist

  • All feature columns are numeric (no text, no dates)
  • The last column is the label (integer: 0, 1, 2, ...)
  • No empty cells or missing values
  • At least 50 rows (500+ recommended)
  • Balanced classes — roughly similar number of rows per label
  • Saved as UTF-8, comma-separated
  • First row is a header with column names
Ready? Check the CSV Format reference for detailed specs, or go straight to Getting Started to upload and train.