Sensor Data Analysis

The Language Machines Speak

Every running machine vibrates, generates heat, draws current, and produces sound. These are not random — they are signals. A healthy bearing produces a specific vibration signature. A misaligned shaft creates a distinct pattern. A degrading motor winding draws current in a characteristic way. Learning to read these signals — or more precisely, teaching AI to read them — is the foundation of modern condition monitoring.

This chapter focuses on the two most common and valuable types of sensor data: vibration and temperature. Together, they cover 70-80% of all detectable mechanical faults. You will learn what the numbers mean, how to interpret them, and how AI processes them to predict failures weeks or months in advance.

Vibration Analysis: The ISO 10816 Standard

Vibration analysis is the single most powerful tool for monitoring rotating machinery — motors, pumps, fans, compressors, turbines, and gearboxes. The international standard ISO 10816 (now largely superseded by ISO 20816, but still widely referenced) defines vibration severity zones for different classes of machines.

Machine Classification

ISO 10816 groups machines into four classes based on size and mounting:

Severity Zones

For each class, vibration velocity (measured in mm/s RMS or in/s peak) falls into four zones:

For a typical Class III machine (large pump or compressor), the thresholds are approximately:

Common Vibration Fault Patterns

Different faults produce vibration at different frequencies relative to the machine's running speed (1X = shaft speed):

Open data/vibration-data.csv in the code panel. You will find time-series vibration data from a 50 HP centrifugal pump over 60 days. The data includes overall velocity (mm/s RMS) and dominant frequency ratios. Watch how the 1X component stays stable while the BPFO component gradually increases — this is a classic outer race bearing defect developing over time.

FFT: Turning Time into Frequency

Raw vibration data is a waveform — amplitude changing over time. This is useful but limited. The real diagnostic power comes from FFT (Fast Fourier Transform), which converts the time-domain signal into a frequency-domain spectrum.

What FFT Tells You

Think of it like music. A time-domain waveform is the overall sound. FFT separates it into individual notes — each frequency component with its own amplitude. In machine vibration:

AI Advantage Over Manual FFT Analysis

A trained vibration analyst can read an FFT spectrum and diagnose faults. But this takes years of experience, and even experts can only review one machine at a time. AI processes FFT spectra from hundreds of machines simultaneously, comparing current spectra to historical baselines, detecting subtle changes that would take a human analyst hours to spot, and trending those changes to predict when they will reach actionable levels.

Temperature Monitoring

Temperature is the second most valuable parameter for machine health. While vibration detects mechanical faults, temperature catches thermal issues — overheating bearings, electrical faults, cooling system failures, and process anomalies.

Key Temperature Monitoring Points

Infrared Thermography

Handheld or fixed IR cameras are standard in North American and European plants for:

Open data/temperature-logs.csv in the code panel. You will find 90 days of temperature data from 8 critical motors in a process plant. Two of the motors show gradual temperature increases starting around Day 45 — one from a failing bearing and one from reduced cooling airflow. The AI model flags both, but only the bearing issue requires immediate attention. This is where multi-parameter correlation becomes essential.

Multi-Parameter Correlation

The real power of AI-driven monitoring is combining multiple data streams. A single parameter change might be noise. Two or three parameters changing together is almost always a real fault developing.

Correlation Examples

Why AI Outperforms Rule-Based Alarms

Traditional monitoring uses simple thresholds: if vibration exceeds 7 mm/s, trigger an alarm. This produces false alarms (a cold start might briefly exceed 7 mm/s) and missed faults (a slow increase from 2 to 5 mm/s might not trigger the alarm but is clearly trending toward failure).

AI models learn the normal relationship between parameters for each specific machine under different operating conditions (load, speed, ambient temperature). They detect when that relationship changes, even if no single parameter has exceeded its absolute threshold. This is why AI-based monitoring catches 30-40% more faults than rule-based systems, with 60-70% fewer false alarms.

Key Takeaways